Recent research has explored LLMs as scalable tools for relevance labeling, but studies indicate they are susceptible to priming effects, where prior relevance judgments influence later ones. Although psychological theories link personality traits to such biases, it is unclear whether simulated personalities in LLMs exhibit similar effects. We investigate how Big Five personality profiles in LLMs influence priming in relevance labeling, using multiple LLMs on TREC 2021 and 2022 Deep Learning Track datasets. Our results show that certain profiles, such as High Openness and Low Neuroticism, consistently reduce priming susceptibility. Additionally, the most effective personality in mitigating priming may vary across models and task types. Based on these findings, we propose personality prompting as a method to mitigate threshold priming, connecting psychological evidence with LLM-based evaluation practices.

Users tend to remember failures of a search session more than its many successes. This observation has led to work on search robustness, where systems are penalized if they perform very poorly on some queries. However, this principle of robustness has been overlooked within a single query. An ambiguous or underspecified query (e.g., ''jaguar'') can have several user intents, where popular intents often dominate the ranking, leaving users with minority intents unsatisfied. Although the diversification literature has long recognized this issue, existing metrics only model the average relevance across intents and provide no robustness guarantees. More surprisingly, we show theoretically and empirically that many well-known diversification algorithms are no more robust than a naive, non-diversified algorithm. To address this critical gap, we propose to frame diversification as a risk-minimization problem. We introduce VRisk, which measures the expected risk faced by the least-served fraction of intents in a query. Optimizing VRisk produces a robust ranking, reducing the likelihood of poor user experiences. We then propose VRisker, a fast greedy re-ranker with provable approximation guarantees. Finally, experiments on NTCIR INTENT-2, TREC Web 2012, and MovieLens show the vulnerability of existing methods. VRisker reduces worst-case intent failures by up to 33% with a minimal 2% drop in average performance.

The development of large language models (LLMs) has achieved superior performance in a range of downstream tasks, including LLM-based retrieval-augmented generation (RAG). The quality of generated content heavily relies on the usefulness of the retrieved information and the capacity of LLMs' internal information processing mechanism to incorporate it in answer generation. It is generally assumed that the retrieved information is relevant to the question. However, the retrieved information may have a variable degree of relevance and usefulness, depending on the question and the document collection. It is important to take into account the relevance of the retrieved information in answer generation. In this paper, we propose OpenDecoder, a new approach that leverages explicit evaluation of the retrieved information as quality indicator features for generation. We aim to build a RAG model that is more robust to varying levels of noisy context. Three types of explicit evaluation information are considered: relevance score, ranking score, and QPP (query performance prediction) score. The experimental results on five benchmark datasets demonstrate the effectiveness and better robustness of OpenDecoder by outperforming various baseline methods. Importantly, this paradigm is flexible to be integrated with the post-training of LLMs for any purposes and incorporated with any type of external indicators.

There is currently a controversy as to whether LLM-based relevance assessment can replace manual relevance assessment for evaluating search engines accurately at least at the run level (e.g., ranking TREC runs by mean nDCG) if not at the individual topic level (e.g., computing an nDCG score for a Search Engine Result Page). This study utilises an NTCIR web search test collection that features highly reliable human relevance labels (reflecting the collective view of eight independent assessors per topic) to complement prior findings from the skeptic camp. Our experiments show that LLM-based assessment (using Llama and Qwen) cannot replace human assessment even for ranking systems in terms of mean nDCG. More importantly, LLM-based assessment lacks discriminative power: it misses many statistically significant differences that manual assessment can detect. Furthermore, LLM-based assessment occasionally yields potential false alarms in terms of statistical significance, which may let researchers reach incorrect conclusions.

When a paired or two-sample t-test shows that a difference is not statistically significant (i.e., the null hypothesis H is "accepted"), all we obtain as a "conclusion" is: " we cannot conclude from the data whether the difference is real or not. " In general, statistical significance tests with a significance level of α give us a proper conclusion with 100(1-α)% confidence only if the H is rejected. Hence, for example, in a reproducibility study in IR, traditional significance tests like the t-test are not adequate if one wishes to claim that the reproduced run is as good as the target run. In this paper, we show that the equivalence test, which has been gaining popularity in the medical domain for comparing new drugs with standard drugs, is applicable to IR research in situations such as above. We also discuss the less ambitious noninferiority test, which can be used, for example, when an IR researcher wants to claim that their system performs no worse than a state-of-the-art system that is computationally much more expensive and therefore earth-unfriendly. The equivalence and noninferiority tests require researchers to pre-define an indifference zone, which defines how much difference is "practically negligible"; hence we also discuss one possible approach to establishing an indifference zone for a given IR effectiveness measure. In addition, we provide recommendations for IR researchers who wish to conduct equivalence/noninferiority tests involving more than two systems. It is hoped that equivalence and noninferiority tests will become standard practices in IR, so that we can avoid making claims based on lack of statistical significance.

Test collections are information retrieval tools that allow researchers to quickly and easily evaluate ranking algorithms. While test collections have become an integral part of IR research, the process of data creation involves significant manual annotation effort, which often makes it very expensive and time-consuming. Consequently, test collections could become too small when the budget is limited, which may lead to unstable evaluations. As a cheaper alternative, recent studies have proposed the use of large language models (LLMs) to completely replace human assessors. However, while LLMs seem to somewhat correlate with human judgments, their predictions are not perfect and often show bias. Thus, a complete replacement with LLMs is argued to be too risky and not fully reliable.

This study addresses (1) the influence of the decoy effect, a cognitive bias where the presence of an inferior item alters preferences between two options, on users’ search interactions and (2) the measurement of information retrieval systems’ vulnerability to the decoy effect. 1 From the perspective of user behavior, this study investigates the influence of the decoy effect in information retrieval (IR) by examining how decoy results affect users’ interaction on search engine result pages (SERPs), particularly in terms of click-through likelihood, browsing dwell time, and perceived document usefulness. We conducted an experiment based upon regression analysis on user interaction logs from three user study datasets which in total encompass 24 topics, 841 unique search sessions, and 2,685 queries. The findings indicate that decoys significantly increase the likelihood of document clicks and perceived usefulness. To investigate whether the influence of the decoy varies across different levels of task difficulty and user knowledge, we ran an additional experiment on one of the three datasets, which encompasses 6 topics, 166 search sessions and 652 queries. The results indicate that when the task is less challenging, users are more likely to click on a document with a decoy. Additionally, they spend more time on the target document and assign it a higher usefulness score. Furthermore, users with lower knowledge levels about the topic tend to give higher usefulness ratings to the target document. Regarding IR system evaluation, this study provides empirical insights into measuring the vulnerability of text retrieval models to potential decoy effect. An evaluation metric, namely DEcoy Judgement and Assessment VUlnerability (DEJA-VU), is proposed to evaluate the possibility of a retrieval model ranking results in a way that could trigger decoy biases. The experiments on the Text REtrieval Conference (TREC) 19 Deep Learning (DL) passage retrieval task and the TREC 20 DL passage retrieval task demonstrate that ColBERT and SPLADE show higher relevance-oriented retrieval effectiveness while also displaying lower vulnerability to decoy effect. Overall, this work advances the understanding of decoy effect, a well-established concept in cognitive psychology and behavioral economics, in a novel application field (i.e., Information Retrieval). It contributes to modeling users’ search behavior in the context of cognitive biases, as well as assessment of the vulnerability of systems and ranking algorithms to the decoy effect.

Personalized content-based recommender systems have become indispensable tools for users to navigate through the vast amount of content available on platforms like daily news websites and book recommendation services. However, existing recommenders face significant challenges in understanding the content of items. Large language models (LLMs), which possess deep semantic comprehension and extensive knowledge from pretraining, have proven to be effective in various natural language processing tasks. In this study, we explore the potential of leveraging both open- and closed-source LLMs to enhance content-based recommendation. With open-source LLMs, we utilize their deep layers as content encoders, enriching the representation of content at the embedding level. For closed-source LLMs, we employ prompting techniques to enrich the training data at the token level. Through comprehensive experiments, we demonstrate the high effectiveness of both types of LLMs and show the synergistic relationship between them. Notably, we observed a significant relative improvement of up to 19.32% compared to existing state-of-the-art recommendation models. These findings highlight the immense potential of both open- and closed-source of LLMs in enhancing content-based recommendation systems. We have made our code and LLM-generated data available (https://github.com/Jyonn/ONCE) for other researchers to reproduce our results.

