Depression places substantial pressure on mental health services, and many people describe their experiences outside clinical settings in high-volume user-generated text (e.g., online forums and social media). Automatically identifying clinical symptom evidence in such text can therefore complement limited clinical capacity and scale to large populations. We address this need through sentence-level classification of 21 depression symptoms from the BDI-II questionnaire, using BDI-Sen, a dataset annotated for symptom relevance. This task is fine-grained and highly imbalanced, and we find that common LLM approaches (zero-shot, in-context learning, and fine-tuning) struggle to apply consistent relevance criteria for most symptoms. We propose Symptom Induction (SI), a novel approach which compresses labeled examples into short, interpretable guidelines that specify what counts as evidence for each symptom and uses these guidelines to condition classification. Across four LLM families and eight models, SI achieves the best overall weighted F1 on BDI-Sen, with especially large gains for infrequent symptoms. Cross-domain evaluation on an external dataset further shows that induced guidelines generalize across other diseases shared symptomatology (bipolar and eating disorders).

Retrieval-Augmented Generation (RAG) systems rely on retrieved documents being concatenated into a model's input context, making both document ordering and context size critical yet controversial design choices. Prior work reports position-based effects such as lost in the middle and related long-context phenomena. However, empirical findings remain inconsistent and hard to reproduce across models, datasets, and evaluation protocols. In this paper, we present a systematic reproducibility study that revisits these claims and examines how they evolve with contemporary LLMs under a controlled evaluation framework. We first show that topic sampling is a major source of variance: small topic sets can mask or exaggerate ordering effects. Based on repeated subset sampling across multiple topic budgets, we provide a practical calibration procedure that identifies topic counts yielding stable trends at feasible cost. Using these fixed topic sets, we then reproduce and extend results on position sensitivity, re-evaluating lost in the middle and positional biases in modern LLMs. Then, we also study a more realistic RAG scenario in which relevance is mediated by a retriever rather than oracle access to ground-truth documents. In this setting, we re-examine a recent industry study and identify discrepancies to evaluation choices such as limited topic coverage and reliance on LLM-based judges. Finally, we conduct an analysis of how retrieval order and context size affect downstream LLM performance under imperfect retrieval. Our results demonstrate that both factors interact strongly with retrieval quality and model choice, and that conclusions drawn from idealised setups do not always transfer to real-world RAG pipelines. We release all code and configurations to support reproducibility and future work on robust RAG evaluation.

Evaluating query suggestion systems in a manner that reflects real-world query formulation remains a persistent challenge. Most offline methodologies adopt static assumptions, such as users accepting all or the top-e suggestions, ignoring the inherently selective and intent-driven nature of interactive search. While online experiments provide realistic behavioural signals, they are costly, difficult to scale, and often irreproducible. To bridge this gap, we introduce SIQSE (Simulation-based Interactive Query Suggestion Evaluation), a framework that models query reformulation as an interactive selection task performed by a simulated user. In SIQSE, a Large Language Model (LLM) acts as a surrogate user that progressively selects suggestions according to contextual relevance and explicit search intent. Unlike static offline protocols, this simulation captures the iterative and selective dynamics of real query formulation. Our contributions are twofold. First, we develop and validate an LLM-based selection model, systematically analysing how varying levels of intent information and selection strategies affect its ability to approximate human selection behaviour. Second, we employ this selector to benchmark multiple query suggestion systems across diverse datasets under interactive conditions. Importantly, while the selector is LLM-based, the final evaluation is computed exclusively through ranking-based effectiveness metrics over the rankings produced by selected expansions, ensuring that system performance reflects retrieval quality rather than alignment with the surrogate user model. By modelling round-based interaction while maintaining metric independence, SIQSE offers a scalable, reproducible evaluation paradigm that brings offline assessment closer to the complexity of real-world search behaviour. To facilitate adoption and reproducibility, we release SIQSE as an open-source Python library.

