Pseudo-Relevance Feedback (PRF) aims to improve performance, but it can suffer from query drift, creating asymmetric impacts across queries. Current Selective PRF (sPRF) strategies attempt to mitigate this by predicting when PRF will help, but they rely on offline metrics that may not reflect actual user preferences. To bridge this gap, we propose a participatory auditing strategy that evaluates PRF's real-world impact through natural user interactions. Specifically, we design a 3×3 grid interface using Latin square ordering to mitigate position bias, and deploy Team Draft Interleaving (TDI) to collect implicit user preferences without disrupting normal search behaviour. We conduct two user studies: in our first user study, we find that only 20.9% of queries benefit from PRF, while 25.6% result in a degraded user experience. Our second study demonstrates that avoiding harmful PRF yields a 68.5% relative improvement, nearly twice the 37.0% gained from amplifying beneficial PRF, highlighting that harm avoidance is more valuable than benefit amplification. Our findings highlight the necessity for reliable sPRF and suggest that future sPRF research could benefit from incorporating interaction-aware participatory auditing when developing sPRF approaches.

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

Learning to Rank (LTR) models learn from historical user interactions, such as user clicks. However, there is an inherent bias in the clicks of users due to position bias, i.e., users are more likely to click highly-ranked documents than low-ranked documents. To address this bias when training LTR models, many approaches from the literature re-weight the users' click data using Inverse Propensity Scoring (IPS). IPS re-weights the user's clicks proportionately to the position in the historical ranking that a document was placed when it was clicked since low-ranked documents are less likely to be seen by a user. In this paper, we argue that low-ranked documents that are similar to highly-ranked relevant documents are also likely to be relevant. Moreover, accounting for the similarity of low-ranked documents to highly ranked relevant documents when calculating IPS can more effectively mitigate the effects of position bias. Therefore, we propose an extension to IPS, called IPSsim, that takes into consideration the similarity of documents when estimating IPS. We evaluate our IPSsim estimator using two large publicly available LTR datasets under a number of simulated user click settings, and with different numbers of training clicks. Our experiments show that our IPSsim estimator is more effective than the existing IPS estimators for learning an unbiased LTR model, particularly in top-n settings when n >= 30. For example, when n = 50, our IPSsim estimator achieves a statistically significant ~3% improvement (p < 0.05) in terms of NDCG compared to the Doubly Robust estimator from the literature.

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

In the first stage of a re-ranking pipeline, an inexpensive ranking model is typically deployed to retrieve a set of documents that are highly likely to be relevant to the user's query. The retrieved documents are then re-ranked by a more effective but expensive ranking model, e.g., a deep neural ranker such as BERT. However, in such a standard pipeline, no new documents are typically discovered after the first stage retrieval. Hence, the amount of exposure that a particular group of documents - e.g., documents from a particular demographic category - can receive is limited by the number of documents that are retrieved in the first stage retrieval. Indeed, if too few documents from a group are retrieved in the first stage retrieval, ensuring that the group receives a fair amount of exposure to the user may become infeasible. Therefore, it is useful to identify more documents from underrepresented groups that are potentially relevant to the query during the re-ranking stage. In this work, we investigate how deploying adaptive re-ranking, which enables the discovery of additional potentially relevant documents in the re-ranking stage, can improve the exposure that a given group of documents receives in the final ranking. We propose six adaptive re-ranking policies that can discover documents from underrepresented groups to increase the disadvantaged groups' exposure in the final ranking. Our experiments on the TREC 2021 and 2022 Fair Ranking Track test collections show that our policies consistently improve the fairness of the exposure distribution in the final ranking, compared to standard adaptive re-ranking approaches, resulting in increases of up to ~13% in Attention Weighted Ranked Fairness (AWRF). Moreover, our best performing policy, Policy 6, consistently maintains and frequently increases the utility of the search results in terms of nDCG.

