How do large language models (LLMs) know what they know? Answering this question has been difficult because pre-training data is often a ''black box'' – unknown or inaccessible. The recent release of nanochat – a family of small LLMs with fully open pre-training data – addresses this as it provides a transparent view into where a model's parametric knowledge comes from. Towards the goal of understanding how knowledge is encoded by LLMs, we release NanoKnow, a benchmark dataset that partitions questions from Natural Questions and SQuAD into splits based on whether their answers are present in nanochat's pre-training corpus. Using these splits, we can now properly disentangle the sources of knowledge that LLMs rely on when producing an output. To demonstrate NanoKnow's utility, we conduct experiments using eight nanochat checkpoints. Our findings show: (1) closed-book accuracy is strongly influenced by answer frequency in the pre-training data, (2) providing external evidence can mitigate this frequency dependence, (3) even with external evidence, models are more accurate when answers were seen during pre-training, demonstrating that parametric and external knowledge are complementary, and (4) non-relevant information is harmful, with accuracy decreasing based on both the position and the number of non-relevant contexts. We release all NanoKnow artifacts at https://github.com/castorini/NanoKnow.

Recent approaches that leverage large language models (LLMs) for pseudo-relevance feedback (PRF) have generally not utilized well-established feedback models like Rocchio and RM3 when expanding queries for BM25. Instead, they often opt for a simple string concatenation of the query and LLM-generated expansion content. In this paper, we revisit and systematically evaluate traditional BM25 feedback models in the context of HyDE, a popular method that enriches query representations using LLM-generated hypothetical answer documents. Experiments demonstrate a mutually beneficial relationship between BM25 feedback models and HyDE. Feedback models are more effective when provided LLM-generated documents versus top-ranked documents retrieved by BM25, while HyDE benefits from feedback algorithms that select and weight expansion terms. In fact, by incorporating BM25 feedback models within the HyDE setup, we can further narrow the gap between BM25 and a strong ''single-shot'' dense retriever, without incurring the costs associated with such embedding-based retrieval methods. Ultimately, our results demonstrate that traditional BM25 feedback models can play an important role within modern LLM PRF methods and provide a simple approach to further enhance the accuracy of HyDE. Our code is available at https://github.com/nourj98/hyde-feedback.

Retrieval benchmarks for large language models (LLMs) should reflect the long, reasoning-intensive queries typical of retrieval-augmented generation (RAG). We present a systematic study of BRIGHT, a reasoning-focused retrieval benchmark, along with strong, reproducible reference methods integrated into Anserini, Pyserini, and RankLLM. We evaluate lexical, sparse, dense, and fusion-based retrievers, as well as LLM rerankers, under long-query settings. In reproducing BRIGHT's lexical baseline, we identify a key under-documented detail: query-side BM25 (BM25Q), which applies BM25 weighting to the query itself. On long, multi-sentence queries, BM25Q consistently outperforms standard BM25, making it the strongest lexical baseline for reasoning-oriented retrieval. We further audit the BRIGHT corpus, uncovering data quality issues that impact evaluation, and offer mitigation. Finally, we study the generalizability of BM25Q across five additional benchmarks, finding its gains largely specific to BRIGHT, while fusion with standard BM25 provides the most consistent improvements across datasets.

Traditional short keyword queries are increasingly being replaced by longer, more detailed queries that reflect complex and nuanced user information needs, especially in conversational assistants equipped with web search capabilities. In this work, we present a methodology for automatically generating such human-style long-form queries (narratives) by clustering raw short queries to form synthetic search sessions, designed to reflect a real user's search behavior. Based on our human interpretation study of a 50-narrative set (comprising both human-written and automated narratives), 44% of the automated narratives are misidentified as human-written, underscoring not only the realism and complexity of the generated content but also its indistinguishability from authentic human narratives. Furthermore, we share a collection of automated narratives as a testbed for evaluating LLMs on long-form question answering (QA), which was used in the TREC 2025 RAG track. Our code is available at https://github.com/castorini/narrative-generation.

While dense retrieval models have been the standard for state-of-the-art information retrieval, their deployment is often constrained by high memory requirements and reliance on GPU accelerators for vector similarity search at scale. Learned sparse retrieval offers a compelling alternative by enabling efficient search via inverted indices, yet it has historically received less attention than dense approaches. In this paper, we introduce LACONIC, a family of learned sparse retrievers based on the Llama3 architecture (1B, 3B, and 8B). We propose a streamlined two-phase training curriculum consisting of (1) weakly supervised pre-finetuning to adapt causal LLMs for bidirectional contextualization and (2) high-signal finetuning using curated hard negatives. Our results demonstrate that LACONIC effectively bridges the performance gap with dense models: the 8B variant achieves a state-of-the-art 60.2 nDCG@10 on the MTEB Retrieval benchmark, ranking 15th on the leaderboard as of February 5th, 2026, while utilizing 74% less index memory than an equivalent dense model. By delivering high retrieval effectiveness on commodity CPU hardware with a fraction of the compute budget required by competing models, LACONIC provides a scalable and efficient solution for real-world search applications. We fully open source our code implementation and trained checkpoints to facilitate reproducibility.

Promptagator demonstrated that Large Language Models (LLMs) with few-shot prompts can be used as task-specific query generators for fine-tuning domain-specialized dense retrieval models. However, the original Promptagator approach relied on proprietary and large-scale LLMs which users may not have access to or may be prohibited from using with sensitive data. In this work, we study the impact of open-source LLMs at accessible scales (≤14B parameters) as an alternative. Our results demonstrate that open-source LLMs as small as 3B parameters can serve as effective Promptagator-style query generators. We hope our work will inform practitioners with reliable alternatives for synthetic data generation and give insights to maximize fine-tuning results for domain-specific applications. Our code is available at https://www.github.com/mitll/promptodile

Given the dominance of dense retrievers which do not generalize well beyond their training dataset distributions, domain-specific test sets are essential in evaluating retrieval performance.Few test datasets are available for retrieval systems intended for use by healthcare providers in a point-of-care setting.To fill this gap we have collaborated with medical professionals to create CURE, a test dataset for passage retrieval composed of 2,000 expert written queries spanning 10 medical domains with a monolingual (English) and two cross-lingual (French/Spanish to English) conditions.In this paper, we describe how CURE was constructed and provide baseline results to showcase its effectiveness as an evaluation tool.CURE is published with a CC BY-NC 4.0 license and can be accessed on Hugging Face 1 and as a retrieval task on MTEB.

Large language models (LLMs) have emerged as powerful listwise rerankers but remain prohibitively slow for many real-world applications. What's more, training on the language modeling (LM) objective is not intrinsically aligned with reranking tasks. To address these challenges, FIRST, a novel approach for listwise reranking, integrates a learning-to-rank objective and leverages only the logits of the first generated token for reranking, significantly reducing computational overhead while preserving effectiveness. We systematically evaluate the capabilities and limitations of FIRST. By extending its evaluation to TREC Deep Learning collections (DL19-23), we show that FIRST achieves robust out-of-domain effectiveness. Through training FIRST on a variety of backbone models, we demonstrate its generalizability across different model architectures, and achieve effectiveness surpassing the original implementation. Further analysis of the interaction between FIRST and various first-stage retrievers reveals diminishing returns akin to traditional LLM rerankers. A comprehensive latency study confirms that FIRST consistently delivers a 40% efficiency gain over traditional rerankers without sacrificing effectiveness. Notably, while LM training implicitly improves zero-shot single-token reranking, our experiments also highlight potential conflicts between LM pre-training and subsequent fine-tuning on the FIRST objective. These findings pave the way for more efficient and effective listwise reranking in future applications. Our code is available at: https://rankllm.ai.

The Anserini IR toolkit has come a long way since efforts began in 2015. Although the goals of the project - to bridge research and practice in information retrieval, and to provide reproducible, easy-to-use baselines - have remained constant, the world has changed quite a bit. We discuss how Anserini has evolved in response to this changing environment, the most significant of which is the advent of transformer-based retrieval models that did not exist when the project started. The bi-encoder architecture provides a framework for understanding retrieval models based on dense and sparse vector representations, and offers a reference for conveying the capabilities of our toolkit. Anserini provides end-to-end first-stage retrieval based on single-vector learned dense and sparse representations, directly building on the open-source Lucene search library and the ONNX runtime. This minimal design accelerates the pace of research and fosters reproducibility, enabling ''two-click reproductions''. By better aligning research and practice, we increase the potential real-world impact of research innovations.

Recent advancements in large language models (LLMs) have driven interest in billion-scale retrieval models with strong generalization across retrieval tasks and languages. Additionally, progress in large vision-language models has created new opportunities for multimodal retrieval. In response, we have updated the Tevatron toolkit, introducing a unified pipeline that enables researchers to explore retriever models at different scales, across multiple languages, and with various modalities. This demo paper highlights the toolkit's key features, bridging academia and industry by supporting efficient training, inference, and evaluation of neural retrievers. We showcase a unified dense retriever achieving strong multilingual and multimodal effectiveness, and conduct a cross-modality zero-shot study to demonstrate its research potential. Alongside, we release OmniEmbed, to the best of our knowledge, the first embedding model that unifies text, image document, video, and audio retrieval, serving as a baseline for future research.

Large Language Models (LLMs) have significantly enhanced the capabilities of information access systems, especially with retrieval-augmented generation (RAG). Nevertheless, the evaluation of RAG systems remains a barrier to continued progress, a challenge we tackle in this work by proposing an automatic evaluation framework that is validated against human annotations. We believe that the nugget evaluation methodology provides a solid foundation for evaluating RAG systems. This approach, originally developed for the TREC Question Answering (QA) Track in 2003, evaluates systems based on atomic facts that should be present in good answers. Our efforts focus on ''refactoring'' this methodology, where we describe the AutoNuggetizer framework that specifically applies LLMs to both automatically create nuggets and automatically assign nuggets to system answers. In the context of the TREC 2024 RAG Track, we calibrate a fully automatic approach against strategies where nuggets are created manually or semi-manually by human assessors and then assigned manually to system answers. Based on results from a community-wide evaluation, we observe strong agreement at the run level between scores derived from fully automatic nugget evaluation and human-based variants. The agreement is stronger when individual framework components such as nugget assignment are automated independently. This suggests that our evaluation framework provides tradeoffs between effort and quality that can be used to guide the development of future RAG systems. However, further research is necessary to refine our approach, particularly in establishing robust per-topic agreement to diagnose system failures effectively.

The adoption of large language models (LLMs) as rerankers in multi-stage retrieval systems has gained significant traction in academia and industry. These models refine a candidate list of retrieved documents, often through carefully designed prompts, and are typically used in applications built on retrieval-augmented generation (RAG). This paper introduces RankLLM, an open-source Python package for reranking that is modular, highly configurable, and supports both proprietary and open-source LLMs in customized reranking workflows. To improve usability, RankLLM features optional integration with Pyserini for retrieval and provides integrated evaluation for multi-stage pipelines. Additionally, RankLLM includes a module for detailed analysis of input prompts and LLM responses, addressing reliability concerns with LLM APIs and non-deterministic behavior in Mixture-of-Experts (MoE) models. This paper presents the architecture of RankLLM, along with a detailed step-by-step guide and sample code. We reproduce results from RankGPT, LRL, RankVicuna, RankZephyr, and other recent models. RankLLM integrates with common inference frameworks and a wide range of LLMs. This compatibility allows for quick reproduction of reported results, helping to speed up both research and real-world applications. The complete repository is available at rankllm.ai, and the package can be installed via PyPI.

Retrieval-augmented generation (RAG) enables large language models (LLMs) to generate answers with citations from source documents containing ''ground truth''. A crucial factor in RAG evaluation is ''support'', or whether the information in the cited documents supports the answer. We conducted a comparative study of submissions to the TREC 2024 RAG Track, evaluating an automatic LLM judge (GPT-4o) against human judges for support assessment. We considered two conditions: (1) fully manual assessments from scratch and (2) manual assessments with post-editing of LLM predictions. Our results indicate good agreement between human and GPT-4o predictions. Further analysis of the disagreements shows that an independent human judge correlates better with GPT-4o than a human judge, suggesting that LLM judges can be a reliable alternative for support assessment. We provide a qualitative analysis of human and GPT-4o errors to help guide future evaluations.

Recent advancements in dense retrieval have introduced vision-language model (VLM)-based retrievers, such as DSE and ColPali, which leverage document screenshots embedded as vectors to enable effective search and offer a simplified pipeline over traditional text-only methods. In this study, we propose three pixel poisoning attack methods designed to compromise VLM-based retrievers and evaluate their effectiveness under various attack settings and parameter configurations. Our empirical results demonstrate that injecting even a single adversarial screenshot into the retrieval corpus can significantly disrupt search results, poisoning the top-10 retrieved documents for 41.9% of queries in the case of DSE and 26.4% for ColPali. These vulnerability rates notably exceed those observed with equivalent attacks on text-only retrievers. Moreover, when targeting a small set of known queries, the attack success rate raises, achieving complete success in certain cases. By exposing the vulnerabilities inherent in vision-language models, this work highlights the potential risks associated with their deployment.

Knowledge-intensive analytical applications retrieve context from both structured tabular data and unstructured free text documents for effective decision-making. Large language models (LLMs) have significantly simplified the prototyping of such retrieval and reasoning data pipelines. However, implementing them efficiently remains challenging and demands significant effort. Developers must often orchestrate heterogeneous systems, manage data movement, and handle low-level concerns such as LLM context management. To address these challenges, we introduce FlockMTL: an extension for DBMSs that integrates LLM capabilities and enables retrieval-augmented generation (RAG) within SQL. FlockMTL provides LLM-powered scalar and aggregate functions, enabling chained predictions over tuples. It further provides data fusion functions to support hybrid search. Drawing inspiration from the relational model, FlockMTL incorporates: (i) seamless optimizations such as batching and meta-prompting; and (ii) resource independence through novel SQL DDL abstractions: PROMPT and MODEL, introduced as first-class schema objects alongside TABLE.

Query expansion has been widely used to improve the search results of first-stage retrievers, yet its influence on second-stage, cross-encoder rankers remains under-explored. A recent study shows that current expansion techniques benefit weaker models but harm stronger rankers. In this paper, we re-examine this conclusion and raise the following question: Can query expansion improve generalization of strong cross-encoder rankers? To answer this question, we first apply popular query expansion methods to different cross-encoder rankers and verify the deteriorated zero-shot effectiveness. We identify two vital steps in the experiment: high-quality keyword generation and minimally-disruptive query modification. We show that it is possible to improve the generalization of a strong neural ranker, by generating keywords through a reasoning chain and aggregating the ranking results of each expanded query via self-consistency, reciprocal rank weighting, and fusion. Experiments on BEIR and TREC Deep Learning 2019/2020 show that the nDCG@10 scores of both MonoT5 and RankT5 following these steps are improved, which points out a direction for applying query expansion to strong cross-encoder rankers.

