The breadth of Information Retrieval (IR) methods has expanded substantially in the past ten years, especially with the introduction of neural networks, pretrained language models, and large language models. PyTerrier was introduced to provide easy access to these methods and to provide "glue" that enables researchers to experiment with them in combination. Since its introduction, PyTerrier has both evolved and spawned a growing ecosystem of plugins for different retrievers, rerankers, and answer generation. This tutorial provides a hands-on overview of using prominent methods in IR through PyTerrier, ranging from efficient lexical retrieval to answer generation. Unlike typical tutorials, this one intersperses traditional lecture-style material with interactive PyTerrier demonstrations, enabling participants to gain practical experience with the methods. https://pyterrier-tutorial.github.io/

Dense retrieval uses embedded dense representations for documents to be obtained from language models ingested in vector stores. To reduce large space requirements and/or efficiency concerns, approximate nearest neighbours algorithms can be deployed, typically at the cost of effectiveness. Among existing approaches, Product Quantisation (PQ) is one well-known algorithm, where embeddings are split along dimensions, and then assigned to similar cluster centroids on a per-split basis. Joint Product Quantisation (JPQ) improves over PQ by further training the obtained centroid embeddings, as well as the query encoder, for enhanced effectiveness. In this replicability study, we reimplement JPQ, comparing its performance on the STAR biencoder model with that reported in the original paper, as well as considering its generalisability in several dimensions: to other biencoders such as TCT and TAS-B; to larger models such as RepLLama; and beyond document retrieval, to a retrieval augmented generation task. Among many interesting results, we find that JPQ does work on more modern biencoder models such as TCT but not on larger models such as RepLLama; however, it does scale to large corpora, allowing dense retrieval for RAG that is as effective from an index that is 32x larger.

Traditional search tasks aim to satisfy user information needs by returning a subset of a collection of documents, ranked by the documents' relevance to a user query. However, some collections that contain useful information also contain sensitive personal information. Recently, there has been increasing interest in the development of Sensitivity-Aware Search (SAS) retrieval models to provide users with effective retrieval results without revealing such sensitive information. To develop such systems, test collections containing both sensitive and non-sensitive information, a set of queries, and query-document relevance assessments are required. The Enron email corpus contains real business-related emails, where some emails also contain sensitive personal information. However, the original Enron collection does not contain queries or query-relevance assessments. To this end, we crowdsource 150 query formulations for 50 different topics and 11,471 query-relevance assessments for a subset of the Enron documents that have been manually labelled for sensitivity. We follow best practices for using large language models (LLMs) in Information Retrieval evaluation to extend the collection further with additional LLM judged query-relevance assessments and sensitivity labels. We present baseline performances for relevance, sensitivity classification, and sensitivity-aware search on the collection. We make the collection available, including through the popular ir_datasets package, and provide pre-built sparse and dense indices on Huggingface to facilitate easy experimentation.

Multi-vector dense retrieval models, such as ColBERT, achieve strong retrieval effectiveness by modelling fine-grained token-level interactions between queries and documents. Methods such as PLAID use centroid-based quantisation of each token's vector to reduce the index size and speed up retrieval while maintaining strong effectiveness. In this work, we introduce PLAID-PRF, a method that performs Pseudo-Relevance Feedback (PRF) over PLAID to reformulate ColBERT's query vectors based on the top-retrieved results. In contrast with prior methods that perform PRF on multi-vector retrieval models, PLAID-PRF keeps computational costs low by leveraging the internal PLAID centroid vectors, treating them similarly to tokens in traditional PRF methods. The method selects a small and diverse set of high-utility expansion vectors and appends them to the original query, rerunning PLAID to refine both candidate generation and final scoring. Extensive experiments on the standard in-domain MSMARCO and four out-of-domain BEIR benchmarks show that PLAID-PRF consistently improves retrieval effectiveness over various baselines. In particular, PLAID-PRF improves over PLAID by up to 4.3% nDCG@10 and 7.3% MRR@10, while introducing substantially less computation overhead than prior PRF methods. The results demonstrate that our proposed centroid-aware PRF method offers an effective and lightweight mechanism to improve the quality of top-ranked retrieved results. Overall, this work enables effective and efficient feedback-aware late-interaction retrieval without expensive query-time document-token clustering.

While question-like queries are gaining popularity, keyphrase search is still the cornerstone of web search and other specialised domains such as academic and professional search. However, current dense retrieval models often fail with keyphrase-like queries, primarily because they are mostly trained on question-like ones. This paper introduces a novel model that employs the ColBERT architecture to enhance document ranking for keyphrase queries. For that, given the lack of large keyphrase-based retrieval datasets, we first explore how Large Language Models can convert question-like queries into keyphrase format. Then, using those keyphrases, we train a keyphrase-based ColBERT ranker ( ColBERTKP Q D ) to improve the performance when working with keyphrase queries. Furthermore, to make the model more flexible, allowing the use of both the question and keyphrase encoders depending on the query type, we investigate the feasibility of training only a keyphrase query encoder while keeping the document encoder weights static ( ColBERTKP Q ). We assess our proposals’ ranking performance using both automatically generated and manually annotated keyphrases. Our results reveal the potential of the late interaction architecture when working under the keyphrase search scenario. This study’s code and generated resources are available at https://github.com/JorgeGabin/ColBERTKP . • We reveal key weaknesses in current dense retrieval models when handling keyphrase queries. • We leverage Large Language Models (LLMs) and manual annotations to create tailored collections of keyphrase search queries for training keyphrase models. • Our experiments on seven query sets demonstrate the superior performance of keyphrase-based models in keyphrase retrieval tasks. • We demonstrate how our training strategy generalises effectively across different retrieval models. • We show how the keyphrase-based models adapt to old-fashioned title query collections.

The use of recommender systems to assist in the provision of financial asset and portfolio recommendations to investors is increasing, spanning a wide range of algorithms and techniques. Several strategies have been devised for the evaluation of financial asset recommendations, with the two most prominent perspectives measuring, respectively, (a) the money customers could obtain if they followed the recommendations (profitability-based evaluation) and (b) the ability of models to predict future customer investments (transaction-based evaluation). If customers are effective investors, we would expect these two perspectives to be positively correlated. In this article, we explore the actual relationship between these two families of metrics. Theoretically, we prove that these perspectives are independent. Furthermore, we perform experiments over a large-scale financial recommendation dataset with real customer investment transactions. Surprisingly, we find that transaction and profitability-based metrics are, in fact, negatively correlated. Moreover, algorithms that actively learn from past customer transactions might lose money in the mid-term. A thorough analysis of model performance and customer transaction patterns over time shows that this is due to customers failing to consistently beat the market with their investments, with time appearing as an important confounding variable—since the point of time where recommendations are provided and the investment horizon largely affect the customer’s investment performance.

Query performance prediction (QPP) estimates the retrieval quality of ranking models without the use of any human-assessed relevance judgements and finds applications in query-specific selective decision making to improve overall retrieval effectiveness. Although unsupervised QPP approaches are effective for lexical retrieval models, they usually perform weaker for neural rankers. Recent work shows that leveraging query variants (QVs), i.e., queries with potentially similar information needs to a given query, can enhance unsupervised QPP accuracy. However, existing QV-based prediction methods rely on QVs generated by term expansion of the input query, which is likely to yield incoherent, hallucinatory and off-topic QVs. In this article, we propose to make use of queries retrieved from a log of past queries as QVs to be subsequently used for QPP. In addition to directly applying retrieved QVs in QPP, we further propose to leverage large language models to generate QVs conditioned on the retrieved QVs, thus mitigating the limitation of relying only on existing queries in a log. Experiments on TREC DL’19 and DL’20 show that QPP enhanced with retrieval-augmented query generation outperforms the best-performing existing QV-based prediction approach by as much as 30% on neural ranking models such as MonoT5.

In the realm of music recommendation, sequential recommenders have shown promise in capturing the dynamic nature of music consumption. A key characteristic of this domain is repetitive listening, where users frequently replay familiar tracks. To capture these repetition patterns, recent research has introduced Personalised Popularity Scores (PPS), which quantify user-specific preferences based on historical frequency. While PPS enhances relevance in recommendation, it often reinforces already-known content, limiting the system’s ability to surface novel or serendipitous items—key elements for fostering long-term user engagement and satisfaction. To address this limitation, we build upon RecJPQ, a Transformer-based framework initially developed to improve scalability in large-item catalogues through sub-item decomposition. We repurpose RecJPQ’s sub-item architecture to model personalised popularity at a finer granularity. This allows us to capture shared repetition patterns across sub-embeddings—latent structures not accessible through item-level popularity alone. We propose a novel integration of sub-ID-level personalised popularity within the RecJPQ framework, enabling explicit control over the trade-off between accuracy and personalised novelty. Our sub-ID-level PPS method (sPPS) consistently outperforms item-level PPS by achieving significantly higher personalised novelty without compromising recommendation accuracy. Code and experiments are publicly available at https://github.com/sisinflab/Sub-id-Popularity.

Search engines often follow a pipeline architecture, where complex but effective reranking components are used to refine the results of an initial retrieval. Retrieval augmented generation (RAG) is an exciting application of the pipeline architecture, where the final component generates a coherent answer for the users from the retrieved documents. In this demo paper, we describe how such RAG pipelines can be formulated in the declarative PyTerrier architecture, and the advantages of doing so. Our PyTerrier-RAG extension for PyTerrier provides easy access to standard RAG datasets and evaluation measures, state-of-the-art LLM readers, and using PyTerrier's unique operator notation, easy-to-build pipelines. We demonstrate the succinctness of indexing and RAG pipelines on standard datasets (including Natural Questions) and how to build on the larger PyTerrier ecosystem with state-of-the-art sparse, learned-sparse, and dense retrievers, and other neural rankers.

The evaluation of Information Retrieval (IR) systems typically uses query-document pairs with corresponding human-labelled relevance assessments (qrels). These qrels are used to determine if one system is better than another based on average retrieval performance. Acquiring large volumes of human relevance assessments is expensive. Therefore, more efficient relevance assessment approaches have been proposed, necessitating comparisons between qrels to ascertain their efficacy. Discriminative power, i.e. the ability to correctly identify significant differences between systems, is important for drawing accurate conclusions on the robustness of qrels. Previous work has measured the proportion of pairs of systems that are identified as significantly different and has quantified Type I statistical errors. Type I errors lead to incorrect conclusions due to false positive significance tests. We argue that also identifying Type II errors (false negatives) is important as they lead science in the wrong direction. We quantify Type II errors and propose that balanced classification metrics, such as balanced accuracy, can be used to portray the discriminative power of qrels. We perform experiments using qrels generated using alternative relevance assessment methods to investigate measuring hypothesis testing errors in IR evaluation. We find that additional insights into the discriminative power of qrels can be gained by quantifying Type II errors, and that balanced classification metrics can be used to give an overall summary of discriminative power in one, easily comparable, number.

A large item catalogue is a major challenge for deploying modern sequential recommender models, since it makes the memory footprint of the model large and increases inference latency. One promising approach to address this is RecJPQ, which replaces item embeddings with sub-item embeddings. However, slow inference remains problematic because finding the top highest-scored items usually requires scoring all items in the catalogue, which may not be feasible for large catalogues. By adapting dynamic pruning concepts from document retrieval, we propose the RecJPQPrune dynamic pruning algorithm to efficiently find the top highest-scored items without computing the scores of all items in the catalogue. Our RecJPQPrune algorithm is safe-up-to-rank K since it theoretically guarantees that no potentially high-scored item is excluded from the final top K recommendation list, thereby ensuring no impact on effectiveness. Our experiments on two large datasets and three recommendation models demonstrate the efficiency achievable using RecJPQPrune: for instance, on the Tmall dataset with 2.2M items, we can reduce the median model scoring time by 64× compared to the Transformer Default baseline, and 5.3× compared to a recent scoring approach called PQTopK. Overall, this paper demonstrates the effective and efficient inference of Transformer-based recommendation models at catalogue scales not previously reported in the literature. Indeed, our RecJPQPrune algorithm can score 2 million items in under 10 milliseconds without GPUs, and without relying on Approximate Nearest Neighbour (ANN) techniques.

Many modern sequential recommender systems use deep neural networks, which can effectively estimate the relevance of items but require a lot of time to train. Slow training increases the costs of training, hinders product development timescales, and prevents the model from being regularly updated to adapt to changing user preferences. The training of such sequential models involves appropriately sampling past user interactions to create a realistic training objective. The existing training objectives have limitations. For instance, next item prediction never uses the beginning of the sequence as a learning target, thereby potentially discarding valuable data. However, the item masking used by the state-of-the-art BERT4Rec recommender model is only weakly related to the goal of the sequential recommendation; therefore, it requires much more time to obtain an effective model. Hence, we propose a novel Recency-based Sampling of Sequences (RSS) training objective (which is parameterized by a choice of recency importance function) that addresses both limitations. We apply our method to various recent and state-of-the-art model architectures—such as GRU4Rec, Caser, and SASRec. We show that the models enhanced with our method can achieve performances exceeding or very close to the effective BERT4Rec, but with much less training time. For example, on the MovieLens-20M dataset, RSS applied to the SASRec model can result in a 60% improvement in NDCG over a vanilla SASRec, and a 16% improvement over a fully trained BERT4Rec model, despite taking 93% less training time than BERT4Rec. We also experiment with two families of recency importance functions and show that they perform similarly. We further empirically demonstrate that RSS-enhanced SASRec successfully learns to distinguish differences between recent and older interactions—a property that the original SASRec model does not exhibit. Overall, we show that RSS is a viable (and frequently better) alternative to the existing training objectives, which is both effective and efficient for training sequential recommender model when the computational resources for training are limited.

A large catalogue size is one of the central challenges in training recommendation models: A large number of items makes them memory- and computationally inefficient to compute scores for all items during training, forcing these models to deploy negative sampling. However, negative sampling increases the proportion of positive interactions in the training data, and therefore, models trained with negative sampling tend to overestimate the probabilities of positive interactions—a phenomenon we call overconfidence . While the absolute values of the predicted scores/probabilities are not important for the ranking of retrieved recommendations, overconfident models may fail to estimate nuanced differences in the top-ranked items, resulting in degraded performance. In this article, we show that overconfidence explains why the popular SASRec model underperforms when compared to BERT4Rec. This is contrary to the BERT4Rec authors’ explanation that the difference in performance is due to the bi-directional attention mechanism. To mitigate overconfidence, we propose a novel Generalised Binary Cross-Entropy Loss function (gBCE) and theoretically prove that it can mitigate overconfidence. We further propose the gSASRec model, an improvement over SASRec that deploys an increased number of negatives and the gBCE loss. Through detailed experiments on three datasets, we show that gSASRec does not exhibit the overconfidence problem. As a result, gSASRec can outperform BERT4Rec (e.g., +9.47% NDCG on the MovieLens-1M dataset) while requiring less training time (e.g., -73% training time on MovieLens-1M). Moreover, in contrast to BERT4Rec, gSASRec is suitable for large datasets that contain more than 1 million items. Finally, we show how addressing overconfidence can improve model calibration—the ability of a model to predict actual interaction probabilities accurately. By applying gBCE to the SASRec model on MovieLens-1M dataset, we reduce the models’ expected calibration error by 98.9% (from 0.966 to 0.01).

In recent years, the rapid growth of online multimedia services, such as e-commerce platforms, has necessitated the development of personalised recommendation approaches that can encode diverse content about each item. Indeed, modern multi-modal recommender systems exploit diverse features obtained from raw images and item descriptions to enhance the recommendation performance. However, the existing multi-modal recommender systems primarily depend on the features extracted individually from different media through pre-trained modality-specific encoders, and exhibit only shallow alignments between different modalities, thereby limiting these systems’ ability to capture the underlying relationships between the modalities. In this article, we enhance the deep alignment of large multi-modal encoders to address the shallow alignment of modalities in multi-modal recommender systems. These encoders have previously demonstrated state-of-the-art effectiveness in ranking items across various domains. Specifically, we investigate the use of three state-of-the-art large multi-modal encoders – CLIP (dual-stream), VLMo and BEiT-3 (unified) – for recommendation tasks. We explore their benefits for recommendation through using a range of strategies, including the use of pre-trained and fine-tuned encoders, as well as the evaluation of the end-to-end training of these encoders. We show that pre-trained large multi-modal encoders generate more aligned and effective user/item representations compared with existing modality-specific encoders across four existing multi-modal recommendation datasets. Furthermore, we show that fine-tuning these encoders further improves the recommendation performance, with end-to-end training emerging as the most effective paradigm, significantly outperforming both pre-trained and fine-tuned encoders with an improved recommendation performance. We also demonstrate the effectiveness of large multi-modal encoders in facilitating modality alignment by evaluating the contribution of each modality separately. Finally, we show that the dual-stream approach, specifically CLIP, is the most effective architecture for these large multi-modal encoders, outperforming the unified approaches (i.e., VLMo and BEiT3) in terms of effectiveness and efficiency.

In the realm of music recommendation, sequential recommender systems have shown promise in capturing the dynamic nature of music consumption. Nevertheless, traditional Transformer-based models, such as SASRec and BERT4Rec, while effective, encounter challenges due to the unique characteristics of music listening habits. In fact, existing models struggle to create a coherent listening experience due to rapidly evolving preferences. Moreover, music consumption is characterized by a prevalence of repeated listening, i.e. users frequently return to their favourite tracks, an important signal that could be framed as individual or personalized popularity. This paper addresses these challenges by introducing a novel approach that incorporates personalized popularity information into sequential recommendation. By combining user-item popularity scores with model-generated scores, our method effectively balances the exploration of new music with the satisfaction of user preferences. Experimental results demonstrate that a Personalized Most Popular recommender, a method solely based on user-specific popularity, outperforms existing state-of-the-art models. Furthermore, augmenting Transformer-based models with personalized popularity awareness yields superior performance, showing improvements ranging from 25.2% to 69.8%. The code for this paper is available at https://github.com/sisinflab/personalized-popularity-awareness.

