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.

Test collections are crucial for evaluating Information Retrieval (IR) systems. Creating a diverse set of user queries for these collections can be challenging, and obtaining relevance judgments, which indicate how well retrieved documents match a query, is often costly and resource-intensive. Recently, generating synthetic datasets using Large Language Models (LLMs) has gained attention in various applications. While previous work has used LLMs to generate synthetic queries or documents to improve ranking models, using LLMs to create synthetic test collections is still relatively unexplored. Previous work showed that synthetic test collections have the potential to be used for system evaluation, however, more analysis is needed to validate this claim. In this paper, we thoroughly investigate the reliability of synthetic test collections constructed using LLMs, where LLMs are used to generate synthetic queries, labels, or both. In particular, we examine the potential biases that might occur when such test collections are used for evaluation. We first empirically show the presence of such bias in evaluation results and analyse the effects it might have on system evaluation. We further validate the presence of such bias using a linear mixed-effects model. Our analysis shows that while the effect of bias present in evaluation results obtained using synthetic test collections could be significant, for e.g. computing absolute system performance, its effect may not be as significant in comparing relative system performance. Codes and data are available at: https://github.com/rahmanidashti/BiasSyntheticData

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.

Large language models (LLMs) have demonstrated increasing task-solving abilities not present in smaller models. Utilizing the capabilities and responsibilities of LLMs for automated evaluation (LLM4Eval) has recently attracted considerable attention in multiple research communities. Building on the success of previous workshops, which established foundations in automated judgments and RAG evaluation, this third iteration aims to address emerging challenges as IR systems become increasingly personalized and interactive. The main goal of the third LLM4Eval workshop is to bring together researchers from industry and academia to explore three critical areas: the evaluation of personalized IR systems while maintaining fairness, the boundaries between automated and human assessment in subjective scenarios, and evaluation methodologies for systems that combine multiple IR paradigms (search, recommendations, and dialogue). By examining these challenges, we seek to understand how evaluation approaches can evolve to match the sophistication of modern IR applications. The format of the workshop is interactive, including roundtable discussion sessions, fostering dialogue about the future of IR evaluation while avoiding one-sided discussions. This is the third iteration of the workshop series, following successful events at SIGIR 2024 and WSDM 2025, with the first iteration attracting over 50 participants.

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.

Large language models (LLMs) have demonstrated increasing task-solving abilities not present in smaller models. Utilizing the capabilities and responsibilities of LLMs for automated evaluation (LLM4Eval) has recently attracted considerable attention in multiple research communities. For instance, LLM4Eval models have been studied in the context of automated judgments, natural language generation, and retrieval augmented generation systems. We believe that the information retrieval community can significantly contribute to this growing research area by designing, implementing, analyzing, and evaluating various aspects of LLMs with applications to LLM4Eval tasks. The main goal of LLM4Eval workshop is to bring together researchers from industry and academia to discuss various aspects of LLMs for evaluation in information retrieval, including automated judgments, retrieval-augmented generation pipeline evaluation, altering human evaluation, robustness, and trustworthiness of LLMs for evaluation in addition to their impact on real-world applications. We also plan to run an automated judgment challenge prior to the workshop, where participants will be asked to generate labels for a given dataset while maximising correlation with human judgments. The format of the workshop is interactive, including roundtable and keynote sessions and tends to avoid the one-sided dialogue of a mini-conference. This is the second iteration of the workshop. The first version was held in conjunction with SIGIR 2024, attracting over 50 participants.

Information retrieval (IR) has a large literature on evaluation, dating back decades and forming a central part of the research culture. The largest proportion of this literature discusses techniques to turn a sequence of relevance labels into a single number, reflecting the system's performance: precision or cumulative gain, for example, or dozens of alternatives. Those techniques-metrics-are themselves evaluated, commonly by reference to sensitivity and validity.

Much of the evaluation and tuning of a search system relies on relevance labels---annotations that say whether a document is useful for a given search and searcher. Ideally these come from real searchers, but it is hard to collect this data at scale, so typical experiments rely on third-party labellers who may or may not produce accurate annotations. Label quality is managed with ongoing auditing, training, and monitoring. We discuss an alternative approach. We take careful feedback from real searchers and use this to select a large language model (LLM), and prompt, that agrees with this feedback; the LLM can then produce labels at scale. Our experiments show LLMs are as accurate as human labellers and as useful for finding the best systems and hardest queries. LLM performance varies with prompt features, but also varies unpredictably with simple paraphrases. This unpredictability reinforces the need for high-quality "gold" labels.

Constructing test collections in Information Retrieval (IR) is vital for evaluating search algorithms. Obtaining a diverse set of user queries for test collection construction can be challenging, and acquiring relevance judgments, which indicate the appropriateness of retrieved documents, is often costly and resource-intensive. Generating synthetic datasets using Large Language Models (LLMs) has recently gained significant attention in various applications. In information retrieval, while previous work exploited the capabilities of LLMs to generate synthetic queries or documents to augment training data and improve the performance of ranking models, using LLMs for constructing synthetic test collections is relatively unexplored. Previous studies demonstrate that LLMs have the potential to generate synthetic relevance judgments for use in the evaluation of information retrieval systems. In this paper, we comprehensively investigate whether it is possible to use LLMs to construct fully synthetic test collections by generating not only synthetic judgments but also synthetic queries. To qualify the efficacy of synthetic queries for examining system ordering, we analyze how these synthetic data are suitable for building a reliable and reusable test collection and the potential risks of bias such test collections may exhibit towards LLM-based models. Our comprehensive experiments indicate that test collections generated using LLMs can effectively and reliably evaluate system performance.

Large language models (LLMs) have demonstrated increasing task-solving abilities not present in smaller models. Utilizing the capabilities and responsibilities of LLMs for automated evaluation (LLM4Eval) has recently attracted considerable attention in multiple research communities. For instance, LLM4Eval models have been studied in the context of automated judgments, natural language generation, and retrieval augmented generation systems. We believe that the information retrieval community can significantly contribute to this growing research area by designing, implementing, analyzing, and evaluating various aspects of LLMs with applications to LLM4Eval tasks. The main goal of LLM4Eval workshop is to bring together researchers from industry and academia to discuss various aspects of LLMs for evaluation in information retrieval, including automated judgments, retrieval-augmented generation pipeline evaluation, altering human evaluation, robustness, and trustworthiness of LLMs for evaluation in addition to their impact on real-world applications. We also plan to run an automated judgment challenge prior to the workshop, where participants will be asked to generate labels for a given dataset while maximising correlation with human judgments. The format of the workshop is interactive, including roundtable and keynote sessions and tends to avoid the one-sided dialogue of a mini-conference.

Users of search systems often reformulate their queries by adding query terms to reflect their evolving information need or to more precisely express their information need when the system fails to surface relevant content. Analyzing these query reformulations can inform us about both system and user behavior. In this work, we study a special category of query reformulations that involve specifying demographic group attributes, such as gender, as part of the reformulated query (e.g., ''olympic 2021 soccer results'' -> ''olympic 2021 women's soccer results"). There are many ways a query, the search results, and a demographic attribute such as gender may relate, leading us to hypothesize different causes for these reformulation patterns, such as under-representation on the original result page or based on the linguistic theory of markedness. This paper reports on an observational study of gender-specializing query reformulations---their contexts and effects---as a lens on the relationship between system results and gender, based on large-scale search log data from Bing. We find that these reformulations sometimes correct for and other times reinforce gender representation on the original result page, but typically yield better access to the ultimately-selected results. The prevalence of these reformulations---and which gender they skew towards---differ by topical context. However, we do not find evidence that either group under-representation or markedness alone adequately explains these reformulations. We hope that future research will use such reformulations as a probe for deeper investigation into gender (and other demographic) representation on the search result page.

When information retrieval systems return a ranked list of results in response to a query, they may be choosing from a large set of candidate results that are equally useful and relevant. This means we might be able to identify a difference between rankers A and B, where ranker A systematically prefers a certain type of relevant results. Ranker A may have this systematic difference (different “vibe”) without having systematically better or worse results according to standard information retrieval metrics. We first show that a vibe difference can exist, comparing two publicly available rankers, where the one that is trained on health-related queries will systematically prefer health-related results, even for non-health queries. We define a vibe metric that lets us see the words that a ranker prefers. We investigate the vibe of search engine clicks vs. human labels. We perform an initial study into correcting for vibe differences to make ranker A more like ranker B via changes in negative sampling during training.