Cross-lingual information retrieval (CLIR) continues to be an actively studied topic in information retrieval (IR), and there have been consistent efforts in curating test collections to support its research. However, there is a lack of high-quality human-annotated CLIR resources for African languages: the few existing collections are mostly curated synthetically or from sources with limited corpora for these languages. We present CIRAL, a test collection for cross-lingual retrieval with English queries and passages in four African languages: Hausa, Somali, Swahili, and Yoruba. CIRAL's corpora are obtained from Indigenous African websites and consist of a total of over 2.5 million passages. We gathered over 1,600 queries and 30k high-quality binary relevance judgments annotated by native speakers of the languages. Additional pools were also obtained at CIRAL's shared task, which was hosted at the Forum for Information Retrieval Evaluation 2023 to encourage community participation in CLIR for African languages. We describe the design and curation process of our test collection and provide reproducible baselines that demonstrate CIRAL's utility in evaluating the effectiveness of systems. CIRAL is available at https://github.com/ciralproject/ciral.

BEIR is a benchmark dataset originally designed for zero-shot evaluation of retrieval models across 18 different domain/task combinations. In recent years, we have witnessed the growing popularity of models based on representation learning, which naturally begs the question: How effective are these models when presented with queries and documents that differ from the training data? While BEIR was designed to answer this question, our work addresses two shortcomings that prevent the benchmark from achieving its full potential: First, the sophistication of modern neural methods and the complexity of current software infrastructure create barriers to entry for newcomers. To this end, we provide reproducible reference implementations that cover learned dense and sparse models. Second, comparisons on BEIR are performed by reducing scores from heterogeneous datasets into a single average that is difficult to interpret. To remedy this, we present meta-analyses focusing on effect sizes across datasets that are able to accurately quantify model differences. By addressing both shortcomings, our work facilitates future explorations in a range of interesting research questions.

Instruction-tuned large language models (LLMs) have been shown to be viable surrogates for the widely used, albeit overly rigid, lexical matching metrics in evaluating question answering (QA) models. However, these LLM-based evaluation methods are invariably based on proprietary LLMs. Despite their remarkable capabilities, proprietary LLMs are costly and subject to internal changes that can affect their output, which inhibits the reproducibility of their results and limits the widespread adoption of LLM-based evaluation. In this demo, we aim to use publicly available LLMs for standardizing LLM-based QA evaluation. However, open-source LLMs lag behind their proprietary counterparts. We overcome this gap by adopting chain-of-thought prompting with self-consistency to build a reliable evaluation framework. We demonstrate that our evaluation framework, based on 750M and 7B open LLMs, correlates competitively with human judgment, compared to most recent GPT-3 and GPT-4 models. Our codebase and data are available at https://github.com/castorini/qa-eval.

While large language models (LLMs) have shown impressive NLP capabilities, existing IR applications mainly focus on prompting LLMs to generate query expansions or generating permutations for listwise reranking. In this study, we leverage LLMs directly to serve as components in the widely used multi-stage text ranking pipeline. Specifically, we fine-tune the open-source LLaMA-2 model as a dense retriever (repLLaMA) and a pointwise reranker (rankLLaMA). This is performed for both passage and document retrieval tasks using the MS MARCO training data. Our study shows that finetuned LLM retrieval models outperform smaller models. They are more effective and exhibit greater generalizability, requiring only a straightforward training strategy. Moreover, our pipeline allows for the fine-tuning of LLMs at each stage of a multi-stage retrieval pipeline. This demonstrates the strong potential for optimizing LLMs to enhance a variety of retrieval tasks. Furthermore, as LLMs are naturally pre-trained with longer contexts, they can directly represent longer documents. This eliminates the need for heuristic segmenting and pooling strategies to rank long documents. On the MS MARCO and BEIR datasets, our repLLaMA-rankLLaMA pipeline demonstrates a high level of effectiveness.

The effectiveness of dense retrieval models trained with multilingual language models as backbones has been demonstrated in multilingual and cross-lingual information retrieval contexts. The optimal choice of a backbone model for a given retrieval task is dependent on the target retrieval domain as well as the pre-training domain of available language models and their generalization capabilities, the availability of relevance judgements, etc. In this work, we study the impact of these factors on retrieval effectiveness for African languages using three multilingual benchmark datasets: Mr. TyDi, MIRACL, and the newly released CIRAL dataset. We compare the effectiveness of mBERT as a backbone for dense retrieval models against multilingual language models such as AfriBERTa and AfroXLMR, which are specialized for African languages. Furthermore, we examine the impact of different training regimes on the effectiveness of dense retrieval in different domains for African languages. Our findings show that the pre-training domain of the backbone LM plays a huge role in retrieval effectiveness, especially in the absence of retrieval training data. Code artifacts are available at https://github.com/castorini/afridpr_backbones.

The zero-shot effectiveness of neural retrieval models is often evaluated on the BEIR benchmark---a combination of different IR evaluation datasets. Interestingly, previous studies found that particularly on the BEIR~subset Touché 2020, an argument retrieval task, neural retrieval models are considerably less effective than BM25. Still, so far, no further investigation has been conducted on what makes argument retrieval so "special''. To more deeply analyze the respective potential limits of neural retrieval models, we run a reproducibility study on the Touché 2020 data. In our study, we focus on two experiments: (i) a black-box evaluation (i.e., no model retraining), incorporating a theoretical exploration using retrieval axioms, and (ii) a data denoising evaluation involving post-hoc relevance judgments. Our black-box evaluation reveals an inherent bias of neural models towards retrieving short passages from the Touché 2020 data, and we also find that quite a few of the neural models' results are unjudged in the Touché 2020 data. As many of the short Touché passages are not argumentative and thus non-relevant per se, and as the missing judgments complicate fair comparison, we denoise the Touché 2020 data by excluding very short passages (less than 20 words) and by augmenting the unjudged data with post-hoc judgments following the Touché guidelines. On the denoised data, the effectiveness of the neural models improves by up to 0.52 in nDCG@10, but BM25 is still more effective. Our code and the augmented Touché 2020 dataset are available at https://github.com/castorini/touche-error-analysis.

The remarkable success of large language models (LLMs) has drawn people's great interest in their deployment in specific domains and downstream applications. In this paper, we present the first systematic study of applying large language models (in our case, GPT-3.5 and GPT-4) for the automatic coding (multi-class classification) problem in market research. Our experimental results show that large language models could achieve a macro F1 score of over 0.5 for all our collected real-world market research datasets in a zero-shot setting. We also provide in-depth analyses of the errors made by the large language models. We hope this study sheds light on the lessons we learn and the open challenges large language models have when adapting to a specific market research domain.

We provide a reproducible, end-to-end demonstration of vector search with OpenAI embeddings using Lucene on the popular MS MARCO passage ranking test collection. The main goal of our work is to challenge the prevailing narrative that a dedicated vector store is necessary to take advantage of recent advances in deep neural networks as applied to search. Quite the contrary, we show that hierarchical navigable small-world network (HNSW) indexes in Lucene are adequate to provide vector search capabilities in a standard bi-encoder architecture. This suggests that, from a simple cost-benefit analysis, there does not appear to be a compelling reason to introduce a dedicated vector store into a modern "AI stack" for search, since such applications have already received substantial investments in existing, widely deployed infrastructure.

Dense retrieval models using a transformer-based bi-encoder architecture have emerged as an active area of research. In this article, we focus on the task of monolingual retrieval in a variety of typologically diverse languages using such an architecture. Although recent work with multilingual transformers demonstrates that they exhibit strong cross-lingual generalization capabilities, there remain many open research questions, which we tackle here. Our study is organized as a “best practices” guide for training multilingual dense retrieval models, broken down into three main scenarios: when a multilingual transformer is available, but training data in the form of relevance judgments are not available in the language and domain of interest (“have model, no data”); when both models and training data are available (“have model and data”); and when training data are available but not models (“have data, no model”). In considering these scenarios, we gain a better understanding of the role of multi-stage fine-tuning, the strength of cross-lingual transfer under various conditions, the usefulness of out-of-language data, and the advantages of multilingual vs. monolingual transformers. Our recommendations offer a guide for practitioners building search applications, particularly for low-resource languages, and while our work leaves open a number of research questions, we provide a solid foundation for future work.

Anserini is a Lucene-based toolkit for reproducible information retrieval research in Java that has been gaining traction in the community. It provides retrieval capabilities for both "traditional" bag-of-words retrieval models such as BM25 as well as retrieval using learned sparse representations such as SPLADE. With Pyserini, which provides a Python interface to Anserini, users gain access to both sparse and dense retrieval models, as Pyserini implements bindings to the Faiss vector search library alongside Lucene inverted indexes in a uniform, consistent interface. Nevertheless, hybrid fusion techniques that integrate sparse and dense retrieval models need to stitch together results from two completely different "software stacks", which creates unnecessary complexities and inefficiencies. However, the introduction of HNSW indexes for dense vector search in Lucene promises the integration of both dense and sparse retrieval within a single software framework. We explore exactly this integration in the context of Anserini. Experiments on the MS MARCO passage and BEIR datasets show that our Anserini HNSW integration supports (reasonably) effective and (reasonably) efficient approximate nearest neighbor search for dense retrieval models, using only Lucene.

Recent rapid advances in deep pre-trained language models and the introduction of large datasets have powered research in embedding-based neural retrieval. While many excellent research papers have emerged, most of them come with their own implementations, which are typically optimized for some particular research goals instead of efficiency or code organization. In this paper, we introduce Tevatron, a neural retrieval toolkit that is optimized for efficiency, flexibility, and code simplicity. Tevatron enables model training and evaluation for a variety of ranking components such as dense retrievers, sparse retrievers, and rerankers. It also provides a standardized pipeline that includes text processing, model training, corpus/query encoding, and search. In addition, Tevatron incorporates well-studied methods for improving retriever effectiveness such as hard negative mining and knowledge distillation. We provide an overview of Tevatron in this paper, demonstrating its effectiveness and efficiency on multiple IR and QA datasets. We highlight Tevatron's flexible design, which enables easy generalization across datasets, model architectures, and accelerator platforms (GPUs and TPUs). Overall, we believe that Tevatron can serve as a solid software foundation for research on neural retrieval systems, including their design, modeling, and optimization.

MMEAD, or MS MARCO Entity Annotations and Disambiguations, is a resource for entity links for the MS MARCO datasets. We specify a format to store and share links for both document and passage collections of MS MARCO. Following this specification, we release entity links to Wikipedia for documents and passages in both MS MARCO collections (v1 and v2). Entity links have been produced by the REL and BLINK systems. MMEAD is an easy-to-install Python package, allowing users to load the link data and entity embeddings effortlessly. Using MMEAD takes only a few lines of code. Finally, we show how MMEAD can be used for IR research that uses entity information. We show how to improve recall@1000 and MRR@10 on more complex queries on the MS MARCO v1 passage dataset by using this resource. We also demonstrate how entity expansions can be used for interactive search applications.

This paper introduces Sparsified Late Interaction for Multi-vector (SLIM) retrieval with inverted indexes. Multi-vector retrieval methods have demonstrated their effectiveness on various retrieval datasets, and among them, ColBERT is the most established method based on the late interaction of contextualized token embeddings of pre-trained language models. However, efficient ColBERT implementations require complex engineering and cannot take advantage of off-the-shelf search libraries, impeding their practical use. To address this issue, SLIM first maps each contextualized token vector to a sparse, high-dimensional lexical space before performing late interaction between these sparse token embeddings. We then introduce an efficient two-stage retrieval architecture that includes inverted index retrieval followed by a score refinement module to approximate the sparsified late interaction, which is fully compatible with off-the-shelf lexical search libraries such as Lucene. SLIM achieves competitive accuracy on MS MARCO Passages and BEIR compared to ColBERT while being much smaller and faster on CPUs. To our knowledge, we are the first to explore using sparse token representations for multi-vector retrieval. Source code and data are integrated into the Pyserini IR toolkit.

Traditionally, sparse retrieval systems relied on lexical representations to retrieve documents, such as BM25, dominated information retrieval tasks. With the onset of pre-trained transformer models such as BERT, neural sparse retrieval has led to a new paradigm within retrieval. Despite the success, there has been limited software supporting different sparse retrievers running in a unified, common environment. This hinders practitioners from fairly comparing different sparse models and obtaining realistic evaluation results. Another missing piece is, that a majority of prior work evaluates sparse retrieval models on in-domain retrieval, i.e. on a single dataset: MS MARCO. However, a key requirement in practical retrieval systems requires models that can generalize well to unseen out-of-domain, i.e. zero-shot retrieval tasks. In this work, we provide SPRINT, a unified python toolkit based on Pyserini and Lucene, supporting a common interface for evaluating neural sparse retrieval. The toolkit currently includes five built-in models: uniCOIL, DeepImpact, SPARTA, TILDEv2 and SPLADEv2. Users can also easily add customized models by defining their term weighting method. Using our toolkit, we establish strong and reproducible zero-shot sparse retrieval baselines across the well-acknowledged benchmark, BEIR. Our results demonstrate that SPLADEv2 achieves the best average score of 0.470 [email protected] on BEIR amongst all neural sparse retrievers. In this work, we further uncover the reasons behind its performance gain. We show that SPLADEv2 produces sparse representations with a majority of tokens outside of the original query and document which is often crucial for its performance gains, i.e. a limitation among its other sparse counterparts. We provide our SPRINT toolkit, models, and data used in our experiments publicly here: https://github.com/thakur-nandan/sprint.

This paper presents the AToMiC (Authoring Tools for Multi media Content) dataset, designed to advance research in image/text cross-modal retrieval. While vision--language pretrained transformers have led to significant improvements in retrieval effectiveness, existing research has relied on image-caption datasets that feature only simplistic image--text relationships and underspecified user models of retrieval tasks. To address the gap between these oversimplified settings and real-world applications for multimedia content creation, we introduce a new approach for building retrieval test collections. We leverage hierarchical structures and diverse domains of texts, styles, and types of images, as well as large-scale image--document associations embedded in Wikipedia. We formulate two tasks based on a realistic user model and validate our dataset through retrieval experiments using baseline models. AToMiC offers a testbed for scalable, diverse, and reproducible multimedia retrieval research. Finally, our dataset provides the basis for a dedicated track at the 2023 Text Retrieval Conference (TREC), and is publicly available at https://github.com/TREC-AToMiC/AToMiC.