Transformer-based recommender systems, such as BERT4Rec or SASRec, achieve state-of-the-art results in sequential recommendation. However, it is challenging to use these models in production environments with catalogues of millions of items: scaling Transformers beyond a few thousand items is problematic for several reasons, including high model memory consumption and slow inference. In this respect, RecJPQ is a state-of-the-art method of reducing the models’ memory consumption; RecJPQ compresses item catalogues by decomposing item IDs into a small number of shared sub-item IDs. Despite reporting the reduction of memory consumption by a factor of up to 50 ×, the original RecJPQ paper did not report inference efficiency improvements over the baseline Transformer-based models. Upon analysing RecJPQ’s scoring algorithm, we find that its efficiency is limited by its use of score accumulators for each item, which prevents parallelisation. In contrast, LightRec (a non-sequential method that uses a similar idea of sub-ids) reported large inference efficiency improvements using an algorithm we call PQTopK. We show that it is also possible to improve RecJPQ-based models’ inference efficiency using the PQTopK algorithm. In particular, we speed up RecJPQ-enhanced SASRec by a factor of 4.5 × compared to the original SASRec’s inference method and by the factor of 1.56 × compared to the method implemented in RecJPQ code on a large-scale Gowalla dataset with more than million items. Further, using simulated data, we show that PQTopK remains efficient with catalogues of up to tens of millions of items, removing one of the last obstacles to using Transformer-based models in production environments with large catalogues.

Neural networks-especially those that use large, pre-trained language models-have improved search engines in various ways. Most prominently, they can estimate the relevance of a passage or document to a user's query. In this work, we depart from this direction by exploring whether neural networks can effectively predict which of a document's passages are unlikely to be relevant to any query submitted to the search engine.We refer to this query-agnostic estimation of passage relevance as a passage's quality.We find that our novel methods for estimating passage quality allow passage corpora to be pruned considerably while maintaining statistically equivalent effectiveness; our best methods can consistently prune >25% of passages in a corpora, across various retrieval pipelines. Such substantial pruning reduces the operating costs of neural search engines in terms of computing resources, power usage, and carbon footprint-both when processing queries (thanks to a smaller index size) and when indexing (lightweight models can prune low-quality passages prior to the costly dense or learned sparse encoding step). This work sets the stage for developing more advanced neural "learning-what-to-index" methods.

Cross-encoders leverage fine-grained interactions between documents and queries for effective relevance ranking. Such ranking models are typically trained to satisfy the single objective of providing relevant information to the users. However, not all information should be made available. For example, documents containing sensitive information, such as personal or confidential information, should not be returned in the search results. Sensitivity-aware search (SAS) aims to develop retrieval models that can satisfy two objectives, namely: (1) providing the user with relevant search results, while (2) ensuring that no documents that contain sensitive information are included in the ranking. In this work, we propose three novel negative sampling strategies that enable cross-encoders to be trained to satisfy the bi-objective task of SAS. Additionally, we investigate and compare with filtering sensitive documents in ranking pipelines. Our experiments on a collection labelled for sensitivity show that our proposed negative sampling strategies lead to a ~37% increase in terms of cost-sensitive nDCG (nCSDCG) for SAS.

Sequential Recommendation is a popular recommendation task that uses the order of user-item interaction to model evolving users' interests and sequential patterns in their behaviour. Current state-of-the-art Transformer-based models for sequential recommendation, such as BERT4Rec, generate sequence embeddings and compute scores for catalogue items, but the increasing catalogue size makes training these models costly. The Joint Product Quantisation (JPQ) method, originally proposed for passage retrieval, markedly reduces the size of the retrieval index with minimal effect on model effectiveness, by replacing passage embeddings with a limited number of shared sub-embeddings. This paper introduces RecJPQ, a novel adaptation of JPQ for sequential recommendations, which replaces each item embedding with a concatenation of a limited number of shared sub-embeddings and, therefore, limits the number of learnable model parameters. The main idea of RecJPQ is to split items into sub-item entities before training the main recommendation model, which is inspired by splitting words into tokens and training tokenizers in language models. We apply RecJPQ to SASRec, BERT4Rec, and GRU models on three large-scale sequential datasets. Our results showed that RecJPQ could notably reduce the model size (e.g., 48× reduction for the Gowalla dataset with no effectiveness degradation). RecJPQ can also improve model performance through a regularization effect (e.g. +0.96% NDCG@10 improvement on the Booking.com dataset). Overall, RecJPQ allows the training of state-of-the-art transformer models in industrial applications, where datasets with millions of items are common.

Recommender systems commonly suffer from the long-standing data sparsity problem where insufficient user-item interaction data limits the systems’ ability to make accurate recommendations. This problem can be alleviated using cross-domain recommendation techniques. In particular, in a cross-domain setting, knowledge sharing between domains permits improved effectiveness on the target domain. While recent cross-domain recommendation techniques used a pre-training configuration, we argue that such techniques lead to a low fine-tuning efficiency, especially when using large neural models. In recent language models, prompts have been used for parameter-efficient and time-efficient tuning of the models on the downstream tasks—these prompts represent a tunable latent vector that permits to freeze the rest of the language model’s parameters. To address the cross-domain recommendation task in an efficient manner, we propose a novel Personalised Graph Prompt-based Recommendation (PGPRec) framework, which leverages the efficiency benefits from prompt-tuning. In such a framework, we develop personalised and item-wise graph prompts based on relevant items to those items the user has interacted with. In particular, we apply Contrastive Learning to generate the pre-trained embeddings, to allow an increased generalisability in the pre-training stage, and to ensure an effective prompt-tuning stage. To evaluate the effectiveness of our PGPRec framework in a cross-domain setting, we conduct an extensive evaluation with the top- k recommendation task and perform a cold-start analysis. The obtained empirical results on four Amazon Review datasets show that our proposed PGPRec framework can reduce up to 74% of the tuned parameters with a competitive performance and achieves an 11.41% improved performance compared to the strongest baseline in a cold-start scenario.

Multi-modal interactive recommender systems (MMIRS) can effectively guide users towards their desired items through multi-turn interactions by leveraging the users’ real-time feedback (in the form of natural-language critiques) on previously recommended items (such as images of fashion products). In this scenario, the users’ preferences can be expressed by both the users’ past interests from their historical interactions and their current needs from the real-time interactions. However, it is typically challenging to make satisfactory personalised recommendations across multi-turn interactions due to the difficulty in balancing the users’ past interests and the current needs for generating the users’ state (i.e., current preferences) representations over time. However, hierarchical reinforcement learning has been successfully applied in various fields by decomposing a complex task into a hierarchy of more easily addressed subtasks. In this journal article, we propose a novel personalised multi-modal interactive recommendation model (PMMIR) using hierarchical reinforcement learning to more effectively incorporate the users’ preferences from both their past and real-time interactions. In particular, PMMIR decomposes the personalised interactive recommendation process into a sequence of two subtasks with hierarchical state representations: a first subtask where a history encoder learns the users’ past interests with the hidden states of history for providing personalised initial recommendations and a second subtask where a state tracker estimates the current needs with the real-time estimated states for updating the subsequent recommendations. The history encoder and the state tracker are jointly optimised with a single objective by maximising the users’ future satisfaction with the recommendations. Following previous work, we train and evaluate our PMMIR model using a user simulator that can generate natural-language critiques about the recommendations as a surrogate for real human users. Experiments conducted on two derived fashion datasets from two well-known public datasets demonstrate that our proposed PMMIR model yields significant improvements in comparison to the existing state-of-the-art baseline models. The datasets and code are publicly available at: https://github.com/yashonwu/pmmir

Passage ranking aims to rank a set of passages based on their relevance to a query. Current state-of-the-art models for this task typically employ a cross-encoder structure. However, these models lack access to background knowledge, i.e., information related to the query that can be helpful in retrieving relevant passages. Knowledge Graphs (KGs) provide a structured way of storing information about entities and their relationships, offering valuable background knowledge about entities. While KGs have been used to augment pretrained language models (LMs) to perform several reasoning tasks such as question answering, it remains an open question of how to utilise the information from KGs to enhance the performance of cross-encoders on the passage ranking task. Therefore, we propose KGPR, a KG-enhanced cross-encoder for the Passage Retrieval task. KGPR is built upon LUKE, an entity-aware pretrained LM, with an additional module that fuses information from KGs into LUKE. By leveraging the background knowledge from KGs, KGPR enhances the model's comprehension of queries and passages, resulting in improved ranking performance. Experimental results demonstrate that using KGs can enhance the performance of LUKE in the passage retrieval task, and KGPR can outperform state-of-the-art monoT5 cross-encoder by 3.32% and 10.77% on the MS MARCO development set and TREC DL-HARD query set respectively, using a model with a similar number of parameters.

Dense retrieval approaches are challenging the prevalence of inverted index-based sparse representation approaches for information retrieval systems. Different families have arisen: single representations for each query or passage (such as ANCE or DPR), or multiple representations (usually one per token) as exemplified by the ColBERT model. While ColBERT is effective, it requires significant storage space for each token's embedding. In this work, we aim to prune the embeddings for tokens that are not important for effectiveness. Indeed, we show that, by adapting standard uniform and document-centric static pruning methods to embedding-based indexes, but retaining their focus on low-IDF tokens, we can attain large improvements in space efficiency while maintaining high effectiveness. Indeed, on experiments conducted on the MSMARCO passage ranking task, by removing all embeddings corresponding to the 100 most frequent BERT tokens, the index size is reduced by 45%, with limited impact on effectiveness (e.g. no statistically significant degradation of NDCG@10 or MAP on the TREC 2020 queryset). Similarly, on TREC Covid, we observed a 1.3% reduction in nDCG@10 for a 38% reduction in total index size.

Interactive recommendation enables users to provide verbal and non-verbal relevance feedback (such as natural-language critiques and likes/dislikes) when viewing a ranked list of recommendations (such as images of fashion products), in order to guide the recommender system towards their desired items (i.e. goals) across multiple interaction turns. Such a multi-modal interactive recommendation (MMIR) task has been successfully formulated with deep reinforcement learning (DRL) algorithms by simulating the interactions between an environment (i.e. a user) and an agent (i.e. a recommender system). However, it is typically challenging and unstable to optimise the agent to improve the recommendation quality associated with implicit learning of multi-modal representations in an end-to-end fashion in DRL. This is known as the coupling of policy optimisation and representation learning. To address this coupling issue, we propose a novel goal-oriented multi-modal interactive recommendation model (GOMMIR) that uses both verbal and non-verbal relevance feedback to effectively incorporate the users’ preferences over time. Specifically, our GOMMIR model employs a multi-task learning approach to explicitly learn the multi-modal representations using a multi-modal composition network when optimising the recommendation agent. Moreover, we formulate the MMIR task using goal-oriented reinforcement learning and enhance the optimisation objective by leveraging non-verbal relevance feedback for hard negative sampling and providing extra goal-oriented rewards to effectively optimise the recommendation agent. Following previous work, we train and evaluate our GOMMIR model by using user simulators that can generate natural-language feedback about the recommendations as a surrogate for real human users. Experiments conducted on four well-known fashion datasets demonstrate that our proposed GOMMIR model yields significant improvements in comparison to the existing state-of-the-art baseline models.

A large catalogue size is one of the central challenges in training recommendation models: a large number of items makes them memory and computationally inefficient to compute scores for all items during training, forcing these models to deploy negative sampling. However, negative sampling increases the proportion of positive interactions in the training data, and therefore models trained with negative sampling tend to overestimate the probabilities of positive interactions – a phenomenon we call overconfidence. While the absolute values of the predicted scores/probabilities are not important for the ranking of retrieved recommendations, overconfident models may fail to estimate nuanced differences in the top-ranked items, resulting in degraded performance. In this paper, we show that overconfidence explains why the popular SASRec model underperforms when compared to BERT4Rec. This is contrary to the BERT4Rec authors’ explanation that the difference in performance is due to the bi-directional attention mechanism. To mitigate overconfidence, we propose a novel Generalised Binary Cross-Entropy Loss function (gBCE) and theoretically prove that it can mitigate overconfidence. We further propose the gSASRec model, an improvement over SASRec that deploys an increased number of negatives and the gBCE loss. We show through detailed experiments on three datasets that gSASRec does not exhibit the overconfidence problem. As a result, gSASRec can outperform BERT4Rec (e.g. +9.47% NDCG on the MovieLens-1M dataset), while requiring less training time (e.g. -73% training time on MovieLens-1M). Moreover, in contrast to BERT4Rec, gSASRec is suitable for large datasets that contain more than 1 million items.

A query performance prediction (QPP) method predicts the effectiveness of an IR system for a given query. While unsupervised approaches have been shown to work well for statistical IR models, it is likely that these approaches would yield limited effectiveness for neural ranking models (NRMs) because the retrieval scores of these models lie within a short range unlike their statistical counterparts. In this work, we propose to leverage a pairwise inference-based NRM's (specifically, DuoT5) output to accumulate evidences on the pairwise believes of one document ranked above the other. We hypothesize that the more consistent these pairwise likelihoods are, the higher is the likelihood of the retrieval to be of better quality, thus yielding a higher QPP score. We conduct our experiments on the TREC-DL dataset leveraging pairwise likelihoods from an auxiliary model DuoT5. Our experiments demonstrate that the proposed method called Pairwise Rank Preference-based QPP (QPP-PRP) leads to significantly better results than a number of standard unsupervised QPP baselines on several NRMs.

Dense multi-representation retrieval models, exemplified as ColBERT, estimate the relevance between a query and a document based on the similarity of their contextualised token-level embeddings. Indeed, by using contextualised token embeddings, dense retrieval, conducted as either exact or semantic matches, can result in increased effectiveness for both in-domain and out-of-domain retrieval tasks, indicating that it is an important model to study. However, the exact role that these semantic matches play is not yet well investigated. For instance, although tokenisation is one of the crucial design choices for various pretrained language models, its impact on the matching behaviour has not been examined in detail. In this work, we inspect the reproducibility and replicability of the contextualised late interaction mechanism by extending ColBERT to Col⋆ which implements the late interaction mechanism across various pretrained models and different types of tokenisers. As different tokenisation methods can directly impact the matching behaviour within the late interaction mechanism, we study the nature of matches occurring in different Col⋆ models, and further quantify the contribution of lexical and semantic matching on retrieval effectiveness. Overall, our experiments successfully reproduce the performance of ColBERT on various query sets, and replicate the late interaction mechanism upon different pretrained models with different tokenisers. Moreover, our experimental results yield new insights, such as: (i) semantic matching behaviour varies across different tokenisers; (ii) more specifically, high-frequency tokens tend to perform semantic matching than other token families; (iii) late interaction mechanism benefits more from lexical matching than semantic matching; (iv) special tokens, such as [CLS], play a very important role in late interaction.

Leveraging the side information associated with entities (i.e., users and items) to enhance recommendation systems has been widely recognized as an essential modeling dimension. Most of the existing approaches address this task by theintegration-based scheme, which incorporates the entity side information by combining the recommendation objective with an extra side information-aware objective. Despite the growing progress made by the existing integration-based approaches, they are largely limited by the potential conflicts between the two objectives. Moreover, the heterogeneous side information among entities is still under-explored in these systems. In this article, we propose a novelpre-training schemeto leverage the entity side information by pre-training entity embeddings using the multi-graph neural network. Instead of jointly training with two objectives, ourpre-training schemefirst pre-trains two representation models under the entity multi/single relational graphs constructed by their side information and then fine-tunes their embeddings under an existing general representation-based recommendation model. Our proposed multi-graph and single-graph neural networks can generate within-entity knowledge-encapsulated embeddings, while capturing the heterogeneity from the entity side information simultaneously, thereby improving the performance of the underlying recommendation model. An extensive evaluation of our pre-training scheme fine-tuned under four general representation-based recommender models, namely, MF, NCF, NGCF, and LightGCN, shows that effectively pre-training embeddings with both the user’s and item’s side information can significantly improve these original models in terms of both effectiveness and stability.

Pseudo-relevance feedback mechanisms, from Rocchio to the relevance models, have shown the usefulness of expanding and reweighting the users’ initial queries using information occurring in an initial set of retrieved documents, known as the pseudo-relevant set. Recently, dense retrieval – through the use of neural contextual language models such as BERT for analysing the documents’ and queries’ contents and computing their relevance scores – has shown a promising performance on several information retrieval tasks still relying on the traditional inverted index for identifying documents relevant to a query. Two different dense retrieval families have emerged: the use of single embedded representations for each passage and query, e.g., using BERT’s [CLS] token, or via multiple representations, e.g., using an embedding for each token of the query and document (exemplified by ColBERT). In this work, we conduct the first study into the potential for multiple representation dense retrieval to be enhanced using pseudo-relevance feedback and present our proposed approach ColBERT-PRF. In particular, based on the pseudo-relevant set of documents identified using a first-pass dense retrieval, ColBERT-PRF extracts the representative feedback embeddings from the document embeddings of the pseudo-relevant set. Among the representative feedback embeddings, the embeddings that most highly discriminate among documents are employed as the expansion embeddings, which are then added to the original query representation. We show that these additional expansion embeddings both enhance the effectiveness of a reranking of the initial query results as well as an additional dense retrieval operation. Indeed, experiments on the MSMARCO passage ranking dataset show that MAP can be improved by up to 26% on the TREC 2019 query set and 10% on the TREC 2020 query set by the application of our proposed ColBERT-PRF method on a ColBERT dense retrieval approach.We further validate the effectiveness of our proposed pseudo-relevance feedback technique for a dense retrieval model on MSMARCO document ranking and TREC Robust04 document ranking tasks. For instance, ColBERT-PRF exhibits up to 21% and 14% improvement in MAP over the ColBERT E2E model on the MSMARCO document ranking TREC 2019 and TREC 2020 query sets, respectively. Additionally, we study the effectiveness of variants of the ColBERT-PRF model with different weighting methods. Finally, we show that ColBERT-PRF can be made more efficient, attaining up to 4.54× speedup over the default ColBERT-PRF model, and with little impact on effectiveness, through the application of approximate scoring and different clustering methods.