A long-standing challenge for search and conversational assistants is query intention detection in ambiguous queries. Asking clarifying questions in conversational search has been widely studied and considered an effective solution to resolve query ambiguity. Existing work have explored various approaches for clarifying question ranking and generation. However, due to the lack of real conversational search data, they have to use artificial datasets for training, which limits their generalizability to real-world search scenarios. As a result, the industry has shown reluctance to implement them in reality, further suspending the availability of real conversational search interaction data. The above dilemma can be formulated as a cold start problem of clarifying question generation and conversational search in general. Furthermore, even if we do have large-scale conversational logs, it is not realistic to gather training data that can comprehensively cover all possible queries and topics in open-domain search scenarios. The risk of fitting bias when training a clarifying question retrieval/generation model on incomprehensive dataset is thus another important challenge.

Like many other researchers, at Microsoft Bing we use external “crowd” judges to label results from a search engine—especially, although not exclusively, to obtain relevance labels for offline evaluation in the Cranfield tradition. Crowdsourced labels are relatively cheap, and hence very popular, but are prone to disagreements, spam, and various biases which appear to be unexplained “noise” or “error”. In this paper, we provide examples of problems we have encountered running crowd labelling at large scale and around the globe, for search evaluation in particular. We demonstrate effects due to the time of day and day of week that a label is given; fatigue; anchoring; exposure; left-side bias; task switching; and simple disagreement between judges. Rather than simple “error”, these effects are consistent with well-known physiological and cognitive factors. “The crowd” is not some abstract machinery, but is made of people. Human factors that affect people’s judgement behaviour must be considered when designing research evaluations and in interpreting evaluation metrics.

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.

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.

The dramatic improvements in core information retrieval tasks engendered by neural rankers create a need for novel evaluation methods. If every ranker returns highly relevant items in the top ranks, it becomes difficult to recognize meaningful differences between them and to build reusable test collections. Several recent papers explore pairwise preference judgments as an alternative to traditional graded relevance assessments. Rather than viewing items one at a time, assessors view items side-by-side and indicate the one that provides the better response to a query, allowing fine-grained distinctions. If we employ preference judgments to identify the probably best items for each query, we can measure rankers by their ability to place these items as high as possible. We frame the problem of finding best items as a dueling bandits problem. While many papers explore dueling bandits for online ranker evaluation via interleaving, they have not been considered as a framework for offline evaluation via human preference judgments. We review the literature for possible solutions. For human preference judgments, any usable algorithm must tolerate ties, since two items may appear nearly equal to assessors, and it must minimize the number of judgments required for any specific pair, since each such comparison requires an independent assessor. Since the theoretical guarantees provided by most algorithms depend on assumptions that are not satisfied by human preference judgments, we simulate selected algorithms on representative test cases to provide insight into their practical utility. Based on these simulations, one algorithm stands out for its potential. Our simulations suggest modifications to further improve its performance. Using the modified algorithm, we collect over 10,000 preference judgments for pools derived from submissions to the TREC 2021 Deep Learning Track, confirming its suitability. We test the idea of best-item evaluation and suggest ideas for further theoretical and practical progress.

Information seeking conversations between users and Conversational Search Agents (CSAs) consist of multiple turns of interaction. While users initiate a search session, ideally a CSA should sometimes take the lead in the conversation by obtaining feedback from the user by offering query suggestions or asking for query clarifications i.e. mixed initiative. This creates the potential for more engaging conversational searches, but substantially increases the complexity of modelling and evaluating such scenarios due to the large interaction space coupled with the trade-offs between the costs and benefits of the different interactions. In this paper, we present a model for conversational search -- from which we instantiate different observed conversational search strategies, where the agent elicits: (i) Feedback-First, or (ii) Feedback-After. Using 49 TREC WebTrack Topics, we performed an analysis comparing how well these different strategies combine with different mixed initiative approaches: (i) Query Suggestions vs. (ii) Query Clarifications. Our analysis reveals that there is no superior or dominant combination, instead it shows that query clarifications are better when asked first, while query suggestions are better when asked after presenting results. We also show that the best strategy and approach depends on the trade-offs between the relative costs between querying and giving feedback, the performance of the initial query, the number of assessments per query, and the total amount of gain required. While this work highlights the complexities and challenges involved in analyzing CSAs, it provides the foundations for evaluating conversational strategies and conversational search agents in batch/offline settings.

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.

The TREC Deep Learning (DL) Track studies ad hoc search in the large data regime, meaning that a large set of human-labeled training data is available. Results so far indicate that the best models with large data may be deep neural networks. This paper supports the reuse of the TREC DL test collections in three ways. First we describe the data sets in detail, documenting clearly and in one place some details that are otherwise scattered in track guidelines, overview papers and in our associated MS MARCO leaderboard pages. We intend this description to make it easy for newcomers to use the TREC DL data. Second, because there is some risk of iteration and selection bias when reusing a data set, we describe the best practices for writing a paper using TREC DL data, without overfitting. We provide some illustrative analysis. Finally we address a number of issues around the TREC DL data, including an analysis of reusability.

An emerging recipe for achieving state-of-the-art effectiveness in neural document re-ranking involves utilizing large pre-trained language models - e.g., BERT - to evaluate all individual passages in the document and then aggregating the outputs by pooling or additional Transformer layers. A major drawback of this approach is high query latency due to the cost of evaluating every passage in the document with BERT. To make matters worse, this high inference cost and latency varies based on the length of the document, with longer documents requiring more time and computation. To address this challenge, we adopt an intra-document cascading strategy, which prunes passages of a candidate document using a less expensive model, called ESM, before running a scoring model that is more expensive and effective, called ETM. We found it best to train ESM (short for Efficient Student Model) via knowledge distillation from the ETM (short for Effective Teacher Model) e.g., BERT. This pruning allows us to only run the ETM model on a smaller set of passages whose size does not vary by document length. Our experiments on the MS MARCO and TREC Deep Learning Track benchmarks suggest that the proposed Intra-Document Cascaded Ranking Model (IDCM) leads to over 400% lower query latency by providing essentially the same effectiveness as the state-of-the-art BERT-based document ranking models.

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]).

Traces of searcher behaviour, such as query reformulation or clicks, are commonly used to evaluate a running search engine. The underlying expectation is that these behaviours are proxies for something more important, such as relevance, utility, or satisfaction. Affective computing technology gives us the tools to help confirm some of these expectations, by examining visceral expressive responses during search sessions. However, work to date has only studied small populations in laboratory settings and with a limited number of contrived search tasks. In this study, we analysed longitudinal, in-situ, search behaviours of 152 information workers, over the course of several weeks while simultaneously tracking their facial expressions. Results from over 20,000 search sessions and 45,000 queries allow us to observe that indeed affective expressions are consistent with, and complementary to, existing "click-based'' metrics. On a query-level, searches that result in a short dwell time are associated with a decrease in smiles (expressions of "happiness'') and that if a query is reformulated the results of the reformulation are associated with an increase in smiling---suggesting a positive outcome as people converge on the information they need. On a session-level, sessions that feature reformulations are more commonly associated with fewer smiles and more furrowed brows (expressions of "anger/frustration''). Similarly, sessions with short-dwell clicks are also associated with fewer smiles. These data provide an insight into visceral aspects of search experience and present a new dimension for evaluating engine performance.