Neural retrievers have been shown to be effective for math-aware search. Their ability to cope with math symbol mismatches, to represent highly contextualized semantics, and to learn effective representations are critical to improving math information retrieval. However, the most effective retriever for math remains impractical as it depends on token-level dense representations for each math token, which leads to prohibitive storage demands, especially considering that math content generally consumes more tokens. In this work, we try to alleviate this efficiency bottleneck while boosting math information retrieval effectiveness via hybrid search. To this end, we propose MABOWDOR, a Math-Aware Bestof-Worlds Domain Optimized Retriever, which has an unsupervised structure search component, a dense retriever, and optionally a sparse retriever on top of a domain-adapted backbone learned by context-enhanced pretraining, each addressing a different need in retrieving heterogeneous data from math documents. Our hybrid search outperforms the previous state-of-the-art math IR system while eliminating efficiency bottlenecks. Our system is available at https://github.com/approach0/pya0.

Lexical and semantic matching capture different successful approaches to text retrieval and the fusion of their results has proven to be more effective and robust than either alone. Prior work performs hybrid retrieval by conducting lexical and semantic matching using different systems (e.g., Lucene and Faiss, respectively) and then fusing their model outputs. In contrast, our work integrates lexical representations with dense semantic representations by densifying high-dimensional lexical representations into what we call low-dimensional dense lexical representations (DLRs). Our experiments show that DLRs can effectively approximate the original lexical representations, preserving effectiveness while improving query latency. Furthermore, we can combine dense lexical and semantic representations to generate dense hybrid representations (DHRs) that are more flexible and yield faster retrieval compared to existing hybrid techniques. In addition, we explore jointly training lexical and semantic representations in a single model and empirically show that the resulting DHRs are able to combine the advantages of the individual components. Our best DHR model is competitive with state-of-the-art single-vector and multi-vector dense retrievers in both in-domain and zero-shot evaluation settings. Furthermore, our model is both faster and requires smaller indexes, making our dense representation framework an attractive approach to text retrieval. Our code is available at https://github.com/castorini/dhr .

Researchers have had much recent success with ranking models based on so-called learned sparse representations generated by transformers. One crucial advantage of this approach is that such models can exploit inverted indexes for top- k retrieval, thereby leveraging decades of work on efficient query evaluation. Yet, there remain many open questions about how these learned representations fit within the existing literature, which our work aims to tackle using four representative learned sparse models. We find that impact weights generated by transformers appear to greatly reduce opportunities for skipping and early exiting optimizations in well-studied document-at-a-time ( DaaT ) approaches. Similarly, “off-the-shelf” application of score-at-a-time ( SaaT ) processing exhibits a mismatch between these weights and assumptions behind accumulator management strategies. Building on these observations, we present solutions to address deficiencies with both DaaT and SaaT approaches, yielding substantial speedups in query evaluation. Our detailed empirical analysis demonstrates that both methods lie on the effectiveness–efficiency Pareto frontier, indicating that the optimal choice for deployment depends on operational constraints.

Current pre-trained language model approaches to information retrieval can be broadly divided into two categories: sparse retrievers (to which belong also non-neural approaches such as bag-of-words methods, e.g., BM25) and dense retrievers. Each of these categories appears to capture different characteristics of relevance. Previous work has investigated how relevance signals from sparse retrievers could be combined with those from dense retrievers via interpolation. Such interpolation would generally lead to higher retrieval effectiveness.

We articulate the design and implementation of the MS MARCO document ranking and passage ranking leaderboards. In contrast to "standard" community-wide evaluations such as those at TREC, which can be characterized as simultaneous games, leaderboards represent sequential games, where every player move is immediately visible to the entire community. The fundamental challenge with this setup is that every leaderboard submission leaks information about the held-out evaluation set, which conflicts with the fundamental tenant in machine learning about separation of training and test data. These "leaks", accumulated over long periods of time, threaten the validity of the insights that can be derived from the leaderboards. In this paper, we share our experiences grappling with this issue over the past few years and how our considerations are operationalized into a coherent submission policy. Our work provides a useful guide to help the community understand the design choices made in the popular MS MARCO leaderboards and offers lessons for designers of future leaderboards.

Over two decades ago, Berger and Lafferty proposed "information retrieval as statistical translation" (IRST), a simple and elegant method for ad hoc retrieval based on the noisy channel model. At the time, they lacked the large-scale human-annotated datasets necessary to properly train their models. In this paper, we ask the simple question: What if Berger and Lafferty had access to datasets such as the MS MARCO passage ranking dataset that we take for granted today? The answer to this question tells us how much of recent improvements in ranking can be solely attributed to having more data available, as opposed to improvements in models (e.g., pretrained transformers) and optimization techniques (e.g., contrastive loss). In fact, Boytsov and Kolter recently began to answer this question with a replication of Berger and Lafferty's model, and this work can be viewed as another independent replication effort, with generalizations to additional conditions not previously explored, including replacing the sum of translation probabilities with ColBERT's MaxSim operator. We confirm that while neural models (particularly pretrained transformers) have indeed led to great advances in retrieval effectiveness, the IRST model proposed decades ago is quite effective if provided sufficient training data.

With doc2query, we train a neural sequence-to-sequence model that, given an input span of text, predicts a natural language query that the text might answer. These predictions can be viewed as document expansions that feed standard bag-of-words term weighting models such as BM25 or neural retrieval models based on learned sparse lexical representations such as uniCOIL. Previous experiments on the MS MARCO datasets have demonstrated the effectiveness of these methods, and they serve as baselines that are widely used by the community today. Following the recent release of the MS MARCO V2 passage and document ranking test collections, we have refreshed our doc2query and uniCOIL models. This work describes a number of resources that support competitive, reproducible baselines for both the MS MARCO V1 and V2 test collections using our Anserini and Pyserini IR toolkits. Together, they provide a solid foundation for future research on neural retrieval models using the MS MARCO datasets and beyond.

In this work, we propose an effective multi-stage neural ranking system for the clinical trial matching problem. First, we introduce NQS, a neural query synthesis method that leverages a zero-shot document expansion model to generate multiple sentence-long queries from lengthy patient descriptions. These queries are independently issued to a search engine and the results are fused. We find that on the TREC 2021 Clinical Trials Track, this method outperforms strong traditional baselines like BM25 and BM25 + RM3 by about 12 points in [email protected], a relative improvement of 34%. This simple method is so effective that even a state-of-the-art neural relevance ranking method trained on the medical subset of MS MARCO passage, when reranking the results of NQS, fails to improve on the ranked list. Second, we introduce a two-stage neural reranking pipeline trained on clinical trial matching data using tailored ranking templates. In this setting, we can train a pointwise reranker using just 1.1k positive examples and obtain effectiveness improvements over NQS by 24 points. This end-to-end multi-stage system demonstrates a 20% relative effectiveness gain compared to the second-best submission at TREC 2021, making it an important step towards better automated clinical trial matching.

Information retrieval has traditionally been framed in terms of searching and extracting information from mostly static resources. Interactive information retrieval (IIR) has widened the scope, with interactive dialogues largely playing the role of clarifying (i.e., making explicit, and/or refining) the information search space. Informed by market research practices, we seek to reframe IIR as a process of eliciting novel information from human interlocutors, with a chatbot-inspired virtual agent playing the role of an interviewer. This reframing flips conventional IIR into what we call an inverse information seeking dialogue, wherein the virtual agent recurrently extracts information from human utterances and poses questions intended to elicit related information. In this work, we introduce and provide a formal definition of an inverse information seeking agent, outline some of its unique challenges, and propose our novel framework to tackle this problem based on techniques from natural language processing (NLP) and IIR.

Document-at-a-time (DaaT) and score-at-a-time (SaaT) query evaluation techniques are different approaches to top-k retrieval with inverted indexes. While modern systems are dominated by DaaT, the academic literature has seen decades of debate about the merits of each. Recently, there has been renewed interest in SaaT methods for learned sparse lexical models, where studies have shown that transformers generate "wacky weights" that appear to reduce opportunities for optimizations in DaaT methods. However, researchers currently lack an easy-to-use SaaT system to support further exploration. This is the gap that our work fills. Starting with a modern SaaT system (JASS), we built Python bindings in order to integrate into the DaaT Pyserini IR toolkit (Lucene). The result is a common frontend to both a DaaT and a SaaT system. We demonstrate how recent experiments with a wide range of learned sparse lexical models can be easily reproduced. Our contribution is a framework that enables future research comparing DaaT and SaaT methods in the context of modern neural retrieval models.

This paper presents the lessons regarding the construction and use of large Cranfield-style test collections learned from the TREC 2021 Deep Learning track. The corpus used in the 2021 edition of the track was much bigger than the corpus used previously and it contains many more relevant documents. The process used to select documents to judge that had been used in earlier years of the track failed to produce a reliable collection because most topics have too many relevant documents. Judgment budgets were exceeded before an adequate sample of the relevant set could be found, so there are likely many unknown relevant documents in the unjudged portion of the corpus. As a result, the collection is not reusable, and furthermore, recall-based measures are unreliable even for the retrieval systems that were used to build the collection. Yet, early-precision measures cannot distinguish among system results because the maximum score is easily obtained for many topics. And since the existing tools for appraising the quality of test collections depend on systems' scores, they also fail when there are too many relevant documents. Collection builders will need new strategies and tools for building reliable test collections for continued use of the Cranfield paradigm on ever-larger corpora. Ensuring that the definition of 'relevant' truly reflects the desired systems' rankings is a provisional strategy for continued collection building.

We demonstrate, via competitive bag-of-words first-stage retrieval baselines for the MS MARCO document ranking task, seamless replicability and interoperability between Elasticsearch and the Pyserini IR toolkit, which are both built on the open-source Lucene search library. This integration highlights the benefits of recent efforts to promote the use of Lucene in information retrieval research to better align the research and practice of building search applications. Closer alignment between academia and industry is mutually beneficial: Academic researchers gain a smoother path to real-world impact because their contributions can be more easily deployed in production applications. Industry practitioners gain an easy way to benchmark their innovations in a rigorous and vendor-neutral manner by exploiting evaluation resources and infrastructure built by the academic community. This two-way exchange between academia and industry allows both parties to "have their cakes and eat them too".

The acceleration of climate change and its impact highlight the need for long-term reliable climate data at high spatiotemporal resolution to answer key science questions in cold regions hydrology. Prior to the digital age, climate records were archived on paper. For example, from the 1950s to the 1990s, solar radiation data from recording stations worldwide were published in booklets by the former Union of Soviet Socialist Republics (USSR) Hydrometeorological Service. As a result, the data are not easily accessible by most researchers. The overarching aim of this research is to develop techniques to convert paper-based climate records into a machine-readable format to support environmental research in cold regions. This study compares the performance of a proprietary optical character recognition (OCR) service with an open-source OCR tool for digitizing hydrometeorological data. We built a digitization pipeline combining different image preprocessing techniques, semantic segmentation, and an open-source OCR engine for extracting data and metadata recorded in the scanned documents. Each page contains blocks of text with station names and tables containing the climate data. The process begins with image preprocessing to reduce noise and to improve quality before the page content is segmented to detect tables and finally run through an OCR engine for text extraction. We outline the digitization process and report on initial results, including different segmentation approaches, preprocessing image algorithms, and OCR techniques to ensure accurate extraction and organization of relevant metadata from thousands of scanned climate records. We evaluated the performance of Tesseract OCR and ABBYY FineReader on text extraction. We find that although ABBY FineReader has better accuracy on the sample data, our custom extraction pipeline using Tesseract is efficient and scalable because it is flexible and allows for more customization.

Evaluation efforts such as TREC, CLEF, NTCIR and FIRE, alongside public leaderboard such as MS MARCO, are intended to encourage research and track our progress, addressing big questions in our field. However, the goal is not simply to identify which run is "best", achieving the top score. The goal is to move the field forward by developing new robust techniques, that work in many different settings, and are adopted in research and practice. This paper uses the MS MARCO and TREC Deep Learning Track as our case study, comparing it to the case of TREC ad hoc ranking in the 1990s. We show how the design of the evaluation effort can encourage or discourage certain outcomes, and raising questions about internal and external validity of results. We provide some analysis of certain pitfalls, and a statement of best practices for avoiding such pitfalls. We summarize the progress of the effort so far, and describe our desired end state of "robust usefulness", along with steps that might be required to get us there.

A vital step towards the widespread adoption of neural retrieval models is their resource efficiency throughout the training, indexing and query workflows. The neural IR community made great advancements in training effective dual-encoder dense retrieval (DR) models recently. A dense text retrieval model uses a single vector representation per query and passage to score a match, which enables low-latency first-stage retrieval with a nearest neighbor search. Increasingly common, training approaches require enormous compute power, as they either conduct negative passage sampling out of a continuously updating refreshing index or require very large batch sizes. Instead of relying on more compute capability, we introduce an efficient topic-aware query and balanced margin sampling technique, called TAS-Balanced. We cluster queries once before training and sample queries out of a cluster per batch. We train our lightweight 6-layer DR model with a novel dual-teacher supervision that combines pairwise and in-batch negative teachers. Our method is trainable on a single consumer-grade GPU in under 48 hours. We show that our TAS-Balanced training method achieves state-of-the-art low-latency (64ms per query) results on two TREC Deep Learning Track query sets. Evaluated on [email protected], we outperform BM25 by 44%, a plainly trained DR by 19%, docT5query by 11%, and the previous best DR model by 5%. Additionally, TAS-Balanced produces the first dense retriever that outperforms every other method on recall at any cutoff on TREC-DL and allows more resource intensive re-ranking models to operate on fewer passages to improve results further.

Leaderboards are a ubiquitous part of modern research in applied machine learning. By design, they sort entries into some linear order, where the top-scoring entry is recognized as the "state of the art" (SOTA). Due to the rapid progress being made today, particularly with neural models, the top entry in a leaderboard is replaced with some regularity. These are touted as improvements in the state of the art. Such pronouncements, however, are almost never qualified with significance testing. In the context of the MS MARCO document ranking leaderboard, we pose a specific question: How do we know if a run is significantly better than the current SOTA? Against the backdrop of recent IR debates on scale types, our study proposes an evaluation framework that explicitly treats certain outcomes as distinct and avoids aggregating them into a single-point metric. Empirical analysis of SOTA runs from the MS MARCO document ranking leaderboard reveals insights about how one run can be "significantly better" than another that are obscured by the current official evaluation metric ([email protected]).