Search systems often employ a re-ranking pipeline, wherein documents (or passages) from an initial pool of candidates are assigned new ranking scores. The process enables the use of highly-effective but expensive scoring functions that are not suitable for use directly in structures like inverted indices or approximate nearest neighbour indices. However, re-ranking pipelines are inherently limited by the recall of the initial candidate pool; documents that are not identified as candidates for re-ranking by the initial retrieval function cannot be identified. We propose a novel approach for overcoming the recall limitation based on the well-established clustering hypothesis. Throughout the re-ranking process, our approach adds documents to the pool that are most similar to the highest-scoring documents up to that point. This feedback process adapts the pool of candidates to those that may also yield high ranking scores, even if they were not present in the initial pool. It can also increase the score of documents that appear deeper in the pool that would have otherwise been skipped due to a limited re-ranking budget. We find that our Graph-based Adaptive Re-ranking (GAR) approach significantly improves the performance of re-ranking pipelines in terms of precision- and recall-oriented measures, is complementary to a variety of existing techniques (e.g., dense retrieval), is robust to its hyperparameters, and contributes minimally to computational and storage costs. For instance, on the MS MARCO passage ranking dataset, GAR can improve the nDCG of a BM25 candidate pool by up to 8% when applying a monoT5 ranker.

Multi-modal interactive recommendation is a type of task that allows users to receive visual recommendations and express natural-language feedback about the recommended items across multiple iterations of interactions. However, such multi-modal dialog sequences (i.e. turns consisting of the system’s visual recommendations and the user’s natural-language feedback) make it challenging to correctly incorporate the users’ preferences across multiple turns. Indeed, the existing formulations of interactive recommender systems suffer from their inability to capture the multi-modal sequential dependencies of textual feedback and visual recommendations because of their use of recurrent neural network-based (i.e., RNN-based) or transformer-based models. To alleviate the multi-modal sequential dependency issue, we propose a novel multi-modal recurrent attention network (MMRAN) model to effectively incorporate the users’ preferences over the long visual dialog sequences of the users’ natural-language feedback and the system’s visual recommendations. Specifically, we leverage a gated recurrent network (GRN) with a feedback gate to separately process the textual and visual representations of natural-language feedback and visual recommendations into hidden states (i.e. representations of the past interactions) for multi-modal sequence combination. In addition, we apply a multi-head attention network (MAN) to refine the hidden states generated by the GRN and to further enhance the model’s ability in dynamic state tracking. Following previous work, we conduct extensive experiments on the Fashion IQ Dresses, Shirts, and Tops & Tees datasets to assess the effectiveness of our proposed model by using a vision-language transformer-based user simulator as a surrogate for real human users. Our results show that our proposed MMRAN model can significantly outperform several existing state-of-the-art baseline models.

Many modern sequential recommender systems use deep neural networks, which can effectively estimate the relevance of items but require a lot of time to train. Slow training increases expenses, hinders product development timescales and prevents the model from being regularly updated to adapt to changing user preferences. Training such sequential models involves appropriately sampling past user interactions to create a realistic training objective. The existing training objectives have limitations. For instance, next item prediction never uses the beginning of the sequence as a learning target, thereby potentially discarding valuable data. On the other hand, the item masking used by BERT4Rec is only weakly related to the goal of the sequential recommendation; therefore, it requires much more time to obtain an effective model. Hence, we propose a novel Recency-based Sampling of Sequences training objective that addresses both limitations. We apply our method to various recent and state-of-the-art model architectures – such as GRU4Rec, Caser, and SASRec. We show that the models enhanced with our method can achieve performances exceeding or very close to state-of-the-art BERT4Rec, but with much less training time.

BERT4Rec is an effective model for sequential recommendation based on the Transformer architecture. In the original publication, BERT4Rec claimed superiority over other available sequential recommendation approaches (e.g. SASRec), and it is now frequently being used as a state-of-the art baseline for sequential recommendation. However, not all subsequent publications confirmed its superiority and have proposed other models that were shown to outperform BERT4Rec in effectiveness. In this paper we systematically review all publications that compare BERT4Rec with another popular Transformer-based model, namely SASRec, and show that BERT4Rec results are not consistent within these publications. To understand the reasons behind this inconsistency, we analyse the available implementations of BERT4Rec and show that we fail to reproduce results of the original BERT4Rec publication when using their default configuration parameters. However, we are able to replicate the reported results with the original code if training for a much longer amount of time (up to 30x) compared to the default configuration. We also propose our own implementation of BERT4Rec based on the HuggingFace Transformers library, which we demonstrate replicates the originally reported results on 3 out 4 datasets, while requiring up to 95% less training time to converge. Overall, from our systematic review and detailed experiments, we conclude that BERT4Rec does indeed exhibit state-of-the-art effectiveness for sequential recommendation, but only when trained for a sufficient amount of time. Additionally, we show that our implementation can further benefit from adapting other Transformer architectures that are available in the HuggingFace Transformers library (e.g. using disentangled attention, as provided by DeBERTa, or larger hidden layer size cf. ALBERT). For example, on the MovieLens-1M dataset, we demonstrate that both these models can improve BERT4Rec performance by up to 9%. Moreover, we show that an ALBERT-based BERT4Rec model achieves better performance on that dataset than state-of-the-art results reported in the most recent publications.

Graph-based recommender systems (GBRSs) have achieved promising performance by incorporating the user-item bipartite graph using the Graph Neural Network (GNN). Among GBRSs, the information from each user and item's multi-hop neighbours is effectively conveyed between nodes through neighbourhood aggregation and message passing. Although effective, existing neighbourhood information aggregation and passing functions are usually computationally expensive. Motivated by the emerging contrastive learning technique, we design a simple neighbourhood construction method in conjunction with the contrastive objective function to simulate the neighbourhood information processing of GNN. In addition, we propose a simple algorithm based on Multilayer Perceptron (MLP) for learning users and items' representations with extra non-linearity while lowering computational burden compared with multi-layers GNNs. Our extensive empirical experiments on three public datasets demonstrate that our proposed model, i.e. MLP-CGRec, can reduce the GPU memory consumption and training time by up to 24.0% and 33.1%, respectively, without significantly degenerating the recommendation accuracy in comparison with competitive baselines.

PISA (Performant Indexes and Search for Academia) provides very efficient implementations of various retrieval algorithms over sparse inverted indices. The highly-optimized C++ implementation, however, has previously only been accessible via command line tools. From indexing to retrieval, 5--6 commands need to be executed in sequence, making the process relatively involved. Further complications when using PISA include a lengthy build process and minimal interoperability with other tools. In this work, we demonstrate a new tool that provides a native Python wrapper around PISA. The wrapper features a simplified interface that adheres to the PyTerrier API, making it easy to use (e.g., via Pandas DataFrames), apply to a multitude of datasets (e.g., those from the ir_datasets package) and combine with other methods (e.g., neural re-ranking and dense retrieval methods).

Dense retrieval approaches are of increasing interest because they can better capture contextualised similarity compared to sparse retrieval models such as BM25. Among the most prominent of these approaches is TCT-ColBERT, which trains a light-weight "student'' model from a more expensive "teacher'' model. In this work, we take a closer look into TCT-ColBERT concerning its reproducibility and replicability. To structure our study, we propose a three-stage perspective on reproducing the training, inference, and evaluation of model-focused papers, each using artefacts produced from different stages in the pipeline. We find that --- perhaps as expected --- precise reproduction is more challenging when the complete training process is conducted, rather than just inference from a released trained model. Each stage provides the opportunity to perform replication and ablation experiments. We are able to replicate (i.e., produce an effective independent implementation) for model inference and dense indexing/retrieval, but are unable to replicate the training process. We conduct several ablations to cover gaps in the original paper, and make the following observations: (1) the model can function as an inexpensive re-ranker, establishing a new Pareto-optimal result; (2) the index size can be reduced by using lower-precision floating point values, but only if ties in scores are handled appropriately; (3) training needs to be conducted for the entire suggested duration to achieve optimal performance; and (4) student initialisation from the teacher is not necessary.

Recently micro-videos have become more popular in social media platforms such as TikTok and Instagram. Engagements in these platforms are facilitated by multi-modal recommendation systems. Indeed, such multimedia content can involve diverse modalities, often represented as visual, acoustic, and textual features to the recommender model. Existing works in micro-video recommendation tend to unify the multi-modal channels, thereby treating each modality with equal importance. However, we argue that these approaches are not sufficient to encode item representations with multiple modalities, since the used methods cannot fully disentangle the users' tastes on different modalities. To tackle this problem, we propose a novel learning method named Multi-Modal Graph Contrastive Learning (MMGCL), which aims to explicitly enhance multi-modal representation learning in a self-supervised learning manner. In particular, we devise two augmentation techniques to generate the multiple views of a user/item: modality edge dropout and modality masking. Furthermore, we introduce a novel negative sampling technique that allows to learn the correlation between modalities and ensures the effective contribution of each modality. Extensive experiments conducted on two micro-video datasets demonstrate the superiority of our proposed MMGCL method over existing state-of-the-art approaches in terms of both recommendation performance and training convergence speed.

Pseudo-relevance feedback (PRF) is a classical technique to improve search engine retrieval effectiveness, by closing the vocabulary gap between users’ query formulations and the relevant documents. While PRF is typically applied on the same target corpus as the final retrieval, in the past, external expansion techniques have sometimes been applied to obtain a high-quality pseudo-relevant feedback set using the external corpus . However, such external expansion approaches have only been studied for sparse (BoW) retrieval methods, and its effectiveness for recent dense retrieval methods remains under-investigated. Indeed, dense retrieval approaches such as ANCE and ColBERT, which conduct similarity search based on encoded contextualised query and document embeddings, are of increasing importance. Moreover, pseudo-relevance feedback mechanisms have been proposed to further enhance dense retrieval effectiveness. In particular, in this work, we examine the application of dense external expansion to improve zero-shot retrieval effectiveness, i.e. evaluation on corpora without further training. Zero-shot retrieval experiments with six datasets, including two TREC datasets and four BEIR datasets, when applying the MSMARCO passage collection as external corpus, indicate that obtaining external feedback documents using ColBERT can significantly improve NDCG@10 for the sparse retrieval (by upto 28%) and the dense retrieval (by upto 12%). In addition, using ANCE on the external corpus brings upto 30% NDCG@10 improvements for the sparse retrieval and upto 29% for the dense retrieval.

Abstract Providing users with relevant search results has been the primary focus of information retrieval research. However, focusing on relevance alone can lead to undesirable side effects. For example, small differences between the relevance scores of documents that are ranked by relevance alone can result in large differences in the exposure that the authors of relevant documents receive, i.e., the likelihood that the documents will be seen by searchers. Therefore, developing fair ranking techniques to try to ensure that search results are not dominated, for example, by certain information sources is of growing interest, to mitigate against such biases. In this work, we argue that generating fair rankings can be cast as a search results diversification problem across a number of assumed fairness groups, where groups can represent the demographics or other characteristics of information sources. In the context of academic search, as in the TREC Fair Ranking Track, which aims to be fair to unknown groups of authors, we evaluate three well-known search results diversification approaches from the literature to generate rankings that are fair to multiple assumed fairness groups, e.g. early-career researchers vs. highly-experienced authors. Our experiments on the 2019 and 2020 TREC datasets show that explicit search results diversification is a viable approach for generating effective rankings that are fair to information sources. In particular, we show that building on xQuAD diversification as a fairness component can result in a significant ( $$p<0.05$$ p < 0.05 ) increase (up to 50% in our experiments) in the fairness of exposure that authors from unknown protected groups receive.

Recommendation systems are often evaluated based on user’s interactions that were collected from an existing, already deployed recommendation system. In this situation, users only provide feedback on the exposed items and they may not leave feedback on other items since they have not been exposed to them by the deployed system. As a result, the collected feedback dataset that is used to evaluate a new model is influenced by the deployed system, as a form of closed loop feedback. In this article, we show that the typical offline evaluation of recommender systems suffers from the so-called Simpson’s paradox. Simpson’s paradox is the name given to a phenomenon observed when a significant trend appears in several different sub-populations of observational data but disappears or is even reversed when these sub-populations are combined together. Our in-depth experiments based on stratified sampling reveal that a very small minority of items that are frequently exposed by the deployed system plays a confounding factor in the offline evaluation of recommendation systems. In addition, we propose a novel evaluation methodology that takes into account the confounder, i.e., the deployed system’s characteristics. Using the relative comparison of many recommendation models as in the typical offline evaluation of recommender systems, and based on the Kendall rank correlation coefficient, we show that our proposed evaluation methodology exhibits statistically significant improvements of 14% and 40% on the examined open loop datasets (Yahoo! and Coat), respectively, in reflecting the true ranking of systems with an open loop (randomised) evaluation in comparison to the standard evaluation.

Many recent recommendation systems leverage the large quantity of reviews placed by users on items. However, it is both challenging and important to accurately measure the usefulness of such reviews for effective recommendation. In particular, users have been shown to exhibit distinct preferences over different types of reviews (e.g., preferring longer versus shorter or recent versus old reviews), indicating that users might differ in their viewpoints on what makes the reviews useful. Yet, there have been limited studies that account for the personalised usefulness of reviews when estimating the users’ preferences. In this article, we propose a novel neural model, called BanditProp, which addresses this gap in the literature. It first models reviews according to both their content and associated properties (e.g., length, sentiment and recency). Thereafter, it constructs a multi-task learning (MTL) framework to model the reviews’ content encoded with various properties.In such an MTL framework, each task corresponds to producing recommendations focusing on an individual property. Next, we address the selection of the features from reviews with different review properties as a bandit problem using multinomial rewards. We propose a neural contextual bandit algorithm (i.e., ConvBandit) and examine its effectiveness in comparison to eight existing bandit algorithms in addressing the bandit problem. Our extensive experiments on two well-known Amazon and Yelp datasets show that BanditProp can significantly outperform one classic and six existing state-of-the-art recommendation baselines. Moreover, BanditProp using ConvBandit consistently outperforms the use of other bandit algorithms over the two used datasets. In particular, we experimentally demonstrate the effectiveness of our proposed customised multinomial rewards in comparison to binary rewards, when addressing our bandit problem.

Dense retrieval, which describes the use of contextualised language models such as BERT to identify documents from a collection by leveraging approximate nearest neighbour (ANN) techniques, has been increasing in popularity. Two families of approaches have emerged, depending on whether documents and queries are represented by single or multiple embeddings. ColBERT, the exemplar of the latter, uses an ANN index and approximate scores to identify a set of candidate documents for each query embedding, which are then re-ranked using accurate document representations. In this manner, a large number of documents can be retrieved for each query, hindering the efficiency of the approach. In this work, we investigate the use of ANN scores for ranking the candidate documents, in order to decrease the number of candidate documents being fully scored. Experiments conducted on the MSMARCO passage ranking corpus demonstrate that, by cutting of the candidate set by using the approximate scores to only 200 documents, we can still obtain an effective ranking without statistically significant differences in effectiveness, and resulting in a 2x speedup in efficiency.

The task of adhoc search is undergoing a renaissance, sparked by advances in natural language processing. In particular, pre-trained contextualized language models (such as BERT and T5) have consistently shown to be a highly-effective foundation upon which to build ranking models. These models are equipped with a far deeper understanding of language than the capabilities of bag-of-words (BoW) models. Applying these techniques to new tasks can be tricky, however, as they require knowledge of deep learning frameworks, and significant scripting and data munging. In this full-day tutorial, we build up from foundational retrieval principles to the latest neural ranking techniques. We first provide foundational background on classical bag-of-words methods. We then show how feature-based Learning to Rank methods can be used to re-rank these results. Finally, we cover contemporary approaches, such as BERT, doc2query, and dense retrieval. Throughout the process, we demonstrate how these can be easily experimentally applied to new search tasks in a declarative style of conducting experiments exemplified by the PyTerrier and OpenNIR search toolkits.

PyTerrier is a Python-based retrieval framework for expressing simple and complex information retrieval (IR) pipelines in a declarative manner. While making use of the long-established Terrier IR platform for basic text indexing and retrieval, its salient utility comes from its expressive Python operators, which allow for individual IR operations to be pipelined and combined in different flexible manners as requested by the search application. Each operation applies a transformation upon a dataframe, while operators are defined with clear semantics in relational algebra. Going further, we have recently expanded the PyTerrier framework to include additional support for state-of-the-art BERT-based text re-rankers (such as EPIC) and dense retrieval implementations (such as ANCE and ColBERT). Transformer pipelines can be tuned and evaluated in a declarative manner. To increase the reusability of this framework as a resource for the IR community, PyTerrier provides easy access to a variety of standard benchmark datasets, including pre-built indices. Finally, we highlight the advantages of such a framework for information retrieval researchers and educators.

Recent advances in dense retrieval techniques have offered the promise of being able not just to re-rank documents using contextualised language models such as BERT, but also to use such models to identify documents from the collection in the first place. However, when using dense retrieval approaches that use multiple embedded representations for each query, a large number of documents can be retrieved for each query, hindering the efficiency of the method. Hence, this work is the first to consider efficiency improvements in the context of a dense retrieval approach (namely ColBERT), by pruning query term embeddings that are estimated not to be useful for retrieving relevant documents. Our proposed query embeddings pruning reduces the cost of the dense retrieval operation, as well as reducing the number of documents that are retrieved and hence require to be fully scored. Experiments conducted on the MSMARCO passage ranking corpus demonstrate that, when reducing the number of query embeddings used from 32 to 3 based on the collection frequency of the corresponding tokens, query embedding pruning results in no statistically significant differences in effectiveness, while reducing the number of documents retrieved by 70%. In terms of mean response time for the end-to-end to end system, this results in a 2.65x speedup.