The Transformer-Kernel (TK) model has demonstrated strong reranking performance on the TREC Deep Learning benchmark---and can be considered to be an efficient (but slightly less effective) alternative to other Transformer-based architectures that employ (i) large-scale pretraining (high training cost), (ii) joint encoding of query and document (high inference cost), and (iii) larger number of Transformer layers (both high training and high inference costs). Since, a variant of the TK model---called TKL---has been developed that incorporates local self-attention to efficiently process longer input sequences in the context of document ranking. In this work, we propose a novel Conformer layer as an alternative approach to scale TK to longer input sequences. Furthermore, we incorporate query term independence and explicit term matching to extend the model to the full retrieval setting. We benchmark our models under the strictly blind evaluation setting of the TREC 2020 Deep Learning track and find that our proposed architecture changes lead to improved retrieval quality over TKL. Our best model also outperforms all non-neural runs ("trad") and two-thirds of the pretrained Transformer-based runs ("nnlm") on [email protected]

Conversational information retrieval is a relatively new and fast-developing research area, but conversation itself has been well studied for decades. Researchers have analysed linguistic phenomena such as structure and semantics but also paralinguistic features such as tone, body language, and even the physiological states of interlocutors. We tend to treat computers as social agents—especially if they have some humanlike features in their design—and so work from human-to-human conversation is highly relevant to how we think about the design of human-to-computer applications. In this article, we summarise some salient past work, focusing on social norms; structures; and affect, prosody, and style. We examine social communication theories briefly as a review to see what we have learned about how humans interact with each other and how that might pertain to agents and robots. We also discuss some implications for research and design of conversational IR systems.

Evaluation metrics for search typically assume items are homoge- neous. However, in the context of web search, this assumption does not hold. Modern search engine result pages (SERPs) are composed of a variety of item types (e.g., news, web, entity, etc.), and their influence on browsing behavior is largely unknown.

Users of Web search engines reveal their information needs through queries and clicks, making click logs a useful asset for information retrieval. However, click logs have not been publicly released for academic use, because they can be too revealing of personally or commercially sensitive information. This paper describes a click data release related to the TREC Deep Learning Track document corpus. After aggregation and filtering, including a k -anonymity requirement, we find 1.4 million of the TREC DL URLs have 18 million connections to 10 million distinct queries. Our dataset of these queries and connections to TREC documents is of similar size to proprietary datasets used in previous papers on query mining and ranking. We perform some preliminary experiments using the click data to augment the TREC DL training data, offering by comparison: 28x more queries, with 49x more connections to 4.4x more URLs in the corpus. We present a description of the dataset's generation process, characteristics, use in ranking and other potential uses.

Search clarification has recently attracted much attention due to its applications in search engines. It has also been recognized as a major component in conversational information seeking systems. Despite its importance, the research community still feels the lack of a large-scale dataset for studying different aspects of search clarification. In this paper, we introduce MIMICS, a collection of search clarification datasets for real web search queries sampled from the Bing query logs. Each clarification in MIMICS is generated by a Bing production algorithm and consists of a clarifying question and up to five candidate answers. MIMICS contains three datasets: (1) MIMICS-Click includes over 400k unique queries, their associated clarification panes, and the corresponding aggregated user interaction signals (i.e., clicks). (2) MIMICS-ClickExplore is an exploration data that includes aggregated user interaction signals for over 60k unique queries, each with multiple clarification panes. (3) MIMICS-Manual includes over 2k unique real search queries. Each query-clarification pair in this dataset has been manually labeled by at least three trained annotators. It contains graded quality labels for the clarifying question, the candidate answer set, and the landing result page for each candidate answer.

Past work in information-seeking conversation has demonstrated that people exhibit different conversational styles---for example, in word choice or prosody---that differences in style lead to poorer conversations, and that partners actively align their styles over time. One might assume that this would also be true for conversations with an artificial agent such as Cortana, Siri, or Alexa; and that agents should therefore track and mimic a user's style. We examine this hypothesis with reference to a lab study, where 24 participants carried out relatively long information-seeking tasks with an embodied conversational agent. The agent combined topical language models with a conversational dialogue engine, style recognition and alignment modules. We see that "style'' can be measured in human-to-agent conversation, although it looks somewhat different to style in human-to-human conversation and does not correlate with self-reported preferences. There is evidence that people align their style to the agent, and that conversations run more smoothly if the agent detects, and aligns to, the human's style as well.

Neural networks, particularly Transformer-based architectures, have achieved significant performance improvements on several retrieval benchmarks. When the items being retrieved are documents, the time and memory cost of employing Transformers over a full sequence of document terms can be prohibitive. A popular strategy involves considering only the first n terms of the document. This can, however, result in a biased system that under retrieves longer documents. In this work, we propose a local self-attention which considers a moving window over the document terms and for each term attends only to other terms in the same window. This local attention incurs a fraction of the compute and memory cost of attention over the whole document. The windowed approach also leads to more compact packing of padded documents in minibatches resulting in additional savings. We also employ a learned saturation function and a two-staged pooling strategy to identify relevant regions of the document. The Transformer-Kernel pooling model with these changes can efficiently elicit relevance information from documents with thousands of tokens. We benchmark our proposed modifications on the document ranking task from the TREC 2019 Deep Learning track and observe significant improvements in retrieval quality as well as increased retrieval of longer documents at moderate increase in compute and memory costs.

As deep learning based models are increasingly being used for information retrieval, a major challenge is to ensure the availability of test collections for measuring their quality. Test collections are usually generated based on pooling results of various retrieval systems, but until recently this did not include deep learning systems. This raises a major challenge for reusable evaluation: Since deep learning based models use external resources (e.g. word embeddings) and advanced representations when compared to traditional methods, they may return different types of relevant document that were not identified in the original pooling. If so, test collections constructed using traditional methods could lead to biased and unfair evaluation results for deep learning systems. This paper uses simulated pooling to test the fairness and reusability of test collections, showing that especially when shallow pools (e.g. depth-10 pools) are used, pooling based on traditional systems only may lead to biased evaluation of deep learning systems.

Conversational information seeking (CIS) has been recognized as a major emerging research area in information retrieval. Such research will require data and tools, to allow the implementation and study of conversational systems. This paper introduces Macaw, an open-source framework with a modular architecture for CIS research. Macaw supports multi-turn, multi-modal, and mixed-initiative interactions, and enables research for tasks such as document retrieval, question answering, recommendation, and structured data exploration. It has a modular design to encourage the study of new CIS algorithms, which can be evaluated in batch mode. It can also integrate with a user interface, which allows user studies and data collection in an interactive mode, where the back end can be fully algorithmic or a wizard of oz setup. Macaw is distributed under the MIT License.

Asking clarifying questions in response to search queries has been recognized as a useful technique for revealing the underlying intent of the query. Clarification has applications in retrieval systems with different interfaces, from the traditional web search interfaces to the limited bandwidth interfaces as in speech-only and small screen devices. Generation and evaluation of clarifying questions have been recently studied in the literature. However, user interaction with clarifying questions is relatively unexplored. In this paper, we conduct a comprehensive study by analyzing large-scale user interactions with clarifying questions in a major web search engine. In more detail, we analyze the user engagements received by clarifying questions based on different properties of search queries, clarifying questions, and their candidate answers. We further study click bias in the data, and show that even though reading clarifying questions and candidate answers does not take significant efforts, there still exist some position and presentation biases in the data. We also propose a model for learning representation for clarifying questions based on the user interaction data as implicit feedback. The model is used for re-ranking a number of automatically generated clarifying questions for a given query. Evaluation on both click data and human labeled data demonstrates the high quality of the proposed method.

This paper studies a new scenario in conversational search, conversational question suggestion, which leads search engine users to more engaging experiences by suggesting interesting, informative, and useful follow-up questions. We first establish a novel evaluation metric, usefulness, which goes beyond relevance and measures whether the suggestions provide valuable information for the next step of a user’s journey, and construct a public benchmark for useful question suggestion. Then we develop two suggestion systems, a BERT based ranker and a GPT-2 based generator, both trained with novel weak supervision signals that convey past users’ search behaviors in search sessions. The weak supervision signals help ground the suggestions to users’ information-seeking trajectories: we identify more coherent and informative sessions using encodings, and then weakly supervise our models to imitate how users transition to the next state of search. Our offline experiments demonstrate the crucial role our “next-turn” inductive training plays in improving usefulness over a strong online system. Our online A/B test in Bing shows that our more useful question suggestions receive 8% more user clicks than the previous system.