To scale the size of Information Retrieval collections, crowdsourcing has become a common way to collect relevance judgments at scale. Crowdsourcing experiments usually employ 100–10,000 workers, but such a number is often decided in a heuristic way. The downside is that the resulting dataset does not have any guarantee of meeting predefined statistical requirements as, for example, have enough statistical power to be able to distinguish in a statistically significant way between the relevance of two documents. We propose a methodology adapted from literature on sound topic set size design, based on t-test and ANOVA, which aims at guaranteeing the resulting dataset to meet a predefined set of statistical requirements. We validate our approach on several public datasets. Our results show that we can reliably estimate the recommended number of workers needed to achieve statistical power, and that such estimation is dependent on the topic, while the effect of the relevance scale is limited. Furthermore, we found that such estimation is dependent on worker features such as agreement. Finally, we describe a set of practical estimation strategies that can be used to estimate the worker set size, and we also provide results on the estimation of document set sizes.

We present a simple and versatile framework for evaluating ranked lists in terms of Group Fairness and Relevance, in which the groups (i.e., possible attribute values) can be either nominal or ordinal in nature. First, we demonstrate that when our framework is applied to a binary hard group membership setting, our Group Fairness and Relevance (GFR) measures can easily quantify the overall polarity of each ranked list. Second, by utilising an existing diversified search test collection and treating each intent as an attribute value, we demonstrate that our framework can also handle soft group membership and that the GFR measures are highly correlated with a diversified information retrieval (IR) measure in this context as well. Third, using real data from a Japanese local search service, we demonstrate how our framework enables researchers to study intersectional group fairness based on multiple attribute sets. We also show that the similarity function for comparing the achieved and target distributions over the attribute values should be chosen carefully when the attribute values are ordinal. For such situations, our recommendation is to use multiple similarity functions with our framework: for example, one based on Jensen-Shannon Divergence (which disregards the ordinal nature of the groups) and another based on Root Normalised Order-aware Divergence (which has been designed specifically for handling ordinal groups). In addition, we highlight the fundamental differences between our framework and Attention-Weighted Rank Fairness (AWRF), a group fairness measure used at the TREC Fair Ranking Track.

The present study leverages a recent opportunity we had to create a new English web search test collection for the NTCIR-16 We Want Web (WWW-4) task, which concluded in June 2022. More specifically, through the test collection construction effort, we examined two factors that may affect the relevance assessments of depth- k pools, which in turn may affect the relative evaluation of different IR systems. The first factor is the document ordering strategy for the assessors, namely, prioritisation (PRI) and randomisation (RND). PRI is a method that has been used in NTCIR tasks for over a decade; it ranks the pooled documents by a kind of pseudorelevance for the assessors. The second factor is assessor type, i.e., Gold or Bronze. Gold assessors are the topic creators and therefore they “know” which documents are (highly) relevant and which are not; Bronze assessors are not the topic creators and may lack sufficient knowledge about the topics. We believe that our study is unique in that the authors of this article served as the Gold assessors when creating the WWW-4 test collection, which enabled us to closely examine why Bronze assessments differ from the Gold ones. Our research questions examine assessor efficiency ( RQ1 ), inter-assessor agreement ( RQ2 ), system ranking similarity with different qrels files ( RQ3 ), system ranking robustness to the choice of test topics ( RQ4 ), and the reasons why Bronze assessors tend to be more liberal than Gold assessors ( RQ5 ). The most remarkable of our results are as follows: First, in the comparisons for RQ1 through RQ4 , it turned out that what may matter more than the document ordering strategy (PRI vs. RND) and the assessor type (Gold vs. Bronze) is how well-motivated and/or well-trained the Bronze assessors are. Second, regarding RQ5 , of the documents originally judged nonrelevant by the Gold assessors contrary to the Bronze assessors in our experiments, almost one half were truly relevant according to the Gold assessors’ own reconsiderations. This result suggests that even Gold assessors are far from perfect; budget permitting, it may be beneficial to hire highly motivated Bronze assessors in addition to Gold assessors so they can complement each other.

Information Extraction (IE) focuses on transforming unstructured data into structured knowledge, of which Named Entity Recognition (NER) is a fundamental component. In the realm of Information Retrieval (IR), effectively recognizing entities can substantially enhance the precision of search and recommendation systems. Existing methods frame NER as a sequence labeling task, which requires extra data and, therefore may be limited in terms of sustainability. One promising solution is to employ a Machine Reading Comprehension (MRC) approach for NER tasks, thereby eliminating the dependence on additional data. This process encounters key challenges, including: (1) Unconventional predictions; (2) Inefficient multi-stream processing; (3) Absence of a proficient reasoning strategy. To this end, we present the Single-Stream Reasoner (SSR), a solution utilizing a reasoning strategy and standardized inputs. This yields a type-agnostic solution for both flat and nested NER tasks, without the need for additional data. On ten NER benchmarks, SSR achieved state-of-the-art results, highlighting its robustness. Furthermore, we illustrated its efficiency through convergence, inference speed, and low-resource scenario performance comparisons. Our architecture displays adaptability and can effortlessly merge with various foundational models and reasoning strategies, fostering advancements in both the IR and IE fields.

One of the most challenging aspects of operating a large-scale web search engine is to accurately evaluate and monitor the search engine's result quality regardless of search types. From a business perspective, in the face of such challenges, it is important to establish a universal search quality metric that can be easily understood by the entire organisation. In this paper, we introduce a model-based quality metric using Explainable Boosting Machine as the classifier and online user behaviour signals as features to predict search quality. The proposed metric takes into account a variety of search types and has good interpretability. To examine the performance of the metric, we constructed a large dataset of user behaviour on search engine results pages (SERPs) with SERP quality ratings from professional annotators. We compared the performance of the model in our metric to those of other black-box machine learning models on the dataset. We also share a few experiences within our company for the org-wide adoption of this metric relevant to metric design.

Previous researches demonstrate that users’ actions in search interaction are associated with relative gains and losses to reference points, known as the reference dependence effect. However, this widely confirmed effect is not represented in most user models underpinning existing search evaluation metrics. In this study, we propose a new evaluation metric framework, namely Reference Dependent Metric (ReDeM), for assessing query-level search by incorporating the effect of reference dependence into the modelling of user search behavior. To test the overall effectiveness of the proposed framework, (1) we evaluate the performance, in terms of correlation with user satisfaction, of ReDeMs built upon different reference points against that of the widely-used metrics on three search datasets; (2) we examine the performance of ReDeMs under different task states, like task difficulty and task urgency; and (3) we analyze the statistical reliability of ReDeMs in terms of discriminative power. Experimental results indicate that: (1) ReDeMs integrated with a proper reference point achieve better correlations with user satisfaction than most of the existing metrics, like Discounted Cumulative Gain (DCG) and Rank-Biased Precision (RBP), even though their parameters have already been well-tuned; (2) ReDeMs reach relatively better performance compared to existing metrics when the task triggers a high-level cognitive load; (3) the discriminative power of ReDeMs is far stronger than Expected Reciprocal Rank (ERR), slightly stronger than Precision and similar to DCG, RBP and INST. To our knowledge, this study is the first to explicitly incorporate the reference dependence effect into the user browsing model and offline evaluation metrics. Our work illustrates a promising approach to leveraging the insights about user biases from cognitive psychology in better evaluating user search experience and enhancing user models.

Models of existing evaluation metrics assume that users are rational decision-makers trying to pursue maximised utility. However, studies in behavioural economics show that people are not always rational when making decisions. Previous studies showed that the anchoring effect can influence the relevance judgement of a document. In this paper, we challenge the rational user assumption and introduce the anchoring effect into user models. We first propose a framework for query-level evaluation metrics by incorporating the anchoring effect into the user model. In the framework, the magnitude of the anchoring effect is related to the quality of the previous document. We then apply our framework to several query-level evaluation metrics and compare them with their vanilla version as the baseline in terms of user satisfaction on a publicly available search dataset. As a result, our Anchoring-aware Metrics (AMs) outperformed their baselines in term of correlation with user satisfaction. The result suggests that we can better predict user query satisfaction feedbacks by incorporating the anchoring effect into user models of existing evaluating metrics. As far as we know, we are the first to introduce the anchoring effect into information retrieval evaluation metrics. Our findings provide a perspective from behavioural economics to better understand user behaviour and satisfaction in search interaction.