A dialog-based interactive recommendation task is where users can express natural-language feedback when interacting with the recommender system. However, the users’ feedback, which takes the form of natural-language critiques about the recommendation at each iteration, can only allow the recommender system to obtain a partial portrayal of the users’ preferences. Indeed, such partial observations of the users’ preferences from their natural-language feedback make it challenging to correctly track the users’ preferences over time, which can result in poor recommendation performances and a less effective satisfaction of the users’ information needs when in presence of limited iterations. Reinforcement learning, in the form of a partially observable Markov decision process (POMDP), can simulate the interactions between a partially observable environment (i.e. a user) and an agent (i.e. a recommender system). To alleviate such a partial observation issue, we propose a novel dialog-based recommendation model, the Estimator-Generator-Evaluator (EGE) model, with Q-learning for POMDP, to effectively incorporate the users’ preferences over time. Specifically, we leverage an Estimator to track and estimate users’ preferences, a Generator to match the estimated preferences with the candidate items to rank the next recommendations, and an Evaluator to judge the quality of the estimated preferences considering the users’ historical feedback. Following previous work, we train our EGE model by using a user simulator which itself is trained to describe the differences between the target users’ preferences and the recommended items in natural language. Thorough and extensive experiments conducted on two recommendation datasets – addressing images of fashion products (namely dresses and shoes) – demonstrate that our proposed EGE model yields significant improvements in comparison to the existing state-of-the-art baseline models.

Biomedical literature retrieval has greatly benefited from recent advances in neural language modeling. In particular, fine-tuning pretrained contextual language models has shown impressive results in recent biomedical retrieval evaluation campaigns. Nevertheless, current approaches neglect the inherent structure available from biomedical abstracts, which are (often explicitly) organised into semantically coherent sections such as background, methods, results, and conclusions. In this paper, we investigate the suitability of leveraging biomedical abstract sections for fine-tuning pretrained contextual language models at a finer granularity. Our results on two TREC biomedical test collections demonstrate the effectiveness of the proposed structured fine-tuning regime in contrast to a standard fine-tuning that does not leverage structure. Through an ablation study, we show that models fine-tuned on individual sections are able to capture potentially useful word contexts that may be otherwise ignored by structure-agnostic models.

Many state-of-the-art recommendation systems leverage explicit item reviews posted by users by considering their usefulness in representing the users’ preferences and describing the items’ attributes. These posted reviews may have various associated properties, such as their length, their age since they were posted, or their rating of the item. However, it remains unclear how these different review properties contribute to the usefulness of their corresponding reviews in addressing the recommendation task. In particular, users show distinct preferences when considering different aspects of the reviews (i.e. properties) for making decisions about the items. Hence, it is important to model the relationship between the reviews’ properties and the usefulness of the reviews while learning the users’ preferences and the items’ attributes. In this paper, we propose to model the reviews with their associated available properties. We introduce a novel review properties-based recommendation model (RPRM) that learns which review properties are more important than others in capturing the usefulness of reviews, thereby enhancing the recommendation results. Furthermore, inspired by the users’ information adoption framework, we integrate two loss functions and a negative sampling strategy into our proposed RPRM model, to ensure that the properties of reviews are correlated with the users’ preferences. We examine the effectiveness of RPRM using the well-known Yelp and Amazon datasets. Our results show that RPRM significantly outperforms a classical and five existing state-of-the-art baselines. Moreover, we experimentally show the advantages of using our proposed loss functions and negative sampling strategy, which further enhance the recommendation performances of RPRM.

Abstract Representation learning has been widely applied in real-world recommendation systems to capture the features of both users and items. Existing grocery recommendation methods only represent each user and item by single deterministic points in a low-dimensional continuous space, which limit the expressive ability of their embeddings, resulting in recommendation performance bottlenecks. In addition, existing representation learning methods for grocery recommendation only consider the items (products) as independent entities, neglecting their other valuable side information, such as the textual descriptions and the categorical data of items. In this paper, we propose the Variational Bayesian Context-Aware Representation (VBCAR) model for grocery recommendation. VBCAR is a novel variational Bayesian model that learns distributional representations of users and items by leveraging basket context information from historical interactions. Our VBCAR model is also extendable to leverage side information by encoding contextual features into representations based on the inference encoder. We conduct extensive experiments on three real-world grocery datasets to assess the effectiveness of our model as well as the impact of different construction strategies for item side information. Our results show that our VBCAR model outperforms the current state-of-the-art grocery recommendation models while integrating item side information (especially the categorical features with the textual information of items) results in further significant performance gains. Furthermore, we demonstrate through analysis that our model is able to effectively encode similarities between product types, which we argue is the primary reason for the observed effectiveness gains.

Abstract Making use of search systems to foster learning is an emerging research trend known as search as learning. Earlier works identified result diversification as a useful technique to support learning‐oriented search, since diversification ensures a comprehensive coverage of various aspects of the queried topic in the result list. Inspired by this finding, first we define a new research problem, multidimensional result diversification, in the context of educational search. We argue that in a search engine for the education domain, it is necessary to diversify results across multiple dimensions, that is, not only for the topical aspects covered by the retrieved documents, but also for other dimensions, such as the type of the document (e.g., text, video, etc.) or its intellectual level (say, for beginners/experts). Second, we propose a framework that extends the probabilistic and supervised diversification methods to take into account the coverage of such multiple dimensions. We demonstrate its effectiveness upon a newly developed test collection based on a real‐life educational search engine. Thorough experiments based on gathered relevance annotations reveal that the proposed framework outperforms the baseline by up to 2.4%. An alternative evaluation utilizing user clicks also yields improvements of up to 2% w.r.t. various metrics.

The incompleteness of positive labels and the presence of many unlabelled instances are common problems in binary classification applications such as in review helpfulness classification. Various studies from the classification literature consider all unlabelled instances as negative examples. However, a classification model that learns to classify binary instances with incomplete positive labels while assuming all unlabelled data to be negative examples will often generate a biased classifier. In this work, we propose a novel Negative Confidence-aware Weakly Supervised approach (NCWS), which customises a binary classification loss function by discriminating the unlabelled examples with different negative confidences during the classifier's training. NCWS allows to effectively, unbiasedly identify and separate positive and negative instances after its integration into various binary classifiers from the literature, including SVM, CNN and BERT-based classifiers. We use the review helpfulness classification as a test case for examining the effectiveness of our NCWS approach. We thoroughly evaluate NCWS by using three different datasets, namely one from Yelp (venue reviews), and two from Amazon (Kindle and Electronics reviews). Our results show that NCWS outperforms strong baselines from the literature including an existing SVM-based approach (i.e. SVM-P), the positive and unlabelled learning-based approach (i.e. C-PU) and the positive confidence-based approach (i.e. P-conf) in addressing the classifier's bias problem. Moreover, we further examine the effectiveness of NCWS by using its classified helpful reviews in a state-of-the-art review-based venue recommendation model (i.e. DeepCoNN) and demonstrate the benefits of using NCWS in enhancing venue recommendation effectiveness in comparison to the baselines.

Effective methodologies for evaluating recommender systems are critical, so that different systems can be compared in a sound manner. A commonly overlooked aspect of evaluating recommender systems is the selection of the data splitting strategy. In this paper, we both show that there is no standard splitting strategy and that the selection of splitting strategy can have a strong impact on the ranking of recommender systems during evaluation. In particular, we perform experiments comparing three common data splitting strategies, examining their impact over seven state-of-the-art recommendation models on two datasets. Our results demonstrate that the splitting strategy employed is an important confounding variable that can markedly alter the ranking of recommender systems, making much of the currently published literature non-comparable, even when the same datasets and metrics are used.

The field of recommender systems has rapidly evolved over the last few years, with significant advances made due to the in-flux of deep learning techniques. However, as a result of this rapid progress, escalating barriers-to-entry for new researchers is emerging. In particular, state-of-the-art approaches have fragmented into a large number of code-bases, often requiring different input formats, pre-processing stages and evaluating with different metric packages. Hence, it is time-consuming for new researchers to reach the point of having both an effective baseline set and a sound comparative environment. As a step towards elevating this problem, we have developed BETA-Rec, an open source project for Building, Evaluating and Tuning Automated Recommender Systems. BETA-Rec aims to provide a practical data toolkit for building end-to-end recommendation systems in a standardized way. It provides means for dataset preparation and splitting using common strategies, a generalized model engine for implementing recommender models using Pytorch with 9 models available out-of-the-box, as well as a unified training, validation, tuning and testing pipeline. Furthermore, BETA-Rec is designed to be both modular and extensible, enabling new models to be quickly added to the framework. It is deployable in a wide range of environments via pre-built docker containers and supports distributed parameter tuning using Ray. In this demo, we will illustrate the deployment and use of BETA-Rec for researchers and practitioners on a number of standard recommendation datasets. The source code of the project is available at github: https://github.com/beta-team/beta-recsys.

Recommendation systems are often trained and evaluated based on users' interactions obtained through the use of an existing, already deployed, recommendation system. Hence the deployed recommendation systems will recommend some items and not others, and items will have varying levels of exposure to users. As a result, the collected feedback dataset (including most public datasets) can be skewed towards the particular items favored by the deployed model. In this manner, training new recommender systems from interaction data obtained from a previous model creates a feedback loop, i.e. a closed loop feedback. In this paper, we first introduce the closed loop feedback and then investigate the effect of closed loop feedback in both the training and offline evaluation of recommendation models, in contrast to a further exploration of the users' preferences (obtained from the randomly presented items). To achieve this, we make use of open loop datasets, where randomly selected items are presented to users for feedback. Our experiments using an open loop Yahoo! dataset reveal that there is a strong correlation between the deployed model and a new model that is trained based on the closed loop feedback. Moreover, with the aid of exploration we can decrease the effect of closed loop feedback and obtain new and better generalizable models.

There exists a natural tension between encouraging a diverse ecosystem of open-source search engines and supporting fair, replicable comparisons across those systems. To balance these two goals, we examine two approaches to providing interoperability between the inverted indexes of several systems. The first takes advantage of internal abstractions around index structures and building wrappers that allow one system to directly read the indexes of another. The second involves sharing indexes across systems via a data exchange specification that we have developed, called the Common Index File Format (CIFF). We demonstrate the first approach with the Java systems Anserini and Terrier, and the second approach with Anserini, JASSv2, OldDog, PISA, and Terrier. Together, these systems provide a wide range of implementations and features, with different research goals. Overall, we recommend CIFF as a low-effort approach to support independent innovation while enabling the types of fair evaluations that are critical for driving the field forward.

The users' historical interactions usually contain their interests and purchase habits based on which personalised recommendations can be made. However, such user interactions are often sparse, leading to the well-known cold-start problem when a user has no or very few interactions. In this paper, we propose a new recommendation model, named Heterogeneous Graph Neural Recommender (HGNR), to tackle the cold-start problem while ensuring effective recommendations for all users. Our HGNR model learns users and items' embeddings by using the Graph Convolutional Network based on a heterogeneous graph, which is constructed from user-item interactions, social links and semantic links predicted from the social network and textual reviews. Our extensive empirical experiments on three public datasets demonstrate that HGNR significantly outperforms competitive baselines in terms of the Normalised Discounted Cumulative Gain and Hit Ratio measures.

Active learning strategies are often deployed in technology-assisted review tasks, such as e-discovery and sensitivity review, to learn a classifier that can assist the reviewers with their task. In particular, an active learning strategy selects the documents that are expected to be the most useful for learning an effective classifier, so that these documents can be reviewed before the less useful ones. However, when reviewing for sensitivity, the order in which the documents are reviewed can impact on the reviewers' ability to perform the review. Therefore, when deploying active learning in technology-assisted sensitivity review, we want to know when a sufficiently effective classifier has been learned, such that the active learning can stop and the reviewing order of the documents can be selected by the reviewer instead of the classifier. In this work, we propose two active learning stopping strategies for technology-assisted sensitivity review. We evaluate the effectiveness of our proposed approaches in comparison with three state-of-the-art stopping strategies from the literature. We show that our best performing approach results in a significantly more effective sensitivity classifier (+6.6% F2) than the best performing stopping strategy from the literature (McNemar's test, p<0.05).

Government documents must be manually reviewed to identify any sensitive information, e.g., confidential information, before being publicly archived. However, human-only sensitivity review is not practical for born-digital documents due to, for example, the volume of documents that are to be reviewed. In this work, we conduct a user study to evaluate the effectiveness of sensitivity classification for assisting human sensitivity reviewers. We evaluate how the accuracy and confidence levels of sensitivity classification affects the number of documents that are correctly judged as being sensitive (reviewer accuracy) and the time that it takes to sensitivity review a document (reviewing speed). In our within-subject study, the participants review government documents to identify real sensitivities while being assisted by three sensitivity classification treatments , namely None (no classification predictions), Medium (sensitivity predictions from a simulated classifier with a balanced accuracy (BAC) of 0.7), and Perfect (sensitivity predictions from a classifier with an accuracy of 1.0). Our results show that sensitivity classification leads to significant improvements (ANOVA, p < 0.05) in reviewer accuracy in terms of BAC (+37.9% Medium , +60.0% Perfect ) and also in terms of F 2 (+40.8% Medium , +44.9% Perfect ). Moreover, we show that assisting reviewers with sensitivity classification predictions leads to significantly increased (ANOVA, p < 0.05) mean reviewing speeds (+72.2% Medium , +61.6% Perfect ). We find that reviewers do not agree with the classifier significantly more as the classifier’s confidence increases. However, reviewing speed is significantly increased when the reviewers agree with the classifier (ANOVA, p < 0.05). Our in-depth analysis shows that when the reviewers are not assisted with sensitivity predictions, mean reviewing speeds are 40.5% slower for sensitive judgements compared to not-sensitive judgements. However, when the reviewers are assisted with sensitivity predictions, the difference in reviewing speeds between sensitive and not-sensitive judgements is reduced by ˜10%, from 40.5% to 30.8%. We also find that, for sensitive judgements, sensitivity classification predictions significantly increase mean reviewing speeds by 37.7% when the reviewers agree with the classifier’s predictions ( t -test, p < 0.05). Overall, our findings demonstrate that sensitivity classification is a viable technology for assisting human reviewers with the sensitivity review of digital documents.

Predicting the query latency by a search engine has important benefits, for instance, in allowing the search engine to adjust its configuration to address long-running queries without unnecessarily sacrificing its effectiveness. However, for the dynamic pruning techniques that underlie many commercial search engines, achieving accurate predictions of query latencies is difficult. We propose the use of index synopses—which are stochastic samples of the full index—for attaining accurate timing predictions. Indeed, we experiment using the TREC ClueWeb09 collection, and a large set of real user queries, and find that using small index synopses it is possible to very accurately estimate properties of the larger index, including sizes of posting list unions and intersections. Thereafter, we demonstrate that index synopses facilitate two key use cases: first, for query efficiency prediction, we show that predicting the query latencies on the full index and classifying long-running queries can be accurately achieved using index synopses; second, for query performance prediction, we show that the effectiveness of queries can be estimated more accurately using a synopsis index post-retrieval predictor than a pre-retrieval predictor. Overall, our experiments demonstrate the value of such a stochastic sample of a larger index at predicting the properties of the larger index.

All government documents that are released to the public must first be manually reviewed to identify and protect any sensitive information, e.g. confidential information. However, the unassisted manual sensitivity review of born-digital documents is not practical due to, for example, the volume of documents that are created. Previous work has shown that sensitivity classification can be effective for predicting if a document contains sensitive information. However, since all of the released documents must be manually reviewed, it is important to know if sensitivity classification can assist sensitivity reviewers in making their sensitivity judgements. Hence, in this paper, we conduct a digital sensitivity review user study, to investigate if the accuracy of sensitivity classification effects the number of documents that a reviewer correctly judges to be sensitive or not (reviewer accuracy) and the time that it takes to sensitivity review a document (reviewing speed). Our results show that providing reviewers with sensitivity classification predictions, from a classifier that achieves 0.7 Balanced Accuracy, results in a 38% increase in mean reviewer accuracy and an increase of 72% in mean reviewing speeds, compared to when reviewers are not provided with predictions. Overall, our findings demonstrate that sensitivity classification is a viable technology for assisting with the sensitivity review of born-digital government documents.

Nowadays, everyone can create and publish news and information anonymously online. However, the credibility of such news and information are not guaranteed. To differentiate fake news from genuine news, one can compare a recent news with earlier posted ones. Identified suspicious news can be debunked to stop the fake news from spreading further. In this paper, we investigate the advantages of recurrent neural networks-based language representations (e.g., BERT, BiLSTM) in order to build ensemble classifiers that can accurately predict if one news title is related to, and, additionally disagrees with an earlier news title. Our experiments, on a dataset of 321k news titles created for the WSDM 2019 challenge, show that the BERT-based models significantly outperform BiLSTM, which in-turn significantly outperforms a simpler embedding-based representation. Furthermore, even the state-of-the-art BERT approach can be enhanced when combined with a simple BM25 feature.

Several previous approaches attempted to predict bursty topics on Twitter. Such approaches have usually reported that the time information (e.g. the topic popularity over time) of hashtag topics contribute the most to the prediction of bursty topics. In this paper, we propose a novel approach to use time features to predict bursty topics on Twitter. We model the popularity of topics as density curves described by the density function of a beta distribution with different parameters. We then propose various approaches to predict/classify the bursty topics by estimating the parameters of topics, using estimators such as Gradient Decent or Likelihood Maximization. In our experiments, we show that the estimated parameters of topics have a positive effect on classifying bursty topics. In particular, our estimators when combined together improve the bursty topic classification by 6.9 in terms of micro F1 compared to a baseline classifier using hashtag content features.