Pyserini is a Python toolkit for reproducible information retrieval research with sparse and dense representations. It aims to provide effective, reproducible, and easy-to-use first-stage retrieval in a multi-stage ranking architecture. Our toolkit is self-contained as a standard Python package and comes with queries, relevance judgments, pre-built indexes, and evaluation scripts for many commonly used IR test collections. We aim to support, out of the box, the entire research lifecycle of efforts aimed at improving ranking with modern neural approaches. In particular, Pyserini supports sparse retrieval (e.g., BM25 scoring using bag-of-words representations), dense retrieval (e.g., nearest-neighbor search on transformer-encoded representations), as well as hybrid retrieval that integrates both approaches. This paper provides an overview of toolkit features and presents empirical results that illustrate its effectiveness on two popular ranking tasks. Around this toolkit, our group has built a culture of reproducibility through shared norms and tools that enable rigorous automated testing.

The COVID-19 pandemic has brought about a proliferation of harmful news articles online, with sources lacking credibility and misrepresenting scientific facts. Misinformation has real consequences for consumer health search, i.e., users searching for health information. In the context of multi-stage ranking architectures, there has been little work exploring whether they prioritize correct and credible information over misinformation. We find that, indeed, training models on standard relevance ranking datasets like MS MARCO passage---which have been curated to contain mostly credible information---yields models that might also promote harmful misinformation. To rectify this, we propose a label prediction technique that can separate helpful from harmful content. Our design leverages pretrained sequence-to-sequence transformer models for both relevance ranking and label prediction. Evaluated at the TREC 2020 Health Misinformation Track, our techniques represent the top-ranked system: Our best submitted run was 19.2 points higher than the second-best run based on the primary metric, a 68% relative improvement. Additional post-hoc experiments show that we can boost effectiveness by another 3.5 points.

The goal of text ranking is to generate an ordered list of texts retrieved from a corpus in response to a query. Although the most common formulation of text ranking is search, instances of the task can also be found in many natural language processing applications. This tutorial, based on a forthcoming book, provides an overview of text ranking with neural network architectures known as transformers, of which BERT is the best-known example. The combination of transformers and self-supervised pretraining has, without exaggeration, revolutionized the fields of natural language processing (NLP), information retrieval (IR), and beyond. We provide a synthesis of existing work as a single point of entry for both researchers and practitioners. Our coverage is grouped into two categories: transformer models that perform reranking in multi-stage ranking architectures and learned dense representations that perform ranking directly. Two themes pervade our treatment: techniques for handling long documents and techniques for addressing the tradeoff between effectiveness (result quality) and efficiency (query latency). Although transformer architectures and pretraining techniques are recent innovations, many aspects of their application are well understood. Nevertheless, there remain many open research questions, and thus in addition to laying out the foundations of pretrained transformers for text ranking, we also attempt to prognosticate the future.

Chatty Goose is an open-source Python conversational search framework that provides strong, reproducible reranking pipelines built on recent advances in neural models. The framework comprises extensible modular components that integrate with popular libraries such as Transformers by HuggingFace and ParlAI by Facebook. Our aim is to lower the barrier of entry for research in conversational search by providing reproducible baselines that researchers can build on top of. We provide an overview of the framework and demonstrate how to instantiate a new system from scratch. Chatty Goose incorporates improvements to components that we introduced in the TREC 2019 Conversational Assistance Track (CAsT), where our submission represented the top-performing system. Using our framework, a comparable run can be reproduced with just a few lines of code.

Mathematical Information Retrieval (MIR) has been actively studied in recent years and many fruitful results have emerged. Among those, the Approach Zero system is one of the few math-aware search engines that is able to perform substructure matching efficiently. Furthermore, it has been deployed in ARQMath2020, the most recent community-wide MIR evaluation, as a strong baseline due to its empirical effectiveness and ability to handle structured math content. However, in order to implement a retrieval model that handles structured queries efficiently, Approach Zero is written in C from the ground up, requiring special pipelines for processing math content and queries. Thus, the system is not conveniently accessible and reusable to the community as a research tool. In this paper, we present PyA0, an easy-to-use Python toolkit built on Approach Zero that improves its accessibility to researchers. We introduce the toolkit interface and report evaluation results on popular MIR datasets to demonstrate the effectiveness and efficiency of our toolkit. We have made PyA0 source code publicly accessible at https://github.com/approach0/pya0, which includes a link to a notebook demo.

The goal of text ranking is to generate an ordered list of texts retrieved from a corpus in response to a query. Although the most common formulation of text ranking is search, instances of the task can also be found in many natural language processing applications. This tutorial, based on a forthcoming book, provides an overview of text ranking with neural network architectures known as transformers, of which BERT is the best-known example. The combination of transformers and self-supervised pretraining has, without exaggeration, revolutionized the fields of natural language processing (NLP), information retrieval (IR), and beyond. We provide a synthesis of existing work as a single point of entry for both researchers and practitioners. Our coverage is grouped into two categories: transformer models that perform reranking in multi-stage ranking architectures and learned dense representations that perform ranking directly. Two themes pervade our treatment: techniques for handling long documents and techniques for addressing the tradeoff between effectiveness (result quality) and efficiency (query latency). Although transformer architectures and pretraining techniques are recent innovations, many aspects of their application are well understood. Nevertheless, there remain many open research questions, and thus in addition to laying out the foundations of pretrained transformers for text ranking, we also attempt to prognosticate the future.

Conversational search plays a vital role in conversational information seeking. As queries in information seeking dialogues are ambiguous for traditional ad hoc information retrieval (IR) systems due to the coreference and omission resolution problems inherent in natural language dialogue, resolving these ambiguities is crucial. In this article, we tackle conversational passage retrieval, an important component of conversational search, by addressing query ambiguities with query reformulation integrated into a multi-stage ad hoc IR system. Specifically, we propose two conversational query reformulation (CQR) methods: (1) term importance estimation and (2) neural query rewriting. For the former, we expand conversational queries using important terms extracted from the conversational context with frequency-based signals. For the latter, we reformulate conversational queries into natural, stand-alone, human-understandable queries with a pretrained sequence-to-sequence model. Detailed analyses of the two CQR methods are provided quantitatively and qualitatively, explaining their advantages, disadvantages, and distinct behaviors. Moreover, to leverage the strengths of both CQR methods, we propose combining their output with reciprocal rank fusion, yielding state-of-the-art retrieval effectiveness, 30% improvement in terms of NDCG@3 compared to the best submission of Text REtrieval Conference (TREC) Conversational Assistant Track (CAsT) 2019.

We live in an era in which the ways that we can make sense of our past are evolving as more artifacts from that past become digital. At the same time, the responsibilities of traditional gatekeepers who have negotiated the ethics of historical data collection and use, such as librarians and archivists, are increasingly being sidelined by the system builders who decide whether and how to provide access to historical digital collections, often without sufficient reflection on the ethical issues at hand. It is our aim to better prepare system builders to grapple with these issues. This paper focuses discussions around one such digital collection from the dawn of the web, asking what sorts of analyses can and should be conducted on archival copies of the GeoCities web hosting platform that dates to 1994.

Real-time summarization systems monitor continuous streams of documents with the goal of delivering relevant, novel, and timely updates to users. These updates can either be sent to users' mobile devices as push notifications or be silently deposited in an inbox to be consumed - the important difference is whether the user is interrupted by the delivery. Previously, a two-year study examining user attention under these different mechanisms revealed interesting findings about users' information consumption behavior, but the conclusions were marred by a few methodological shortcomings. We present a reproducibility study that follows the same design as the original evaluation, but corrects its flaws. We find that most conclusions from the original study are confirmed, although there are some surprising differences as well. Overall, the magnitude of the observed effects are not as strong as in the original study.

While a number of recent open-source toolkits for training and using neural information retrieval models have greatly simplified experiments with neural reranking methods, they essentially hard code a "search-then-rerank'' experimental pipeline. These pipelines consist of an efficient first-stage ranking method, like BM25, followed by a neural reranking method. Deviations from this setup often require hacks; some improvements, like adding a second reranking step that uses a more expensive neural method, are infeasible without major code changes. In order to improve the flexibility of such toolkits, we propose implementing experimental pipelines as dependency graphs of functional "IR primitives,'' which we call modules, that can be used and combined as needed.

Personalized recommendation products at Twitter target a multitude of heterogeneous items: Tweets, Events, Topics, Hashtags, and users. Each of these targets varies in their cardinality (which affects the scale of the problem) and their "shelf life'' (which constrains the latency of generating the recommendations). Although Twitter has built a variety of recommendation systems before dating back a decade, solutions to the broader problem were mostly tackled piecemeal. In this paper, we present SimClusters, a general-purpose representation layer based on overlapping communities into which users as well as heterogeneous content can be captured as sparse, interpretable vectors to support a multitude of recommendation tasks. We propose a novel algorithm for community discovery based on Metropolis-Hastings sampling, which is both more accurate and significantly faster than off-the-shelf alternatives. SimClusters scales to networks with billions of users and has been effective across a variety of deployed applications at Twitter.

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.

Tools, computing environments, and datasets form the three critical ingredients for teaching and learning the practical aspects of experimental IR research. Assembling these ingredients can often be challenging, particularly in the context of short courses that cannot afford large startup costs. As an initial attempt to address these issues, we describe materials that we have developed for the "Introduction to IR" session at the ACM SIGIR/SIGKDD Africa Summer School on Machine Learning for Data Mining and Search (AFIRM 2020), which builds on three components: the open-source Lucene search library, cloud-based notebooks, and the MS MARCO dataset. We offer a self-reflective evaluation of our efforts and hope that our lessons shared can benefit future efforts.

We present Capreolus, a toolkit designed to facilitate end-to-end it ad hoc retrieval experiments with neural networks by providing implementations of prominent neural ranking models within a common framework. Our toolkit adopts a standard reranking architecture via tight integration with the Anserini toolkit for candidate document generation using standard bag-of-words approaches. Using Capreolus, we are able to reproduce Yang et al.'s recent SIGIR 2019 finding that, in a reranking scenario on the test collection from the TREC 2004 Robust Track, many neural retrieval models do not significantly outperform a strong query expansion baseline. Furthermore, we find that this holds true for five additional models implemented in Capreolus. We describe the architecture and design of our toolkit, which includes a Web interface to facilitate comparisons between rankings returned by different models.

We tackle the problem of question answering directly on a large document collection, combining simple “bag of words” passage retrieval with a BERT-based reader for extracting answer spans. In the context of this architecture, we present a data augmentation technique using distant supervision to automatically annotate paragraphs as either positive or negative examples to supplement existing training data, which are then used together to fine-tune BERT. We explore a number of details that are critical to achieving high accuracy in this setup: the proper sequencing of different datasets during fine-tuning, the balance between “difficult” vs. “easy” examples, and different approaches to gathering negative examples. Experimental results show that, with the appropriate settings, we can achieve large gains in effectiveness on two English and two Chinese QA datasets. We are able to achieve results at or near the state of the art without any modeling advances, which once again affirms the cliché “there’s no data like more data”.

The importance of repeatability, replicability, and reproducibility is broadly recognized in the computational sciences, both in supporting desirable scientific methodology as well as sustaining empirical progress. This workshop tackles the replicability challenge for ad hoc document retrieval, via a common Docker interface specification to support images that capture systems performing ad hoc retrieval experiments on standard test collections.

Anserini is an open-source information retrieval toolkit built around Lucene to facilitate replicable research. In this demonstration, we examine different architectures for Solr integration in order to address two current limitations of the system: the lack of an interactive search interface and support for distributed retrieval. Two architectures are explored: In the first approach, Anserini is used as a frontend to index directly into a running Solr instance. In the second approach, Lucene indexes built directly with Anserini can be copied into a Solr installation and placed under its management. We discuss the tradeoffs associated with each architecture and report the results of a performance evaluation comparing indexing throughput. To illustrate the additional capabilities enabled by Anserini/Solr integration, we present a search interface built using the open-source Blacklight discovery interface.

Despite the broad adoption of both Apache Spark and Apache Solr, there is little integration between these two platforms to support scalable, end-to-end text analytics. We believe this is a missed opportunity, as there is substantial synergy in building analytical pipelines where the results of potentially complex faceted queries feed downstream text processing components. This demonstration explores exactly such an integration: we evaluate performance under different analytical scenarios and present three simple case studies that illustrate the range of possible analyses enabled by seamlessly connecting Spark to Solr.

Document ranking experiments should be repeatable. However, the interaction between multi-threaded indexing and score ties during retrieval may yield non-deterministic rankings, making repeatability not as trivial as one might imagine. In the context of the open-source Lucene search engine, score ties are broken by internal document ids, which are assigned at index time. Due to multi-threaded indexing, which makes experimentation with large modern document collections practical, internal document ids are not assigned consistently between different index instances of the same collection, and thus score ties are broken unpredictably. This short paper examines the effectiveness impact of such score ties, quantifying the variability that can be attributed to this phenomenon. The obvious solution to this non-determinism and to ensure repeatable document ranking is to break score ties using external collection document ids. This approach, however, comes with measurable efficiency costs due to the necessity of consulting external identifiers during query evaluation.

Millions of consumers issue voice queries through television-based entertainment systems such as the Comcast X1, the Amazon Fire TV, and Roku TV. Automatic speech recognition (ASR) systems are responsible for transcribing these voice queries into text to feed downstream natural language understanding modules. However, ASR is far from perfect, often producing incorrect transcriptions and forcing users to take corrective action. To better understand their impact on sessions, this paper characterizes speech recognition errors as well as subsequent user responses. We provide both quantitative and qualitative analyses, examining the acoustic as well as lexical attributes of the utterances. This work represents, to our knowledge, the first analysis of speech recognition errors from real users on a widely-deployed entertainment system.

Modern in-home entertainment platforms---representing the evolution of the humble television of yesteryear---are packed with features and content: they offer a dizzying array of programs spanning hundreds of channels as well as a catalog of on-demand programs offering tens of thousands of options. Furthermore, the entertainment platform may serve as an in-home hub, providing capabilities ranging from playing music to controlling the home security system. At a high level, our goal is to provide natural speech-based access to these myriad features as an alternative to physical button entry on a remote control.

Is neural IR mostly hype? In a recent SIGIR Forum article, Lin expressed skepticism that neural ranking models were actually improving ad hoc retrieval effectiveness in limited data scenarios. He provided anecdotal evidence that authors of neural IR papers demonstrate "wins" by comparing against weak baselines. This paper provides a rigorous evaluation of those claims in two ways: First, we conducted a meta-analysis of papers that have reported experimental results on the TREC Robust04 test collection. We do not find evidence of an upward trend in effectiveness over time. In fact, the best reported results are from a decade ago and no recent neural approach comes close. Second, we applied five recent neural models to rerank the strong baselines that Lin used to make his arguments. A significant improvement was observed for one of the models, demonstrating additivity in gains. While there appears to be merit to neural IR approaches, at least some of the gains reported in the literature appear illusory.