Search queries are often short, and the underlying user intent may be ambiguous. This makes it challenging for search engines to predict possible intents, only one of which may pertain to the current user. To address this issue, search engines often diversify the result list and present documents relevant to multiple intents of the query. An alternative approach is to ask the user a question to clarify her information need. Asking clarifying questions is particularly important for scenarios with “limited bandwidth” interfaces, such as speech-only and small-screen devices. In addition, our user studies and large-scale online experiments show that asking clarifying questions is also useful in web search. Although some recent studies have pointed out the importance of asking clarifying questions, generating them for open-domain search tasks remains unstudied and is the focus of this paper. Lack of training data even within major search engines for this task makes it challenging. To mitigate this issue, we first identify a taxonomy of clarification for open-domain search queries by analyzing large-scale query reformulation data sampled from Bing search logs. This taxonomy leads us to a set of question templates and a simple yet effective slot filling algorithm. We further use this model as a source of weak supervision to automatically generate clarifying questions for training. Furthermore, we propose supervised and reinforcement learning models for generating clarifying questions learned from weak supervision data. We also investigate methods for generating candidate answers for each clarifying question, so users can select from a set of pre-defined answers. Human evaluation of the clarifying questions and candidate answers for hundreds of search queries demonstrates the effectiveness of the proposed solutions.

Modern web search engines use many signals to select and rank results in response to queries. However, searchers’ mental models of search are relatively unsophisticated, hindering their ability to use search engines efficiently and effectively. Annotating results with more in-depth explanations could help, but search engine providers need to know what to explain. To this end, we report on a study of searchers’ mental models of web selection and ranking, with more than 400 respondents to an online survey and 11 face-to-face interviews. Participants volunteered a range of factors and showed good understanding of important concepts such as popularity, wording, and personalization. However, they showed little understanding of recency or diversity and incorrect ideas of payment for ranking. Where there are already explanatory annotations on the results page—such as “ad” markers and keyword highlighting—participants were familiar with ranking concepts. This suggests that further explanatory annotations may be useful.

Emotions are an essential part of most human activities, including decision-making. Emotions arise in response to information, e.g., presented in web pages, and are also expressed in the words used to convey that information in the first place. In this paper, we study the emotion profile of retrieved and clicked web search results towards the goal of better understanding the role of emotions in web search. Using click logs from a four-month period, up to the end of January 2019, we examine the emotions associated with search results and contrast them to the emotions of clicked results, taking rank and relevance into account. Emotions are assigned to web pages based on two lexicons: SentiWordNet (positive, negative and objective sentiments) and EmoLexData (afraid, amused, angry, annoyed, don't care, happy, inspired, and sad emotions). We look at the sentiment/emotion profiles of search results grouped around a set of controversial and mundane topics and hypothesise that users are more likely to click emotionally charged results than emotionless results, both in general, and in particular when their query relates to controversial topics.

Axiomatic information retrieval (IR) seeks a set of principle properties desirable in IR models. These properties when formally expressed provide guidance in the search for better relevance estimation functions. Neural ranking models typically contain many learnable parameters. The training of these models involves a search for appropriate parameter values based on large quantities of labeled examples. Intuitively, axioms that can guide the search for better traditional IR models should also help in better parameter estimation for machine learning based rankers. This work explores the use of IR axioms to augment the direct supervision from labeled data for training neural ranking models. We modify the documents in our dataset along the lines of well-known axioms during training and add a regularization loss based on the agreement between the ranking model and the axioms on which version of the document---the original or the perturbed---should be preferred. Our experiments show that the neural ranking model achieves faster convergence and better generalization with axiomatic regularization.

This paper presents GEneric iNtent Encoder (GEN Encoder) which learns a distributed representation space for user intent in search. Leveraging large scale user clicks from Bing search logs as weak supervision of user intent, GEN Encoder learns to map queries with shared clicks into similar embeddings end-to-end and then fine-tunes on multiple paraphrase tasks. Experimental results on an intrinsic evaluation task - query intent similarity modeling - demonstrate GEN Encoder's robust and significant advantages over previous representation methods. Ablation studies reveal the crucial role of learning from implicit user feedback in representing user intent and the contributions of multi-task learning in representation generality. We also demonstrate that GEN Encoder alleviates the sparsity of tail search traffic and cuts down half of the unseen queries by using an efficient approximate nearest neighbor search to effectively identify previous queries with the same search intent. Finally, we demonstrate distances between GEN encodings reflect certain information seeking behaviors in search sessions.

Analysis of casual chit-chat indicates that differences in conversational style---the way things are said---can significantly impact a participants» impressions of the conversation and of each other. However, prior work has not systematically analyzed how important style is in task-oriented, information-seeking exchanges of the sort we might have with a conversational search agent. We examine recordings from the MISC data set, where pairs of "users" and "intermediaries" collaborate on information-seeking tasks, and look for indications of style which can be computed at scale. We find that stylistic markers identified by Tannen in casual chat do exist in information-seeking dialogue, and that participants can be arranged along a single stylistic dimension: "considerate" to "involved". This labelling for style needs no manual intervention. Furthermore, we find that there is no clear best style; but that differences in style, previously thought to impede communication, are only a problem for shorter tasks. This result is likely due to alignment of conversational style over the course of an interaction.

Relevance judgments are essential for designing information retrieval systems. Traditionally, judgments have been gathered via desktop interfaces. However, with the rise in popularity of smaller devices for information access, it has become imperative to investigate whether desktop based judgments are different from mobile judgments. Recently, user effort and document usefulness have also emerged as important dimensions to optimize and evaluate information retrieval systems. Since existing work is limited to desktops, it remains to be seen how these judgments are affected by user»s search device. In this paper, we address these shortcomings by collecting and analyzing relevance, usefulness and effort judgments on mobiles and desktops. Analysis of these judgments shows high agreement rate between desktop and mobile judges for relevance, followed by usefulness and findability. We also found that desktop judges are likely to spend more time and examine non-relevant/not-useful/difficult documents in greater depth compared to mobile judges. Based on our findings, we suggest that relevance judgments should be gathered via desktops and effort judgments should be collected on each device independently.

Web Search Engine Result Pages (SERPs) are complex responses to queries, containing many heterogeneous result elements (web results, advertisements, and specialised "answers'') positioned in a variety of layouts. This poses numerous challenges when trying to measure the quality of a SERP because standard measures were designed for homogeneous ranked lists. In this paper, we aim to measure the utility and cost of SERPs. To ground this work we adopt the \CWL framework which enables a direct comparison between different measures in the same units of measurement, i.e. expected (total) utility and cost. Within this framework, we propose a new measure based on information foraging theory, which can account for the heterogeneity of elements, through different costs, and which naturally motivates the development of a user stopping model that adapts behaviour depending on the rate of gain. This directly connects models of how people search with how we measure search, providing a number of new dimensions in which to investigate and evaluate user behaviour and performance. We perform an analysis over~1000 popular queries issued to a major search engine, and report the aggregate utility experienced by users over time. Then in an comparison against common measures, we show that the proposed foraging based measure provides a more accurate reflection of the utility and of observed behaviours (stopping rank and time spent).

In recent years, machine learning approaches, and in particular deep neural networks, have yielded significant improvements on several natural language processing and computer vision tasks; however, such breakthroughs have not yet been observed in the area of information retrieval. Besides the complexity of IR tasks, such as understanding the user's information needs, a main reason is the lack of high-quality and/or large-scale training data for many IR tasks. This necessitates studying how to design and train machine learning algorithms where there is no large-scale or high-quality data in hand. Therefore, considering the quick progress in development of machine learning models, this is an ideal time for a workshop that especially focuses on learning in such an important and challenging setting for IR tasks. The goal of this workshop is to bring together researchers from industry---where data is plentiful but noisy---with researchers from academia---where data is sparse but clean to discuss solutions to these related problems.

Deep neural networks have recently shown promise in the ad-hoc retrieval task. However, such models have often been based on one field of the document, for example considering document title only or document body only. Since in practice documents typically have multiple fields, and given that non-neural ranking models such as BM25F have been developed to take advantage of document structure, this paper investigates how neural models can deal with multiple document fields. We introduce a model that can consume short text fields such as document title and long text fields such as document body. It can also handle multi-instance fields with variable number of instances, for example where each document has zero or more instances of incoming anchor text. Since fields vary in coverage and quality, we introduce a masking method to handle missing field instances, as well as a field-level dropout method to avoid relying too much on any one field. As in the studies of non-neural field weighting, we find it is better for the ranker to score the whole document jointly, rather than generate a per-field score and aggregate. We find that different document fields may match different aspects of the query and therefore benefit from comparing with separate representations of the query text. The combination of techniques introduced here leads to a neural ranker that can take advantage of full document structure, including multiple instance and missing instance data, of variable length. The techniques significantly enhance the performance of the ranker, and outperform a learning to rank baseline with hand-crafted features.