Fine-grained geolocation of tweets has become an important feature for reliably performing a wide range of tasks such as real-time event detection, topic detection or disaster and emergency analysis. Recent work adopted a ranking approach to return a predicted location based on content-based similarity to already available individual geotagged tweets. However, this work made use of the IDF weighting model to compute the ranking, which can diminish the quality of the Top-N retrieved tweets. In this work, we adopt a learning to rank approach towards improving the effectiveness of the ranking and increasing the accuracy of fine-grained geolocalisation. To this end we propose a set of features extracted from pairs of geotagged tweets generated within the same fine-grained geographical area (squared areas of size 1 km). Using geotagged tweets from two cities (Chicago and New York, USA), our experimental results show that our learning to rank approach significantly outperforms previous work based on IDF ranking, and improves accuracy of tweet geolocalisation at a fine-grained level.

Experimentation using IR systems has traditionally been a procedural and laborious process. Queries must be run on an index, with any parameters of the retrieval models suitably tuned. With the advent of learning-to-rank, such experimental processes (including the appropriate folding of queries to achieve cross-fold validation) have resulted in complicated experimental designs and hence scripting. At the same time, machine learning platforms such as Scikit Learn and Apache Spark have pioneered the notion of an experimental pipeline , which naturally allows a supervised classification experiment to be expressed a series of stages, which can be learned or transformed. In this demonstration, we detail Terrier-Spark, a recent adaptation to the Terrier Information Retrieval platform which permits it to be used within the experimental pipelines of Spark. We argue that this (1) provides an agile experimental platform for information retrieval, comparable to that enjoyed by other branches of data science; (2) aids research reproducibility in information retrieval by facilitating easily-distributable notebooks containing conducted experiments; and (3) facilitates the teaching of information retrieval experiments in educational environments.

Venue recommendation systems aim to effectively rank a list of interesting venues users should visit based on their historical feedback (e.g. checkins). Such systems are increasingly deployed by Location-based Social Networks (LBSNs) such as Foursquare and Yelp to enhance their usefulness to users. Recently, various RNN architectures have been proposed to incorporate contextual information associated with the users' sequence of checkins (e.g. time of the day, location of venues) to effectively capture the users' dynamic preferences. However, these architectures assume that different types of contexts have an identical impact on the users' preferences, which may not hold in practice. For example, an ordinary context such as the time of the day reflects the user's current contextual preferences, whereas a transition context - such as a time interval from their last visited venue - indicates a transition effect from past behaviour to future behaviour. To address these challenges, we propose a novel Contextual Attention Recurrent Architecture (CARA) that leverages both sequences of feedback and contextual information associated with the sequences to capture the users' dynamic preferences. Our proposed recurrent architecture consists of two types of gating mechanisms, namely 1) a contextual attention gate that controls the influence of the ordinary context on the users' contextual preferences and 2) a time- and geo-based gate that controls the influence of the hidden state from the previous checkin based on the transition context. Thorough experiments on three large checkin and rating datasets from commercial LBSNs demonstrate the effectiveness of our proposed CARA architecture by significantly outperforming many state-of-the-art RNN architectures and factorisation approaches.

The development of automatic systems that can produce timeline summaries by filtering high-volume streams of text documents, retaining only those that are relevant to a particular information need (e.g. topic or event), remains a very challenging task. To advance the field of automatic timeline generation, robust and reproducible evaluation methodologies are needed. To this end, several evaluation metrics and labeling methodologies have recently been developed - focusing on information nugget or cluster-based ground truth representations, respectively. These methodologies rely on human assessors manually mapping timeline items (e.g. tweets) to an explicit representation of what information a 'good' summary should contain. However, while these evaluation methodologies produce reusable ground truth labels, prior works have reported cases where such labels fail to accurately estimate the performance of new timeline generation systems due to label incompleteness. In this paper, we first quantify the extent to which timeline summary ground truth labels fail to generalize to new summarization systems, then we propose and evaluate new automatic solutions to this issue. In particular, using a depooling methodology over 21 systems and across three high-volume datasets, we quantify the degree of system ranking error caused by excluding those systems when labeling. We show that when considering lower-effectiveness systems, the test collections are robust (the likelihood of systems being miss-ranked is low). However, we show that the risk of systems being miss-ranked increases as the effectiveness of systems held-out from the pool increases. To reduce the risk of miss-ranking systems, we also propose two different automatic ground truth label expansion techniques. Our results show that our proposed expansion techniques can be effective for increasing the robustness of the TREC-TS test collections, markedly reducing the number of miss-rankings by up to 50% on average among the scenarios tested.

Typically, techniques that benefit effectiveness of information retrieval (IR) systems have a negative impact on efficiency. Yet, with the large scale of Web search engines, there is a need to deploy efficient query processing techniques to reduce the cost of the infrastructure required. This tutorial aims to provide a detailed overview of the infrastructure of an IR system devoted to the efficient yet effective processing of user queries. This tutorial guides the attendees through the main ideas, approaches and algorithms developed in the last 30 years in query processing. In particular, we illustrate, with detailed examples and simplified pseudo-code, the most important query processing strategies adopted in major search engines, with a particular focus on dynamic pruning techniques. Moreover, we present and discuss the state-of-the-art innovations in query processing, such as impact-sorted and blockmax indexes. We also describe how modern search engines exploit such algorithms with learning-to-rank (LtR) models to produce effective results, exploiting new approaches in LtR query processing. Finally, this tutorial introduces query efficiency predictors for dynamic pruning, and discusses their main applications to scheduling, routing, selective processing and parallelisation of query processing, as deployed by a major search engine.

During major events, such as emergencies and disasters, a large volume of information is reported on newswire and social media platforms. Temporal summarization (TS) approaches are used to automatically produce concise overviews of such events by extracting text snippets from related articles over time. Current TS approaches rely on a combination of event relevance and textual novelty for snippet selection. However, for events that span multiple days, textual novelty is often a poor criterion for selecting snippets, since many snippets are textually unique but are semantically redundant or non-informative. In this article, we propose a framework for the diversification of snippets using explicit event aspects, building on recent works in search result diversification. In particular, we first propose two techniques to identify explicit aspects that a user might want to see covered in a summary for different types of event. We then extend a state-of-the-art explicit diversification framework to maximize the coverage of these aspects when selecting summary snippets for unseen events. Through experimentation over the TREC TS 2013, 2014, and 2015 datasets, we show that explicit diversification for temporal summarization significantly outperforms classical novelty-based diversification, as the use of explicit event aspects reduces the amount of redundant and off-topic snippets returned, while also increasing summary timeliness.

Large popular events are nowadays well reflected in social media fora (e.g. Twitter), where people discuss their interest in participating in the events. In this paper we propose to exploit the content of non-geotagged posts in social media to build machine-learned classifiers able to infer users' attendance of large events in three temporal periods: before, during and after an event. The categories of features used to train the classifier reflect four different dimensions of social media: textual, temporal, social, and multimedia content. We detail the approach followed to design the feature space and report on experiments conducted on two large music festivals in the UK, namely the VFestival and Creamfields events. Our attendance classifier attains very high accuracy with the highest result observed for the Creamfields dataset ~87% accuracy to classify users that will participate in the event.

Global events such as terrorist attacks are commented upon in social media, such as Twitter, in different languages and from different parts of the world. Most prior studies have focused on monolingual sentiment analysis, and therefore excluded an extensive proportion of the Twitter userbase. In this paper, we perform a multilingual comparative sentiment analysis study on the terrorist attack in Paris, during November 2015. In particular, we look at targeted sentiment, investigating opinions on specific entities, not simply the general sentiment of each tweet. Given the potentially inflammatory and polarizing effect that these types of tweets may have on attitudes, we examine the sentiments expressed about different targets and explore whether disproportionate reaction was expressed about such targets across different languages. Specifically, we assess whether the sentiment for French speaking Twitter users during the Paris attack differs from English-speaking ones. We identify disproportionately negative attitudes in the English dataset over the French one towards some entities and, via a crowdsourcing experiment, illustrate that this also extends to forming an annotator bias.

Both politicians and citizens are increasingly embracing social media as a means to disseminate information and comment on various topics, particularly during significant political events, such as elections. Such commentary during elections is also of interest to social scientists and pollsters. To facilitate the study of social media during elections, there is a need to automatically identify posts that are topically related to those elections. However, current studies have focused on elections within English-speaking regions, and hence the resultant election content classifiers are only applicable for elections in countries where the predominant language is English. On the other hand, as social media is becoming more prevalent worldwide, there is an increasing need for election classifiers that can be generalised across different languages, without building a training dataset for each election. In this paper, based upon transfer learning, we study the development of effective and reusable election classifiers for use on social media across multiple languages. We combine transfer learning with different classifiers such as Support Vector Machines (SVM) and state-of-the-art Convolutional Neural Networks (CNN), which make use of word embedding representations for each social media post. We generalise the learned classifier models for cross-language classification by using a linear translation approach to map the word embedding vectors from one language into another. Experiments conducted over two election datasets in different languages show that without using any training data from the target language, linear translations outperform a classical transfer learning approach, namely Transfer Component Analysis (TCA), by 80% in recall and 25% in F1 measure.

Venue recommendation is an important application for Location-Based Social Networks (LBSNs), such as Yelp, and has been extensively studied in recent years. Matrix Factorisation (MF) is a popular Collaborative Filtering (CF) technique that can suggest relevant venues to users based on an assumption that similar users are likely to visit similar venues. In recent years, deep neural networks have been successfully applied to tasks such as speech recognition, computer vision and natural language processing. Building upon this momentum, various approaches for recommendation have been proposed in the literature to enhance the effectiveness of MF-based approaches by exploiting neural network models such as: word embeddings to incorporate auxiliary information (e.g. textual content of comments); and Recurrent Neural Networks (RNN) to capture sequential properties of observed user-venue interactions. However, such approaches rely on the traditional inner product of the latent factors of users and venues to capture the concept of collaborative filtering, which may not be sufficient to capture the complex structure of user-venue interactions. In this paper, we propose a Deep Recurrent Collaborative Filtering framework (DRCF) with a pairwise ranking function that aims to capture user-venue interactions in a CF manner from sequences of observed feedback by leveraging Multi-Layer Perception and Recurrent Neural Network architectures. Our proposed framework consists of two components: namely Generalised Recurrent Matrix Factorisation (GRMF) and Multi-Level Recurrent Perceptron (MLRP) models. In particular, GRMF and MLRP learn to model complex structures of user-venue interactions using element-wise and dot products as well as the concatenation of latent factors. In addition, we propose a novel sequence-based negative sampling approach that accounts for the sequential properties of observed feedback and geographical location of venues to enhance the quality of venue suggestions, as well as alleviate the cold-start users problem. Experiments on three large checkin and rating datasets show the effectiveness of our proposed framework by outperforming various state-of-the-art approaches.

Recommending a ranked list of interesting venues to users based on their preferences has become a key functionality in Location-Based Social Networks (LBSNs) such as Yelp and Gowalla. Bayesian Personalised Ranking (BPR) is a popular pairwise recommendation technique that is used to generate the ranked list of venues of interest to a user, by leveraging the user's implicit feedback such as their check-ins as instances of positive feedback, while randomly sampling other venues as negative instances. To alleviate the sparsity that affects the usefulness of recommendations by BPR for users with few check-ins, various approaches have been proposed in the literature to incorporate additional sources of information such as the social links between users, the textual content of comments, as well as the geographical location of the venues. However, such approaches can only readily leverage one source of additional information for negative sampling. Instead, we propose a novel Personalised Ranking Framework with Multiple sampling Criteria (PRFMC) that leverages both geographical influence and social correlation to enhance the effectiveness of BPR. In particular, we apply a multi-centre Gaussian model and a power-law distribution method, to capture geographical influence and social correlation when sampling negative venues, respectively. Finally, we conduct comprehensive experiments using three large-scale datasets from the Yelp, Gowalla and Brightkite LBSNs. The experimental results demonstrate the effectiveness of fusing both geographical influence and social correlation in our proposed PRFMC framework and its superiority in comparison to BPR-based and other similar ranking approaches. Indeed, our PRFMC approach attains a 37% improvement in MRR over a recently proposed approach that identifies negative venues only from social links.

To enhance effectiveness, a user's query can be rewritten internally by the search engine in many ways, for example by applying proximity, or by expanding the query with related terms. However, approaches that benefit effectiveness often have a negative impact on efficiency, which has impacts upon the user satisfaction, if the query is excessively slow. In this paper, we propose a novel framework for using the predicted execution time of various query rewritings to select between alternatives on a per-query basis, in a manner that ensures both effectiveness and efficiency. In particular, we propose the prediction of the execution time of ephemeral (e.g., proximity) posting lists generated from uni-gram inverted index posting lists, which are used in establishing the permissible query rewriting alternatives that may execute in the allowed time. Experiments examining both the effectiveness and efficiency of the proposed approach demonstrate that a 49% decrease in mean response time (and 62% decrease in 95th-percentile response time) can be attained without significantly hindering the effectiveness of the search engine.

Freedom of Information (FOI) laws legislate that government documents should be opened to the public. However, many government documents contain sensitive information, such as confidential information, that is exempt from release. Therefore, government documents must be sensitivity reviewed prior to release, to identify and close any sensitive information. With the adoption of born-digital documents, such as email, there is a need for automatic sensitivity classification to assist digital sensitivity review. SVM classifiers and Part-of-Speech sequences have separately been shown to be promising for sensitivity classification. However, sequence classification methodologies, and specifically SVM kernel functions, have not been fully investigated for sensitivity classification. Therefore, in this work, we present an evaluation of five SVM kernel functions for sensitivity classification using POS sequences. Moreover, we show that an ensemble classifier that combines POS sequence classification with text classification can significantly improve sensitivity classification effectiveness (+6.09% F2) compared with a text classification baseline, according to McNemar's test of significance.

Venue recommendation is an important capability of Location-Based Social Networks such as Yelp and Foursquare. Matrix Factorisation (MF) is a collaborative filtering-based approach that can effectively recommend venues that are relevant to the users' preferences, by training upon either implicit or explicit feedbacks (e.g. check-ins or venue ratings) that these users express about venues. However, MF suffers in that users may only have rated very few venues. To alleviate this problem, recent literature have leveraged additional sources of evidence, e.g. using users' social friendships to reduce the complexity of - or regularise - the MF model, or identifying similar venues based on their comments. This paper argues for a combined regularisation model, where the venues suggested for a user are influenced by friends with similar tastes (as defined by their comments). We propose a MF regularisation technique that seamlessly incorporates both social network information and textual comments, by exploiting word embeddings to estimate a semantic similarity of friends based on their explicit textual feedback, to regularise the complexity of the factorised model. Experiments on a large existing dataset demonstrate that our proposed regularisation model is promising, and can enhance the prediction accuracy of several state-of-the-art matrix factorisation-based approaches.

A robust retrieval system ensures that user experience is not damaged by the presence of poorly-performing queries. Such robustness can be measured by risk-sensitive evaluation measures, which assess the extent to which a system performs worse than a given baseline system. However, using a particular, single system as the baseline suffers from the fact that retrieval performance highly varies among IR systems across topics. Thus, a single system would in general fail in providing enough information about the real baseline performance for every topic under consideration, and hence it would in general fail in measuring the real risk associated with any given system. Based upon the Chi-squared statistic, we propose a new measure ZRisk that exhibits more promise since it takes into account multiple baselines when measuring risk, and a derivative measure called GeoRisk, which enhances ZRisk by also taking into account the overall magnitude of effectiveness. This paper demonstrates the benefits of ZRisk and GeoRisk upon TREC data, and how to exploit GeoRisk for risk-sensitive learning to rank, thereby making use of multiple baselines within the learning objective function to obtain effective yet risk-averse/robust ranking systems. Experiments using 10,000 topics from the MSLR learning to rank dataset demonstrate the efficacy of the proposed Chi-square statistic-based objective function.

Topic modelling approaches help scholars to examine the topics discussed in a corpus. Due to the popularity of Twitter, two distinct methods have been proposed to accommodate the brevity of tweets: the tweet pooling method and Twitter LDA. Both of these methods demonstrate a higher performance in producing more interpretable topics than the standard Latent Dirichlet Allocation (LDA) when applied on tweets. However, while various metrics have been proposed to estimate the coherence of the generated topics from tweets, the coherence of the top ranked topics, those that are most likely to be examined by users, has not been investigated. In addition, the effect of the number of generated topics K on the topic coherence scores has not been studied. In this paper, we conduct large-scale experiments using three topic modelling approaches over two Twitter datasets, and apply a state-of-the-art coherence metric to study the coherence of the top ranked topics and how K affects such coherence. Inspired by ranking metrics such as precision at n, we use coherence at n to assess the coherence of a topic model. To verify our results, we conduct a pairwise user study to obtain human preferences over topics. Our findings are threefold: we find evidence that Twitter LDA outperforms both LDA and the tweet pooling method because the top ranked topics it generates have more coherence; we demonstrate that a larger number of topics (K) helps to generate topics with more coherence; and finally, we show that coherence at n is more effective when evaluating the coherence of a topic model than the average coherence score.

Scholars often seek to understand topics discussed on Twitter using topic modelling approaches. Several coherence metrics have been proposed for evaluating the coherence of the topics generated by these approaches, including the pre-calculated Pointwise Mutual Information (PMI) of word pairs and the Latent Semantic Analysis (LSA) word representation vectors. As Twitter data contains abbreviations and a number of peculiarities (e.g. hashtags), it can be challenging to train effective PMI data or LSA word representation. Recently, Word Embedding (WE) has emerged as a particularly effective approach for capturing the similarity among words. Hence, in this paper, we propose new Word Embedding-based topic coherence metrics. To determine the usefulness of these new metrics, we compare them with the previous PMI/LSA-based metrics. We also conduct a large-scale crowdsourced user study to determine whether the new Word Embedding-based metrics better align with human preferences. Using two Twitter datasets, our results show that the WE-based metrics can capture the coherence of topics in tweets more robustly and efficiently than the PMI/LSA-based ones.