In the context of depth- k pooling for constructing web search test collections, we compare two approaches to ordering pooled documents for relevance assessors: The prioritisation strategy (PRI) used widely at NTCIR, and the simple randomisation strategy (RND). In order to address research questions regarding PRI and RND, we have constructed and released the WWW3E8 dataset, which contains eight independent relevance labels for 32,375 topic-document pairs, i.e., a total of 259,000 labels. Four of the eight relevance labels were obtained from PRI-based pools; the other four were obtained from RND-based pools. Using WWW3E8, we compare PRI and RND in terms of inter-assessor agreement, system ranking agreement, and robustness to new systems that did not contribute to the pools. We also utilise an assessor activity log we obtained as a byproduct of WWW3E8 to compare the two strategies in terms of assessment efficiency. Our main findings are: (a) The presentation order has no substantial impact on assessment efficiency; (b) While the presentation order substantially affects which documents are judged (highly) relevant, the difference between the inter-assessor agreement under the PRI condition and that under the RND condition is of no practical significance; (c) Different system rankings under the PRI condition are substantially more similar to one another than those under the RND condition; and (d) PRI-based relevance assessment files (qrels) are substantially and statistically significantly more robust to new systems than RND-based ones. Finding (d) suggests that PRI helps the assessors identify relevant documents that affect the evaluation of many existing systems, including those that did not contribute to the pools. Hence, if researchers need to evaluate their current IR systems using legacy IR test collections, we recommend the use of those constructed using the PRI approach unless they have a good reason to believe that their systems retrieve relevant documents that are vastly different from the pooled documents. While this robustness of PRI may also mean that the PRI-based pools are biased against future systems that retrieve highly novel relevant documents, one should note that there is no evidence that RND is any better in this respect.

While batch evaluation plays a central part in Information Retrieval (IR) research, most evaluation metrics are based on user models which mainly focus on browsing and clicking behaviors. As users' perceived satisfaction may also be impacted by their search intent, constructing different user models across various search intent may help design better evaluation metrics. However, user intents are usually unobservable in practice. As query reformulating behaviors may reflect their search intents to a certain extent and highly correlate with users' perceived satisfaction for a specific query, these observable factors may be beneficial for the design of evaluation metrics. How to incorporate the search intent behind query reformulation into user behavior and satisfaction models remains under-investigated. To investigate the relationships among query reformulations, search intent, and user satisfaction, we explore a publicly available web search dataset and find that query reformulations can be a good proxy for inferring user intent, and therefore, reformulating actions may be beneficial for designing better web search effectiveness metrics. A group of Reformulation-Aware Metrics (RAMs) is then proposed to improve existing click model-based metrics. Experimental results on two public session datasets have shown that RAMs have significantly higher correlations with user satisfaction than existing evaluation metrics. In the robustness test, we have found that RAMs can achieve good performance when only a small proportion of satisfaction training labels are available. We further show that RAMs can be directly applied in a new dataset for offline evaluation once trained. This work shows the possibility of designing better evaluation metrics by incorporating fine-grained search context factors.

Relevance judgments play an essential role in the evaluation of information retrieval systems. As many different relevance judgment settings have been proposed in recent years, an evaluation metric to compare relevance judgments in different annotation settings has become a necessity. Traditional metrics, such as ĸ, Krippendorff's α and Φ have mainly focused on the inter-assessor consistency to evaluate the quality of relevance judgments. They encounter "reliable but useless" problem when employed to compare different annotation settings (e.g. binary judgment v.s. 4-grade judgment). Meanwhile, other existing popular metrics such as discriminative power (DP) are not designed to compare relevance judgments across different annotation settings, they therefore suffer from limitations, such as the requirement of result ranking lists from different systems. Therefore, how to design an evaluation metric to compare relevance judgments under different grade settings needs further investigation. In this work, we propose a novel metric named pairwise discriminative power (PDP) to evaluate the quality of relevance judgment collections. By leveraging a small amount of document-level preference tests, PDP estimates the discriminative ability of relevance judgments on separating ranking lists with various qualities. With comprehensive experiments on both synthetic and real-world datasets, we show that PDP maintains a high degree of consistency with annotation quality in various grade settings. Compared with existing metrics (e.g., Krippendorff's α, Φ, DP, etc), it provides reliable evaluation results with affordable additional annotation efforts.

While previous work in comparing statistical significance tests for IR system evaluation have focused on paired data tests (e.g., for evaluating two systems using a common test collection), two-sample tests must be used when the reproducibility of IR experiments across different test collections must be examined. Using real runs and a test collection from the NTCIR-15 WWW-3 Task, the present study compares the properties of three two-sample significance tests for comparing two systems: Student's t-test (i.e., the classical parametric test), the Wilcoxon rank sum test (i.e., the classical nonparametric test), and the randomisation test (i.e., a population-free method that utilises modern computational power). In terms of the false positive rate (i.e., the chance of detecting a statistical significance even though the two samples of evaluation measure scores come from the same system), the three tests behave similarly, although the Wilcoxon rank sum test appears to be slightly more robust than the other two for very small topic set sizes (e.g., 10 topics each) with a large significance level (e.g., α=0.10). On the other hand, the t-test and the Wilcoxon rank sum test are very similar to each other from the following two viewpoints: "How often do they both detect a nonexistent difference?" and "How often do they both overlook a true difference?" Compared to the two classical significance tests, the randomisation test behaves markedly differently in terms of the above two viewpoints. Hence, we suggest that researchers should at least be aware of the above properties of the three two-sample tests when choosing from them.

In IR evaluation based on depth-k pooling, there are several strategies to order the pooled documents for relevance assessors. Among them, the simplest approach is to completely randomise the order "so assessors cannot tell if a document was highly ranked by some system or how many systems (or which systems) retrieved the document." An approach that is in sharp contrast to the above is the prioritisation approach taken by NTCIRPOOL, a tool widely used at NTCIR. NTCIRPOOL sorts the pooled documents by "pseudorelevance," a statistic that reflects the popularity of each document within the depth-k pools. Although these two strategies have coexisted for over two decades, the IR research community has yet to reach a consensus as to what advantages each of these two strategies actually offer. To help researchers directly address this question using their favourite methods of analysis, we have released a large-scale data set called WWW3E8. It comprises eight independent sets of qrels for the 160 English topics of the NTCIR-15 WWW-3 task: four qrels files constructed using the randomisation approach, and another four constructed using the prioritisation approach of NTCIRPOOL. Each qrels file covers 32,375 topic-document pairs; hence, WWW3E8 contains a total of 259,000 relevance labels. Moreover, the data set contains the raw English subtask run files from the WWW-3 task, the randomised and prioritised pool files, and topic-by-run score matrices of the official measures used in the task. Hence, researchers interested in the above research question regarding document ordering can utilise WWW3E8 as a common ground to directly compare the two strategies.

Learning from implicit feedback is challenging because of the difficult nature of the one-class problem: we can observe only positive examples. Most conventional methods use a pairwise ranking approach and negative samplers to cope with the one-class problem. However, such methods have two main drawbacks particularly in large-scale applications; (1) the pairwise approach is severely inefficient due to the quadratic computational cost; and (2) even recent model-based samplers (e.g. IRGAN) cannot achieve practical efficiency due to the training of an extra model.

We examine the “goodness” of ranked retrieval evaluation measures in terms of how well they align with users’ Search Engine Result Page (SERP) preferences for web search. The SERP preferences cover 1,127 topic-SERP-SERP triplets extracted from the NTCIR-9 INTENT task, reflecting the views of 15 different assessors. Each assessor made two SERP preference judgements for each triplet: one in terms of relevance and the other in terms of diversity. For each evaluation measure, we compute the Agreement Rate (AR) of each triplet: the proportion of assessors that agree with the measure’s SERP preference. We then compare the mean ARs of the measures as well as those of best/median/worst assessors using Tukey HSD tests. Our first experiment compares traditional ranked retrieval measures based on the SERP relevance preferences: we find that normalised Discounted Cumulative Gain (nDCG) and intentwise Rank-biased Utility (iRBU) perform best in that they are the only measures that are statistically indistinguishable from our best assessor; nDCG also statistically significantly outperforms our median assessor. Our second experiment utilises 119,646 document preferences that we collected for a subset of the above topic-SERP-SERP triplets (containing 894 triplets) to compare preference-based evaluation measures as well as traditional ones. Again, we evaluate them based on the SERP relevance preferences. The results suggest that measures such as wpref5 are the most promising among the preference-based measures considered, although they underperform the best traditional measures such as nDCG on average. Our third experiment compares diversified search measures based on the SERP diversity preferences as well as the SERP relevance preferences, and it shows that D♯-measures are clearly the most reliable: in particular, D♯-nDCG and D♯-RBP statistically significantly outperform the median assessor and all intent-aware measures; they also outperform the recently proposed RBU on average. Also, in terms of agreement with SERP diversity preferences, D♯-nDCG statistically significantly outperforms RBU. Hence, if IR researchers want to use evaluation measures that align well with users’ SERP preferences, then we recommend nDCG and iRBU for traditional search, and D♯-measures such as D♯-nDCG for diversified search. As for document preference-based measures that we have examined, we do not have a strong reason to recommended them over traditional measures like nDCG, since they align slightly less well with users’ SERP preferences despite their quadratic assessment cost.