High-quality relevance judgements over large query sets are essential for evaluating Information Retrieval (IR) systems, yet manual annotation remains costly and time-consuming. Large Language Models (LLMs) have recently shown promise as automatic relevance assessors, but their reliability is still limited. Most existing approaches rely on zero-shot prompting or in-context learning (ICL) with a small number of labelled examples. However, standard ICL treats examples as independent instances and fails to explicitly capture the underlying relevance criteria of a topic, restricting its ability to generalise to unseen query-document pairs. To address this limitation, we introduce Relevance Context Learning (RCL), a novel framework that leverages human relevance judgements to explicitly model topic-specific relevance criteria. Rather than directly using labelled examples for in-context prediction, RCL first prompts an LLM (Instructor LLM) to analyse sets of judged query-document pairs and generate explicit narratives that describe what constitutes relevance for a given topic. These relevance narratives are then used as structured prompts to guide a second LLM (Assessor LLM) in producing relevance judgements. To evaluate RCL in a realistic data collection setting, we propose a hybrid pooling strategy in which a shallow depth-k pool from participating systems is judged by human assessors, while the remaining documents are labelled by LLMs. Experimental results demonstrate that RCL substantially outperforms zero-shot prompting and consistently improves over standard ICL. Overall, our findings indicate that transforming relevance examples into explicit, context-aware relevance narratives is a more effective way of exploiting human judgements for LLM-based IR dataset construction.

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

Abstract Topic modeling has emerged as a crucial tool in the field of natural language processing, enabling the automatic discovery of latent structures in large textual corpora. However, determining the quality of the topics remains a significant challenge, particularly in measuring the coherence of the top words of the extracted topics. Early efforts relied on human judgments, but these approaches are resource-intensive. Automated coherence metrics have since been developed. For example, some measures exploit word co-occurrence, while other methods are grounded in distributional semantics (e.g., employing word embeddings). In this study, we thoroughly explore the application of embedded representations to evaluate the quality of topics. While a number of isolated studies have analyzed the role of specific word representation techniques for measuring topic coherence, a complete picture of their effectiveness is still lacking. This work brings together different embedding-based approaches, including Word2Vec, FastText, GloVe, and BERT, which had been studied separately, and extends prior research by incorporating additional models, such as RoBERTa, ALBERT and MPNET. Topic coherence is measured by computing similarity scores between word embeddings, thus obtaining rich semantic associations that traditional measures may overlook. Our analysis demonstrates that these methods are as effective as, and often surpass, classical coherence measures. Our results contribute to a growing body of research advocating for advanced semantic representations as robust alternatives to traditional approaches in evaluating topic model coherence.

The increasing demand for mental health services has outpaced the availability of real training data to develop clinical professionals, leading to limited support for the diagnosis of depression. This shortage has motivated the development of simulated or virtual patients to assist in training and evaluation, but existing approaches often fail to generate clinically valid, natural, and diverse symptom presentations. In this work, we embrace the recent advanced language models as the backbone and propose a novel clinician-in-the-loop patient simulation pipeline, TalkDep, with access to diversified patient profiles to develop simulated patients. By conditioning the model on psychiatric diagnostic criteria, symptom severity scales, and contextual factors, our goal is to create authentic patient responses that can better support diagnostic model training and evaluation. We verify the reliability of these simulated patients with thorough assessments conducted by clinical professionals. The availability of validated simulated patients offers a scalable and adaptable resource for improving the robustness and generalisability of automatic depression diagnosis systems.

Depression is a pervasive mental health condition that affects hundreds of millions of individuals worldwide, yet many cases remain undiagnosed due to barriers in traditional clinical access and pervasive stigma. Social media platforms, and Reddit in particular, offer rich, user-generated narratives that can reveal early signs of depressive symptomatology. However, existing computational approaches often label entire posts simply as depressed or not depressed, without linking language to specific criteria from the DSM-5, the standard clinical framework for diagnosing depression. This limits both clinical relevance and interpretability. To address this gap, we introduce ReDSM5, a novel Reddit corpus comprising 1484 long-form posts, each exhaustively annotated at the sentence level by a licensed psychologist for the nine DSM-5 depression symptoms. For each label, the annotator also provides a concise clinical rationale grounded in DSM-5 methodology. We conduct an exploratory analysis of the collection, examining lexical, syntactic, and emotional patterns that characterize symptom expression in social media narratives. Compared to prior resources, ReDSM5 uniquely combines symptom-specific supervision with expert explanations, facilitating the development of models that not only detect depression but also generate human-interpretable reasoning. We establish baseline benchmarks for both multi-label symptom classification and explanation generation, providing reference results for future research on detection and interpretability.