Email continues to be an important form of communication as well as a way to manage tasks and archive personal information. As the volume of email grows, organizing and finding relevant email remains challenging. In this paper, we present a large-scale log analysis of the activities that people perform on email mes-sages (accessing external information via links or attachments, responding to messages, and organizing messages), their search behavior, and their organizational practices in a popular web email client.

This paper studies conversational approaches to information retrieval, presenting a theory and model of information interaction in a chat setting. In particular, we consider the question of what properties would be desirable for a conversational information retrieval system so that the system can allow users to answer a variety of information needs in a natural and efficient manner. We study past work on human conversations, and propose a small set of properties that taken together could measure the extent to which a system is conversational. Following this, we present a theoretical model of a conversational system that implements the properties. We describe how this system could be implemented, making the action space of an conversational search agent explicit. Our analysis of this model shows that while theoretical, the model could be practically implemented to satisfy the desirable properties presented. In doing so, we show that the properties are also feasible.

Email has been a dominant form of communication for many years, and email search is an important problem. In contrast to other search setting, such as web search, there have been few studies of user behavior and models of email search success. Research in email search is challenging for many reasons including the personal and private nature of the collection. Third party judges can not look at email search queries or email message content requiring new modeling techniques.

In recent years, deep neural networks have yielded significant performance improvements in application areas such as speech recognition, computer vision, and machine translation. This has led to expectations in the information retrieval (IR) community that these novel machine learning approaches are likely to demonstrate a similar scale of breakthroughs on IR tasks within the next couple of years. In the Neu-IR (pronounced "new IR") 2016 workshop, however, there was a growing concern that the lack of availability of large scale training and evaluation datasets may be hindering the research community from making adequate progress in this area. It was also highlighted that the community would benefit from establishing a shared public repository of neural IR models and shared evaluation resources for better reproducibility and speed of experimentation. After the first successful Neu-IR workshop at SIGIR 2016, our goal this year will be to host a highly interactive full-day workshop to bring the neural IR community together to specifically address these key challenges facing this line of research. The workshop will request the community to submit proposals on generating large scale benchmark collections, building a shared model repository, and standardizing frameworks appropriate for evaluating deep neural network models. In addition, the workshop will provide a forum for the growing community of IR researchers to present their recent (published and unpublished) work involving (shallow or deep) neural network based approaches in an interactive poster session.

In recent years, the information retrieval (IR) community has witnessed the first successful applications of deep neural network models to short-text matching and ad-hoc retrieval tasks. However, the two communities - focused on deep neural networks and on IR - have less in common when it comes to the choice of programming languages. Indri, an indexing framework popularly used by the IR community, is written in C++, while Torch, a popular machine learning library for deep learning, is written in the light-weight scripting language Lua. To bridge this gap, we introduce Luandri (pronounced "laundry"), a simple interface for exposing the search capabilities of Indri to Torch models implemented in Lua.

A fundamental concern in search engines is to determine which documents have the best content for satisfying the user, based on analysis of the user's query and the text of documents. For this query-content match, many learning to rank systems make use of IR features developed in the 1990s in the TREC framework. Such features are still important in a variety of search tasks, and particularly in the long tail where clicks, links and social media signals become sparse. I will present our current progress, in particular three different neural models, with the goal of surpassing the 1990s models in full text search. This will include evidence, using proprietary Bing datasets, that large-scale training data can be useful. I will also argue that for the field to make progress on query-content relevance modeling, it may be valuable to set up a shared blind evaluation similar to 1990s TREC, possibly with large-scale training data.

In the last few years, neural representation learning approaches have achieved very good performance on many natural language processing tasks, such as language modelling and machine translation. This suggests that neural models will also achieve good performance on information retrieval (IR) tasks, such as relevance ranking, addressing the query-document vocabulary mismatch problem by using a semantic rather than lexical matching. Although initial iterations of neural models do not outperform traditional lexical-matching baselines, the level of interest and effort in this area is increasing, potentially leading to a breakthrough. The popularity of the recent SIGIR 2016 workshop on Neural Information Retrieval provides evidence to the growing interest in neural models for IR. While recent tutorials have covered some aspects of deep learning for retrieval tasks, there is a significant scope for organizing a tutorial that focuses on the fundamentals of representation learning for text retrieval. The goal of this tutorial will be to introduce state-of-the-art neural embedding models and bridge the gap between these neural models with early representation learning approaches in IR (e.g., LSA). We will discuss some of the key challenges and insights in making these models work in practice, and demonstrate one of the toolsets available to researchers interested in this area.

As the number of email users and messages continues to grow, search is becoming more important for finding information in personal archives. In spite of its importance, email search is much less studied than web search, particularly using large-scale behavioral log analysis. In this paper we report the results of a large-scale log analysis of email search and complement this with a survey to better understand email search intent and success. We characterize email search behaviors and highlight differences from web search. When searching for email, people know many attributes about what they are looking for; they often look for specific known items; their queries are shorter and they click on fewer items than in web search. Although repeat queries are common in both email and web search, repeat visits to the same search result are much less common in email search suggesting that the same query is used for different search intents over time. We consider search intent from multiple angles. In email search logs, we find that people use email search not just to find information but also to perform tasks such as cleanup or organization, and that the distribution of actions they perform depends on the type of query. In our survey, people reported that they looked for specific information in both email search and web search, but they were much less likely to search for general information on a topic in email. The differences in overall behavior, re-finding patterns and search intents we observed between email and web search have important implications for the design of email search algorithms and interfaces.

Models such as latent semantic analysis and those based on neural embeddings learn distributed representations of text, and match the query against the document in the latent semantic space. In traditional information retrieval models, on the other hand, terms have discrete or local representations, and the relevance of a document is determined by the exact matches of query terms in the body text. We hypothesize that matching with distributed representations complements matching with traditional local representations, and that a combination of the two is favourable. We propose a novel document ranking model composed of two separate deep neural networks, one that matches the query and the document using a local representation, and another that matches the query and the document using learned distributed representations. The two networks are jointly trained as part of a single neural network. We show that this combination or 'duet' performs significantly better than either neural network individually on a Web page ranking task, and significantly outperforms traditional baselines and other recently proposed models based on neural networks.

Why do people start a search? Why do they stop? Why do they do what they do in-between? Our goal in this paper is to provide a simple yet general explanation for these acts that has its basis in neuropsychology and observed user behavior. We coin the term "ingram", as an information counterpart to Richard Semon's ?engram? or "memory trace". People search to create ingrams. People stop searching because they have created sufficient ingrams, or given up. We describe these acts through a pair of user models and use it to explain various user behaviors in search activity. Understanding people?s search acts in terms of ingrams may help us predict or model the interaction of people?s information needs, the queries they issue, and the information they consume. If we could observe certain decision-making acts within these activities, we might also gain new insight into the relationships between textual information and knowledge representation.

In recent years, deep neural networks have yielded significant performance improvements on speech recognition and computer vision tasks, as well as led to exciting breakthroughs in novel application areas such as automatic voice translation, image captioning, and conversational agents. Despite demonstrating good performance on natural language processing (NLP) tasks (e.g., language modelling and machine translation, the performance of deep neural networks on information retrieval (IR) tasks has had relatively less scrutiny. Recent work in this area has mainly focused on word embeddings and neural models for short text similarity.

Gathering evidence about whether a search result is relevant is a core concern in the evaluation and improvement of information retrieval systems. Two common sources of evidence for establishing relevance are judgements from trained assessors and logs of online user behavior. However, both are limited; it is hard for a trained assessor to know exactly what users want to find, and user behavior only provides an implicit and ambiguous signal. In this paper, we aim to address these limitations by collecting explicit feedback on web search results from users in situ as they search. When users return to the search result page via the browser back button after having clicked on a result, we ask them to provide a binary thumbs up or thumbs down judgment and text feedback. We collect in situ feedback from a large commercial search engine, and compare this feedback with the judgments provided by trained assessors. We find that in situ feedback differs significantly from traditional relevance judgments, and that it suggests a different interpretation of behavior signals, with the dwell time threshold between negative and positive in situ feedback being 87 seconds, longer than the more common heuristic of 30 seconds. Using text feedback from users, we discuss why user feedback may differ from editorial judgments.