Product image search is required to deal with large target image datasets which are frequently updated, and therefore it is not always practical to maintain exhaustive and up-to-date relevance assessments for tuning and evaluating the search engine. Moreover, in similar product image search where the query is also an image, it is difficult to identify the possible search intents behind it and thereby verbalise the relevance criteria for the assessors, especially if graded relevance assessments are required. In this study, we focus on similar product image search within a given product category (e.g., shoes), wherein each image is iconic (i.e., the image clearly shows what the product looks like and basically nothing else), and propose an initial approach to evaluating the task without relying on manual relevance assessments. More specifically, we build a simple probabilistic model that assumes that an image is generated from latent intents representing shape, texture, and colour, which enables us to estimate the relevance score of each image and thereby compute graded relevance measures for any image search engine result page. Through large-scale crowdsourcing experiments, we demonstrate that our proposed measures, InDCG (which is based on per-intent binary relevance) and D-InDCG (which is based on per-intent graded relevance), align reasonably well with human SERP preferences and with human image preferences. Hence, our automatic measures may be useful at least for rough tuning and evaluation of similar product image search.

Replicability and reproducibility of experimental results are primary concerns in all the areas of science and IR is not an exception. Besides the problem of moving the field towards more reproducible experimental practices and protocols, we also face a severe methodological issue: we do not have any means to assess when reproduced is reproduced. Moreover, we lack any reproducibility-oriented dataset, which would allow us to develop such methods.

For offline evaluation of IR systems, some researchers have proposed to utilise pairwise document preference assessments instead of relevance assessments of individual documents, as it may be easier for assessors to make relative decisions rather than absolute ones. Simple preference-based evaluation measures such as ppref and wpref have been proposed, but the past decade did not see any wide use of such measures. One reason for this may be that, while these new measures have been reported to behave more or less similarly to traditional measures based on absolute assessments, whether they actually align with the users' perception of search engine result pages (SERPs) has been unknown. The present study addresses exactly this question, after formally defining two classes of preference-based measures called Pref measures and Δ-measures. We show that the best of these measures perform at least as well as an average assessor in terms of agreement with users' SERP preferences, and that implicit document preferences (i.e., those suggested by a SERP that retrieves one document but not the other) play a much more important role than explicit preferences (i.e., those suggested by a SERP that retrieves one document above the other). We have released our data set containing 119,646 document preferences, so that the feasibility of document preferenced-based evaluation can be further pursued by the IR community.

In product-to-product search and recommendation, the product image often plays a pivotal role for the user to determine the relevance of that product. The present study investigates the relationship between the users' visual intents (in terms of colour, texture and material, and design) and the amount of user feedback (namely, clicks, likes, and purchases) using real product data and crowdsourcing. Through the analysis, we found that visual relevance (i.e., relevance of a target product with respect to a particular visual intent) correlates with the amount of user feedback, and that visual relevance can be the cause of user feedback.

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

This study evaluates 30 IR evaluation measures or their instances, of which nine are for adhoc IR and 21 are for diversified IR, primarily from the viewpoint of whether their preferences of one SERP (search engine result page) over another actually align with users' preferences. The gold preferences were contructed by hiring 15 assessors, who independently examined 1,127 SERP pairs and made preference assessments. Two sets of preference assessments were obtained: one based on a relevance question "Which SERP is more relevant?'' and the other based on a diversity question "Which SERP is likely to satisfy a higher number of users?'' To our knowledge, our study is the first to have collected diversity preference assessments in this way and evaluated diversity measures successfully. Our main results are that (a) Popular adhoc IR measures such as nDCG actually align quite well with the gold relevance preferences; and that (b) While the ♯-measures align well with the gold diversity preferences, intent-aware measures perform relatively poorly. Moreover, as by-products of our analysis of existing evaluation measures, we define new adhoc measures called iRBU (intentwise Rank-Biased Utility) and EBR (Expected Blended Ratio); we demonstrate that an instance of iRBU performs as well as nDCG when compared to the gold relevance preferences. On the other hand, the original RBU, a recently-proposed diversity measure, underperforms the best ♯-measures when compared to the gold diversity preferences.

This hands-on half-day tutorial consists of two sessions. Part~I covers the following topics: Preliminaries; Paired and two-sample t-tests, confidence intervals; One-way ANOVA and two-way ANOVA without replication; Familiwise error rate. Part~II covers the following topics: Tukey's HSD test, simultaneous confidence intervals; Randomisation test and randomised Tukey HSD test; What's wrong with statistical significance tests?; Effect sizes, statistical power; Topic set size design and power analysis; Summary: how to report your results. Participants should have some prior knowledge about the very basics of statistical significance testing and are strongly encouraged to bring a laptop with R already installed. They will learn how to design and conduct statistical significance tests for comparing the mean effectiveness scores of two or more systems appropriately, and to report on the test results in an informative manner.

We tackle Attitude Detection, which we define as the task of extracting the replier's attitude, i.e., a target-polarity pair, from a given one-round conversation. While previous studies considered Target Extraction and Polarity Classification separately, we regard them as subtasks of Attitude Detection. Our experimental results show that treating the two subtasks independently is not the optimal solution for Attitude Detection, as achieving high performance in each subtask is not sufficient for obtaining correct target-polarity pairs. Our jointly trained model AD-NET substantially outperforms the separately trained models by alleviating the target-polarity mismatch problem. Moreover, we proposed a method utilising the attitude detection model to improve retrieval-based chatbots by re-ranking the response candidates with attitude features. Human evaluation indicates that with attitude detection integrated, the new responses to the sampled queries from are statistically significantly more consistent, coherent, engaging and informative than the original ones obtained from a commercial chatbot.

Personalized recommendation has received a lot of attention as a highly practical research topic. However, existing recommender systems provide the recommendations with a generic statement such as “Customers who bought this item also bought…”. Explainable recommendation, which makes a user aware of why such items are recommended, is in demand. The goal of our research is to make the users feel as if they are receiving recommendations from their friends. To this end, we formulate a new challenging problem called personalized reason generation for explainable recommendation for songs in conversation applications and propose a solution that generates a natural language explanation of the reason for recommending a song to that particular user. For example, if the user is a student, our method can generate an output such as “Campus radio plays this song at noon every day, and I think it sounds wonderful,” which the student may find easy to relate to. In the offline experiments, through manual assessments, the gain of our method is statistically significant on the relevance to songs and personalization to users comparing with baselines. Large-scale online experiments show that our method outperforms manually selected reasons by 8.2% in terms of click-through rate. Evaluation results indicate that our generated reasons are relevant to songs and personalized to users, and they attract users to click the recommendations.

Some modern information access tasks such as natural language dialogue tasks are difficult to evaluate, for often there is no such thing as the ground truth: different users may have different opinions about the system's output. A few task designs for dialogue evaluation have been implemented and/or proposed recently, where both the ground truth data and the system's output are represented as a distribution of users' votes over bins on a non-nominal scale. The present study first points out that popular bin-by-bin measures such as Jensen-Shannon divergence and Sum of Squared Errors are clearly not adequate for such tasks, and that cross-bin measures should be used. Through experiments using artificial distributions as well as real ones from a dialogue evaluation task, we demonstrate that two cross-bin measures, namely, the Normalised Match Distance (NMD; a special case of the Earth Mover's Distance) and the Root Symmetric Normalised Order-aware Divergence (RSNOD), are indeed substantially different from the bin-by-bin measures.Furthermore, RSNOD lies between the popular bin-by-bin measures and NMD in terms of how it behaves. We recommend using both of these measures in the aforementioned type of evaluation tasks.

This hands-on half-day tutorial consists of two 90-minute sessions. Part I covers the following topics: paired and two-sample t -tests, confidence intervals (with Excel and R); familywise error rate, multiple comparison procedures; ANOVA (with Excel and R); Tukey's HSD test, simultaneous confidence intervals (with R). Part II covers the following topics: randomised Tukey HSD test (with Discpower); what's wrong with statistical significance tests?; effect sizes, statistical power; topic set size design (with Excel); power analysis (with R); summary: how to report your results. Participants should have some prior knowledge about the very basics of statistical significance testing and are strongly encouraged to bring a laptop with R already installed. The tutorial participants will be able to design and conduct statistical significance tests for comparing the mean effectiveness scores of two or more systems appropriately, and to report on the test results in an informative manner.