Social media has great potential as a means to enable civil protection and law enforcement agencies to more effectively tackle disasters and emergencies. However, there is currently a lack of tools that enable civil protection agencies to easily make use of social media. The Emergency Analysis Identification and Management System (EAIMS) is a prototype service that provides real-time detection of emergency events, related information finding and credibility analysis tools for use over social media during emergencies. This system exploits machine learning over data gathered from past emergencies and disasters to build effective models for identifying new events as they occur, tracking developments within those events and analyzing those developments for the purposes of enhancing the decision making processes of emergency response agencies.

In this paper we show how word embeddings can be used to increase the effectiveness of a state-of-the art Locality Sensitive Hashing (LSH) based first story detection (FSD) system over a standard tweet corpus. Vocabulary mismatch, in which related tweets use different words, is a serious hindrance to the effectiveness of a modern FSD system. In this case, a tweet could be flagged as a first story even if a related tweet, which uses different but synonymous words, was already returned as a first story. In this work, we propose a novel approach to mitigate this problem of lexical variation, based on tweet expansion. In particular, we propose to expand tweets with semantically related paraphrases identified via automatically mined word embeddings over a background tweet corpus. Through experimentation on a large data stream comprised of 50 million tweets, we show that FSD effectiveness can be improved by 9.5% over a state-of-the-art FSD system.

Sensitive documents are those that cannot be made public, e.g., for personal or organizational privacy reasons. For instance, documents requested through Freedom of Information mechanisms must be manually reviewed for the presence of sensitive information before their actual release. Hence, tools that can assist human reviewers in spotting sensitive information are of great value to government organizations subject to Freedom of Information laws. We look at sensitivity identification in terms of semi-automated text classification (SATC), the task of ranking automatically classified documents so as to optimize the cost-effectiveness of human post-checking work. We use a recently proposed utility-theoretic approach to SATC that explicitly optimizes the chosen effectiveness function when ranking the documents by sensitivity; this is especially useful in our case, since sensitivity identification is a recall-oriented task, thus requiring the use of a recall-oriented evaluation measure such as F2. We show the validity of this approach by running experiments on a multi-label multi-class dataset of government documents manually annotated according to different types of sensitivity.

Location-based social networks (LBSNs), such as Foursquare, fostered the emergence of new tasks such as recommending venues a user might wish to visit. In the literature, recommending venues has typically been addressed using user-centric recommendation approaches relying on collaborative filtering techniques. Such approaches not only require many users with detailed profiles to be effective, but they also cannot recommend venues to users who are not actually members of the LBSN. In contrast, in this paper, we introduce a venue-centric yet personalised probabilistic approach that suggests personalised and popular venues for users to visit in the near future. In our approach, we probabilistically incorporate two components, a popularity component for predicting the popularity of a venue at a given point in time, as estimated from the attendance of the venue in the LBSN (i.e. number of check-ins), and a personalisation component for identifying its interestingness with respect to the estimated preferences of the user. The popularity of each venue is predicted using time series forecasting models that are trained on the recent attendance trends of the venue, while the users' interests are modelled from the entity pages that they like on Facebook. Using three major cities, we conduct a user study to evaluate the effectiveness of the two components of our approach in suggesting venues for different types of users at different times of the day. Our experimental results show that an approach that combines the popularity and personalisation components is able to consistently outperform the recommendation service of the leading Foursquare LBSN. We also find that combining popularity and personalisation is effective for both new visitors and residents, while former visitors prefer popular venues.

Interleaving is an online evaluation method that compares two ranking functions by mixing their results and interpreting the users' click feedback. An important property of an interleaving method is its sensitivity, i.e. the ability to obtain reliable comparison outcomes with few user interactions. Several methods have been proposed so far to improve interleaving sensitivity, which can be roughly divided into two areas: (a) methods that optimize the credit assignment function (how the click feedback is interpreted), and (b) methods that achieve higher sensitivity by controlling the interleaving policy (how often a particular interleaved result page is shown).

Dealing with the medical terminology is a challenge when searching for patients based on the relevance of their medical records towards a given query. Existing work used query expansion (QE) to extract expansion terms from different document collections to improve query representation. However, the usefulness of particular document collections for QE was not measured and taken into account during retrieval. In this work, we investigate two automatic approaches that measure and leverage the usefulness of document collections when exploiting multiple document collections to improve query representation. These two approaches are based on resource selection and learning to rank techniques, respectively. We evaluate our approaches using the TREC Medical Records track's test collection. Our results show the potential of the proposed approaches, since they can effectively exploit 14 different document collections, including both domain-specific (e.g. MEDLINE abstracts) and generic (e.g. blogs and webpages) collections, and significantly outperform existing effective baselines, including the best systems participating at the TREC Medical Records track. Our analysis shows that the different collections are not equally useful for QE, while our two approaches can automatically weight the usefulness of expansion terms extracted from different document collections effectively.

Urban Informatics aims to exploit the large quantities of information produced by modern cities in order to gain insights into how they function. These insights lay the foundation for improving the lives of citizens, by improving the efficacy and efficiency of public services, and satisfying complex information needs arising within this context. The goal of the workshop is to provide a multidisciplinary forum which brings together researchers in Big Data (BD), Information Retrieval (IR), Data Mining, and Urban Studies, to explore novel solutions to the numerous theoretical, practical and ethical challenges arising in this context. These include difficulties in collecting city data, creating data management infrastructures, and providing new effective and efficient information access techniques to as many users as possible in the context of a smart city. To foster the development of new BD and IR approaches in Urban Informatics, the workshop makes available a representative dataset of city data, including Internet-based visual (Flickr) and textual (Tweets and News) media collections. The workshop provides enormous opportunities for data scientists who wish to understand the complexities of working with city data, conduct innovative research within Urban Informatics, and build a long-term community in this emerging research area.

In this paper, we study the emerging Information Retrieval (IR) task of local event retrieval using sensor metadata streams. Sensor metadata streams include information such as the crowd density from video processing, audio classifications, and social media activity. We propose to use these metadata streams to identify the topics of local events within a city, where each event topic corresponds to a set of terms representing a type of events such as a concert or a protest. We develop a supervised approach that is capable of mapping sensor metadata observations to an event topic. In addition to using a variety of sensor metadata observations about the current status of the environment as learning features, our approach incorporates additional background features to model cyclic event patterns. Through experimentation with data collected from two locations in a major Spanish city, we show that our approach markedly outperforms an alternative baseline. We also show that modelling background information improves event topic identification.

Web search engine companies require power-hungry data centers with thousands of servers to efficiently perform searches on a large scale. This permits the search engines to serve high arrival rates of user queries with low latency, but poses economical and environmental concerns due to the power consumption of the servers. Existing power saving techniques sacrifice the raw performance of a server for reduced power absorption, by scaling the frequency of the server's CPU according to its utilization. For instance, current Linux kernels include frequency governors i.e., mechanisms designed to dynamically throttle the CPU operational frequency. However, such general-domain techniques work at the operating system level and have no knowledge about the querying operations of the server. In this work, we propose to delegate CPU power management to search engine-specific governors. These can leverage knowledge coming from the querying operations, such as the query server utilization and load. By exploiting such additional knowledge, we can appropriately throttle the CPU frequency thereby reducing the query server power consumption. Experiments are conducted upon the TREC ClueWeb09 corpus and the query stream from the MSN 2006 query log. Results show that we can reduce up to ~24% a server power consumption, with only limited drawbacks in effectiveness w.r.t. a system running at maximum CPU frequency to promote query processing quality.

In the recent Scottish Independence Referendum (hereafter, IndyRef), Twitter offered a broad platform for people to express their opinions, with millions of IndyRef tweets posted over the campaign period. In this paper, we aim to classify people's voting intentions by the content of their tweets---their short messages communicated on Twitter. By observing tweets related to the IndyRef, we find that people not only discussed the vote, but raised topics related to an independent Scotland including oil reserves, currency, nuclear weapons, and national debt. We show that the views communicated on these topics can inform us of the individuals' voting intentions ("Yes"--in favour of Independence vs. "No"--Opposed). In particular, we argue that an accurate classifier can be designed by leveraging the differences in the features' usage across different topics related to voting intentions. We demonstrate improvements upon a Naive Bayesian classifier using the topics enrichment method. Our new classifier identifies the closest topic for each unseen tweet, based on those topics identified in the training data. Our experiments show that our Topics-Based Naive Bayesian classifier improves accuracy by 7.8% over the classical Naive Bayesian baseline.

Modern search engines increasingly rely on online evaluation methods such as A/B tests and interleaving. These online evaluation methods make use of interactions by the search engine's users to test various changes in the search engine. However, since the number of the user sessions per unit of time is limited, the number of simultaneously running on-line evaluation experiments is bounded. In an extreme case, it might be impossible to deploy all experiments since they arrive faster than are proccessed. Consequently, it is very important to efficiently use the limited resource of the user's interactions. In this paper, we formulate the novel problem of schedule optimisation for the queue of the online experiments: given a limited number of the user interactions available for experimentation, we want to re-order the queue so that the number of successful experiments is maximised. In order to build a schedule optimisation algorithm, we start by formulating a model of an online experimentation pipeline. Next, we propose to reduce the task of finding the optimal schedule to a learning-to-rank problem, where we require the most promising experiments to be ranked first in the schedule. To evaluate the proposed approach, we perform an evaluation study using two datasets containing 82 interleaving and 35 A/B test experiments, performed by a commercial search engine. We measure the quality of a schedule as the number of successful experiments executed under the limited resource of the user interactions. Our proposed schedulers obtain improvements of up to 342% compared to the un-optimised baseline schedule on the dataset of interleaving experiments and up to 43% on the dataset of A/B tests.

Online evaluation methods, such as A/B and interleaving experiments, are widely used for search engine evaluation. Since they rely on noisy implicit user feedback, running each experiment takes a considerable time. Recently, the problem of reducing the duration of online experiments has received substantial attention from the research community. However, the possibility of using sequential statistical testing procedures for reducing the time required for the evaluation experiments remains less studied. Such sequential testing procedures allow an experiment to stop early, once the data collected is sufficient to make a conclusion. In this work, we study the usefulness of sequential testing procedures for both interleaving and A/B testing. We propose modified versions of the O'Brien & Fleming and MaxSPRT sequential tests that are applicable for testing in the interleaving scenario. Similarly, for A/B experiments, we assess the usefulness of the O'Brien & Fleming test, as well as that of our proposed MaxSPRT-based sequential testing procedure. In our experiments on datasets containing 115 interleaving and 41 A/B testing experiments, we observe that considerable reductions in the average experiment duration can be achieved by using our proposed tests. In particular, for A/B experiments, the average experiment durations can be reduced by up to 66% in comparison with a single step test procedure, and by up to 44% in comparison with the O'Brien & Fleming test. Similarly, a marked relative reduction of 63% in the duration of the interleaving experiments can be achieved.

Within a search engine, query auto-completion aims to predict the final query the user wants to enter as they type, with the aim of reducing query entry time and potentially preparing the search results in advance of query submission. There are a large number of approaches to automatically rank candidate queries for the purposes of auto-completion. However, no study exists that compares these approaches on a single dataset. Hence, in this paper, we present a comparison study between current approaches to rank candidate query completions for the user query as it is typed. Using a query-log and document corpus from a commercial medical search engine, we study the performance of 11 candidate query ranking approaches from the literature and analyze where they are effective. We show that the most effective approaches to query auto-completion are largely dependent on the number of characters that the user has typed so far, with the most effective approach differing for short and long prefixes. Moreover, we show that if personalized information is available about the searcher, this additional information can be used to more effectively rank query candidate completions, regardless of the prefix length.

We present SmartVenues, a system that recommends nearby venues to a user who visits or lives in a city. SmartVenues models the variation over time of each venue's level of attendance, and uses state-of-the-art time series forecasting algorithms to predict the future attendance of these venues. We use the predicted levels of attendance to infer the popularity of a venue at future points in time, and to provide the user with recommendations at different times of the day. If the users log in with their Facebook account, the recommendations are personalised using the pages they "like". In this demonstrator, we detail the architecture of the system and the data that we collect in real-time to be able to perform the predictions. We also present two different interfaces that build upon our system to display the recommendations: a web-based application and a mobile application.

Suggesting venues to a user in a given geographic context is an emerging task that is currently attracting a lot of attention. Existing studies in the literature consist of approaches that rank candidate venues based on different features of the venues and the user, which either focus on modeling the preferences of the user or the quality of the venue. However, while providing insightful results and conclusions, none of these studies have explored the relative effectiveness of these different features. In this paper, we explore a variety of user-dependent and venue-dependent features and apply state-of-the-art learning to rank approaches to the problem of contextual suggestion in order to find what makes a venue relevant for a given context. Using the test collection of the TREC 2013 Contextual Suggestion track, we perform a number of experiments to evaluate our approach. Our results suggest that a learning to rank technique can significantly outperform a Language Modelling baseline that models the positive and negative preferences of the user. Moreover, despite the fact that the contextual suggestion task is a personalisation task (i.e. providing the user with personalised suggestions of venues), we surprisingly find that user-dependent features are less effective than venue-dependent features for estimating the relevance of a suggestion.

In the medical domain, information retrieval systems can be used for identifying cohorts (i.e. patients) required for clinical studies. However, a challenge faced by such search systems is to retrieve the cohorts whose medical histories cover the inclusion criteria specified in a query, which are often complex and include multiple medical conditions. For example, a query may aim to find patients with both 'lupus nephritis' and 'thrombotic thrombocytopenic purpura'. In a typical best-match retrieval setting, any patient exhibiting all of the inclusion criteria should naturally be ranked higher than a patient that only exhibits a subset, or none, of the criteria. In this work, we extend the two main existing models for ranking patients to take into account the coverage of the inclusion criteria by adapting techniques from recent research into coverage-based diversification. We propose a novel approach for modelling the coverage of the query inclusion criteria within the records of a particular patient, and thereby rank highly those patients whose medical records are likely to cover all of the specified criteria. In particular, our proposed approach estimates the relevance of a patient, based on the mixture of the probability that the patient is retrieved by a patient ranking model for a given query, and the likelihood that the patient's records cover the query criteria. The latter is measured using the relevance towards each of the criteria stated in the query, represented in the form of sub-queries. We thoroughly evaluate our proposed approach using the test collection provided by the TREC 2011 and 2012 Medical Records track. Our results show significant improvements over existing strong baselines.

The automatic summarization of long-running events from news steams is a challenging problem. A long-running event can contain hundreds of unique 'nuggets' of information to summarize, spread-out over its lifetime. Meanwhile, information reported about it can rapidly become outdated and is often highly redundant. Incremental update summarization (IUS) aims to select sentences from news streams to issue as updates to the user, summarising that event over time. The updates issued should cover all of the key nuggets concisely and before the information contained in those nuggets becomes outdated. Prior summarization approaches when applied to IUS can fail, since they define a fixed summary length that cannot effectively account for the different magnitudes and varying rate of development of such events. In this paper, we propose a novel IUS approach that adaptively alters the volume of content issued as updates over time with respect to the prevalence and novelty of discussions about the event. It incorporates existing state-of-the-art summarization techniques to rank candidate sentences, followed by a supervised regression model that balances novelty, nugget coverage and timeliness when selecting sentences from the top ranks. We empirically evaluate our approach using the TREC 2013 Temporal Summarization dataset extended with additional assessments. Our results show that by adaptively adjusting the number of sentences to select over time, our approach can nearly double the performance of effective summarization baselines.

The aim of risk-sensitive evaluation is to measure when a given information retrieval (IR) system does not perform worse than a corresponding baseline system for any topic. This paper argues that risk-sensitive evaluation is akin to the underlying methodology of the Student's t test for matched pairs. Hence, we introduce a risk-reward tradeoff measure TRisk that generalises the existing URisk measure (as used in the TREC 2013 Web track's risk-sensitive task) while being theoretically grounded in statistical hypothesis testing and easily interpretable. In particular, we show that TRisk is a linear transformation of the t statistic, which is the test statistic used in the Student's t test. This inherent relationship between TRisk and the t statistic, turns risk-sensitive evaluation from a descriptive analysis to a fully-fledged inferential analysis. Specifically, we demonstrate using past TREC data, that by using the inferential analysis techniques introduced in this paper, we can (1) decide whether an observed level of risk for an IR system is statistically significant, and thereby infer whether the system exhibits a real risk, and (2) determine the topics that individually lead to a significant level of risk. Indeed, we show that the latter permits a state-of-the-art learning to rank algorithm (LambdaMART) to focus on those topics in order to learn effective yet risk-averse ranking systems.

The LSDS-IR'14 workshop aims to bring together information retrieval practitioners from industry and academic researchers concerned with efficient and distributed IR systems. The workshop also welcomes contributions that propose different ways of leveraging diversity and multiplicity of resources available in distributed systems. The main goal of the workshop is to attract people from industry and academia to present and discuss ideas, problems, and results related to the efficiency of large scale and distributed information retrieval systems.

For many search settings, distributed/replicated search engines deploy a large number of machines to ensure efficient retrieval. This paper investigates how the power consumption of a replicated search engine can be automatically reduced when the system has low contention, without compromising its efficiency. We propose a novel self-adapting model to analyse the trade-off between latency and power consumption for distributed search engines. When query volumes are high and there is contention for the resources, the model automatically increases the necessary number of active machines in the system to maintain acceptable query response times. On the other hand, when the load of the system is low and the queries can be served easily, the model is able to reduce the number of active machines, leading to power savings. The model bases its decisions on examining the current and historical query loads of the search engine. Our proposal is formulated as a general dynamic decision problem, which can be quickly solved by dynamic programming in response to changing query loads. Thorough experiments are conducted to validate the usefulness of the proposed adaptive model using historical Web search traffic submitted to a commercial search engine. Our results show that our proposed self-adapting model can achieve an energy saving of 33% while only degrading mean query completion time by 10 ms compared to a baseline that provisions replicas based on a previous day's traffic.