Offline evaluation of search systems depends on test collections. These benchmarks provide the researchers with a corpus of documents, topics and relevance judgements indicating which documents are relevant for each topic. While test collections are an integral part of Information Retrieval (IR) research, their creation involves significant efforts in manual annotation. Large language models (LLMs) are gaining much attention as tools for automatic relevance assessment. Recent research has shown that LLM-based assessments yield high systems ranking correlation with human-made judgements. These correlations are helpful in large-scale experiments but less informative if we want to focus on top-performing systems. Moreover, these correlations ignore whether and how LLM-based judgements impact the statistically significant differences among systems with respect to human assessments. In this work, we look at how LLM-generated judgements preserve ranking differences among top-performing systems and also how they preserve pairwise significance evaluation as human judgements. Our results show that LLM-based judgements are unfair at ranking top-performing systems. Moreover, we observe an exceedingly high rate of false positives regarding statistical differences.

Hate speech represents a pervasive and detrimental form of online discourse, often manifested through an array of slurs, from hateful tweets to defamatory posts. As such speech proliferates, it connects people globally and poses significant social, psychological, and occasionally physical threats to targeted individuals and communities. Current computational linguistic approaches for tackling this phenomenon rely on labelled social media datasets for training. For unifying efforts, our study advances in the critical need for a comprehensive meta-collection, advocating for an extensive dataset to help counteract this problem effectively. We scrutinized over 60 datasets, selectively integrating those pertinent into MetaHate. This paper offers a detailed examination of existing collections, highlighting their strengths and limitations. Our findings contribute to a deeper understanding of the existing datasets, paving the way for training more robust and adaptable models. These enhanced models are essential for effectively combating the dynamic and complex nature of hate speech in the digital realm.

Creating test collections for offline retrieval evaluation requires human effort to judge documents' relevance. This expensive activity motivated much work in developing methods for constructing benchmarks with fewer assessment costs. In this respect, adjudication methods actively decide both which documents and the order in which experts review them, in order to better exploit the assessment budget or to lower it. Researchers evaluate the quality of those methods by measuring the correlation between the known gold ranking of systems under the full collection and the observed ranking of systems under the lower-cost one. This traditional analysis ignores whether and how the low-cost judgements impact on the statistically significant differences among systems with respect to the full collection. We fill this void by proposing a novel methodology to evaluate how the low-cost adjudication methods preserve the pairwise significant differences between systems as the full collection. In other terms, while traditional approaches look for stability in answering the question "is system A better than system B?", our proposed approach looks for stability in answering the question "is system A significantly better than system B?", which is the ultimate questions researchers need to answer to guarantee the generalisability of their results. Among other results, we found that the best methods in terms of ranking of systems correlation do not always match those preserving statistical significance.

People tend to consider social platforms as convenient media for expressing their concerns and emotional struggles. With their widespread use, researchers could access and analyze user-generated content related to mental states. Computational models that exploit that data show promising results in detecting at-risk users based on engineered features or deep learning models. However, recent works revealed that these approaches have a limited capacity for generalization and interpretation when considering clinical settings. Grounding the models' decisions on clinical and recognized symptoms can help to overcome these limitations. In this paper, we introduce BDI-Sen, a symptom-annotated sentence dataset for depressive disorder. BDI-Sen covers all the symptoms present in the Beck Depression Inventory-II (BDI-II), a reliable questionnaire used for detecting and measuring depression. The annotations in the collection reflect whether a statement about the specific symptom is informative (i.e., exposes traces about the individual's state regarding that symptom). We thoroughly analyze this resource and explore linguistic style, emotional attribution, and other psycholinguistic markers. Additionally, we conduct a series of experiments investigating the utility of BDI-Sen for various tasks, including the detection and severity classification of symptoms. We also examine their generalization when considering symptoms from other mental diseases. BDI-Sen may aid the development of future models that consider trustworthy and valuable depression markers.