Document relevance has been the primary focus in the design, optimization and evaluation of retrieval systems. Traditional testcollections are constructed by asking judges the relevance grade for a document with respect to an input query. Recent work of Yilmaz et al. found an evidence that effort is another important factor in determining document utility, suggesting that more thought should be given into incorporating effort into information retrieval. However, that work did not ask judges to directly assess the level of effort required to consume a document or analyse how effort judgements relate to traditional relevance judgements.

Web search via voice is becoming increasingly popular, taking advantage of recent advances in automatic speech recognition. Speech recognition systems are trained using audio transcripts, which can be generated by a paid annotator listening to some audio and manually transcribing it. This paper considers an alternative source of training data for speech recognition, called implicit transcription. This is based on Web search clicks and reformulations, which can be interpreted as validating or correcting the recognition done during a real Web search. This can give a large amount of free training data that matches the exact characteristics of real incoming voice searches and the implicit transcriptions can better reflect the needs of real users because they come from the user who generated the audio. On an overall basis we demonstrate that the new training data has value in improving speech recognition. We further show that the in-context feedback from real users can allow the speech recognizer to exploit contextual signals, and reduce the recognition error rate further by up to 23%.

Query auto-completion (QAC) systems typically suggest queries that have previously been observed in search logs. Given a partial user query, the system looks up this query prefix against a precomputed set of candidates, then orders them using ranking signals such as popularity. Such systems can only recommend queries for prefixes that have been previously seen by the search engine with adequate frequency. They fail to recommend if the prefix is sufficiently rare such that it has no matches in the precomputed candidate set.

When searching for place entities such as businesses or points of interest, the desired place may be close (finding the nearest ATM) or far away (finding a hotel in another city). Understanding the role of distance in predicting user interests can guide the design of location search and recommendation systems. We analyze a large dataset of location searches on GPS-enabled mobile devices with 15 location categories. We model user-location distance based on raw geographic distance (kilometers) and intervening opportunities (nth closest). Both models are helpful in predicting user interests, with the intervening opportunity model performing somewhat better. We find significant inter-category variation. For instance, the closest movie theater is selected in 17.7% of cases, while the closest restaurant in only 2.1% of cases. Overall, we recommend taking category information into account when modeling location preferences of users in search and recommendation systems.

Test collections play an important role in adhoc and diversity retrieval evaluation. Constructing a test collection for adhoc evaluation involves (1) selecting a set of queries to be judged, (2) selecting an intent (topic) description for that query, and (3) obtaining relevance judgments with respect to the specific intent description for that particular query. Recent work showed that the selection of intents play an important role in the relative performance of retrieval systems for diversity evaluation. However, no previous work has analysed how the choice of a specific intent description may affect adhoc evaluation. We show that intent descriptions have a significant impact in adhoc evaluation and that special care should be given as to how the intent descriptions are selected. We further show that it is better to have very general intent descriptions or no intent descriptions at all when constructing test collections for adhoc evaluation. We then focus on diversity evaluation and identify the effect intent descriptions have on diversity based retrieval evaluation. We quantify this effect and discuss experimental design decisions for the optimal distribution of judgment effort across different intents for a query vs. different queries.

In this paper, we study one important source of the mis-match between user data and relevance judgments, those due to the high degree of effort required by users to identify and consume the information in a document. Information retrieval relevance judges are trained to search for evidence of relevance when assessing documents. For complex documents, this can lead to judges' spending substantial time considering each document. However, in practice, search users are often much more impatient: if they do not see evidence of relevance quickly, they tend to give up.

WSCD 2014 is the fourth workshop on Web Search Click Data, following WSCD 2009, WSCD 2011 and WSCD 2012. It is a forum for new research relating to Web search usage logs and for discussing desirable properties of publicly released search log datasets. Research relating to search logs has been hampered by the limited availability of click datasets. This series of workshops comes with new datasets based on logged user search behaviour and accompanying data mining challenges. This year the challenge and the workshop are focused on the tasks of personalization using logs.

To understand whether a user is satisfied with the current search results, implicit behavior is a useful data source, with clicks being the best-known implicit signal. However, it is possible for a non-clicking user to be satisfied and a clicking user to be dissatisfied. Here we study additional implicit signals based on the relationship between the user's current query and the next query, such as their textual similarity and the inter-query time. Using a large unlabeled dataset, a labeled dataset of queries and a labeled dataset of user tasks, we analyze the relationship between these signals. We identify an easily-implemented rule that indicates dissatisfaction: that a similar query issued within a time interval that is short enough (such as five minutes) implies dissatisfaction. By incorporating additional query-based features in the model, we show that a query-based model (with no click information) can indicate satisfaction more accurately than click-based models. The best model uses both query and click features. In addition, by comparing query sequences in successful tasks and unsuccessful tasks, we observe that search success is an incremental process for successful tasks with multiple queries.

Preference based methods for collecting relevance data for information retrieval (IR) evaluation have been shown to lead to better inter-assessor agreement than the traditional method of judging individual documents. However, little is known as to why preference judging reduces assessor disagreement and whether better agreement among assessors also means better agreement with user satisfaction, as signaled by user clicks. In this paper, we examine the relationship between assessor disagreement and various click based measures, such as click preference strength and user intent similarity, for judgments collected from editorial judges and crowd workers using single absolute, pairwise absolute and pairwise preference based judging methods. We find that trained judges are significantly more likely to agree with each other and with users than crowd workers, but inter-assessor agreement does not mean agreement with users. Switching to a pairwise judging mode improves crowdsourcing quality close to that of trained judges. We also find a relationship between intent similarity and assessor-user agreement, where the nature of the relationship changes across judging modes. Overall, our findings suggest that the awareness of different possible intents, enabled by pairwise judging, is a key reason of the improved agreement, and a crucial requirement when crowdsourcing relevance data.

Query rewriting algorithms can be used as a form of query expansion, by combining the user's original query with automatically generated rewrites. Rewriting algorithms bring linguistic datasets to bear without the need for iterative relevance feedback, but most studies of rewriting have used proprietary datasets such as large-scale search logs. By contrast this paper uses readily available data, particularly ClueWeb09 link text with over 1.2 billion anchor phrases, to generate rewrites. To avoid overfitting, our initial analysis is performed using Million Query Track queries, leading us to identify three algorithms which perform well. We then test the algorithms on Web and newswire data. Results show good properties in terms of robustness and early precision.

Interleaving is an online evaluation technique for comparing the relative quality of information retrieval functions by combining their result lists and tracking clicks. A sequence of such algorithms have been proposed, each being shown to address problems in earlier algorithms. In this paper, we formalize and generalize this process, while introducing a formal model: We identify a set of desirable properties for interleaving, then show that an interleaving algorithm can be obtained as the solution to an optimization problem within those constraints. Our approach makes explicit the parameters of the algorithm, as well as assumptions about user behavior. Further, we show that our approach leads to an unbiased and more efficient interleaving algorithm than any previous approach, using a novel log-based analysis of user search behavior.

WSCD 2013 is the third workshop on Web Search Click Data, following WSCD 2009 and WSCD 2012. It is a forum for new research relating to Web search usage logs and for discussing desirable properties of publicly released search log datasets. Research relating to search logs has been hampered by the limited availability of click datasets. This workshop comes with a new dataset based on logged user search behaviour and an accompanying challenge to predict switches between search engines within a given search session.

Topic modeling provides a powerful way to analyze the content of a collection of documents. It has become a popular tool in many research areas, such as text mining, information retrieval, natural language processing, and other related fields. In real-world applications, however, the usefulness of topic modeling is limited due to scalability issues. Scaling to larger document collections via parallelization is an active area of research, but most solutions require drastic steps, such as vastly reducing input vocabulary. In this article we introduce Regularized Latent Semantic Indexing (RLSI)---including a batch version and an online version, referred to as batch RLSI and online RLSI, respectively---to scale up topic modeling. Batch RLSI and online RLSI are as effective as existing topic modeling techniques and can scale to larger datasets without reducing input vocabulary. Moreover, online RLSI can be applied to stream data and can capture the dynamic evolution of topics. Both versions of RLSI formalize topic modeling as a problem of minimizing a quadratic loss function regularized by ℓ1 and/or ℓ2 norm. This formulation allows the learning process to be decomposed into multiple suboptimization problems which can be optimized in parallel, for example, via MapReduce. We particularly propose adopting ℓ1 norm on topics and ℓ2 norm on document representations to create a model with compact and readable topics and which is useful for retrieval. In learning, batch RLSI processes all the documents in the collection as a whole, while online RLSI processes the documents in the collection one by one. We also prove the convergence of the learning of online RLSI. Relevance ranking experiments on three TREC datasets show that batch RLSI and online RLSI perform better than LSI, PLSI, LDA, and NMF, and the improvements are sometimes statistically significant. Experiments on a Web dataset containing about 1.6 million documents and 7 million terms, demonstrate a similar boost in performance.

