Pierre-Antoine Lequeu, Léo Labat, Laurène Cave, Gaël Lejeune, François Yvon, Benjamin Piwowarski. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

International audience

Transformer-based ranking models, such as MonoBERT, are central to Information Retrieval; yet their inner workings remain largely opaque. This hinders not only our understanding of the systems implementing them, but also our ability to improve them. To alleviate this limitation, we introduce IR Lens, a new interpretability tool tailored to cross-encoders based on two key components: 1) Neuron Integrated Gradients to expose the contributions of model parts at multiple levels, and 2) targeted ablations to support hypothesis tracking. With its interactive graphical interface, IR Lens enables IR practitioners to explore, analyze, and manipulate neuron-level mechanisms in cross-encoders, facilitating a deeper understanding of neural ranking models. By extending the reach of existing interpretability methods, we believe IR Lens has the potential to support the improvement of cross-encoders.

Late-interaction models such as ColBERT offer competitive performance across various retrieval tasks but require storing a dense embedding for each document token, leading to a substantial index storage overhead. Past works address this by attempting to prune low-importance token embeddings based on statistical and empirical measures, but they often either lack formal grounding or are ineffective. To address these shortcomings, we introduce a framework grounded in hyperspace geometry and cast token pruning as a Voronoi cell estimation problem in the embedding space. By interpreting each token's influence as a measure of its Voronoi region, our approach enables principled pruning that retains retrieval quality while reducing index size. Through our experiments, we demonstrate that this approach serves not only as a competitive pruning strategy but also as a valuable tool for improving and interpreting token-level behavior within dense retrieval systems.

Learned Sparse IR models, such as SPLADE, offer an excellent efficiency-effectiveness tradeoff. However, they rely on the underlying backbone vocabulary, which might hinder performance (polysemicity and synonymy) and pose a challenge for multi-lingual and multi-modal usages. To solve this limitation, we propose to replace the backbone vocabulary with a latent space of semantic concepts learned using Sparse Auto-Encoders (SAE). Throughout this paper, we study the compatibility of these 2 concepts, explore training approaches, and analyze the differences between our SAE-SPLADE model and traditional SPLADE models. Our experiments demonstrate that SAE-SPLADE achieves retrieval performance comparable to SPLADE on both in-domain and out-of-domain tasks while offering improved efficiency.

Cross-lingual sentence encoders (CLSE) create fixed-size sentence representations with aligned translations.Current pre-trained CLSE approaches use sentence-level objectives only.This can lead to loss of information, especially for tokens, which then degrades the sentence representation.We propose MEXMA, a novel approach that integrates both sentence-level and token-level objectives.The sentence representation in one language is used to predict masked tokens in another language, with both the sentence representation and all tokens directly updating the encoder.We show that adding token-level objectives greatly improves the sentence representation quality across several tasks.Our approach outperforms current pre-trained cross-lingual sentence encoders on bitext mining as well as several downstream tasks.We also analyse the information encoded in our tokens, and how the sentence representation is built from them.

João Maria Janeiro, Belen Alastruey, Francisco Massa, Maha Elbayad, Benjamin Piwowarski, Patrick Gallinari, Loic Barrault. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Current sparse neural information retrieval (IR) methods, and to a lesser extent more traditional models such as BM25, do not take into account the document collection and the complex interplay between different term weights when representing a single document. In this paper, we show how the Rational Speech Acts (RSA), a linguistics framework used to minimize the number of features to be communicated when identifying an object in a set, can be adapted to the IR case - and in particular to the high number of potential features (here, tokens). RSA dynamically modulates token-document interactions by considering the influence of other documents in the dataset, better contrasting document representations. Experiments show that incorporating RSA consistently improves multiple sparse retrieval models and achieves state-of-the-art performance on out-of-domain datasets from the BEIR benchmark.

Late interaction neural IR models like ColBERT offer a competitive effectiveness-efficiency trade-off across many benchmarks. However, they require a huge memory space to store the contextual representation for all the document tokens. Some works have proposed using either heuristics or statistical-based techniques to prune tokens from each document. This however doesn't guarantee that the removed tokens have no impact on the retrieval score. Our work uses a principled approach to define how to prune tokens without impacting the score between a document and a query. We introduce three regularization losses, that induce a solution with high pruning ratios, as well as two pruning strategies. We study them experimentally (in and out-domain), showing that we can preserve ColBERT's performance while using only 30% of the tokens.