Recommender systems evaluation has evolved rapidly in recent years. However, for offline evaluation, accuracy is the de facto standard for assessing the superiority of one method over another, with most research comparisons focused on tasks ranging from rating prediction to ranking metrics for top-n recommendation. Simultaneously, recommendation diversity and novelty have become recognized as critical to users’ perceived utility, with several new metrics recently proposed for evaluating these aspects of recommendation lists. Consequently, the accuracy-diversity dilemma frequently shows up as a choice to make when creating new recommendation algorithms.

Personalized recommender systems rely on knowledge of user preferences to produce recommendations. While those preferences are often obtained from past user interactions with the recommendation catalog, in some situations such observations are insufficient or unavailable. The most widely studied case is with new users, although other similar situations arise where explicit preference elicitation is valuable. At the same time, a seemingly disparate challenge is that there is a well-known popularity bias in many algorithmic approaches to recommender systems. The most common way of addressing this challenge is diversification, which tends to be applied to the output of a recommender algorithm, prior to items being presented to users. We tie these two problems together, showing a tight relationship. Our results show that popularity bias in preference elicitation contributes to popularity bias in recommendation. In particular, most elicitation methods directly optimize only for the relevance of recommendations that would result from collected preferences. This focus on recommendation accuracy biases the preferences collected. We demonstrate how diversification can instead be applied directly at elicitation time. Our model diversifies the preferences elicited using Multi-Armed Bandits, a classical exploration-exploitation framework from reinforcement learning. This leads to a broader understanding of users' preferences, and improved diversity and serendipity of recommendations, without necessitating post-hoc debiasing corrections.

Statistical significance tests can provide evidence that the observed difference in performance between 2 methods is not due to chance. In information retrieval (IR), some studies have examined the validity and suitability of such tests for comparing search systems. We argue here that current methods for assessing the reliability of statistical tests suffer from some methodological weaknesses, and we propose a novel way to study significance tests for retrieval evaluation. Using Score Distributions, we model the output of multiple search systems, produce simulated search results from such models, and compare them using various significance tests. A key strength of this approach is that we assess statistical tests under perfect knowledge about the truth or falseness of the null hypothesis. This new method for studying the power of significance tests in IR evaluation is formal and innovative. Following this type of analysis, we found that both the sign test and Wilcoxon signed test have more power than the permutation test and the t‐test. The sign test and Wilcoxon signed test also have good behavior in terms of type I errors. The bootstrap test shows few type I errors, but it has less power than the other methods tested.

In information retrieval evaluation, pooling is a well‐known technique to extract a sample of documents to be assessed for relevance. Given the pooled documents, a number of studies have proposed different prioritization methods to adjudicate documents for judgment. These methods follow different strategies to reduce the assessment effort. However, there is no clear guidance on how many relevance judgments are required for creating a reliable test collection. In this article we investigate and further develop methods to determine when to stop making relevance judgments. We propose a highly diversified set of stopping methods and provide a comprehensive analysis of the usefulness of the resulting test collections. Some of the stopping methods introduced here combine innovative estimates of recall with time series models used in Financial Trading. Experimental results on several representative collections show that some stopping methods can reduce up to 95% of the assessment effort and still produce a robust test collection. We demonstrate that the reduced set of judgments can be reliably employed to compare search systems using disparate effectiveness metrics such as Average Precision, NDCG, P@100, and Rank Biased Precision. With all these measures, the correlations found between full pool rankings and reduced pool rankings is very high.

The evaluation of Recommender Systems is still an open issue in the field. Despite its limitations, offline evaluation usually constitutes the first step in assessing recommendation methods due to its reduced costs and high reproducibility. Selecting the appropriate metric is a critical and ranking accuracy usually attracts the most attention nowadays. In this paper, we aim to shed light on the advantages of different ranking metrics which were previously used in Information Retrieval and are now used for assessing top-N recommenders. We propose methodologies for comparing the robustness and the discriminative power of different metrics. On the one hand, we study cut-offs and we find that deeper cut-offs offer greater robustness and discriminative power. On the other hand, we find that precision offers high robustness and Normalised Discounted Cumulative Gain provides the best discriminative power.