We tackle the challenge of understanding voice queries posed against the Comcast Xfinity X1 entertainment platform, where consumers direct speech input at their "voice remotes". Such queries range from specific program navigation (i.e., watch a movie) to requests with vague intents and even queries that have nothing to do with watching TV. We present successively richer neural network architectures to tackle this challenge based on two key insights: The first is that session context can be exploited to disambiguate queries and recover from ASR errors, which we operationalize with hierarchical recurrent neural networks. The second insight is that query understanding requires evidence integration across multiple related tasks, which we identify as program prediction, intent classification, and query tagging. We present a novel multi-task neural architecture that jointly learns to accomplish all three tasks. Our initial model, already deployed in production, serves millions of queries daily with an improved customer experience. The novel multi-task learning model, first described here, is evaluated through carefully-controlled laboratory experiments, which demonstrates further gains in effectiveness and increased system capabilities.

We present a broad overview of personalized content recommendations at Twitter, discussing how our approach has evolved over the years, represented by several generations of systems. Historically, content analysis of Tweets has not been a priority, and instead engineering efforts have focused on graph-based recommendation techniques that exploit structural properties of the follow graph and engagement signals from users. These represent "low hanging fruits" that have enabled high-quality recommendations using simple algorithms. As deployed systems have grown in maturity and our understanding of the problem space has become more refined, we have begun to look for other opportunities to further improve recommendation quality. We overview recent investments in content analysis, particularly named-entity recognition techniques built around recurrent neural networks, and discuss how they integrate with existing graph-based capabilities to open up the design space of content recommendation algorithms.

Real-time summarization systems that monitor document streams to identify relevant content have a few options for delivering system updates to users. In a mobile context, systems could send push notifications to users' mobile devices, hoping to grab their attention immediately. Alternatively, systems could silently deposit updates into "inboxes" that users can access at their leisure. We refer to these mechanisms as push-based vs. pull-based, and present a two-year contrastive study that attempts to understand the effects of the delivery mechanism on mobile user behavior, in the context of the TREC Real-Time Summarization Tracks. Through a cluster analysis, we are able to identify three distinct and coherent patterns of behavior. As expected, we find that users are likely to ignore push notifications, but for those updates that users do pay attention to, content is consumed within a short amount of time. Interestingly, users bombarded with push notifications are less likely to consume updates on their own initiative and less likely to engage in long reading sessions---which is a common pattern for users who pull content from their inboxes. We characterize users as exhibiting "eager" or "apathetic" information consumption behavior as an explanation of these observations, and attempt to operationalize our findings into design recommendations.

A recently-introduced product of Comcast, a large cable company in the United States, is a "voice remote" that accepts spoken queries from viewers. We present an analysis of a large query log from this service to answer the question: "What do viewers say to their TVs?" In addition to a descriptive characterization of queries and sessions, we describe two complementary types of analyses to support query understanding. First, we propose a domain-specific intent taxonomy to characterize viewer behavior: as expected, most intents revolve around watching programs---both direct navigation as well as browsing---but there is a non-trivial fraction of non-viewing intents as well. Second, we propose a domain-specific tagging scheme for labeling query tokens, that when combined with intent and program prediction, provides a multi-faceted approach to understand voice queries directed at entertainment systems.

Twitter's data engineering team is faced with the challenge of processing billions of events every day in batch and in real time, and we have built various tools to meet these demands. In this paper, we describe TSAR (TimeSeries AggregatoR), a robust, scalable, real-time event time series aggregation framework built primarily for engagement monitoring: aggregating interactions with Tweets, segmented along a multitude of dimensions such as device, engagement type, etc. TSAR is built on top of Summingbird, an open-source framework for integrating batch and online MapReduce computations, and removes much of the tedium associated with building end-to-end aggregation pipelines---from the ingestion and processing of events to the publication of results in heterogeneous datastores. Clients are provided a query interface that powers dashboards and supports downstream ad hoc analytics.

Large scale retrieval systems often employ cascaded ranking architectures, in which an initial set of candidate documents are iteratively refined and re-ranked by increasingly sophisticated and expensive ranking models. In this paper, we propose a unified framework for predicting a range of performance-sensitive parameters based on minimizing end-to-end effectiveness loss. The framework does not require relevance judgments for training, is amenable to predicting a wide range of parameters, allows for fine tuned efficiency-effectiveness trade-offs, and can be easily deployed in large scale search systems with minimal overhead. As a proof of concept, we show that the framework can accurately predict a number of performance parameters on a query-by-query basis, allowing efficient and effective retrieval, while simultaneously minimizing the tail latency of an early-stage candidate generation system. On the 50 million document ClueWeb09B collection, and across 25,000 queries, our hybrid system can achieve superior early-stage efficiency to fixed parameter systems without loss of effectiveness, and allows more finely-grained efficiency-effectiveness trade-offs across the multiple stages of the retrieval system.

There is growing interest in systems that generate timeline summaries by filtering high-volume streams of documents to retain only those that are relevant to a particular event or topic. Continued advances in algorithms and techniques for this task depend on standardized and reproducible evaluation methodologies for comparing systems. However, timeline summary evaluation is still in its infancy, with competing methodologies currently being explored in international evaluation forums such as TREC. One area of active exploration is how to explicitly represent the units of information that should appear in a "good" summary. Currently, there are two main approaches, one based on identifying nuggets in an external "ground truth", and the other based on clustering system outputs. In this paper, by building test collections that have both nugget and cluster annotations, we are able to compare these two approaches. Specifically, we address questions related to evaluation effort, differences in the final evaluation products, and correlations between scores and rankings generated by both approaches. We summarize advantages and disadvantages of nuggets and clusters to offer recommendations for future system evaluations.

We tackle the novel problem of navigational voice queries posed against an entertainment system, where viewers interact with a voice-enabled remote controller to specify the TV program to watch. This is a difficult problem for several reasons: such queries are short, even shorter than comparable voice queries in other domains, which offers fewer opportunities for deciphering user intent. Furthermore, ambiguity is exacerbated by underlying speech recognition errors. We address these challenges by integrating word- and character-level query representations and by modeling voice search sessions to capture the contextual dependencies in query sequences. Both are accomplished with a probabilistic framework in which recurrent and feedforward neural network modules are organized in a hierarchical manner. From a raw dataset of 32M voice queries from 2.5M viewers on the Comcast Xfinity X1 entertainment system, we extracted data to train and test our models. We demonstrate the benefits of our hybrid representation and context-aware model, which significantly outperforms competitive baselines that use learning to rank as well as neural networks.

As an empirical discipline, information access and retrieval research requires substantial software infrastructure to index and search large collections. This workshop is motivated by the desire to better align information retrieval research with the practice of building search applications from the perspective of open-source information retrieval systems. Our goal is to promote the use of Lucene for information access and retrieval research.

We present a system for identifying interesting social media posts on Twitter and delivering them to users' mobile devices in real time as push notifications. In our problem formulation, users are interested in broad topics such as politics, sports, and entertainment: our system processes tweets in real time to identify relevant, novel, and salient content. There are three interesting aspects to our work: First, instead of attempting to tame the cacophony of unfiltered tweets, we exploit a smaller, but still sizeable, collection of curated tweet streams corresponding to the Twitter accounts of different media outlets. Second, we apply distant supervision to extract topic labels from curated streams that have a specific focus, which can then be leveraged to build high-quality topic classifiers essentially "for free". Finally, our system delivers content via Twitter direct messages, supporting in situ interactions modeled after conversations with intelligent agents. These ideas are demonstrated in an end-to-end working prototype.

In recent years, neural networks have been applied to many text processing problems. One example is learning a similarity function between pairs of text, which has applications to paraphrase extraction, plagiarism detection, question answering, and ad hoc retrieval. Within the information retrieval community, the convolutional neural network model proposed by Severyn and Moschitti in a SIGIR 2015 paper has gained prominence. This paper focuses on the problem of answer selection for question answering: we attempt to replicate the results of Severyn and Moschitti using their open-source code as well as to reproduce their results via a de novo (i.e., from scratch) implementation using a completely different deep learning toolkit. Our de novo implementation is instructive in ascertaining whether reported results generalize across toolkits, each of which have their idiosyncrasies. We were able to successfully replicate and reproduce the reported results of Severyn and Moschitti, albeit with minor differences in effectiveness, but affirming the overall design of their model. Additional ablation experiments break down the components of the model to show their contributions to overall effectiveness. Interestingly, we find that removing one component actually increases effectiveness and that a simplified model with only four word overlap features performs surprisingly well, even better than convolution feature maps alone.

Real-time push notification systems monitor continuous document streams such as social media posts and alert users to relevant content directly on their mobile devices. We describe a user study of such systems in the context of the TREC 2016 Real-Time Summarization Track, where system updates are immediately delivered as push notifications to the mobile devices of a cohort of users. Our study represents, to our knowledge, the first deployment of an interleaved evaluation framework for prospective information needs, and also provides an opportunity to examine user behavior in a realistic setting. Results of our online in-situ evaluation are correlated against the results a more traditional post-hoc batch evaluation. We observe substantial correlations between many online and batch evaluation metrics, especially for those that share the same basic design (e.g., are utility-based). For some metrics, we observe little correlation, but are able to identify the volume of messages that a system pushes as one major source of differences.

Due to Twitter's terms of service that forbid redistribution of content, creating publicly downloadable collections of tweets for research purposes has been a perpetual problem for the research community. Some collections are distributed by making available the ids of the tweets that comprise the collection and providing tools to fetch the actual content; this approach has scalability limitations. In other cases, evaluation organizers have set up APIs that provide access to collections for specific tasks, without exposing the underlying content. This is a workable solution, but difficult to sustain over the long term since someone has to maintain the APIs. We have noticed that the non-profit Internet Archive has been making available for public download captures of the so-called Twitter "spritzer" stream, which is the same source as the Tweets2013 collection used in the TREC 2013 and 2014 Microblog Tracks. We analyzed both datasets in terms of content overlap and retrieval baselines to show that the Internet Archive data can serve as a drop-in replacement for the Tweets2013 collection, thereby providing the research community with, finally, a downloadable collection of tweets. Beyond this finding, we also study the impact of tweet deletions over time and how they affect the test collections.

Information retrieval test collections are typically built using data from large-scale evaluations in international forums such as TREC, CLEF, and NTCIR. Previous validation studies on pool-based test collections for ad hoc retrieval have examined their reusability to accurately assess the effectiveness of systems that did not participate in the original evaluation. To our knowledge, the reusability of test collections derived from "living labs" evaluations, based on logs of user activity, has not been explored. In this paper, we performed a "leave-one-out" analysis of human judgment data derived from the TREC 2016 Real-Time Summarization Track and show that those judgments do not appear to be reusable. While this finding is limited to one specific evaluation, it does call into question the reusability of test collections built from living labs in general, and at the very least suggests the need for additional work in validating such experimental instruments.

Software toolkits play an essential role in information retrieval research. Most open-source toolkits developed by academics are designed to facilitate the evaluation of retrieval models over standard test collections. Efforts are generally directed toward better ranking and less attention is usually given to scalability and other operational considerations. On the other hand, Lucene has become the de facto platform in industry for building search applications (outside a small number of companies that deploy custom infrastructure). Compared to academic IR toolkits, Lucene can handle heterogeneous web collections at scale, but lacks systematic support for evaluation over standard test collections. This paper introduces Anserini, a new information retrieval toolkit that aims to provide the best of both worlds, to better align information retrieval practice and research. Anserini provides wrappers and extensions on top of core Lucene libraries that allow researchers to use more intuitive APIs to accomplish common research tasks. Our initial efforts have focused on three functionalities: scalable, multi-threaded inverted indexing to handle modern web-scale collections, streamlined IR evaluation for ad hoc retrieval on standard test collections, and an extensible architecture for multi-stage ranking. Anserini ships with support for many TREC test collections, providing a convenient way to replicate competitive baselines right out of the box. Experiments verify that our system is both efficient and effective, providing a solid foundation to support future research.

We propose a heuristic called "one answer per document" for automatically extracting high-quality negative examples for answer selection in question answering. Starting with a collection of question-answer pairs from the popular TrecQA dataset, we identify the original documents from which the answers were drawn. Sentences from these source documents that contain query terms (aside from the answers) are selected as negative examples. Training on the original data plus these negative examples yields improvements in effectiveness by a margin that is comparable to successive recent publications on this dataset. Our technique is completely unsupervised, which means that the gains come essentially for free. We confirm that the improvements can be directly attributed to our heuristic, as other approaches to extracting comparable amounts of training data are not effective. Beyond the empirical validation of this heuristic, we also share our improved TrecQA dataset with the community to support further work in answer selection.

This demonstration explores the novel and unconventional idea of implementing an analytical RDBMS in pure JavaScript so that it runs completely inside a browser with no external dependencies. Our prototype, called Afterburner, generates compiled query plans that exploit two JavaScript features: typed arrays and asm.js. On the TPC-H benchmark, we show that Afterburner achieves comparable performance to MonetDB running natively on the same machine. This is an interesting finding in that it shows how far JavaScript has come as an efficient execution platform. Beyond a mere technical curiosity, we demonstrate how our techniques can support interactive data exploration by automatically generating materialized views from a backend that is then shipped to the browser to facilitate subsequent interactions seamlessly and efficiently.

We present an empirical comparison between document-at-a-time (DaaT) and score-at-a-time (SaaT) document ranking strategies within a common framework. Although both strategies have been extensively explored, the literature lacks a fair, direct comparison: such a study has been difficult due to vastly different query evaluation mechanics and index organizations. Our work controls for score quantization, document processing, compression, implementation language, implementation effort, and a number of details, arriving at an empirical evaluation that fairly characterizes the performance of three specific techniques: WAND (DaaT), BMW (DaaT), and JASS (SaaT). Experiments reveal a number of interesting findings. The performance gap between WAND and BMW is not as clear as the literature suggests, and both methods are susceptible to tail queries that may take orders of magnitude longer than the median query to execute. Surprisingly, approximate query evaluation in WAND and BMW does not significantly reduce the risk of these tail queries. Overall, JASS is slightly slower than either WAND or BMW, but exhibits much lower variance in query latencies and is much less susceptible to tail query effects. Furthermore, JASS query latency is not particularly sensitive to the retrieval depth, making it an appealing solution for performance-sensitive applications where bounds on query latencies are desirable.