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

Due to the massive volume of articles produced online every day, it is challenging for online platforms (e.g., news agencies) to present the information about an event, activity or discussion to their users in an easily digestible format. Therefore, there is a need for automatic methods to extract related and time-ordered information about events (i.e., information threads) from large unstructured collections of documents. In this work, we propose a novel unsupervised hierarchical agglomerative clustering (HAC) based information threading approach to generate chronological and coherent threads of information in a collection. Unlike, the well-known tasks of topic detection and tracking or event threading that focus on grouping information by important keywords and/or entities, our proposed approach identifies threads based on temporal relations and diverse information about an event, i.e., who did what, why, where, when and how (aka the 5W1H questions). In particular, our proposed approach, deploys a tailored similarity function for HAC by leveraging extracted answers to 5W1H questions along with time decay between documents. We evaluate our proposed HAC 5W1H information threading approach on two large expert-annotated collections of news articles, i.e., NewSHead and Multi-News (over 112k and 32k articles, respectively). Our experiments show that HAC 5W1H markedly improves the number of, and quality of, threads that are generated compared to existing state-of-the-art approaches from the literature, e.g., 100.98% more threads and +213.39% improvement in Normalised Mutual Information compared to the best evaluated baseline on the larger NewSHead collection. We also conducted a user study that shows that our proposed HAC 5W1H information threading approach is significantly (p<0.05) preferred by users in terms of coherence, diversity and chronological correctness compared to the existing state-of-the-art approaches.

Government documents must be manually sensitivity reviewed to identify and protect any sensitive information (e.g. personal information) in the documents before the documents can be opened to the public. However, due to the large volume of born-digital documents that need to be reviewed, there is a growing need for technologies to assist human reviewers and improve the efficiency of the review process. For example, in sensitivity review, a reviewer needs to be able to quickly find documents that belong to specific latent semantic categories (e.g., documents about criminality that contain the personal details of victims). However, manually identifying such document categories is a challenging task when reviewing digital documents, due to the size of, and lack of structure in the collections. We hypothesise that reviewing documents that are clustered by their latent semantic categories will increase the efficiency of the human reviewers, since the reviewers will be able to review related documents in sequence. In this work, we conduct a user study to evaluate the effectiveness of different clustering techniques, document metadata and automatic sensitivity classification, for grouping and prioritising documents for review, to increase the efficiency of the review process. Our study shows that reviewing documents in semantic clusters can significantly improve the efficiency (i.e., speed) of the sensitivity reviewers (+15.65%, T-Test, p<0.05) while maintaining the reviewers’ accuracy. Moreover, we propose a novel strategy for prioritising document clusters for review to maximise the number of documents that are opened to the public within a fixed reviewing time budget. Our proposed prioritisation strategy results in a significant increase in the number of documents that are opened to the public (+37.99%, T-Test, p<0.05) compared to prioritising documents without clusters.

With the massive increase in the volume of digitally produced documents, government departments face a logistical issue when conducting the manual sensitivity review of documents that should be opened to the public. When reviewing a document, sensitivity reviewers often need to quickly access related information from other documents in the collection. For example, documents that mention the same topic or event can provide the reviewers with useful contextual information and assist the reviewers to make consistent sensitivity judgements more quickly. However, it is infeasible to manually identify groups of such related documents in large unstructured collections. In this work, we present a sensitivity review system that automatically identifies groups of related documents to assist reviewers and increase the efficiency of sensitivity review. In particular, our system groups the documents that are to be sensitivity reviewed based on the documents' semantic categories (e.g., criminality). Moreover, the system identifies chronological and coherent information threads to describe the full context of an event, activity or discussion that may be spread across multiple documents. Additionally, the system prioritises the identified semantic categories and information threads for review by leveraging automatic sensitivity classification to maximise the number of documents that can be opened to the public in a limited reviewing time-budget.

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

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

Many government and public organisations have a requirement to release their official documents to the public and therefore need to review such documents to identify and protect any sensitive information that they contain. When reviewing a document for sensitivity, reviewers often use information from other documents within the collection to assist in their decisions. It can be difficult for the reviewers to find related documents in large digital collections when they are performing sensitivity review. Receptor is a new solution that aims to provide sensitivity reviewers with the ability to explore a collection of documents to discover latent relations, between for example entities and events, that can be a reliable indicator of sensitive information. The system provides novel scalable graph search and exploration functionalities as well as interactive visualisations of the latent relations between related entities, events, and documents to enable users to identify hidden patterns of sensitivity.

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

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

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