Social media streams, such as Twitter, have shown themselves to be useful sources of real-time information about what is happening in the world. Automatic detection and tracking of events identified in these streams have a variety of real-world applications, e.g. identifying and automatically reporting road accidents for emergency services. However, to be useful, events need to be identified within the stream with a very low latency. This is challenging due to the high volume of posts within these social streams. In this paper, we propose a novel event detection approach that can both effectively detect events within social streams like Twitter and can scale to thousands of posts every second. Through experimentation on a large Twitter dataset, we show that our approach can process the equivalent to the full Twitter Firehose stream, while maintaining event detection accuracy and outperforming an alternative distributed event detection system.

In this paper, we approach the problem of real-time filtering in the Twitter Microblogging platform. We adapt an effective traditional news filtering technique, which uses a text classifier inspired by Rocchio's relevance feedback algorithm, to build and dynamically update a profile of the user's interests in real-time. In our adaptation, we tackle two challenges that are particularly prevalent in Twitter: sparsity and drift. In particular, sparsity stems from the brevity of tweets, while drift occurs as events related to the topic develop or the interests of the user change. First, to tackle the acute sparsity problem, we apply query expansion to derive terms or related tweets for a richer initialisation of the user interests within the profile. Second, to deal with drift, we modify the user profile to balance between the importance of the short-term interests, i.e. emerging subtopics, and the long-term interests in the overall topic. Moreover, we investigate an event detection method from Twitter and newswire streams to predict times at which drift may happen. Through experiments using the TREC Microblog track 2012, we show that our approach is effective for a number of common filtering metrics such as the user's utility, and that it compares favourably with state-of-the-art news filtering baselines. Our results also uncover the impact of different factors on handling topic drifting.

A search engine infrastructure must be able to provide the same quality of service to all queries received during a day. During normal operating conditions, the demand for resources is considerably lower than under peak conditions, yet an oversized infrastructure would result in an unnecessary waste of computing power. A possible solution adopted in this situation might consist of defining a maximum threshold processing time for each query, and dropping queries for which this threshold elapses, leading to disappointed users. In this paper, we propose and evaluate a different approach, where, given a set of different query processing strategies with differing efficiency, each query is considered by a framework that sets a maximum query processing time and selects which processing strategy is the best for that query, such that the processing time for all queries is kept below the threshold. The processing time estimates used by the scheduler are learned from past queries. We experimentally validate our approach on 10,000 queries from a standard TREC dataset with over 50 million documents, and we compare it with several baselines. These experiments encompass testing the system under different query loads and different maximum tolerated query response times. Our results show that, at the cost of a marginal loss in terms of response quality, our search system is able to answer 90% of queries within half a second during times of high query volume.

Interleaving is an online evaluation method to compare two alternative ranking functions based on the users' implicit feedback. In an interleaving experiment, the results from two ranking functions are merged in a single result list and presented to the users. The users' click feedback on the merged result list is analysed to derive preferences over the ranking functions. An important property of interleaving methods is their sensitivity, i.e. their ability to reliably derive the comparison outcome with a relatively small amount of user behaviour data. This allows testing of changes in the search engine ranking functions frequently and, as a result, rapid iterations in developing search quality improvements can be achieved. In this paper we propose a novel approach to further improve interleaving sensitivity by using pre-experimental user behaviour data. In particular, the click history is used to train a click model, which is then used to predict which interleaved result pages are likely to contribute to the experiment outcome. The probabilities of presenting these interleaved result pages to the users are then optimised, such that the sensitivity of interleaving is maximised. In order to evaluate the proposed approach, we re-use data from six actual interleaving experiments, previously performed by a commercial search engine. Our results demonstrate that the proposed approach outperforms a state-of-the-art baseline, achieving up to a median of 48% reduction in the number of impressions for the same level of confidence.

The query suggestion or auto-completion mechanisms help users to type less while interacting with a search engine. A basic approach that ranks suggestions according to their frequency in query logs is suboptimal. Firstly, many candidate queries with the same prefix can be removed as redundant. Secondly, the suggestions can also be personalised based on the user's context. These two directions to improve the mechanisms' quality can be in opposition: while the latter aims to promote suggestions that address search intents that a user is likely to have, the former aims to diversify the suggestions to cover as many intents as possible. We introduce a contextualisation framework that utilises a short-term context using the user's behaviour within the current search session, such as the previous query, the documents examined, and the candidate query suggestions that the user has discarded. This short-term context is used to contextualise and diversify the ranking of query suggestions, by modelling the user's information need as a mixture of intent-specific user models. The evaluation is performed offline on a set of approximately 1.0M test user sessions. Our results suggest that the proposed approach significantly improves query suggestions compared to the baseline approach.

Negated language is frequently used by medical practitioners to indicate that a patient does not have a given medical condition. Traditionally, information retrieval systems do not distinguish between the positive and negative contexts of terms when indexing documents. For example, when searching for patients with angina, a retrieval system might wrongly consider a patient with a medical record stating ``no evidence of angina" to be relevant. While it is possible to enhance a retrieval system by taking into account the context of terms within the indexing representation of a document, some non-relevant medical records can still be ranked highly, if they include some of the query terms with the intended context. In this paper, we propose a novel learning framework that effectively handles negated language. Based on features related to the positive and negative contexts of a term, the framework learns how to appropriately weight the occurrences of the opposite context of any query term, thus preventing documents that may not be relevant from being retrieved. We thoroughly evaluate our proposed framework using the TREC 2011 and 2012 Medical Records track test collections. Our results show significant improvements over existing strong baselines. In addition, in combination with a traditional query expansion and a conceptual representation approach, our proposed framework could achieve a retrieval effectiveness comparable to the performance of the best TREC 2011 and 2012 systems, while not addressing other challenges in medical records search, such as the exploitation of semantic relationships between medical terms.

Two main approaches have emerged in the literature for the effective deployment of a search system to rank patients having a medical history relevant to a query. The first approach is to directly rank patients based on the relevance of their medical history, represented as a concatenation of their associated medical records. Instead, the second approach initially retrieves the relevant medical records of patients, and then ranks the patients based on the relevance of their retrieved medical records. However, these two approaches may be useful for different queries. In this work, we propose a novel supervised approach that can effectively identify when to use either of the two aforementioned patient ranking approaches to attain effective retrieval performance. In particular, our approach deploys a classifier to learn to select a ranking approach when ranking patients, by using query difficulty measures, such as query performance predictors and the number of medical concepts detected in a query, as learning features. We thoroughly evaluate our approach using the standard test collections provided by the TREC Medical Records track. Our results show significant improvements over existing strong baselines.

Query suggestion or auto-completion mechanisms are widely used by search engines and are increasingly attracting interest from the research community. However, the lack of commonly accepted evaluation methodology and metrics means that it is not possible to compare results and approaches from the literature. Moreover, often the metrics used to evaluate query suggestions tend to be an adaptation from other domains without a proper justification. Hence, it is not necessarily clear if the improvements reported in the literature would result in an actual improvement in the users' experience. Inspired by the cascade user models and state-of-the-art evaluation metrics in the web search domain, we address the query suggestion evaluation, by first studying the users behaviour from a search engine's query log and thereby deriving a new family of user models describing the users interaction with a query suggestion mechanism. Next, assuming a query log-based evaluation approach, we propose two new metrics to evaluate query suggestions, pSaved and eSaved. Both metrics are parameterised by a user model. pSaved is defined as the probability of using the query suggestions while submitting a query. eSaved equates to the expected relative amount of effort (keypresses) a user can avoid due to the deployed query suggestion mechanism. Finally, we experiment with both metrics using four user model instantiations as well as metrics previously used in the literature on a dataset of 6.1M sessions. Our results demonstrate that pSaved and eSaved show the best alignment with the users satisfaction amongst the considered metrics.

The complexity of medical terminology raises challenges when searching medical records. For example, 'cancer', 'tumour', and 'neoplasms', which are synonyms, may prevent a traditional search system from retrieving relevant records that contain only synonyms of the query terms. Prior works use bag-of-concepts approaches, to deal with this by representing medical terms sharing the same meanings using concepts from medical resources (e.g. MeSH). The relevance scores are then combined with a traditional bag-of-words representation, when inferring the relevance of medical records. Even though the existing approaches are effective, the predicted retrieval effectiveness of either the bag-of-words or bag-of-concepts representation, which may be used to effectively model the score combination and hence improve retrieval performance, is not taken into account. In this paper, we propose a novel learning framework that models the importance of the bag-of-words and the bag-of-concepts representations, combining their scores on a per-query basis. Our proposed framework leverages retrieval performance predictors, such as the clarity score and AvIDF, calculated on both representations as learning features. We evaluate our proposed framework using the TREC Medical Records track's test collections. As our proposed framework can significantly outperform an existing approach that linearly merges the relevance scores, we conclude that retrieval performance predictors can be effectively leveraged when combining the relevance scores.

News reporting has seen a shift toward fast-paced online reporting in new sources such as social media. Web Search engines that support a news vertical have historically relied upon articles published by major newswire providers when serving news-related queries. In this paper, we investigate to what extent real-time content from newswire, blogs, Twitter and Wikipedia sources are useful to return to the user in the current fast-paced news search setting. In particular, we perform a detailed user study using the emerging medium of crowdsourcing to determine when and where integrating news-related content from these various sources can better serve the user's news need. We sampled approximately 300 news-related search queries using Google Trends and Bitly data in real-time for two time periods. For these queries, we have crowdsourced workers compare Web search rankings for each, with similar rankings integrating real-time news content from sources such as Twitter or the blogosphere. Our results show that users exhibited a preference for rankings integrating newswire articles for only half of our queries, indicating that relying solely on newswire providers for news-related content is now insufficient. Moreover, our results show that users preferred rankings that integrate tweets more often than those that integrate newswire articles, showing the potential of using social media to better serve news queries.

The LSDS-IR'13 workshop aims to bring together both information retrieval practitioners from industry, as well as academic researchers concerned with efficient and distributed IR systems. The workshop also welcomes contributions that propose different ways of leveraging diversity and multiplicity of resources available in distributed systems. The main goal of the workshop is to attract people from industry and academia to present and discuss ideas, problems and results in efficiency of large scale and distributed information retrieval systems, and to foster their participation to the WSDM conference.

Retrieval can be made more efficient by deploying dynamic pruning strategies such as WAND, which do not degrade effectiveness up to a given rank. It is possible to increase the efficiency of such techniques by pruning more 'aggressively'. However, this may reduce effectiveness. In this work, we propose a novel selective framework that determines the appropriate amount of pruning aggressiveness on a per-query basis, thereby increasing overall efficiency without significantly reducing overall effectiveness. We postulate two hypotheses about the queries that should be pruned more aggressively, which generate two approaches within our framework, based on query performance predictors and query efficiency predictors, respectively. We thoroughly experiment to ascertain the efficiency and effectiveness impacts of the proposed approaches, as part of a search engine deploying state-of-the-art learning to rank techniques. Our results on 50 million documents of the TREC ClueWeb09 collection show that by using query efficiency predictors to target inefficient queries, we observe that a 36% reduction in mean response time and a 50% reduction of the response times experienced by the slowest 10% of queries can be achieved while still ensuring effectiveness.

Several questions remain unanswered by the existing literature concerning the deployment of query-dependent features within learning to rank. In this work, we investigate three research questions in order to empirically ascertain best practices for learning-to-rank deployments. (i) Previous work in data fusion that pre-dates learning to rank showed that while different retrieval systems could be effectively combined, the combination of multiple models within the same system was not as effective. In contrast, the existing learning-to-rank datasets (e.g., LETOR), often deploy multiple weighting models as query-dependent features within a single system, raising the question as to whether such a combination is needed. (ii) Next, we investigate whether the training of weighting model parameters, traditionally required for effective retrieval, is necessary within a learning-to-rank context. (iii) Finally, we note that existing learning-to-rank datasets use weighting model features calculated on different fields (e.g., title, content, or anchor text), even though such weighting models have been criticized in the literature. Experiments addressing these three questions are conducted on Web search datasets, using various weighting models as query-dependent and typical query-independent features, which are combined using three learning-to-rank techniques. In particular, we show and explain why multiple weighting models should be deployed as features. Moreover, we unexpectedly find that training the weighting model's parameters degrades learned model's effectiveness. Finally, we show that computing a weighting model separately for each field is less effective than more theoretically-sound field-based weighting models.

Learning to rank studies have mostly focused on query-dependent and query-independent document features, which enable the learning of ranking models of increased effectiveness. Modern learning to rank techniques based on regression trees can support query features, which are document-independent, and hence have the same values for all documents being ranked for a query. In doing so, such techniques are able to learn sub-trees that are specific to certain types of query. However, it is unclear which classes of features are useful for learning to rank, as previous studies leveraged anonymised features. In this work, we examine the usefulness of four classes of query features, based on topic classification, the history of the query in a query log, the predicted performance of the query, and the presence of concepts such as persons and organisations in the query. Through experiments on the ClueWeb09 collection, our results using a state-of-the-art learning to rank technique based on regression trees show that all four classes of query features can significantly improve upon an effective learned model that does not use any query feature.

On the Web, content farms produce articles engineered such that search engines rank them highly, in order to turn a profit from online advertising. Recently, content farms have increasingly been the target of demotion strategies by Web search engines, since content farm articles are often considered to be of suspect quality. In this paper, we study the prevalence of content farms in the results returned by three major Web search engines over time. In particular, we develop a crowdsourced approach to identify content farm articles from the results returned by these search engines. Our results show that between the period of March and August 2011, the number of content farm articles observed on a number of indicative queries was reduced by up to 55% in the top ranks.

For increased efficiency, an information retrieval system can split its index into multiple shards, and then replicate these shards across many query servers. For each new query, an appropriate replica for each shard must be selected, such that the query is answered as quickly as possible. Typically, the replica with the lowest number of queued queries is selected. However, not every query takes the same time to execute, particularly if a dynamic pruning strategy is applied by each query server. Hence, the replica's queue length is an inaccurate indicator of the workload of a replica, and can result in inefficient usage of the replicas. In this work, we propose that improved replica selection can be obtained by using query efficiency prediction to measure the expected workload of a replica. Experiments are conducted using 2.2k queries, over various numbers of shards and replicas for the large GOV2 collection. Our results show that query waiting and completion times can be markedly reduced, showing that accurate response time predictions can improve scheduling accuracy and attesting the benefit of the proposed scheduling algorithm.

In medical records, negative qualifiers, e.g. no or without, are commonly used by health practitioners to identify the absence of a medical condition. Without considering whether the term occurs in a negative or positive context, the sole presence of a query term in a medical record is insufficient to imply that the record is relevant to the query. In this paper, we show how to effectively handle such negation within a medical records information retrieval system. In particular, we propose a term representation that tackles negated language in medical records, which is further extended by considering the dependence of negated query terms. We evaluate our negation handling technique within the search task provided by the TREC Medical Records 2011 track. Our results, which show a significant improvement upon a system that does not consider negated context within records, attest the importance of handling negation.

Dynamic pruning strategies permit efficient retrieval by not fully scoring all postings of the documents matching a query -- without degrading the retrieval effectiveness of the top-ranked results. However, the amount of pruning achievable for a query can vary, resulting in queries taking different amounts of time to execute. Knowing in advance the execution time of queries would permit the exploitation of online algorithms to schedule queries across replicated servers in order to minimise the average query waiting and completion times. In this work, we investigate the impact of dynamic pruning strategies on query response times, and propose a framework for predicting the efficiency of a query. Within this framework, we analyse the accuracy of several query efficiency predictors across 10,000 queries submitted to in-memory inverted indices of a 50-million-document Web crawl. Our results show that combining multiple efficiency predictors with regression can accurately predict the response time of a query before it is executed. Moreover, using the efficiency predictors to facilitate online scheduling algorithms can result in a 22% reduction in the mean waiting time experienced by queries before execution, and a 7% reduction in the mean completion time experienced by users.

A dynamic pruning strategy, such as WAND, enhances retrieval efficiency without degrading effectiveness to a given rank K, known as safe-to-rank-K. However, it is also possible for WAND to obtain more efficient but unsafe retrieval without actually significantly degrading effectiveness. On the other hand, in a modern search engine setting, dynamic pruning strategies can be used to efficiently obtain the set of documents to be re-ranked by the application of a learned model in a learning to rank setting. No work has examined the impact of safeness on the effectiveness of the learned model. In this work, we investigate the impact of WAND safeness through experiments using 150 TREC Web track topics. We find that unsafe WAND is biased towards documents with lower docids, thereby impacting effectiveness.

In this demo, we present CrowdTerrier, an infrastructure extension to the open source Terrier IR platform that enables the semi-automatic generation of relevance assessments for a variety of document ranking tasks using crowdsourcing. The aim of CrowdTerrier is to reduce the time and expertise required to effectively Crowdsource relevance assessments by abstracting away from the complexities of the crowdsourcing process. It achieves this by automating the assessment process as much as possible, via a close integration of the IR system that ranks the documents (Terrier) and the crowdsourcing marketplace that is used to assess those documents (Amazon's Mechanical Turk).

The Twitter real-time information network is the subject of research for information retrieval tasks such as real-time search. However, so far, reproducible experimentation on Twitter data has been impeded by restrictions imposed by the Twitter terms of service. In this paper, we detail a new methodology for legally building and distributing Twitter corpora, developed through collaboration between the Text REtrieval Conference (TREC) and Twitter. In particular, we detail how the first publicly available Twitter corpus - referred to as Tweets2011 - was distributed via lists of tweet identifiers and specialist tweet crawling software. Furthermore, we analyse whether this distribution approach remains robust over time, as tweets in the corpus are removed either by users or Twitter itself. Tweets2011 was successfully used by 58 participating groups for the TREC 2011 Microblog track, while our results attest to the robustness of the crawling methodology over time.