The basic idea behind selective search is to partition a collection into topical clusters, and for each query, consider only a subset of the clusters that are likely to contain relevant documents. Previous work on web collections has shown that it is possible to retain high-quality results while considering only a small fraction of the collection. These studies, however, assume static collections where it is feasible to run batch clustering algorithms for partitioning. In this work, we consider the novel formulation of selective search on document streams (specifically, tweets), where partitioning must be performed incrementally. In our approach, documents are partitioned into temporal segments and selective search is performed within each segment: these segments can either be clustered using batch or online algorithms, and at different temporal granularities. For efficiency, we take advantage of word embeddings to reduce the dimensionality of the document vectors. Experiments with test collections from the TREC Microblog Tracks show that we are able to achieve precision indistinguishable from exhaustive search while considering only around 5% of the collection. Interestingly, we observe no significant effectiveness differences between batch vs. online clustering and between hourly vs. daily temporal segments, despite them being very different index organizations. This suggests that architectural choices should be primarily guided by efficiency considerations.

This paper explores a simple question: How would we provide a high-quality search experience on Mars, where the fundamental physical limit is speed-of-light propagation delays on the order of tens of minutes? On Earth, users are accustomed to nearly instantaneous responses from web services. Is it possible to overcome orders-of-magnitude longer latency to provide a tolerable user experience on Mars? In this paper, we formulate the searching from Mars problem as a tradeoff between "effort" (waiting for responses from Earth) and "data transfer" (pre-fetching or caching data on Mars). The contribution of our work is articulating this design space and presenting two case studies that explore the effectiveness of baseline techniques, using publicly available data from the TREC Total Recall and Sessions Tracks. We intend for this research problem to be aspirational as well as inspirational---even if one is not convinced by the premise of Mars colonization, there are Earth-based scenarios such as searching from rural villages in India that share similar constraints, thus making the problem worthy of exploration and attention from researchers.

Graph processing is becoming increasingly prevalent across many application domains. In spite of this prevalence, there is little research about how graphs are actually used in practice. We conducted an online survey aimed at understanding: (i) the types of graphs users have; (ii) the graph computations users run; (iii) the types of graph software users use; and (iv) the major challenges users face when processing their graphs. We describe the participants' responses to our questions highlighting common patterns and challenges. We further reviewed user feedback in the mailing lists, bug reports, and feature requests in the source repositories of a large suite of software products for processing graphs. Through our review, we were able to answer some new questions that were raised by participants' responses and identify specific challenges that users face when using different classes of graph software. The participants' responses and data we obtained revealed surprising facts about graph processing in practice. In particular, real-world graphs represent a very diverse range of entities and are often very large, and scalability and visualization are undeniably the most pressing challenges faced by participants. We hope these findings can guide future research.

Nugget-based evaluations, such as those deployed in the TREC Temporal Summarization and Question Answering tracks, require human assessors to determine whether a nugget is present in a given piece of text. This process, known as nugget annotation, is labor-intensive. In this paper, we present two active learning techniques that prioritize the sequence in which candidate nugget/sentence pairs are presented to an assessor, based on the likelihood that the sentence contains a nugget. Our approach builds on the recognition that nugget annotation is similar to high-recall retrieval, and we adapt proven existing solutions. Simulation experiments with four existing TREC test collections show that our techniques yield far more matches for a given level of effort than baselines that are typically deployed in previous nugget-based evaluations.

We study answer selection for question answering, in which given a question and a set of candidate answer sentences, the goal is to identify the subset that contains the answer. Unlike previous work which treats this task as a straightforward pointwise classification problem, we model this problem as a ranking task and propose a pairwise ranking approach that can directly exploit existing pointwise neural network models as base components. We extend the Noise-Contrastive Estimation approach with a triplet ranking loss function to exploit interactions in triplet inputs over the question paired with positive and negative examples. Experiments on TrecQA and WikiQA datasets show that our approach achieves state-of-the-art effectiveness without the need for external knowledge sources or feature engineering.

This work presents RTTBurst, an end-to-end system for ingesting descriptions of user interest profiles and discovering new and relevant tweets based on those interest profiles using a simple model for identifying bursts in token usage. Our approach differs from standard retrieval-based techniques in that it primarily focuses on identifying noteworthy moments in the tweet stream, and ?summarizes? those moments using selected tweets. We lay out the architecture of RTTBurst, our participation in and performance at the TREC 2015 Microblog track, and a method for combining and potentially improving existing TREC systems. Official results and post hoc experiments show that our simple targeted burst detection technique is competitive with existing systems. Furthermore, we demonstrate that our burst detection mechanism can be used to improve the performance of other systems for the same task.

We propose and validate a novel interleaved evaluation methodology for two complementary information seeking tasks on document streams: retrospective summarization and prospective notification. In the first, the user desires relevant and non-redundant documents that capture important aspects of an information need. In the second, the user wishes to receive timely, relevant, and non-redundant update notifications for a standing information need. Despite superficial similarities, interleaved evaluation methods for web ranking cannot be directly applied to these tasks; for example, existing techniques do not account for temporality or redundancy. Our proposed evaluation methodology consists of two components: a temporal interleaving strategy and a heuristic for credit assignment to handle redundancy. By simulating user interactions with interleaved results on submitted runs to the TREC 2014 tweet timeline generation (TTG) task and the TREC 2015 real-time filtering task, we demonstrate that our methodology yields system comparisons that accurately match the result of batch evaluations. Analysis further reveals weaknesses in current batch evaluation methodologies to suggest future directions for research.

We present an assessment platform for gathering online relevance judgments for mobile push notifications that will be deployed in the newly-created TREC 2016 Real-Time Summarization (RTS) track. There is emerging interest in building systems that filter social media streams such as tweets to identify interesting and novel content in real time, putatively for delivery to users' mobile phones. In our evaluation design, all participants subscribe to the Twitter streaming API to identify relevant tweets with respect to a set of interest profiles. As the systems generate results, they are pushed in real time to our evaluation broker via a REST API. The broker then "routes" the tweets to assessors who have installed a custom app on their mobile phones. We detail the design of this platform and discuss a number of challenges that need to be tackled in this type of "Living Labs" setup. It is our goal that such an evaluation design will mitigate any issues that have arisen in traditional batch-style evaluations of this type of task.

Push notifications from social media provide a method to keep up-to-date on topics of personal interest. To be effective, notifications must achieve a balance between pushing too much and pushing too little. Push too little and the user misses important updates; push too much and the user is overwhelmed by unwanted information. Using data from the TREC 2015 Microblog track, we explore simple dynamic emission strategies for microblog push notifications. The key to effective notifications lies in establishing and maintaining appropriate thresholds for pushing updates. We explore and evaluate multiple threshold setting strategies, including purely static thresholds, dynamic thresholds without user feedback, and dynamic thresholds with daily feedback. Our best technique takes advantage of daily feedback in a simple yet effective manner, achieving the best known result reported in the literature to date.

How do we evaluate systems that filter social media streams and send users updates via push notifications on their mobile phones? Such notifications must be relevant, timely, and novel. In this paper, we explore various evaluation metrics for this task, focusing specifically on measuring relevance. We begin with an analysis of metrics deployed at the TREC 2015 Microblog evaluations. A simple change to the metrics, reflecting a different assumption, dramatically alters system rankings. Applying another metric, previously used in the TREC Microblog evaluations, again yields different system rankings. We find little correlation between a number of "reasonable" evaluation metrics, which suggests that system effectiveness depends on how you measure it---an undesirable state in IR evaluation. However, we argue that existing evaluation metrics can be generalized into a framework that uses the same underlying contingency table, but places different weights and penalties. Although we stop short of proposing the "one true metric", this framework can guide the future development of a family of metrics that more accurately models user needs.

This paper tackles the challenge of accurately and efficiently estimating the number of relevant documents in a collection for a particular topic. One real-world application is estimating the volume of social media posts (e.g., tweets) pertaining to a topic, which is fundamental to tracking the popularity of politicians and brands, the potential sales of a product, etc. Our insight is to leverage active learning techniques to find all the "easy" documents, and then to use sampling techniques to infer the number of relevant documents in the residual collection. We propose a simple yet effective technique for determining this "switchover" point, which intuitively can be understood as the "knee" in an effort vs. recall gain curve, as well as alternative sampling strategies beyond the knee. We show on several TREC datasets and a collection of tweets that our best technique yields more accurate estimates (with the same effort) than several alternatives.

There is currently a large amount of publicly accessible structured data available as RDF data sets. For example, the Linked Open Data (LOD) cloud now consists of thousands of RDF data sets with over 30 billion triples, and the number and size of the data sets is continuously growing. Many of the data sets in the LOD cloud provide public SPARQL endpoints to allow issuing queries over them. These end-points enable users to retrieve data using precise and highly expressive SPARQL queries. However, in order to do so, the user must have sufficient knowledge about the data sets that she wishes to query, that is, the structure of data, the vocabulary used within the data set, the exact values of literals, their data types, etc. Thus, while SPARQL is powerful, it is not easy to use. An alternative to SPARQL that does not require as much prior knowledge of the data is some form of keyword search over the structured data. Keyword search queries are easy to use, but inherently ambiguous in describing structured queries. This demonstration introduces Sapphire, a system for querying RDF data that strikes a middle ground between ambiguous keyword search and difficult-to-use SPARQL. Our system does not replace either, but utilizes both where they are most effective. Sapphire helps the user construct expressive SPARQL queries that represent her information needs without requiring detailed knowledge about the queried data sets. These queries are then executed over public SPARQL endpoints from the LOD cloud. Sapphire guides the user in the query writing process by showing suggestions of query terms based on the queried data, and by recommending changes to the query based on a predictive user model.

This paper presents GraphJet, a new graph-based system for generating content recommendations at Twitter. As motivation, we trace the evolution of our formulation and approach to the graph recommendation problem, embodied in successive generations of systems. Two trends can be identified: supplementing batch with real-time processing and a broadening of the scope of recommendations from users to content. Both of these trends come together in Graph-Jet, an in-memory graph processing engine that maintains a real-time bipartite interaction graph between users and tweets. The storage engine implements a simple API, but one that is sufficiently expressive to support a range of recommendation algorithms based on random walks that we have refined over the years. Similar to Cassovary, a previous graph recommendation engine developed at Twitter, GraphJet assumes that the entire graph can be held in memory on a single server. The system organizes the interaction graph into temporally-partitioned index segments that hold adjacency lists. GraphJet is able to support rapid ingestion of edges while concurrently serving lookup queries through a combination of compact edge encoding and a dynamic memory allocation scheme that exploits power-law characteristics of the graph. Each GraphJet server ingests up to one million graph edges per second, and in steady state, computes up to 500 recommendations per second, which translates into several million edge read operations per second.

No abstract available.

In information retrieval evaluation, when presented with an effectiveness difference between two systems, there are three relevant questions one might ask. First, are the differences statistically significant? Second, is the comparison stable with respect to assessor differences? Finally, is the difference actually meaningful to a user? This paper tackles the last two questions about assessor differences and user preferences in the context of the newly-introduced tweet timeline generation task in the TREC 2014 Microblog track, where the system's goal is to construct an informative summary of non-redundant tweets that addresses the user's information need. Central to the evaluation methodology is human-generated semantic clusters of tweets that contain substantively similar information. We show that the evaluation is stable with respect to assessor differences in clustering and that user preferences generally correlate with effectiveness metrics even though users are not explicitly aware of the semantic clustering being performed by the systems. Although our analyses are limited to this particular task, we believe that lessons learned could generalize to other evaluations based on establishing semantic equivalence between information units, such as nugget-based evaluations in question answering and temporal summarization.

The underlying assumption behind Hadoop and, more generally, the need for distributed processing is that the data to be analyzed cannot be held in memory on a single machine. Today, this assumption needs to be re-evaluated. Although petabyte-scale datastores are increasingly common, it is unclear whether “typical” analytics tasks require more than a single high-end server. Additionally, we are seeing increased sophistication in analytics, e.g., machine learning, where we process smaller and more refined datasets. To address these trends, we propose “scaling down” Hadoop to run on multi-core, shared-memory machines. This paper presents a prototype runtime called Hone (“Hadoop One”) that is API compatible with Hadoop. With Hone, we can take an existing Hadoop application and run it efficiently on a single server. This allows us to take existing MapReduce algorithms and find the most suitable runtime environment for execution on datasets of varying sizes. For dataset sizes that fit into memory on a single machine, our experiments show that Hone is substantially faster than Hadoop running in pseudo-distributed mode. In some cases, Hone running on a single machine outperforms a 16-node Hadoop cluster.

This paper investigates the temporal cluster hypothesis: in search tasks where time plays an important role, do relevant documents tend to cluster together in time? We explore this question in the context of tweet search and temporal feedback: starting with an initial set of results from a baseline retrieval model, we estimate the temporal density of relevant documents, which is then used for result reranking. Our contributions lie in a method to characterize this temporal density function using kernel density estimation, with and without human relevance judgments, and an approach to integrating this information into a standard retrieval model. Experiments on TREC datasets confirm that our temporal feedback formulation improves search effectiveness, thus providing support for our hypothesis. Our approach out-performs both a standard baseline and previous temporal retrieval models. Temporal feedback improves over standard lexical feedback (with and without human judgments), illus- trating that temporal relevance signals exist independently of document content.

We make the suggestion that instead of implementing custom index structures and query evaluation algorithms, IR researchers should simply store document representations in a column-oriented relational database and implement ranking models using SQL. For rapid prototyping, this is particularly advantageous since researchers can explore new scoring functions and features by simply issuing SQL queries, without needing to write imperative code. We demonstrate the feasibility of this approach by an implementation of conjunctive BM25 using two modern column stores. Experiments on a web collection show that a retrieval engine built in this manner achieves effectiveness and efficiency on par with custom-built retrieval engines, but provides many additional advantages, including cleaner query semantics, a simpler architecture, built-in support for error analysis, and the ability to exploit advances in database technology "for free".

"Evaluation as a service" (EaaS) is a new methodology that enables community-wide evaluations and the construction of test collections on documents that cannot be distributed. The basic idea is that evaluation organizers provide a service API through which the evaluation task can be completed. However, this concept violates some of the premises of traditional pool-based collection building and thus calls into question the quality of the resulting test collection. In particular, the service API might restrict the diversity of runs that contribute to the pool: this might hamper innovation by researchers and lead to incomplete judgment pools that affect the reusability of the collection. This paper shows that the distinctiveness of the retrieval runs used to construct the first test collection built using EaaS, the TREC 2013 Microblog collection, is not substantially different from that of the TREC-8 ad hoc collection, a high-quality collection built using traditional pooling. Further analysis using the `leave out uniques' test suggests that pools from the Microblog 2013 collection are less complete than those from TREC-8, although both collections benefit from the presence of distinctive and effective manual runs. Although we cannot yet generalize to all EaaS implementations, our analyses reveal no obvious flaws in the test collection built using the methodology in the TREC 2013 Microblog track.