Test collections are powerful mechanisms for the evaluation and optimization of information retrieval systems. However, there is reported evidence that experiment outcomes can be affected by changes to the judging guidelines or changes in the judge population. This paper examines such effects in a web search setting, comparing the judgments of four groups of judges: NIST Web Track judges, untrained crowd workers and two groups of trained judges of a commercial search engine. Our goal is to identify systematic judging errors by comparing the labels contributed by the different groups, working under the same or different judging guidelines. In particular, we focus on detecting systematic differences in judging depending on specific characteristics of the queries and URLs. For example, we ask whether a given population of judges, working under a given set of judging guidelines, are more likely to consistently overrate Wikipedia pages than another group judging under the same instructions. Our approach is to identify judging errors with respect to a consensus set, a judged gold set and a set of user clicks. We further demonstrate how such biases can affect the training of retrieval systems.

In information retrieval, relevance judgments play an important role as they are required both for evaluating the quality of retrieval systems and for training learning to rank algorithms. In recent years, numerous papers have been published using judgments obtained from a commercial search engine by researchers in industry. As typically no information is provided about the quality of these judgments, their reliability for evaluating/training retrieval systems remains questionable. In this paper, we analyze the reliability of such judgments for evaluating the quality of retrieval systems by comparing them to judgments by NIST judges at TREC.

WSCD2012 is the second workshop on Web Search Click Data, following WSCD2009. It is a forum for new research relating to Web search usage logs and for discussing desirable properties of publicly released search log datasets. Research relating to search logs has been hampered by the limited availability of click datasets. This workshop comes with a new click dataset based on click logs and an accompanying challenge to predict the relevance of documents based on clicks.

Topic modeling can boost the performance of information retrieval, but its real-world application is limited due to scalability issues. Scaling to larger document collections via parallelization is an active area of research, but most solutions require drastic steps such as vastly reducing input vocabulary. We introduce Regularized Latent Semantic Indexing (RLSI), a new method which is designed for parallelization. It is as effective as existing topic models, and scales to larger datasets without reducing input vocabulary. RLSI formalizes topic modeling as a problem of minimizing a quadratic loss function regularized by l₂ and/or l₁ norm. This formulation allows the learning process to be decomposed into multiple sub-optimization problems which can be optimized in parallel, for example via MapReduce. We particularly propose adopting l₂ norm on topics and l₁ norm on document representations, to create a model with compact and readable topics and useful for retrieval. Relevance ranking experiments on three TREC datasets show that RLSI performs better than LSI, PLSI, and LDA, and the improvements are sometimes statistically significant. Experiments on a web dataset, containing about 1.6 million documents and 7 million terms, demonstrate a similar boost in performance on a larger corpus and vocabulary than in previous studies.

Traditional editorial effectiveness measures, such as nDCG, remain standard for Web search evaluation. Unfortunately, these traditional measures can inappropriately reward redundant information and can fail to reflect the broad range of user needs that can underlie a Web query. To address these deficiencies, several researchers have recently proposed effectiveness measures for novelty and diversity. Many of these measures are based on simple cascade models of user behavior, which operate by considering the relationship between successive elements of a result list. The properties of these measures are still poorly understood, and it is not clear from prior research that they work as intended. In this paper we examine the properties and performance of cascade measures with the goal of validating them as tools for measuring effectiveness. We explore their commonalities and differences, placing them in a unified framework; we discuss their theoretical difficulties and limitations, and compare the measures experimentally, contrasting them against traditional measures and against other approaches to measuring novelty. Data collected by the TREC 2009 Web Track is used as the basis for our experimental comparison. Our results indicate that these measures reward systems that achieve an balance between novelty and overall precision in their result lists, as intended. Nonetheless, other measures provide insights not captured by the cascade measures, and we suggest that future evaluation efforts continue to report a variety of measures.

Search engines can automatically reformulate user queries in a variety of ways, often leading to multiple queries that are candidates to replace the original. However, selecting a replacement can be risky: a reformulation may be more effective than the original or significantly worse, depending on the nature of the query, the source of reformulation candidates, and the corpus. In this paper, we explore methods to mitigate this risk by issuing several versions of the query (including the original) and merging their results. We focus on reformulations generated by random walks on the click graph, a method that can produce very good reformulations but is also variable and prone to topic drift. Our primary contribution is λ-Merge, a supervised merging method that is trained to directly optimize a retrieval metric (such as NDCG or MAP) using features that describe both the reformulations and the documents they return. In experiments on Bing data and GOV2, λ-Merge outperforms the original query and several unsupervised merging methods. λ-Merge also outperforms a supervised method to predict and select the best single formulation, and is competitive with an oracle that always selects the best formulation.

Most information retrieval evaluation metrics are designed to measure the satisfaction of the user given the results returned by a search engine. In order to evaluate user satisfaction, most of these metrics have underlying user models, which aim at modeling how users interact with search engine results. Hence, the quality of an evaluation metric is a direct function of the quality of its underlying user model. This paper proposes EBU, a new evaluation metric that uses a sophisticated user model tuned by observations over many thousands of real search sessions. We compare EBU with a number of state of the art evaluation metrics and show that it is more correlated with real user behavior captured by clicks.

Whole page relevance defines how well the surface-level representation of all elements on a search result page and the corresponding holistic attributes of the presentation respond to users’ information needs. We introduce a method for evaluating the whole-page relevance of Web search engine results pages. Our key contribution is that the method allows us to investigate aspects of component relevance that are difficult or impossible to judge in isolation. Such aspects include component-level information redundancy and cross-component coherence. The method we describe complements traditional document relevance measurement, affords comparative relevance assessment across multiple search engines, and facilitates the study of important factors such as brand presentation effects and component-level quality.

Information retrieval effectiveness is usually evaluated using measures such as Normalized Discounted Cumulative Gain (NDCG), Mean Average Precision (MAP) and Precision at some cutoff ([email protected]) on a set of judged queries. Recent research has suggested an alternative, evaluating information retrieval systems based on user behavior. Particularly promising are experiments that interleave two rankings and track user clicks. According to a recent study, interleaving experiments can identify large differences in retrieval effectiveness with much better reliability than other click-based methods.

To evaluate the diversity of search results, test collections have been developed that identify multiple intents for each query. Intents are the different meanings or facets that should be covered in a search results list. This means that topic development involves proposing a set of intents for each query. We propose four measurable properties of query-to-intent mappings, allowing for more principled topic development for such test collections.

Many researchers have noted that web search queries are often ambiguous or unclear. We present an approach for identifying the popular meanings of queries using web search logs and user click behavior. We show our approach to produce more complete and user-centric intents than expert judges by evaluating on TREC queries. This approach was also used by the TREC 2009 Web Track judges to obtain more representative topic descriptions from real queries.

Crawl selection policy has a direct influence on Web search effectiveness, because a useful page that is not selected for crawling will also be absent from search results. Yet there has been little or no work on measuring this effect. We introduce an evaluation framework, based on relevance judgments pooled from multiple search engines, measuring the maximum potential NDCG that is achievable using a particular crawl. This allows us to evaluate different crawl policies and investigate important scenarios like selection stability over multiple iterations. We conduct two sets of crawling experiments at the scale of 1~billion and 100~million pages respectively. These show that crawl selection based on PageRank, indegree and trans-domain indegree all allow better retrieval effectiveness than a simple breadth-first crawl of the same size. PageRank is the most reliable and effective method. Trans-domain indegree can outperform PageRank, but over multiple crawl iterations it is less effective and more unstable. Finally we experiment with combinations of crawl selection methods and per-domain page limits, which yield crawls with greater potential NDCG than PageRank.