Search result diversification aims at returning diversified document lists to cover different user intents of a query. Existing diversity measures assume that the intents of a query are disjoint, and do not consider their relationships. In this paper, we introduce intent hierarchies to model the relationships between intents, and present four weighing schemes. Based on intent hierarchies, we propose several hierarchical measures that take into account the relationships between intents. We demonstrate the feasibility of hierarchical measures by using a new test collection based on TREC Web Track 2009-2013 diversity test collections and by using NTCIR-11 IMine test collection. Our main experimental findings are: (1) Hierarchical measures are more discriminative and intuitive than existing measures. In terms of intuitiveness, it is preferable for hierarchical measures to use the whole intent hierarchies than to use only the leaf nodes. (2) The types of intent hierarchies used affect the discriminative power and intuitiveness of hierarchical measures. We suggest the best type of intent hierarchies to be used according to whether the nonuniform weights are available. (3) To measure the benefits of the diversification algorithms which use automatically mined hierarchical intents, it is important to use hierarchical measures instead of existing measures.

Due to the tremendous economic value of search result pages, search engine companies have invested a lot to improve their quality. Recently, much effort has been made to directly model key aspects of users' interactions with search system, for example, Benefit and Cost. Time has been widely adopted in both of the two aspects since benefit and cost must be expressed in meaningful units in practical application. Psychological studies have demonstrated that the subjectively perceived time might be different from the objective time measured by timing device and the time perception process of human beings is affected by some psychological factors, such as motivation and interest, which are closely related to the search process. Considering that time is usually used to describe the subject experience of search users, it is necessary to investigate the difference between perceived time and objective time in search process. In psychology, there is a temporal illusion effect named Vierordt's law, i.e. shorter intervals tend to be overestimated while longer intervals tend to be underestimated. In this work, we carefully designed a lab-study to examine the impact of duration length on user's time perception in the context of search. Experimental results show that Vierordt's law is consistently observed in Web search environment. This work could help us to correct the estimation of users' perceived time and provide insights about the mechanism of satisfaction.

We consider the problem of ranking rich verticals, which we call "cards," for a given mobile search query. Examples of card types include "SHOP" (showing access and contact information of a shop), "WEATHER" (showing a weather forecast for a particular location), and "TV" (showing information about a TV programme). These cards can be highly visual and/or concise, and may often satisfy the user's information need without making her click on them. While this "good abandonment" of the search engine result page is ideal especially for mobile environments where the interaction between the user and the search engine should be minimal, it poses a challenge for search engine companies whose ranking algorithms rely heavily on click data. In order to provide the right card types to the user for a given query, we propose a graph-based approach which extends a click-based automatic relevance estimation algorithm of Agrawal et al., by incorporating an abandonment-based preference rule. Using a real mobile query log from a commercial search engine, we constructed a data set containing 2,472 pairwise card type preferences covering 992 distinct queries, by hiring three independent assessors. Our proposed method outperforms a click-only baseline by 53-68% in terms of card type preference accuracy. The improvement is also statistically highly significant, with p ≈ 0.0000 according to the paired randomisation test.

Recent advances in neural networks, along with the growth of rich and diverse community question answering (cQA) data, have enabled researchers to construct robust open-domain question answering (QA) systems. It is often claimed that such state-of-the-art QA systems far outperform traditional IR baselines such as BM25. However, most such studies rely on relatively small data sets, e.g., those extracted from the old TREC QA tracks. Given massive training data plus a separate corpus of Q&A pairs as the target knowledge source, how well would such a system really perform? How fast would it respond? In this demonstration, we provide the attendees of SIGIR 2017 an opportunity to experience a live comparison of two open-domain QA systems, one based on a long short-term memory (LSTM) architecture with over 11 million Yahoo! Chiebukuro (i.e., Japanese Yahoo! Answers) questions and over 27.4 million answers for training, and the other based on BM25. Both systems use the same Q&A knowledge source for answer retrieval. Our core demonstration system is a pair of Japanese monolingual QA systems, but we leverage machine translation for letting the SIGIR attendees enter English questions and compare the Japanese responses from the two systems after translating them into English.

Mobile search engine result pages (SERPs) are becoming highly visual and heterogenous. Unlike the traditional ten-blue-link SERPs for desktop search, different verticals and cards occupy different amounts of space within the small screen. Hence, traditional retrieval measures that regard the SERP as a ranked list of homogeneous items are not adequate for evaluating the overall quality of mobile SERPs. Specifically, we address the following new problems in mobile search evaluation: (1) Different retrieved items have different heights within the scrollable SERP, unlike a ten-blue-link SERP in which results have similar heights with each other. Therefore, the traditional rank-based decaying functions are not adequate for mobile search metrics. (2) For some types of verticals and cards, the information that the user seeks is already embedded in the snippet, which makes clicking on those items to access the landing page unnecessary. (3) For some results with complex sub-components (and usually a large height), the total gain of the results cannot be obtained if users only read part of their contents. The benefit brought by the result is affected by user's reading behavior and the internal gain distribution (over the height) should be modeled to get a more accurate estimation. To tackle these problems, we conduct a lab-based user study to construct suitable user behavior model for mobile search evaluation. From the results, we find that the geometric heights of user's browsing trails can be adopted as a good signal of user effort. Based on these findings, we propose a new evaluation metric, Height-Biased Gain, which is calculated by summing up the product of gain distribution and discount factors that are both modeled in terms of result height. To evaluate the effectiveness of the proposed metric, we compare the agreement of evaluation metrics with side-by-side user preferences on a test collection composed of four mobile search engines. Experimental results show that HBG agrees with user preferences 85.33% of the time, which is better than all existing metrics.

Using classical statistical significance tests, researchers can only discuss P(D+|H), the probability of observing the data D at hand or something more extreme, under the assumption that the hypothesis H is true (i.e., the p-value). But what we usually want is P(H|D), the probability that a hypothesis is true, given the data. If we use Bayesian statistics with state-of-the-art Markov Chain Monte Carlo (MCMC) methods for obtaining posterior distributions, this is no longer a problem. That is, instead of the classical p-values and 95% confidence intervals, which are often misinterpreted respectively as "probability that the hypothesis is (in)correct" and "probability that the true parameter value drops within the interval is 95%," we can easily obtain P(H|D) and credible intervals which represent exactly the above. Moreover, with Bayesian tests, we can easily handle virtually any hypothesis, not just "equality of means," and obtain an Expected A Posteriori (EAP) value of any statistic that we are interested in. We provide simple tools to encourage the IR community to take up paired and unpaired Bayesian tests for comparing two systems. Using a variety of TREC and NTCIR data, we compare P(H|D) with p-values, credible intervals with confidence intervals, and Bayesian EAP effect sizes with classical ones. Our results show that (a) p-values and confidence intervals can respectively be regarded as approximations of what we really want, namely, P(H|D) and credible intervals; and (b) sample effect sizes from classical significance tests can differ considerably from the Bayesian EAP effect sizes, which suggests that the former can be poor estimates of population effect sizes. For both paired and unpaired tests, we propose that the IR community report the EAP, the credible interval, and the probability of hypothesis being true, not only for the raw difference in means but also for the effect size in terms of Glass's Δ.

Time plays an essential role in multiple areas of Information Retrieval (IR) studies such as search evaluation, user behavior analysis, temporal search result ranking and query understanding. Especially, in search evaluation studies, time is usually adopted as a measure to quantify users' efforts in search processes. Psychological studies have reported that the time perception of human beings can be affected by many stimuli, such as attention and motivation, which are closely related to many cognitive factors in search. Considering the fact that users' search experiences are affected by their subjective feelings of time, rather than the objective time measured by timing devices, it is necessary to look into the different factors that have impacts on search users' perception of time. In this work, we make a first step towards revealing the time perception mechanism of search users with the following contributions: (1) We establish an experimental research framework to measure the subjective perception of time while reading documents in search scenario, which originates from but is also different from traditional time perception measurements in psychological studies. (2) With the framework, we show that while users are reading result documents, document relevance has small yet visible effect on search users' perception of time. By further examining the impact of other factors, we demonstrate that the effect on relevant documents can also be influenced by individuals and tasks. (3) We conduct a preliminary experiment in which the difference between perceived time and dwell time is taken into consideration in a search evaluation task. We found that the revised framework achieved a better correlation with users' satisfaction feedbacks. This work may help us better understand the time perception mechanism of search users and provide insights in how to better incorporate time factor in search evaluation studies.