Summingbird is an open-source domain-specific language implemented in Scala and designed to integrate online and batch MapReduce computations in a single framework. Summingbird programs are written using dataflow abstractions such as sources, sinks, and stores, and can run on different execution platforms: Hadoop for batch processing (via Scalding/Cascading) and Storm for online processing. Different execution modes require different bindings for the dataflow abstractions (e.g., HDFS files or message queues for the source) but do not require any changes to the program logic. Furthermore, Summingbird can operate in a hybrid processing mode that transparently integrates batch and online results to efficiently generate up-to-date aggregations over long time spans. The language was designed to improve developer productivity and address pain points in building analytics solutions at Twitter where often, the same code needs to be written twice (once for batch processing and again for online processing) and indefinitely maintained in parallel. Our key insight is that certain algebraic structures provide the theoretical foundation for integrating batch and online processing in a seamless fashion. This means that Summingbird imposes constraints on the types of aggregations that can be performed, although in practice we have not found these constraints to be overly restrictive for a broad range of analytics tasks at Twitter.

We describe a production Twitter system for generating relevant, personalized, and timely recommendations based on observing the temporally-correlated actions of each user's followings. The system currently serves millions of recommendations daily to tens of millions of mobile users. The approach can be viewed as a specific instance of the novel problem of online motif detection in large dynamic graphs. Our current solution partitions the graph across a number of machines, and with the construction of appropriate data structures, motif detection can be translated into the lookup and intersection of adjacency lists in each partition. We conclude by discussing a generalization of the problem that perhaps represents a new class of data management systems.

There is an increasing interest in executing rich and complex analysis tasks over large-scale graphs, many of which require processing and reasoning about a large number of multi-hop neighborhoods or subgraphs in the graph. Examples of such tasks include ego network analysis, motif counting in biological networks, finding social circles, personalized recommendations, link prediction, anomaly detection , analyzing influence cascades , and so on. These tasks are not well served by existing vertex-centric graph processing frameworks whose computation and execution models limit the user program to directly access the state of a single vertex , resulting in high communication, scheduling, and memory overheads in executing such tasks. Further, most existing graph processing frameworks also typically ignore the challenges in extracting the relevant portions of the graph that an analysis task is interested in, and loading it onto distributed memory. In this demonstration proposal, we describe NScale, a novel end-to-end graph processing framework that enables the distributed execution of complex neighborhood-centric analytics over large-scale graphs in the cloud. NScale enables users to write programs at the level of neighborhoods or subgraphs. NScale uses Apache YARN for efficient and fault-tolerant distribution of data and computation; it features GEL, a novel graph extraction and loading phase, that extracts the relevant portions of the graph and loads them into distributed memory using as few machines as possible. NScale utilizes novel techniques for the distributed execution of user computation that minimize memory consumption by exploiting overlap among the neighborhoods of interest. A comprehensive experimental evaluation shows orders-of-magnitude improvements in performance and total cost over vertex-centric approaches.

Tree-based models have proven to be an effective solution for web ranking as well as other machine learning problems in diverse domains. This paper focuses on optimizing the runtime performance of applying such models to make predictions, specifically using gradient-boosted regression trees for learning to rank. Although exceedingly simple conceptually, most implementations of tree-based models do not efficiently utilize modern superscalar processors. By laying out data structures in memory in a more cache-conscious fashion, removing branches from the execution flow using a technique called predication, and micro-batching predictions using a technique called vectorization, we are able to better exploit modern processor architectures. Experiments on synthetic data and on three standard learning-to-rank datasets show that our approach is significantly faster than standard implementations.

This work explores how internal representations of modern statistical machine translation systems can be exploited for cross-language information retrieval. We tackle two core issues that are central to query translation: how to exploit context to generate more accurate translations and how to preserve ambiguity that may be present in the original query, thereby retaining a diverse set of translation alternatives. These two considerations are often in tension since ambiguity in natural language is typically resolved by exploiting context, but effective retrieval requires striking the right balance. We propose two novel query translation approaches: the grammar-based approach extracts translation probabilities from translation grammars, while the decoder-based approach takes advantage of n -best translation hypotheses. Both are context-sensitive , in contrast to a baseline context-insensitive approach that uses bilingual dictionaries for word-by-word translation. Experimental results show that by “opening up” modern statistical machine translation systems, we can access intermediate representations that yield high retrieval effectiveness. By combining evidence from multiple sources, we demonstrate significant improvements over competitive baselines on standard cross-language information retrieval test collections. In addition to effectiveness, the efficiency of our techniques are explored as well.

We explore a real-time Twitter search application where tweets are arriving at a rate of several thousands per second. Real-time search demands that they be indexed and searchable immediately, which leads to a number of implementation challenges. In this paper, we focus on one aspect: dynamic postings allocation policies for index structures that are completely held in main memory. The core issue can be characterized as a "Goldilocks Problem". Because memory remains today a scare resource, an allocation policy that is too aggressive leads to inefficient utilization, while a policy that is too conservative is slow and leads to fragmented postings lists. We present a dynamic postings allocation policy that allocates memory in increasingly-larger "slices" from a small number of large, fixed pools of memory. With an analytical model and experiments, we explore different settings that balance time (query evaluation speed) and space (memory utilization).

This paper examines a multi-stage retrieval architecture consisting of a candidate generation stage, a feature extraction stage, and a reranking stage using machine-learned models. Given a fixed set of features and a learning-to-rank model, we explore effectiveness/efficiency tradeoffs with three candidate generation approaches: postings intersection with SvS, conjunctive query evaluation with WAND, and disjunctive query evaluation with WAND. We find no significant differences in end-to-end effectiveness as measured by NDCG between conjunctive and disjunctive WAND, but conjunctive query evaluation is substantially faster. Postings intersection with SvS, while fast, yields substantially lower end-to-end effectiveness, suggesting that document and term frequencies remain important in the initial ranking stage. These findings show that conjunctive WAND is the best overall candidate generation strategy of those we examined.

Although context-independent word-based approaches remain popular for cross-language information retrieval, many recent studies have shown that integrating insights from modern statistical machine translation systems can lead to substantial improvements in effectiveness. In this paper, we compare flat and hierarchical phrase-based translation models for query translation. Both approaches yield significantly better results than either a token-based or a one-best translation baseline on standard test collections. The choice of model manifests interesting tradeoffs in terms of effectiveness, efficiency, and model compactness.

We present the architecture behind Twitter's real-time related query suggestion and spelling correction service. Although these tasks have received much attention in the web search literature, the Twitter context introduces a real-time "twist": after significant breaking news events, we aim to provide relevant results within minutes. This paper provides a case study illustrating the challenges of real-time data processing in the era of "big data". We tell the story of how our system was built twice: our first implementation was built on a typical Hadoop-based analytics stack, but was later replaced because it did not meet the latency requirements necessary to generate meaningful real-time results. The second implementation, which is the system deployed in production today, is a custom in-memory processing engine specifically designed for the task. This experience taught us that the current typical usage of Hadoop as a "big data" platform, while great for experimentation, is not well suited to low-latency processing, and points the way to future work on data analytics platforms that can handle "big" as well as "fast" data.

WTF ("Who to Follow") is Twitter's user recommendation service, which is responsible for creating millions of connections daily between users based on shared interests, common connections, and other related factors. This paper provides an architectural overview and shares lessons we learned in building and running the service over the past few years. Particularly noteworthy was our design decision to process the entire Twitter graph in memory on a single server, which significantly reduced architectural complexity and allowed us to develop and deploy the service in only a few months. At the core of our architecture is Cassovary, an open-source in-memory graph processing engine we built from scratch for WTF. Besides powering Twitter's user recommendations, Cassovary is also used for search, discovery, promoted products, and other services as well. We describe and evaluate a few graph recommendation algorithms implemented in Cassovary, including a novel approach based on a combination of random walks and SALSA. Looking into the future, we revisit the design of our architecture and comment on its limitations, which are presently being addressed in a second-generation system under development.

The underlying assumption behind Hadoop and, more generally, the need for distributed processing is that the data to be analyzed cannot be held in memory on a single machine. Today, this assumption needs to be re-evaluated. Although petabyte-scale data-stores are increasingly common, it is unclear whether "typical" analytics tasks require more than a single high-end server. Additionally, we are seeing increased sophistication in analytics, e.g., machine learning, which generally operates over smaller and more refined datasets. To address these trends, we propose "scaling down" Hadoop to run on shared-memory machines. This paper presents a prototype runtime called Hone, intended to be both API and binary compatible with standard (distributed) Hadoop. That is, Hone can take an existing Hadoop jar and efficiently execute it, without modification, on a multi-core shared memory machine. This allows us to take existing Hadoop algorithms and find the most suitable run-time environment for execution on datasets of varying sizes. Our experiments show that Hone can be an order of magnitude faster than Hadoop pseudo-distributed mode (PDM); on dataset sizes that fit into memory, Hone can outperform a fully-distributed 15-node Hadoop cluster in some cases as well.

The rise of social media and other forms of user-generated content have created the demand for real-time search: against a high-velocity stream of incoming documents, users desire a list of relevant results at the time the query is issued. In the context of real-time search on tweets, this work explores candidate generation in a two-stage retrieval architecture where an initial list of results is processed by a second-stage rescorer to produce the final output. We introduce Bloom filter chains, a novel extension of Bloom filters that can dynamically expand to efficiently represent an arbitrarily long and growing list of monotonically-increasing integers with a constant false positive rate. Using a collection of Bloom filter chains, a novel approximate candidate generation algorithm called BWand is able to perform both conjunctive and disjunctive retrieval. Experiments show that our algorithm is many times faster than competitive baselines and that this increased performance does not require sacrificing end-to-end effectiveness. Our results empirically characterize the trade-off space defined by output quality, query evaluation speed, and memory footprint for this particular search architecture.

Most modern web search engines employ a two-phase ranking strategy: a candidate list of documents is generated using a "cheap" but low-quality scoring function, which is then reranked by an "expensive" but high-quality method (usually machine-learned). This paper focuses on the problem of candidate generation for conjunctive query processing in this context. We describe and evaluate a fast, approximate postings list intersection algorithms based on Bloom filters. Due to the power of modern learning-to-rank techniques and emphasis on early precision, significant speedups can be achieved without loss of end-to-end retrieval effectiveness. Explorations reveal a rich design space where effectiveness and efficiency can be balanced in response to specific hardware configurations and application scenarios.

The web today is increasingly characterized by social and real-time signals, which we believe represent two frontiers in information retrieval. In this paper, we present Early bird, the core retrieval engine that powers Twitter's real-time search service. Although Early bird builds and maintains inverted indexes like nearly all modern retrieval engines, its index structures differ from those built to support traditional web search. We describe these differences and present the rationale behind our design. A key requirement of real-time search is the ability to ingest content rapidly and make it searchable immediately, while concurrently supporting low-latency, high-throughput query evaluation. These demands are met with a single-writer, multiple-reader concurrency model and the targeted use of memory barriers. Early bird represents a point in the design space of real-time search engines that has worked well for Twitter's needs. By sharing our experiences, we hope to spur additional interest and innovation in this exciting space.

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.

It is well known that anchor text plays an important role in search, providing signals that are often not present in the source document itself. The paper reports results of a preliminary investigation on the value of tweets and tweet conversations as anchor text. We show that using tweets as anchors improves significantly over using HTML anchors, and significantly increases recall of news item retrieval.

Cross-language information retrieval (CLIR) today is dominated by techniques that use token-to-token mappings from bilingual dictionaries. Yet, state-of-the-art statistical translation models (e.g., using Synchronous Context-Free Grammars) are far richer, capturing multi-term phrases, term dependencies, and contextual constraints on translation choice. We present a novel CLIR framework that is able to reach inside the translation "black box" and exploit these sources of evidence. Experiments on the TREC-5/6 English-Chinese test collection show this approach to be promising.

The success of data-driven solutions to difficult problems, along with the dropping costs of storing and processing massive amounts of data, has led to growing interest in large-scale machine learning. This paper presents a case study of Twitter's integration of machine learning tools into its existing Hadoop-based, Pig-centric analytics platform. We begin with an overview of this platform, which handles "traditional" data warehousing and business intelligence tasks for the organization. The core of this work lies in recent Pig extensions to provide predictive analytics capabilities that incorporate machine learning, focused specifically on supervised classification. In particular, we have identified stochastic gradient descent techniques for online learning and ensemble methods as being highly amenable to scaling out to large amounts of data. In our deployed solution, common machine learning tasks such as data sampling, feature generation, training, and testing can be accomplished directly in Pig, via carefully crafted loaders, storage functions, and user-defined functions. This means that machine learning is just another Pig script, which allows seamless integration with existing infrastructure for data management, scheduling, and monitoring in a production environment, as well as access to rich libraries of user-defined functions and the materialized output of other scripts.

In recent years, there has been a substantial amount of work on large-scale data analytics using Hadoop-based platforms running on large clusters of commodity machines. A less-explored topic is how those data, dominated by application logs, are collected and structured to begin with. In this paper, we present Twitter's production logging infrastructure and its evolution from application-specific logging to a unified "client events" log format, where messages are captured in common, well-formatted, flexible Thrift messages. Since most analytics tasks consider the user session as the basic unit of analysis, we pre-materialize "session sequences", which are compact summaries that can answer a large class of common queries quickly. The development of this infrastructure has streamlined log collection and data analysis, thereby improving our ability to rapidly experiment and iterate on various aspects of the service.

Previous research has shown that features based on term proximity are important for effective retrieval. However, they incur substantial costs in terms of larger inverted indexes and slower query execution times as compared to term-based features. This paper explores whether term proximity features based on approximate term positions are as effective as those based on exact term positions. We introduce the novel notion of approximate positional indexes based on dividing documents into coarse-grained buckets and recording term positions with respect to those buckets. We propose different approaches to defining the buckets and compactly encoding bucket ids. In the context of linear ranking functions, experimental results show that features based on approximate term positions are able to achieve effectiveness comparable to exact term positions, but with smaller indexes and faster query evaluation.