Given the diversity of recommendation algorithms, choosing one technique is becoming increasingly difficult. In this paper, we explore methods for combining multiple recommendation approaches. We studied rank aggregation methods that have been proposed for the metasearch task (i.e., fusing the outputs of different search engines) but have never been applied to merge top-N recommender systems. These methods require no training data nor parameter tuning. We analysed two families of methods: voting-based and score-based approaches. These rank aggregation techniques yield significant improvements over state-of-the-art top-N recommenders. In particular, score-based methods yielded good results; however, some voting techniques were also competitive without using score information, which may be unavailable in some recommendation scenarios. The studied methods not only improve the state of the art of recommendation algorithms but they are also simple and efficient.

Probabilistic modelling of recommender systems naturally introduces the concept of prior probability into the recommendation task. Relevance-Based Language Models, a principled probabilistic query expansion technique in Information Retrieval, has been recently adapted to the item recommendation task with success. In this paper, we study the effect of the item and user prior probabilities under that framework. We adapt two priors from the document retrieval field and then we propose other two new probabilistic priors. Evidence gathered from experimentation indicates that a linear prior for the neighbour and a probabilistic prior based on Dirichlet smoothing for the items improve the quality of the item recommendation ranking.

Relevance-Based Language Models are an effective IR approach which explicitly introduces the concept of relevance in the statistical Language Modelling framework of Information Retrieval. These models have shown to achieve state-of-the-art retrieval performance in the pseudo relevance feedback task. In this paper we propose a novel adaptation of this language modeling approach to rating-based Collaborative Filtering. In a memory-based approach, we apply the model to the formation of user neighbourhoods, and the generation of recommendations based on such neighbourhoods. We report experimental results where our method outperforms other standard memory-based algorithms in terms of ranking precision.

Spectral clustering techniques have become one of the most popular clustering algorithms, mainly because of their simplicity and effectiveness. In this work, we make use of one of these techniques, Normalised Cut, in order to derive a cluster-based collaborative filtering algorithm which outperforms other standard techniques in the state-of-the-art in terms of ranking precision. We frame this technique as a method for neighbour selection, and we show its effectiveness when compared with other cluster-based methods. Furthermore, the performance of our method could be improved if standard similarity metrics -- such as Pearson's correlation -- are also used when predicting the user's preferences.

Novelty detection is a difficult task, particularly at sentence level. Most of the approaches proposed in the past consist of re-ordering all sentences following their novelty scores. However, this re-ordering has usually little value. In fact, a naive baseline with no novelty detection capabilities yields often better performance than any state-of-the-art novelty detection mechanism. We argue here that this is because current methods initiate too early the novelty detection process. When few sentences have been seen, it is unlikely that the user is negatively affected by redundancy. Therefore, re-ordering the first sentences may be harmful in terms of performance. We propose here a query-dependent method based on cluster analysis to determine where we must start filtering redundancy.

In the last years Blog Search has been a new exciting task in Information Retrieval. The presence of user generated information with valuable opinions makes this field of huge interest. In this poster we use part of this information, the readers' comments, to improve the quality of post snippets with the objective of enhancing the user access to the relevant posts in a result list. We propose a simple method for snippet generation based on sentence selection, using the comments to guide the selection process. We evaluated our approach with standard TREC methodology in the Blogs06 collection showing significant improvements up to 32% in terms of MAP over the baseline.

The inclusion of document length factors has been a major topic in the development of retrieval models. We believe that current models can be further improved by more refined estimations of the document's scope. In this poster we present a new document length prior that uses the size of the compressed document. This new prior is introduced in the context of Language Modeling with Dirichlet smoothing. The evaluation performed on several collections shows significant improvements in effectiveness.

We present an approach to document clustering based on winnowing fingerprints that achieved good values of effectiveness with considerable save in memory space and computation time.