When an ambiguous query is received, a sensible approach is for the information retrieval (IR) system to diversify the results retrieved for this query, in the hope that at least one of the interpretations of the query intent will satisfy the user. Diversity is an increasingly important topic, of interest to both academic researchers (such as participants in the TREC Web and Blog track diversity tasks, or the NTCIR INTENT task), as well as to search engines professionals. In the 2nd edition of the Diversity in Document Retrieval workshop (DDR 2012), we solicited submissions both on approaches and models for diversity, the evaluation of diverse search results, and on applications of diverse search results. This workshop builds upon a successful 1st edition of DDR which was held at ECIR 2011 in Dublin, Ireland.

Modern Web crawlers seek to visit quality documents first, and re-visit them more frequently than other documents. As a result, the first-tier crawl of a Web corpus is typically of higher quality compared to subsequent crawls. In this paper, we investigate the impact of first-tier documents on adhoc retrieval performance. In particular, we analyse the retrieval performance of runs submitted to the adhoc task of the TREC 2009 Web track in terms of how they rank first-tier documents and how these documents contribute to the performance of each run. Our results show that the performance of these runs is heavily dependent on their ability to rank first-tier documents. Moreover, we show that, different from leading Web search engines, their attempt to go beyond the first tier almost always results in decreased performance. Finally, we show that selectively removing spam from different tiers can be a direction for fully exploiting documents beyond the first tier.

Expanding a query with acronyms or their corresponding 'long-forms' has not been shown to provide consistent improvements in the biomedical IR literature. The major open issue with expanding acronyms in a query is their inherent ambiguity, as an acronym can refer to multiple long-forms. At the same time, a long-form identified in a query can be expanded with its acronym(s); however, some of these may be also ambiguous and lead to poor retrieval performance. In this work, we propose the use of the EMIM (Expected Mutual Information Measure) between a long-form and its abbreviated acronym to measure ambiguity. We experiment with expanding both acronyms and long-forms identified in the queries from the adhoc task of the TREC 2004 Genomics track. Our preliminary analysis shows the potential of both acronym and long-form expansions for biomedical IR.

Search result diversification has gained momentum as a way to tackle ambiguous queries. An effective approach to this problem is to explicitly model the possible aspects underlying a query, in order to maximise the estimated relevance of the retrieved documents with respect to the different aspects. However, such aspects themselves may represent information needs with rather distinct intents (e.g., informational or navigational). Hence, a diverse ranking could benefit from applying intent-aware retrieval models when estimating the relevance of documents to different aspects. In this paper, we propose to diversify the results retrieved for a given query, by learning the appropriateness of different retrieval models for each of the aspects underlying this query. Thorough experiments within the evaluation framework provided by the diversity task of the TREC 2009 and 2010 Web tracks show that the proposed approach can significantly improve state-of-the-art diversification approaches.

An optimally diverse ranking should achieve the maximum coverage of the aspects underlying an ambiguous or under-specified query, with minimum redundancy with respect to the covered aspects. Although evaluation metrics that re-ward coverage and penalise redundancy provide intuitive ob-jective functions for learning a diverse ranking, it is unclear whether they are the most effective. In this paper, we con-trast the suitability of relevance and diversity metrics as ob-jective functions for learning a diverse ranking. Our results in the context of the diversity task of the TREC 2009 and 2010 Web tracks show that diversity metrics are not neces-sarily better suited for guiding a learning approach. More-over, the suitability of these metrics is compromised as they try to penalise redundancy during the learning process.

Search result diversification has recently gained attention as a means to tackle ambiguous queries. While query ambiguity is of particular concern for the short queries commonly observed in a Web search scenario, it is unclear how much diversity is actually promoted by Web search engines (WSEs). In this paper, we assess the diversification performance of two leading WSEs in the context of the diversity task of the TREC 2009 and 2010 Web tracks. Our results show that these WSEs perform effectively for queries with multiple interpretations, but not for those open to multiple aspects related to a single interpretation. Moreover, by deploying a state-of-the-art diversification approach based on query suggestions from these WSEs themselves, we show that their diversification performance can be further improved.

Document-at-a-time (DAAT) dynamic pruning strategies for information retrieval systems such as MaxScore and Wand can increase querying efficiency without decreasing effectiveness. Both work on posting lists sorted by ascending document identifier (docid). The order in which docids are assigned -- and hence the order of postings in the posting lists -- is known to have a noticeable impact on posting list compression. However, the resulting impact on dynamic pruning strategies is not well understood. In this poster, we examine the impact on the efficiency of these strategies across different docid orderings, by experimenting using the TREC ClueWeb09 corpus. We find that while the number of postings scored by dynamic pruning strategies do not markedly vary for different docid orderings, the ordering still has a marked impact on mean query response time. Moreover, when docids are assigned by lexicographical URL ordering, the benefit to response time for is more pronounced for Wand than for MaxScore.

Dynamic pruning strategies for information retrieval systems can increase querying efficiency without decreasing effectiveness by using upper bounds to safely omit scoring documents that are unlikely to make the final retrieved set. Often, such upper bounds are pre-calculated at indexing time for a given weighting model. However, this precludes changing, adapting or training the weighting model without recalculating the upper bounds. Instead, upper bounds should be approximated at querying time from various statistics of each term to allow on-the-fly adaptation of the applied retrieval strategy. This article, by using uniform notation, formulates the problem of determining a term upper-bound given a weighting model and discusses the limitations of existing approximations. Moreover, we propose an upper-bound approximation using a constrained nonlinear maximization problem. We prove that our proposed upper-bound approximation does not impact the retrieval effectiveness of several modern weighting models from various different families. We also show the applicability of the approximation for the Markov Random Field proximity model. Finally, we empirically examine how the accuracy of the upper-bound approximation impacts the number of postings scored and the resulting efficiency in the context of several large Web test collections.

Search result diversification is a natural approach for tackling ambiguous queries. Nevertheless, not all queries are equally ambiguous, and hence different queries could benefit from different diversification strategies. A more lenient or more aggressive diversification strategy is typically encoded by existing approaches as a trade-off between promoting relevance or diversity in the search results. In this paper, we propose to learn such a trade-off on a per-query basis. In particular, we examine how the need for diversification can be learnt for each query - given a diversification approach and an unseen query, we predict an effective trade-off between relevance and diversity based on similar previously seen queries. Thorough experiments using the TREC ClueWeb09 collection show that our selective approach can significantly outperform a uniform diversification for both classical and state-of-the-art diversification approaches.

When a Web user's underlying information need is not clearly specified from the initial query, an effective approach is to diversify the results retrieved for this query. In this paper, we introduce a novel probabilistic framework for Web search result diversification, which explicitly accounts for the various aspects associated to an underspecified query. In particular, we diversify a document ranking by estimating how well a given document satisfies each uncovered aspect and the extent to which different aspects are satisfied by the ranking as a whole. We thoroughly evaluate our framework in the context of the diversity task of the TREC 2009 Web track. Moreover, we exploit query reformulations provided by three major Web search engines (WSEs) as a means to uncover different query aspects. The results attest the effectiveness of our framework when compared to state-of-the-art diversification approaches in the literature. Additionally, by simulating an upper-bound query reformulation mechanism from official TREC data, we draw useful insights regarding the effectiveness of the query reformulations generated by the different WSEs in promoting diversity.

The retrieval effectiveness of the underlying document search component of an expert search engine can have an important impact on the effectiveness of the generated expert search results. In this large-scale study, we perform novel experiments in the context of the document search and expert search tasks of the TREC Enterprise track, to measure the influence that the performance of the document ranking has on the ranking of candidate experts. In particular, we show, using real and simulated document rankings, that while the expert search system performance is related to the relevance of the retrieved documents, surprisingly, it is not always the case that increasing document search effectiveness causes an increase in expert search performance.

Enterprise intranets are often sparse in nature, with limited use of alternative lexical representations between authors, making query expansion (QE) ineffective. Hence, for some enterprise search queries, it can be advantageous to instead use the well-known collection enrichment (CE) method to gather higher quality pseudo-feedback documents from a more diverse external resource. However, it is not always clear for which queries the collection enrichment technique should be applied. In this paper, we study two different approaches, namely a predictor-based approach and a divergence-based approach, to decide on when to apply CE. We thoroughly evaluate both approaches on the TREC Enterprise track CERC test collection and its corresponding topic sets, in combination with three different external resources and nine different query performance predictors. Our results show that both approaches are effective to selectively apply CE for enterprise search. In particular, the divergence-based approach leads to consistent and marked retrieval improvements over the systematic application of QE or CE on all external resources.

Expert search systems often employ a document search component to identify on-topic documents, which are then used to identify people likely to have relevant expertise. This work investigates the impact of the retrieval effectiveness of the underlying document search component. It has been previously shown that applying techniques to the underlying document search component that normally improve the effectiveness of a document search engine also have a positive impact on the retrieval effectiveness of the expert search engine. In this work, we experiment with fictitious perfect document rankings, to attempt to identify an upper-bound in expert search system performance. Our surprising results infer that non-relevant documents can bring useful expertise evidence, and that removing these does not lead to an upper-bound in retrieval performance.

In opinion-finding, the retrieval system is tasked with retrieving not just relevant documents, but those that also express an opinion towards the query target entity. This task has been studied in the context of the blogosphere by groups participating in the 2006-2008 TREC Blog tracks. Spam blogs (splogs) are thought to be a problem on the blogosphere. In this paper, we investigate the extent to which spam has affected the participating groups' retrieval systems over the three years of the TREC Blog track opinion-finding task. Our results show that spam can be an issue, with most systems retrieving some spam for every topic. However, removing spam from the rankings does not markedly change the relative performance of opinion-finding approaches.

The task in expert finding is to identify members of an organisation with relevant expertise on a given topic. Typically, an expert search engine uses evidence from the authors of on-topic documents found in the organisation's intranet by search engines. The search result click-through behaviour of many intranet search engine users provides an additional source of evidence to identify topically-relevant documents, and via document authorship, experts. In this poster, we assess the usefulness of click-through log data for expert finding. We find that ranking authors based solely on the clicks their documents receive is reasonably effective at correctly identifying relevant experts. Moreover, we show that this evidence can successfully be integrated with an existing expert search engine to increase its retrieval effectiveness.

Indexing is an important Information Retrieval (IR) operation, which must be parallelised to support large-scale document corpora. We propose a novel adaptation of the state-of-the-art single-pass indexing algorithm in terms of the MapReduce programming model. We then experiment with this adaptation, in the context of the Hadoop MapReduce implementation. In particular, we explore the scale of improvements that can be achieved when using firstly more processing hardware and secondly larger corpora. Our results show that indexing speed increases in a close to linear fashion when scaling corpus size or number of processing machines. This suggests that the proposed indexing implementation is viable to support upcoming large-scale corpora.

No abstract available.

Finding opinionated blog posts is still an open problem in information retrieval, as exemplified by the recent TREC blog tracks. Most of the current solutions involve the use of external resources and manual efforts in identifying subjective features. In this paper, we propose a novel and effective dictionary-based statistical approach, which automatically derives evidence for subjectivity from the blog collection itself, without requiring any manual effort. Our experiments show that the proposed approach is capable of achieving remarkable and statistically significant improvements over robust baselines, including the best TREC baseline run. In addition, with relatively little computational costs, our proposed approach provides an effective performance in retrieving opinionated blog posts, which is as good as a computationally expensive approach using Natural Language Processing techniques.

Searchers on the blogosphere often have a need to identify other key bloggers with similar interests to their own. However, a main difference of this blog distillation task from normal adhoc or Web document retrieval is that each blog can be seen as an aggregate of its constituent posts. On the other hand, we show that the task is similar to the expert search task, where a person's expertise is derived from the aggregate of their publications or emails. In this paper, we investigate several aspects of blog retrieval: Firstly, we experiment whether a blog should be represented as a whole unit, or as by considering each of its posts as indicators of its relevance, showing that expert search techniques can be adapted for blog search; Secondly, we examine whether indexing only the XML feed provided by each blog (and which is often incomplete) is sufficient, or whether the full-text of each blog post should be downloaded; Lastly, we use approaches to detect the central or recurring interests of each blog to increase the retrieval effectiveness of the system. Using the TREC 2007 Blog dataset, the results show that our proposed expert search paradigm is indeed useful in identifying key bloggers, achieving high retrieval effectiveness.

Various measures, such as binary preference (bpref), inferred average precision (infAP), and binary normalised discounted cumulative gain (nDCG) have been proposed as alternatives to mean average precision (MAP) for being less sensitive to the relevance judgements completeness. As the primary aim of any system building is to train the system to respond to user queries in a more robust and stable manner, in this paper, we investigate the importance of the choice of the evaluation measure for training, under different levels of evaluation incompleteness. We simulate evaluation incompleteness by sampling from the relevance assessments. Through large-scale experiments on two standard TREC test collections, we examine retrieval sensitivity when training - i.e. if a training process, based on any of the four discussed measures has an impact on the final retrieval performance. Experimental results show that training by bpref, infAP and nDCG provides significantly better retrieval performance than training by MAP when relevance judgements completeness is extremely low. When relevance judgements completeness increases, the measures behave more similarly.

The aim of an opinion finding system is not just to retrieve relevant documents, but to also retrieve documents that express an opinion towards the query target entity. In this work, we propose a way to use and integrate an opinion-identification toolkit, OpinionFinder, into the retrieval process of an Information Retrieval (IR) system, such that opinionated, relevant documents are retrieved in response to a query. In our experiments, we vary the number of top-ranked documents that must be parsed in response to a query, and investigate the effect on opinion retrieval performance and required parsing time. We find that opinion finding retrieval performance is improved by integrating OpinionFinder into the retrieval system, and that retrieval performance grows as more posts are parsed by OpinionFinder. However, the benefit eventually tails off at a deep rank, suggesting that an optimal setting for the system has been achieved.

In opinion-finding, the retrieval system is tasked with retrieving not just relevant documents, but which also express an opinion towards the query target entity. Most opinion-finding systems are based on a two-stage approach, where initially the system aims to retrieve relevant documents, which are then re-ranked according to the extent to which they are detected to be of an opinionated nature. In this work, we investigate how the underlying 'baseline' retrieval system performance affects the overall opinion-finding performance. We apply two effective opinion-finding techniques to all the baseline runs submitted to the TREC 2007 Blog track, and draw new insights and conclusions.

Document prior features, such as Pagerank and URL depth, can improve the retrieval effectiveness of Web Information Retrieval (IR) systems. However, not all queries equally benefit from the application of a document prior feature. This paper aims to investigate whether the retrieval performance can be further enhanced by selecting the best document prior feature on a per-query basis. We present a novel method for selecting the best document prior feature on a per-query basis. We evaluate our technique on the TREC .GOV Web test collection and its associated TREC 2003 Web search tasks. Our experiments demonstrate the effectiveness and robustness of our proposed selection method.

Pseudo-relevance feedback, or query expansion, has been shown to improve retrieval performance in the adhoc retrieval task. In such a scenario, a few top-ranked documents are assumed to be relevant, and these are then used to expand and refine the initial user query, such that it retrieves a higher quality ranking of documents. However, there has been little work in applying query expansion in the expert search task. In this setting, query expansion is applied by assuming a few top-ranked candidates have relevant expertise, and using these to expand the query. Nevertheless, retrieval is not improved as expected using such an approach. We show that the success of the application of query expansion is hindered by the presence of topic drift within the profiles of experts that the system considers. In this work, we demonstrate how topic drift occurs in the expert profiles, and moreover, we propose three measures to predict the amount of drift occurring in an expert's profile. Finally, we suggest and evaluate ways of enhancing query expansion in expert search using our new insights. Our results show that, once topic drift has been anticipated, query expansion can be successfully applied in a general manner in the expert search task.

Term dependency, or co-occurrence, has been studied in language modelling, for instance by Metzler & Croft who showed that retrieval performance could be significantlyenhanced using term dependency information. In this work, weshow how term dependency can be modelled within the Divergence From Randomness (DFR) framework. We evaluate our term dependency model on the two adhoc retrieval tasks using the TREC .GOV2 Terabyte collection. Furthermore, we examine the effect of varying the term dependency window size on the retrieval performance of the proposed model. Our experiments show that term dependency can indeed besuccessfully incorporated within the DFR framework.

In an expert search task, the users' need is to identify people who have relevant expertise to a topic of interest. An expert search system predicts and ranks the expertise of a set of candidate persons with respect to the users' query. In this paper, we propose a novel approach for predicting and ranking candidate expertise with respect to a query. We see the problem of ranking experts as a voting problem, which we model by adapting eleven data fusion techniques.We investigate the effectiveness of the voting approach and the associated data fusion techniques across a range of document weighting models, in the context of the TREC 2005 Enterprise track. The evaluation results show that the voting paradigm is very effective, without using any collection specific heuristics. Moreover, we show that improving the quality of the underlying document representation can significantly improve the retrieval performance of the data fusion techniques on an expert search task. In particular, we demonstrate that applying field-based weighting models improves the ranking of candidates. Finally, we demonstrate that the relative performance of the adapted data fusion techniques for the proposed approach is stable regardless of the used weighting models.

Emails are examples of structured documents with various fields. These fields can be exploited to enhance the retrieval effectiveness of an Information Retrieval (IR) system that mailing list archives. In recent experiments of the TREC2005 Enterprise track, various fields were applied to varying degrees of success by the participants. In his work, using a field-based weighting model, we investigate the retrieval performance attainable by each field, and examine when fields evidence should be combined or not.

No abstract available.