We are interested in the problem of tracking broad topics such as "baseball" and "fashion" in continuous streams of short texts, exemplified by tweets from the microblogging service Twitter. The task is conceived as a language modeling problem where per-topic models are trained using hashtags in the tweet stream, which serve as proxies for topic labels. Simple perplexity-based classifiers are then applied to filter the tweet stream for topics of interest. Within this framework, we evaluate, both intrinsically and extrinsically, smoothing techniques for integrating "foreground" models (to capture recency) and "background" models (to combat sparsity), as well as different techniques for retaining history. Experiments show that unigram language models smoothed using a normalized extension of stupid backoff and a simple queue for history retention performs well on the task.

Test collections are the primary drivers of progress in information retrieval. They provide yardsticks for assessing the effectiveness of ranking functions in an automatic, rapid, and repeatable fashion and serve as training data for learning to rank models. However, manual construction of test collections tends to be slow, labor-intensive, and expensive. This paper examines the feasibility of constructing web search test collections in a completely unsupervised manner given only a large web corpus as input. Within our proposed framework, anchor text extracted from the web graph is treated as a pseudo query log from which pseudo queries are sampled. For each pseudo query, a set of relevant and non-relevant documents are selected using a variety of web-specific features, including spam and aggregated anchor text weights. The automatically mined queries and judgments form a pseudo test collection that can be used for training ranking functions. Experiments carried out on TREC web track data show that learning to rank models trained using pseudo test collections outperform an unsupervised ranking function and are statistically indistinguishable from a model trained using manual judgments, demonstrating the usefulness of our approach in extracting reasonable quality training data "for free".

No abstract available.

This work explores the problem of cross-lingual pairwise similarity, where the task is to extract similar pairs of documents across two different languages. Solutions to this problem are of general interest for text mining in the multilingual context and have specific applications in statistical machine translation. Our approach takes advantage of cross-language information retrieval (CLIR) techniques to project feature vectors from one language into another, and then uses locality-sensitive hashing (LSH) to extract similar pairs. We show that effective cross-lingual pairwise similarity requires working with similarity thresholds that are much lower than in typical monolingual applications, making the problem quite challenging. We present a parallel, scalable MapReduce implementation of the sort-based sliding window algorithm, which is compared to a brute-force approach on German and English Wikipedia collections. Our central finding can be summarized as "no free lunch": there is no single optimal solution. Instead, we characterize effectiveness-efficiency tradeoffs in the solution space, which can guide the developer to locate a desirable operating point based on application- and resource-specific constraints.

There is a fundamental tradeoff between effectiveness and efficiency when designing retrieval models for large-scale document collections. Effectiveness tends to derive from sophisticated ranking functions, such as those constructed using learning to rank, while efficiency gains tend to arise from improvements in query evaluation and caching strategies. Given their inherently disjoint nature, it is difficult to jointly optimize effectiveness and efficiency in end-to-end systems. To address this problem, we formulate and develop a novel cascade ranking model, which unlike previous approaches, can simultaneously improve both top k ranked effectiveness and retrieval efficiency. The model constructs a cascade of increasingly complex ranking functions that progressively prunes and refines the set of candidate documents to minimize retrieval latency and maximize result set quality. We present a novel boosting algorithm for learning such cascades to directly optimize the tradeoff between effectiveness and efficiency. Experimental results show that our cascades are faster and return higher quality results than comparable ranking models.

This paper introduces the notion of temporally constrained ranked retrieval, which, given a query and a time constraint, produces the best possible ranked list within the specified time limit. Naturally, more time should translate into better results, but the ranking algorithm should always produce some results. This property is desirable from a number of perspectives: to cope with diverse users and information needs, as well as to better manage system load and variance in query execution times. We propose two temporally constrained ranking algorithms based on a class of probabilistic prediction models that can naturally incorporate efficiency constraints: one that makes independent feature selection decisions, and the other that makes joint feature selection decisions. Experiments on three different test collections show that both ranking algorithms are able to satisfy imposed time constraints, although the joint model outperforms the independent model in being able to deliver more effective results, especially under tight time constraints, due to its ability to capture feature dependencies.

It has been shown that learning to rank approaches are capable of learning highly effective ranking functions. However, these approaches have mostly ignored the important issue of efficiency. Given that both efficiency and effectiveness are important for real search engines, models that are optimized for effectiveness may not meet the strict efficiency requirements necessary to deploy in a production environment. In this work, we present a unified framework for jointly optimizing effectiveness and efficiency. We propose new metrics that capture the tradeoff between these two competing forces and devise a strategy for automatically learning models that directly optimize the tradeoff metrics. Experiments indicate that models learned in this way provide a good balance between retrieval effectiveness and efficiency. With specific loss functions, learned models converge to familiar existing ones, which demonstrates the generality of our framework. Finally, we show that our approach naturally leads to a reduction in the variance of query execution times, which is important for query load balancing and user satisfaction.

This paper explores the problem of computing pairwise similarity on document collections, focusing on the application of "more like this" queries in the life sciences domain. Three MapReduce algorithms are introduced: one based on brute force, a second where the problem is treated as large-scale ad hoc retrieval, and a third based on the Cartesian product of postings lists. Each algorithm supports one or more approximations that trade effectiveness for efficiency, the characteristics of which are studied experimentally. Results show that the brute force algorithm is the most efficient of the three when exact similarity is desired. However, the other two algorithms support approximations that yield large efficiency gains without significant loss of effectiveness.

Transaction logs from online search engines are valuable for two reasons: First, they provide insight into human information-seeking behavior. Second, log data can be used to train user models, which can then be applied to improve retrieval systems. This article presents a study of logs from PubMed((R)), the public gateway to the MEDLINE((R)) database of bibliographic records from the medical and biomedical primary literature. Unlike most previous studies on general Web search, our work examines user activities with a highly-specialized search engine. We encode user actions as string sequences and model these sequences using n-gram language models. The models are evaluated in terms of perplexity and in a sequence prediction task. They help us better understand how PubMed users search for information and provide an enabler for improving users' search experience.

Abstract This work explores computational models of multi‐party discourse, using transcripts from U.S. Supreme Court oral arguments. The turn‐taking behavior of participants is treated as a supervised sequence‐labeling problem and modeled using first‐ and second‐order conditional random fields (CRFs). We specifically explore the hypothesis that discourse markers and personal references provide important features in such models. Results from a sequence prediction experiment demonstrate that incorporating these two types of features yields significant improvements in accuracy. Our experiments are couched in the broader context of developing tools to support legal scholarship, although we see other natural language processing applications as well.

In the context of document retrieval in the biomedical domain, this paper explores the complex relationship between the quality of initial query results and the overall utility of an interactive retrieval system. We demonstrate that a content-similarity browsing tool can compensate for poor retrieval results, and that the relationship between retrieval performance and overall utility is non-linear. Arguments are advanced with user simulations, which characterize the relevance of documents that a user might encounter with different browsing strategies. With broader implications to IR, this work provides a case study of how user simulations can be exploited as a formative tool for automatic utility evaluation. Simulation-based studies provide researchers with an additional evaluation tool to complement interactive and Cranfield-style experiments.

This work explores the hypothesis that interactions between a trained human search intermediary and an information seeker can inform the design of interactive IR systems. We discuss results from a controlled Wizard-of-Oz case study, set in the context of the TREC 2005 HARD track evaluation, in which a trained intermediary executed an integrated search and interaction strategy based on conceptual facet analysis and informed by need negotiation techniques common in reference interviews. Having a human “in the loop” yielded large improvements over fully automated systems as measured by standard ranked-retrieval metrics, demonstrating the value of mediated search. We present a detailed analysis of the intermediary's actions to gain a deeper understanding of what worked and why. One contribution is a taxonomy of clarification types informed both by empirical results and existing theories in library and information science. We discuss how these findings can guide the development of future systems. Overall, this work illustrates how studying human information-seeking processes can lead to better information retrieval applications.

A methodology based on "information nuggets" has recently emerged as the de facto standard by which answers to complex questions are evaluated. After several implementations in the TREC question answering tracks, the community has gained a better understanding of its many characteristics. This paper focuses on one particular aspect of the evaluation: the human assignment of nuggets to answer strings, which serves as the basis of the F-score computation. As a byproduct of the TREC 2006 ciQA task, identical answer strings were independently evaluated twice, which allowed us to assess the consistency of human judgments. Based on these results, we explored simulations of assessor behavior that provide a method to quantify scoring variations. Understanding these variations in turn lets researchers be more confident in their comparisons of systems.

We explore a syntactic approach to sentence compression in the biomedical domain, grounded in the context of result presentation for related article search in the PubMed search engine. By automatically trimming inessential fragments of article titles, a system can effectively display more results in the same amount of space. Our implemented prototype operates by applying a sequence of syntactic trimming rules over the parse trees of article titles. Two separate studies were conducted using a corpus of manually compressed examples from MEDLINE: an automatic evaluation using B leu and a summative evaluation involving human assessors. Experiments show that a syntactic approach to sentence compression is effective in the biomedical domain and that the presentation of compressed article titles supports accurate “interest judgments”, decisions by users as to whether an article is worth examining in more detail.

The so-called “redundancy-based” approach to question answering represents a successful strategy for mining answers to factoid questions such as “Who shot Abraham Lincoln?” from the World Wide Web. Through contrastive and ablation experiments with Aranea, a system that has performed well in several TREC QA evaluations, this work examines the underlying assumptions and principles behind redundancy-based techniques. Specifically, we develop two theses: that stable characteristics of data redundancy allow factoid systems to rely on external “black box” components, and that despite embodying a data-driven approach, redundancy-based methods encode a substantial amount of knowledge in the form of heuristics. Overall, this work attempts to address the broader question of “what really matters” and to provide guidance for future researchers.

Despite its intuitive appeal, the hypothesis that retrieval at the level of "concepts" should outperform purely term-based approaches remains unverified empirically. In addition, the use of "knowledge" has not consistently resulted in performance gains. After identifying possible reasons for previous negative results, we present a novel framework for "conceptual retrieval" that articulates the types of knowledge that are important for information seeking. We instantiate this general framework in the domain of clinical medicine based on the principles of evidence-based medicine (EBM). Experiments show that an EBM-based scoring algorithm dramatically outperforms a state-of-the-art baseline that employs only term statistics. Ablation studies further yield a better understanding of the performance contributions of different components. Finally, we discuss how other domains can benefit from knowledge-based approaches.

Single-iteration clarification dialogs, as implemented in the TREC HARD track, represent an attempt to introduce interaction into ad hoc retrieval, while preserving the many benefits of large-scale evaluations. Although previous experiments have not conclusively demonstrated performance gains resulting from such interactions, it is unclear whether these findings speak to the nature of clarification dialogs, or simply the limitations of current systems. To probe the limits of such interactions, we employed a human intermediary to formulate clarification questions and exploit user responses. In addition to establishing a plausible upper bound on performance, we were also able to induce an "ontology of clarifications" to characterize human behavior. This ontology, in turn, serves as the input to a regression model that attempts to determine which types of clarification questions are most helpful. Our work can serve to inform the design of interactive systems that initiate user dialogs.

We present a novel language modeling approach to capturing the query reformulation behavior of Web search users. Based on a framework that categorizes eight different types of "user moves" (adding/removing query terms, etc.), we treat search sessions as sequence data and build n-gram language models to capture user behavior. We evaluated our models in a prediction task. The results suggest that useful patterns of activity can be extracted from user histories. Furthermore, by examining prediction performance under different order n-gram models, we gained insight into the amount of history/context that is associated with different types of user actions. Our work serves as the basis for more refined user models.

Abstract In contrast to traditional information retrieval systems, which return ranked lists of documents that users must manually browse through, a question answering system attempts to directly answer natural language questions posed by the user. Although such systems possess language‐processing capabilities, they still rely on traditional document retrieval techniques to generate an initial candidate set of documents. In this article, the authors argue that document retrieval for question answering represents a task different from retrieving documents in response to more general retrospective information needs. Thus, to guide future system development, specialized question answering test collections must be constructed. They show that the current evaluation resources have major shortcomings; to remedy the situation, they have manually created a small, reusable question answering test collection for research purposes. In this article they describe their methodology for building this test collection and discuss issues they encountered regarding the notion of “answer correctness.”

Controlled and reproducible laboratory experiments, enabled by reusable test collections, represent a well-established methodology in modern information retrieval research. In order to confidently draw conclusions about the performance of different retrieval methods using test collections, their reliability and trustworthiness must first be established. Although such studies have been performed for ad hoc test collections, currently available resources for evaluating question answering systems have not been similarly analyzed. This study evaluates the quality of answer patterns and lists of relevant documents currently employed in automatic question answering evaluation, and concludes that they are not suitable for post-hoc experimentation. These resources, created from runs submitted by TREC QA track participants, do not produce fair and reliable assessments of systems that did not participate in the original evaluations. Potential solutions for addressing this evaluation gap and their shortcomings are discussed.

When applying blind relevance feedback for ad hoc document retrieval, is it possible to identify, a priori, the set of query terms that will most improve retrieval performance? Can this complex problem be reduced into the simpler one of making independent decisions about the performance effects of each query term? Our experiments suggest that, for the selection of terms for blind relevance feedback, the term independence assumption may be empirically justified.

We present a strategy for answering fact-based natural language questions that is guided by a characterization of real-world user queries. Our approach, implemented in a system called Aranea, extracts answers from the Web using two different techniques: knowledge annotation and knowledge mining. Knowledge annotation is an approach to answering large classes of frequently occurring questions by utilizing semi\-structured and structured Web sources. Knowledge mining is a statistical approach that leverages massive amounts of Web data to overcome many natural language processing challenges. We have integrated these two different paradigms into a question answering system capable of providing users with concise answers that directly address their information needs.

Passage retrieval is an important component common to many question answering systems. Because most evaluations of question answering systems focus on end-to-end performance, comparison of common components becomes di#cult. To address this shortcoming, we present a quantitative evaluation of various passage retrieval algorithms for question answering, implemented in a framework called Pauchok. We present three important findings: Boolean querying schemes perform well in the question answering task. The performance di#erences between various passage retrieval algorithms vary with the choice of document retriever, which suggests significant interactions between document retrieval and passage retrieval. The best algorithms in our evaluation employ density-based measures for scoring query terms. Our results reveal future directions for passage retrieval and question answering.

This paper describes a question answering system that is designed to capitalize on the tremendous amount of data that is now available online. Most question answering systems use a wide variety of linguistic resources. We focus instead on the redundancy available in large corpora as an important resource. We use this redundancy to simplify the query rewrites that we need to use, and to support answer mining from returned snippets. Our system performs quite well given the simplicity of the techniques being utilized. Experimental results show that question answering accuracy can be greatly improved by analyzing more and more matching passages. Simple passage ranking and n-gram extraction techniques work well in our system making it efficient to use with many backend retrieval engines.