A lot of the world's knowledge is stored in books, which, as a result of recent mass-digitisation efforts, are increasingly available online. Search engines, such as Google Books, provide mechanisms for searchers to enter this vast knowledge space using queries as entry points. In this paper, we view Wikipedia as a summary of this world knowledge and aim to use this resource to guide users to relevant books. Thus, we investigate possible ways of using Wikipedia as an intermediary between the user's query and a collection of books being searched. We experiment with traditional query expansion techniques, exploiting Wikipedia articles as rich sources of information that can augment the user's query. We then propose a novel approach based on link distance in an extended Wikipedia graph: we associate books with Wikipedia pages that cite these books and use the link distance between these nodes and the pages that match the user query as an estimation of a book's relevance to the query. Our results show that a) classical query expansion using terms extracted from query pages leads to increased precision, and b) link distance between query and book pages in Wikipedia provides a good indicator of relevance that can boost the retrieval score of relevant books in the result ranking of a book search engine.

SALSA is a link-based ranking algorithm that takes the result set of a query as input, extends the set to include additional neighboring documents in the web graph, and performs a random walk on the induced subgraph. The stationary probability distribution of this random walk, used as a relevance score, is significantly more effective for ranking purposes than popular query-independent link-based ranking algorithms such as PageRank. Unfortunately, this requires significant effort at query-time, to access the link graph and compute the stationary probability distribution. In this paper, we explore whether it is possible to perform most of the computation off-line, prior to the arrival of any queries. The off-line phase of our approach computes a "score map" for each node in the web graph by performing a SALSA-like algorithm on the neighborhood of that node and retaining the scores of the most promising nodes in the neighborhood graph. The on-line phase takes the results to a query, retrieves the score map of each result, and returns for each result a score that is the sum of the matching scores from each score map. We evaluated this algorithm on a collection of about 28,000 queries with partially labeled results, and found that it is significantly more effective than PageRank, although not quite as effective as SALSA. We also studied the trade-off between ranking effectiveness and space requirements.

We investigate to what extent people making relevance judgements for a reusable IR test collection are exchangeable. We consider three classes of judge: "gold standard" judges, who are topic originators and are experts in a particular information seeking task; "silver standard" judges, who are task experts but did not create topics; and "bronze standard" judges, who are those who did not define topics and are not experts in the task.

Previous scalability experiments found that early precision improves as collection size increases. However, that was under the assumption that a collection's documents are all sampled with uniform probability from the same population. We contrast this to a large breadth-first web crawl, an important scenario in real-world Web search, where the early documents have quite different characteristics from the later documents.

Search engine click logs provide an invaluable source of relevance information, but this information is biased. A key source of bias is presentation order: the probability of click is influenced by a document's position in the results page. This paper focuses on explaining that bias, modelling how probability of click depends on position. We propose four simple hypotheses about how position bias might arise. We carry out a large data-gathering effort, where we perturb the ranking of a major search engine, to see how clicks are affected. We then explore which of the four hypotheses best explains the real-world position effects, and compare these to a simple logistic regression model. The data are not well explained by simple position models, where some users click indiscriminately on rank 1 or there is a simple decay of attention over ranks. A 'cascade' model, where users view results from top to bottom and leave as soon as they see a worthwhile document, is our best explanation for position bias in early ranks

A bipartite query-URL graph, where an edge indicates that a document was clicked for a query, is a useful construct for finding groups of related queries and URLs. Here we use this behavior graph for classification. We choose a click graph sampled from two weeks of image search activity, and the task of "adult" filtering: identifying content in the graph that is inappropriate for minors. We show how to perform classification using random walks on this graph, and two methods for estimating classifier parameters.

Search engines can record which documents were clicked for which query, and use these query-document pairs as "soft" relevance judgments. However, compared to the true judgments, click logs give noisy and sparse relevance information. We apply a Markov random walk model to a large click log, producing a probabilistic ranking of documents for a given query. A key advantage of the model is its ability to retrieve relevant documents that have not yet been clicked for that query and rank those effectively. We conduct experiments on click logs from image search, comparing our ("backward") random walk model to a different ("forward") random walk, varying parameters such as walk length and self-transition probability. The most effective combination is a long backward walk with high self-transition probability.

Optimising the parameters of ranking functions with respect to standard IR rank-dependent cost functions has eluded satisfactory analytical treatment. We build on recent advances in alternative differentiable pairwise cost functions, and show that these techniques can be successfully applied to tuning the parameters of an existing family of IR scoring functions (BM25), in the sense that we cannot do better using sensible search heuristics that directly optimize the rank-based cost function NDCG. We also demonstrate how the size of training set affects the number of parameters we can hope to tune this way.

Subject-specific search facilities on health sites are usually built using manual inclusion and exclusion rules. These can be expensive to maintain and often provide incomplete coverage of Web resources. On the other hand, health information obtained through whole-of-Web search may not be scientifically based and can be potentially harmful.To address problems of cost, coverage and quality, we built a focused crawler for the mental health topic of depression, which was able to selectively fetch higher quality relevant information. We found that the relevance of unfetched pages can be predicted based on link anchor context, but the quality cannot. We therefore estimated quality of the entire linking page, using a learned IR-style query of weighted single words and word pairs, and used this to predict the quality of its links. The overall crawler priority was determined by the product of link relevance and source quality.We evaluated our crawler against baseline crawls using both relevance judgments and objective site quality scores obtained using an evidence-based rating scale. Both a relevance focused crawler and the quality focused crawler retrieved twice as many relevant pages as a breadth-first control. The quality focused crawler was quite effective in reducing the amount of low quality material fetched while crawling more high quality content, relative to the relevance focused crawler.Analysis suggests that quality of content might be improved by post-filtering a very big breadth-first crawl, at the cost of substantially increased network traffic.

A query independent feature, relating perhaps to document content, linkage or usage, can be transformed into a static, per-document relevance weight for use in ranking. The challenge is to find a good function to transform feature values into relevance scores. This paper presents FLOE, a simple density analysis method for modelling the shape of the transformation required, based on training data and without assuming independence between feature and baseline. For a new query independent feature, it addresses the questions: is it required for ranking, what sort of transformation is appropriate and, after adding it, how successful was the chosen transformation? Based on this we apply sigmoid transformations to PageRank, indegree, URL Length and ClickDistance, tested in combination with a BM25 baseline.

Okapi BM25 scoring of anchor text surrogate documents has been shown to facilitate effective ranking in navigational search tasks over web data. We hypothesize that even better ranking can be achieved in certain important cases, particularly when anchor scores must be fused with content scores, by avoiding length normalisation and by reducing the attentuation of scores associated with high tf. Preliminary results are presented.

Hyperlink recommendation evidence, that is, evidence based on the structure of a web's link graph, is widely exploited by commercial Web search systems. However there is little published work to support its popularity. Another form of query-independent evidence, URL-type, has been shown to be beneficial on a home page finding task. We compared the usefulness of these types of evidence on the home page finding task, combined with both content and anchor text baselines. Our experiments made use of five query sets spanning three corpora---one enterprise crawl, and the WT10g and VLC2 Web test collections.We found that, in optimal conditions, all of the query-independent methods studied (in-degree, URL-type, and two variants of PageRank) offered a better than random improvement on a content-only baseline. However, only URL-type offered a better than random improvement on an anchor text baseline. In realistic settings, for either baseline, only URL-type offered consistent gains. In combination with URL-type the anchor text baseline was more useful for finding popular home pages, but URL-type with content was more useful for finding randomly selected home pages. We conclude that a general home page finding system should combine evidence from document content, anchor text, and URL-type classification.

Link-based ranking methods have been described in the literature and applied in commercial Web search engines. However, according to recent TREC experiments, they are no better than traditional content-based methods. We conduct a different type of experiment, in which the task is to find the main entry point of a specific Web site. In our experiments, ranking based on link anchor text is twice as effective as ranking based on document content, even though both methods used the same BM25 formula. We obtained these results using two sets of 100 queries on a 18.5 million document set and another set of 100 on a 0.4 million document set. This site finding effectiveness begins to explain why many search engines have adopted link methods. It also opens a rich new area for effectiveness improvement, where traditional methods fail.