We conducted a systematic review of 840 SIGIR full papers and 215 TOIS papers published between 2006 and 2015. The original objective of the study was to identify IR effectiveness experiments that are seriously underpowered (i.e., the sample size is far too small so that the probability of missing a real difference is extremely high) or overpowered (i.e., the sample size is so large that a difference will be considered statistically significant even if the actual effect size is extremely small). However, it quickly became clear to us that many IR effectiveness papers either lack significance testing or fail to report p-values and/or test statistics, which prevents us from conducting power analysis. Hence we first report on how IR researchers (fail to) report on significance test results, what types of tests they use, and how the reporting practices may have changed over the last decade. From those papers that reported enough information for us to conduct power analysis, we identify extremely overpowered and underpowered experiments, as well as appropriate sample sizes for future experiments. The raw results of our systematic survey of 1,055 papers and our R scripts for power analysis are available online. Our hope is that this study will help improve the reporting practices and experimental designs of future IR effectiveness studies.

There are two well-known versions of the t-test for comparing means from unpaired data: Student's t-test and Welch's t-test. While Welch's t-test does not assume homoscedasticity (i.e., equal variances), nit involves approximations. A classical textbook recommendation would be to use Student's t-test if either the two sample sizes are similar or the two sample variances are similar, and to use Welch's t-test only when both of the above conditions are violated. However, a more recent recommendation seems to be to use Welch's t-test unconditionally. Using past data from both TREC and NTCIR, the present study demonstrates that the latter advice should not be followed blindly in the context of IR system evaluation. More specifically, our results suggest that if the sample sizes differ substantially and if the larger sample has a substantially larger variance,Welch's t-test may not be reliable.

Search result diversification aims at returning diversified document lists to cover different user intents for ambiguous or broad queries. Existing diversity measures assume that user intents are independent or exclusive, and do not consider the relationships among the intents. In this paper, we introduce intent hierarchies to model the relationships among intents. Based on intent hierarchies, we propose several hierarchical measures that can consider the relationships among intents. We demonstrate the feasibility of hierarchical measures by using a new test collection based on TREC Web Track 2009-2013 diversity test collections. Our main experimental findings are: (1) Hierarchical measures are generally more discriminative and intuitive than existing measures using flat lists of intents; (2) When the queries have multilayer intent hierarchies, hierarchical measures are less correlated to existing measures, but can get more improvement in discriminative power; (3) Hierarchical measures are more intuitive in terms of diversity or relevance. The hierarchical measures using the whole intent hierarchies are more intuitive than only using the leaf nodes in terms of diversity and relevance.

Traditional pooling-based information retrieval (IR) test collections typically have $$n= 50$$ –100 topics, but it is difficult for an IR researcher to say why the topic set size should really be n. The present study provides details on principled ways to determine the number of topics for a test collection to be built, based on a specific set of statistical requirements. We employ Nagata’s three sample size design techniques, which are based on the paired t test, one-way ANOVA, and confidence intervals, respectively. These topic set size design methods require topic-by-run score matrices from past test collections for the purpose of estimating the within-system population variance for a particular evaluation measure. While the previous work of Sakai incorrectly used estimates of the total variances, here we use the correct estimates of the within-system variances, which yield slightly smaller topic set sizes than those reported previously by Sakai. Moreover, this study provides a comparison across the three methods. Our conclusions nevertheless echo those of Sakai: as different evaluation measures can have vastly different within-system variances, they require substantially different topic set sizes under the same set of statistical requirements; by analysing the tradeoff between the topic set size and the pool depth for a particular evaluation measure in advance, researchers can build statistically reliable yet highly economical test collections.

Collaborative Information Seeking/Retrieval (CIS/CIR) has given rise to several challenges in terms of search behavior analysis, retrieval model formalization as well as interface design. However, the major issue of evaluation in CIS/CIR is still underexplored. The goal of this workshop is to investigate the evaluation challenges in CIS/CIR with the hope of building standardized evaluation frameworks, methodologies, and task specifications that would foster and grow the research area (in a collaborative fashion).

A large percentage of queries issued to search engines are broad or ambiguous. Search result diversification aims to solve this problem, by returning diverse results that can fulfill as many different information needs as possible. Most existing intent-aware search result diversification algorithms formulate user intents for a query as a flat list of subtopics. In this paper, we introduce a new hierarchical structure to represent user intents and propose two general hierarchical diversification models to leverage hierarchical intents. Experimental results show that our hierarchical diversification models outperform state-of-the-art diversification methods that use traditional flat subtopics.

A researcher decides to build a test collection for comparing her new information retrieval (IR) systems with several state-of-the-art baselines. She wants to know the number of topics (n) she needs to create in advance, so that she can start looking for (say) a query log large enough for sampling n good topics, and estimating the relevance assessment cost. We provide practical solutions to researchers like her using power analysis and sample size design techniques, and demonstrate its usefulness for several IR tasks and evaluation measures. We consider not only the paired t-test but also one-way analysis of variance (ANOVA) for significance testing to accommodate comparison of m(≥ 2) systems under a given set of statistical requirements (α: the Type I error rate, ß: the Type II error rate, and minD: the minimum detectable difference between the best and the worst systems). Using our simple Excel tools and some pooled variance estimates from past data, researchers can design statistically well-designed test collections. We demonstrate that, as different evaluation measures have different variances across topics, they inevitably require different topic set sizes. This suggests that the evaluation measures should be chosen at the test collection design phase. Moreover, through a pool depth reduction experiment with past data, we show how the relevance assessment cost can be reduced dramatically while freezing the set of statistical requirements. Based on the cost analysis and the available budget, researchers can determine the right balance between n and the pool depth pd. Our techniques and tools are applicable to test collections for non-IR tasks as well.

It has long been recognized that search queries are often broad and ambiguous. Even when submitting the same query, different users may have different search intents. Moreover, the intents are dynamically evolving. Some intents are constantly popular with users, others are more bursty. We propose a method for mining dynamic query intents from search query logs. By regarding the query logs as a data stream, we identify constant intents while quickly capturing new bursty intents. To evaluate the accuracy and efficiency of our method, we conducted experiments using 50 topics from the NTCIR INTENT-9 data and additional five popular topics, all supplemented with six-month query logs from a commercial search engine. Our results show that our method can accurately capture new intents with short response time.

Aggregating search results from a variety of diverse verticals such as news, images, videos and Wikipedia into a single interface is a popular web search presentation paradigm. Although several aggregated search (AS) metrics have been proposed to evaluate AS result pages, their properties remain poorly understood. In this paper, we compare the properties of existing AS metrics under the assumptions that (1) queries may have multiple preferred verticals; (2) the likelihood of each vertical preference is available; and (3) the topical relevance assessments of results returned from each vertical is available. We compare a wide range of AS metrics on two test collections. Our main criteria of comparison are (1) discriminative power, which represents the reliability of a metric in comparing the performance of systems, and (2) intuitiveness, which represents how well a metric captures the various key aspects to be measured (i.e. various aspects of a user's perception of AS result pages). Our study shows that the AS metrics that capture key AS components (e.g., vertical selection) have several advantages over other metrics. This work sheds new lights on the further developments and applications of AS metrics.

Building test collections based on nuggets is useful evaluating systems that return documents, answers, or summaries. However, nugget construction requires a lot of manual work and is not feasible for large query sets. Towards an efficient and scalable nugget-based evaluation, we study the applicability of semi-automatic nugget extraction in the context of the ongoing NTCIR One Click Access (1CLICK) task. We compare manually-extracted and semi-automatically-extracted Japanese nuggets to demonstrate the coverage and efficiency of the semi-automatic nugget extraction. Our findings suggest that the manual nugget extraction can be replaced with a direct adaptation of the English semi-automatic nugget extraction system, especially for queries for which the user desires broad answers from free-form text.

The One Click Access Task (1CLICK) of NTCIR requires systems to return a concise multi-document summary of web pages in response to a query which is assumed to have been submitted in a mobile context. Systems are evaluated based on information units (or iUnits), and are required to present important pieces of information first and to minimise the amount of text the user has to read. Using the official Japanese results of the second round of the 1CLICK task from NTCIR-10, we discuss our task setting and evaluation framework. Our analyses show that: (1) Simple baseline methods that leverage search engine snippets or Wikipedia are effective for 'lookup' type queries but not necessarily for other query types; (2) There is still a substantial gap between manual and automatic runs; and (3) Our evaluation metrics are relatively robust to the incompleteness of iUnits.