Table Question-Answering involves both understanding the natural language query and grounding it in the context of the input table to extract relevant information.In this context, many methods have highlighted the benefits of intermediate pre-training using SQL queries.However, while most approaches aim at generating final answers directly from inputs, we claim that there is better to do with SQL queries during training.By learning to imitate a restricted subset of SQL-like algebraic operations, we demonstrate that their execution flow provides intermediate supervision steps that allow for increased generalization and structural reasoning compared to classical approaches.Our method, bridges the gap between semantic parsing and direct answering methods, offering valuable insights into which types of operations should be predicted by a generative architecture and which should be executed by an external algorithm.

Sparse representation learning based on Pre-trained Language Models has seen a growing interest in Information Retrieval. Such approaches can take advantage of the proven efficiency of inverted indexes and inherit desirable IR priors such as explicit lexical matching or some degree of interpretability. In this work, we thoroughly develop the framework of sparse representation learning in IR, which unifies term weighting and expansion in a supervised setting. We then build on SPLADE—a sparse expansion-based retriever—and show to which extent it is able to benefit from the same training improvements as dense bi-encoders by studying the effect of distillation, hard negative mining, as well as the Pre-trained Language Model’s initialization on itseffectiveness, leading to state-of-the-art results in both in- and out-of-domain evaluation settings (SPLADE++). We furthermore proposeefficiencyimprovements, allowing us to reach latency requirements on par with traditional keyword-based approaches (Efficient-SPLADE).

International audience

During past years, several frameworks for (Neural) Information Retrieval have been proposed. However, while they allow reproducing already published results, it is still very hard to re-use some parts of the learning pipelines, such as for instance the pre-training, sampling strategy, or a loss in newly developed models. It is also difficult to use new training techniques with old models, which makes it more difficult to assess the usefulness of ideas on various neural IR models. This slows the adoption of new techniques, and in turn, the development of the IR field. In this paper, we present XpmIR, a Python library defining a reusable set of experimental components. The library already contains state-of-the-art models and indexation techniques and is integrated with the HuggingFace hub.

Generative Adversarial Networks (GANs) have known a tremendous success for many continuous generation tasks, especially in the field of image generation. However, for discrete outputs such as language, optimizing GANs remains an open problem with many instabilities, as no gradient can be properly back-propagated from the discriminator output to the generator parameters. An alternative is to learn the generator network via reinforcement learning, using the discriminator signal as a reward, but such a technique suffers from moving rewards and vanishing gradient problems. Finally, it often falls short compared to direct maximum-likelihood approaches. In this paper, we introduce Generative Cooperative Networks, in which the discriminator architecture is cooperatively used along with the generation policy to output samples of realistic texts for the task at hand. We give theoretical guarantees of convergence for our approach, and study various efficient decoding schemes to empirically achieve state-of-the-art results in two main NLG tasks.

Language models generate texts by successively predicting probability distributions for next tokens given past ones. A growing field of interest tries to leverage external information in the decoding process so that the generated texts have desired properties, such as being more natural, non toxic, faithful, or having a specific writing style. A solution is to use a classifier at each generation step, resulting in a cooperative environment where the classifier guides the decoding of the language model distribution towards relevant texts for the task at hand. In this paper, we examine three families of (transformer-based) discriminators for this specific task of cooperative decoding: bidirectional, left-to-right and generative ones. We evaluate the pros and cons of these different types of discriminators for cooperative generation, exploring respective accuracy on classification tasks along with their impact on the resulting sample quality and computational performances. We also provide the code of a batched implementation of the powerful cooperative decoding strategy used for our experiments, the Monte Carlo Tree Search, working with each discriminator for Natural Language Generation.

Neural retrievers based on dense representations combined with Approximate Nearest Neighbors search have recently received a lot of attention, owing their success to distillation and/or better sampling of examples for training -- while still relying on the same backbone architecture. In the meantime, sparse representation learning fueled by traditional inverted indexing techniques has seen a growing interest, inheriting from desirable IR priors such as explicit lexical matching. While some architectural variants have been proposed, a lesser effort has been put in the training of such models. In this work, we build on SPLADE -- a sparse expansion-based retriever -- and show to which extent it is able to benefit from the same training improvements as dense models, by studying the effect of distillation, hard-negative mining as well as the Pre-trained Language Model initialization. We furthermore study the link between effectiveness and efficiency, on in-domain and zero-shot settings, leading to state-of-the-art results in both scenarios for sufficiently expressive models.