Most commercial search engines have a query suggestion feature, which is designed to capture various possible search intents behind the user's original query. However, even though different search intents behind a given query may have been popular at different time periods in the past, existing query suggestion methods neither utilize nor present such information. In this study, we propose Time-aware Structured Query Suggestion (TaSQS) which clusters query suggestions along a timeline so that the user can narrow down his search from a temporal point of view. Moreover, when a suggested query is clicked, TaSQS presents web pages from query-URL bipartite graphs after ranking them according to the click counts within a particular time period. Our experiments using data from a commercial search engine log show that the time-aware clustering and the time-aware document ranking features of TaSQS are both effective.

We introduce a general information access evaluation framework that can potentially handle summaries, ranked document lists and even multi query sessions seamlessly. Our framework first builds a trailtext which represents a concatenation of all the texts read by the user during a search session, and then computes an evaluation metric called U-measure over the trailtext. Instead of discounting the value of a retrieved piece of information based on ranks, U-measure discounts it based on its position within the trailtext. U-measure takes the document length into account just like Time-Biased Gain (TBG), and has the diminishing return property. It is therefore more realistic than rank-based metrics. Furthermore, it is arguably more flexible than TBG, as it is free from the linear traversal assumption (i.e., that the user scans the ranked list from top to bottom), and can handle information access tasks other than ad hoc retrieval.

To construct a diversified search test collection, a set of possible subtopics (or intents) needs to be determined for each topic, in one way or another, and perintent relevance assessments need to be obtained. In the TREC Web Track Diversity Task, subtopics are manually developed at NIST, based on results of automatic click log analysis; in the NTCIR INTENT Task, intents are determined by manually clustering 'subtopics strings' returned by participating systems. In this study, we address the following research question: Does the choice of intents for a test collection affect relative performances of diversified search systems? To this end, we use the TREC 2012 Web Track Diversity Task data and the NTCIR-10 INTENT-2 Task data, which share a set of 50 topics but have different intent sets. Our initial results suggest that the choice of intents may affect relative performances, and that this choice may be far more important than how many intents are selected for each topic

The NTCIR INTENT task comprises two subtasks: {\em Subtopic Mining}, where systems are required to return a ranked list of {\em subtopic strings} for each given query; and {\em Document Ranking}, where systems are required to return a diversified web search result for each given query. This paper summarises the novel features of the Second INTENT task at NTCIR-10 and its main findings, and poses some questions for future diversified search evaluation.

This paper tackles the problem of mining subgoals of a given search goal from data. For example, when a searcher wants to travel to London, she may need to accomplish several subtasks such as "book flights," "book a hotel," "find good restaurants" and "decide which sightseeing spots to visit." As another example, if a searcher wants to lose weight, there may exist several alternative solutions such as "do physical exercise," "take diet pills," and "control calorie intake." In this paper, we refer to such subtasks or solutions as subgoals, and propose to utilize sponsored search data for finding subgoals of a given query by means of query clustering. Advertisements (ads) reflect advertisers' tremendous efforts in trying to match a given query with implicit user needs. Moreover, ads are usually associated with a particular action or transaction. We therefore hypothesized that they are useful for subgoal mining. To our knowledge, our work is the first to use sponsored search data for this purpose. Our experimental results show that sponsored search data is a good resource for obtaining related queries and for identifying subgoals via query clustering. In particular, our method that combines ad impressions from sponsored search data and query co-occurrences from session data outperforms a state-of-the-art query clustering method that relies on document clicks rather than ad impressions in terms of purity, NMI, Rand Index, F1-measure and subgoal recall.

A diversified search result for an underspecified query generally contains web pages in which there are answers that are relevant to different aspects of the query. In order to help the user locate such relevant answers, we propose a simple extension to the standard Search Engine Result Page (SERP) interface, called AspecTiles. In addition to presenting a ranked list of URLs with their titles and snippets, AspecTiles visualizes the relevance degree of a document to each aspect by means of colored squares ("tiles"). To compare AspecTiles with the standard SERP interface in terms of usefulness, we conducted a user study involving 30 search tasks designed based on the TREC web diversity task topics as well as 32 participants. Our results show that AspecTiles has some advantages in terms of search performance, user behavior, and user satisfaction. First, AspecTiles enables the user to gather relevant information significantly more efficiently than the standard SERP interface for tasks where the user considers several different aspects of the query to be important at the same time (multi-aspect tasks). Second, AspecTiles affects the user's information seeking behavior: with this interface, we observed significantly fewer query reformulations, shorter queries and deeper examinations of ranked lists in multi-aspect tasks. Third, participants of our user study found AspecTiles significantly more useful for finding relevant information and easy to use than the standard SERP interface. These results suggest that simple interfaces like AspecTiles can enhance the search performance and search experience of the user when their queries are underspecified.

Query suggestion is a useful tool to help users express their information needs by supplying alternative queries. When evaluating the effectiveness of query suggestion algorithms, many previous studies focus on measuring whether a suggestion query is relevant or not to the input query. This assessment criterion is too simple to describe users' requirements. In this paper, we introduce two scenarios of query suggestion. The first scenario represents cases where the search result of the input query is unsatisfactory. The second scenario represents cases where the search result is satisfactory but the user may be looking for alternative solutions. Based on the two scenarios, we propose two assessment criteria. Our labeling results indicate that the new assessment criteria provide finer distinctions among query suggestions than the traditional relevance-based criterion.

In this poster, we propose two evaluation tasks for mobile information access. The first task evaluates the system's ability to guess what the user's query should be given a context ("Knowing What"). The second task evaluates the system's ability to decide when to proactively deploy a given query ("Knowing When"). We conduct a preliminary manual analysis of a mobile query log to limit the space of possible queries so as to design feasible and practical evaluation tasks.

Query suggestion, which enables the user to revise a query with a single click, has become one of the most fundamental features of Web search engines. However, it is often difficult for the user to choose from a list of query suggestions, and to understand the relation between an input query and suggested ones. In this paper, we propose a new method to present query suggestions to the user, which has been designed to help two popular query reformulation actions, namely, specialization (e.g. from "nikon" to "nikon camera" ) and parallel movement (e.g. from "nikon camera" to "canon camera"). Using a query log collected from a popular commercial Web search engine, our prototype called SParQS classifies query suggestions into automatically generated categories and generates a label for each category. Moreover, SParQS presents some new entities as alternatives to the original query (e.g. "canon" in response to the query "nikon"), together with their query suggestions classified in the same way as the original query's suggestions. We conducted a task-based user study to compare SParQS with a traditional "flat list" query suggestion interface. Our results show that the SParQS interface enables subjects to search more successfully than the flat list case, even though query suggestions presented were exactly the same in the two interfaces. In addition, the subjects found the query suggestions more helpful when they were presented in the SParQS interface rather than in a flat list.

Given an ambiguous or underspecified query, search result diversification aims at accomodating different user intents within a single "entry-point" result page. However, some intents are informational, for which many relevant pages may help, while others are navigational, for which only one web page is required. We propose new evaluation metrics for search result diversification that considers this distinction, as well as a simple method for comparing the intuitiveness of a given pair of metrics quantitatively. Our main experimental findings are: (a) In terms of discriminative power which reflects statistical reliability, the proposed metrics, DIN#-nDCG and P+Q#, are comparable to intent recall and D#-nDCG, and possibly superior to α-nDCG; (b) In terms of preference agreement with intent recall, P+Q# is superior to other diversity metrics and therefore may be the most intuitive as a metric that emphasises diversity; and (c) In terms of preference agreement with effective precision, DIN#-nDCG is superior to other diversity metrics and therefore may be the most intuitive as a metric that emphasises relevance. Moreover, DIN#-nDCG may be the most intuitive as a metric that considers both diversity and relevance. In addition, we demonstrate that the randomised Tukey's Honestly Significant Differences test that takes the entire set of available runs into account is substantially more conservative than the paired bootstrap test that only considers one run pair at a time, and therefore recommend the former approach for significance testing when a set of runs is available for evaluation.

We define Direct Information Access as a type of information access where there is no user operation such as clicking or scrolling between the user's click on the search button and the user's information acquisition; we define Immediate Information Access as a type of information access where the user can locate the relevant information within the system output very quickly. Hence, a Direct and Immediate Information Access (DIIA) system is expected to satisfy the user's information need very quickly with its very first response. We propose a nugget-based evaluation framework for DIIA, which takes nugget positions into account in order to evaluate the ability of a system to present important nuggets first and to minimise the amount of text the user has to read. To demonstrate the integrity, usefulness and limitations of our framework, we built a Japanese DIIA test collection with 60 queries and over 2,800 nuggets as well as an offset-based nugget match evaluation interface, and conducted experiments with manual and automatic runs. The results suggest our proposal is a useful complement to traditional ranked retrieval evaluation based on document relevance.

Search queries are often ambiguous and/or underspecified. To accomodate different user needs, search result diversification has received attention in the past few years. Accordingly, several new metrics for evaluating diversification have been proposed, but their properties are little understood. We compare the properties of existing metrics given the premises that (1) queries may have multiple intents; (2) the likelihood of each intent given a query is available; and (3) graded relevance assessments are available for each intent. We compare a wide range of traditional and diversified IR metrics after adding graded relevance assessments to the TREC 2009 Web track diversity task test collection which originally had binary relevance assessments. Our primary criterion is discriminative power, which represents the reliability of a metric in an experiment. Our results show that diversified IR experiments with a given number of topics can be as reliable as traditional IR experiments with the same number of topics, provided that the right metrics are used. Moreover, we compare the intuitiveness of diversified IR metrics by closely examining the actual ranked lists from TREC. We show that a family of metrics called D#-measures have several advantages over other metrics such as α-nDCG and Intent-Aware metrics.

Community Question Answering (CQA) sites such as Yahoo! Answers have emerged as rich knowledge resources for information seekers. However, answers posted to CQA sites can be irrelevant, incomplete, redundant, incorrect, biased, ill-formed or even abusive. Hence, automatic selection of "good" answers for a given posted question is a practical research problem that will help us manage the quality of accumulated knowledge. One way to evaluate answer selection systems for CQA would be to use the Best Answers (BAs) that are readily available from the CQA sites. However, BAs may be biased, and even if they are not, there may be other good answers besides BAs. To remedy these two problems, we propose system evaluation methods that involve multiple answer assessors and graded-relevance information retrieval metrics. Our main findings from experiments using the NTCIR-8 CQA task data are that, using our evaluation methods, (a) we can detect many substantial differences between systems that would have been overlooked by BA-based evaluation; and (b) we can better identify hard questions (i.e. those that are handled poorly by many systems and therefore require focussed investigation) compared to BAbased evaluation. We therefore argue that our approach is useful for building effective CQA answer selection systems despite the cost of manual answer assessments.

We analyse the query log of a click-oriented Japanese search engine that utilises the link structures of Wikipedia for encouraging the user to change his information need and to perform repeated, serendipitous, exploratory search. Our results show that users tend to make transitions within the same query type: from person names to person names, from place names to place names, and so on.

Test collections are growing larger, and relevance data constructed through pooling are suspected of becoming more and more incomplete and biased. Several studies have used evaluation metrics specifically designed to handle this problem, but most of them have only examined the metrics under incomplete but unbiased conditions, using random samples of the original relevance data. This paper examines nine metrics in a more realistic setting, by reducing the number of pooled systems. Even though previous work has shown that metrics based on a condensed list, obtained by removing all unjudged documents from the original ranked list, are effective for handling very incomplete but unbiased relevance data, we show that these results do not hold in the presence of system bias. In our experiments using TREC and NTCIR data, we first show that condensed-list metrics overestimate new systems while traditional metrics underestimate them, and that the overestimation tends to be larger than the underestimation. We then show that, when relevance data is heavily biased towards a single team or a few teams, the condensed-list versions of Average Precision (AP), Q-measure (Q) and normalised Discounted Cumulative Gain (nDCG), which we call AP', Q' and nDCG', are not necessarily superior to the original metrics in terms of discriminative power, i.e., the overall ability to detect pairwise statistical significance. Nevertheless, even under system bias, AP' and Q' are generally more discriminative than bpref and the condensed-list version of Rank-Biased Precision (RBP), which we call RBP'.

No abstract available.

Information retrieval systems are compared using evaluation metrics, with researchers commonly reporting results for simple metrics such as precision-at-10 or reciprocal rank together with more complex ones such as average precision or discounted cumulative gain. In this paper, we demonstrate that complex metrics are as good as or better than simple metrics at predicting the performance of the simple metrics on other topics. Therefore, reporting of results from simple metrics alongside complex ones is redundant.

Modern information retrieval (IR) test collections have grown in size, but the available manpower for relevance assessments has more or less remained constant. Hence, how to reliably evaluate and compare IR systems using incomplete relevance data, where many documents exist that were never examined by the relevance assessors, is receiving a lot of attention. This article compares the robustness of IR metrics to incomplete relevance assessments, using four different sets of graded-relevance test collections with submitted runs—the TREC 2003 and 2004 robust track data and the NTCIR-6 Japanese and Chinese IR data from the crosslingual task. Following previous work, we artificially reduce the original relevance data to simulate IR evaluation environments with extremely incomplete relevance data. We then investigate the effect of this reduction on discriminative power , which we define as the proportion of system pairs with a statistically significant difference for a given probability of Type I Error, and on Kendall’s rank correlation , which reflects the overall resemblance of two system rankings according to two different metrics or two different relevance data sets. According to these experiments, Q′, nDCG′ and AP′ proposed by Sakai are superior to bpref proposed by Buckley and Voorhees and to Rank-Biased Precision proposed by Moffat and Zobel. We also point out some weaknesses of bpref and Rank-Biased Precision by examining their formal definitions.

Recently, a number of TREC tracks have adopted a retrieval effectiveness metric called bpref which has been designed for evaluation environments with incomplete relevance data. A graded-relevance version of this metric called rpref has also been proposed. However, we show that the application of Q-measure, normalised Discounted Cumulative Gain (nDCG) or Average Precision (AveP)to condensed lists, obtained by ?ltering out all unjudged documents from the original ranked lists, is actually a better solution to the incompleteness problem than bpref. Furthermore, we show that the use of graded relevance boosts the robustness of IR evaluation to incompleteness and therefore that Q-measure and nDCG based on condensed lists are the best choices. To this end, we use four graded-relevance test collections from NTCIR to compare ten different IR metrics in terms of system ranking stability and pairwise discriminative power.

This paper describes how the Bootstrap approach to statistics can be applied to the evaluation of IR effectiveness metrics. First, we argue that Bootstrap Hypothesis Tests deserve more attention from the IR community, as they are based on fewer assumptions than traditional statistical significance tests. We then describe straightforward methods for comparing the sensitivity of IR metrics based on Bootstrap Hypothesis Tests. Unlike the heuristics-based "swap" method proposed by Voorhees and Buckley, our method estimates the performance difference required to achieve a given significance level directly from Bootstrap Hypothesis Test results. In addition, we describe a simple way of examining the accuracy of rank correlation between two metrics based on the Bootstrap Estimate of Standard Error. We demonstrate the usefulness of our methods using test collections and runs from the NTCIR CLIR track for comparing seven IR metrics, including those that can handle graded relevance and those based on the Geometric Mean.

We introduce an evaluation metric called P-measure for the task of retrieving <ione highly relevant document. It models user behaviour in practical tasks such as known-item search, and is more stable and sensitive than Reciprocal Rank which cannot handle graded relevance.

No abstract available.

No abstract available.

No abstract available.

No abstract available.

This paper examines the use of generic summaries for indexing in information retrieval. Our main observations are that: (1) With or without pseudo-relevance feedback, a summary index may be as effective as the corresponding fulltext index forprecision-oriented search of highly relevant documents. %43 But a reasonably sophisticated summarizer, using a compression ratio of 10-30%, is desirable for this purpose. (2) In pseudo-relevance feedback, using a summary index at initial search and a fulltext index at final search is possibly effective for precision-oriented search, regardless of relevance levels. This strategy is significantly more effective than the one using the summary index only and probably more effective than using summaries as mere term selection filters. %the use of summaries as mere term selection filters. %The summary quality is probably not a critical factor for this strategy, For this strategy, the summary quality is probably not a critical factor, and a compression ratio of 5-10% appears best.

This paper describes a structured investigation into the retrieval of Japanese text.The study includes a comparison of different indexing strategies for documents and queries, investigation of term weighting strategies principally derived for use with English texts, and the application of relevance feedback for query expansion.Results on the standard BMIR-Jl and BMIR-J2 Japanese retrieval collections indicate that term weighting transfers well to Japanese text.Indexing using dictionary based morphological analysis and character strings are both shown to be individually effective, but marginally better in combination.We also demonstrate that relevance feedback can be used effectively for query expansion in Japanese routing applications.