When conducting a search task, users may find it difficult to articulate their need, even more so when the task is complex. To help them complete their search, search engine usually provide query suggestions. A good query suggestion system requires to model user behavior during the search session. In this article, we study multiple Transformer architectures applied to the query suggestion task and compare them with recurrent neural network (RNN)-based models. We experiment Transformer models with different tokenizers, with different Encoders (large pretrained models or fully trained ones), and with two kinds of architectures (flat or hierarchic). We study the performance and the behaviors of these various models, and observe that Transformer-based models outperform RNN-based ones. We show that while the hierarchical architectures exhibit very good performances for query suggestion, the flat models are more suitable for complex and long search tasks. Finally, we investigate the flat models behavior and demonstrate that they indeed learn to recover the hierarchy of a search session.

Clement Rebuffel, Thomas Scialom, Laure Soulier, Benjamin Piwowarski, Sylvain Lamprier, Jacopo Staiano, Geoffrey Scoutheeten, Patrick Gallinari. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.

Thomas Scialom, Paul-Alexis Dray, Sylvain Lamprier, Benjamin Piwowarski, Jacopo Staiano, Alex Wang, Patrick Gallinari. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.

Due to the discrete nature of words, language GANs require to be optimized from rewards provided by discriminator networks, via reinforcement learning methods. This is a much harder setting than for continuous tasks, which enjoy gradient flows from discriminators to generators, usually leading to dramatic learning instabilities. However, we claim that this can be solved by making discriminator and generator networks cooperate to produce output sequences during training. These cooperative outputs, inherently built to obtain higher discrimination scores, not only provide denser rewards for training but also form a more compact artificial set for discriminator training, hence improving its accuracy and stability.In this paper, we show that our SelfGAN framework, built on this cooperative principle, outperforms Teacher Forcing and obtains state-of-the-art results on two challenging tasks, Summarization and Question Generation.

In neural Information Retrieval, ongoing research is directed towards improving the first retriever in ranking pipelines. Learning dense embeddings to conduct retrieval using efficient approximate nearest neighbors methods has proven to work well. Meanwhile, there has been a growing interest in learning sparse representations for documents and queries, that could inherit from the desirable properties of bag-of-words models such as the exact matching of terms and the efficiency of inverted indexes. In this work, we present a new first-stage ranker based on explicit sparsity regularization and a log-saturation effect on term weights, leading to highly sparse representations and competitive results with respect to state-of-the-art dense and sparse methods. Our approach is simple, trained end-to-end in a single stage. We also explore the trade-off between effectiveness and efficiency, by controlling the contribution of the sparsity regularization.

We present MLSUM, the first large-scale Mul-tiLingual SUMmarization dataset.Obtained from online newspapers, it contains 1.5M+ article/summary pairs in five different languages -namely, French, German, Spanish, Russian, Turkish.Together with English news articles from the popular CNN/Daily mail dataset, the collected data form a large scale multilingual dataset which can enable new research directions for the text summarization community.We report cross-lingual comparative analyses based on state-of-the-art systems.These highlight existing biases which motivate the use of a multi-lingual dataset.

We introduce a novel approach for sequence decoding, Discriminative Adversarial Search (DAS), which has the desirable properties of alleviating the effects of exposure bias without requiring external metrics. Inspired by Generative Adversarial Networks (GANs), wherein a discriminator is used to improve the generator, our method differs from GANs in that the generator parameters are not updated at training time and the discriminator is used to drive sequence generation at inference time. We investigate the effectiveness of the proposed approach on the task of Abstractive Summarization: the results obtained show that a naive application of DAS improves over the state-of-the-art methods, with further gains obtained via discriminator retraining. Moreover, we show how DAS can be effective for cross-domain adaptation. Finally, all results reported are obtained without additional rule-based filtering strategies, commonly used by the best performing systems available: this indicates that DAS can effectively be deployed without relying on post-hoc modifications of the generated outputs.

Training regimes based on Maximum Likelihood Estimation (MLE) suffer from known limitations, often leading to poorly generated text sequences that lack of coherence, factualness, and are prone to repetitions. At the root of these limitations is the mismatch between training and inference, i.e. the so-called exposure bias. Another problem lies in considering only the reference text as correct, while in practice several alternative formulations could be as good. Generative Adversarial Networks (GANs) could mitigate those limitations. Nonetheless, the discrete nature of text has hindered their application to language generation: the approaches proposed so far, based on Reinforcement Learning, have been shown to under-perform MLE. In this context, the exploration is known to be critical, while surprisingly being under-studied. In this work, we show how the most popular sampling method results in unstable training for language GANs. We propose alternative exploration strategies that we named Cold-GANs. By forcing the sampling to be close to the distribution mode, the learning dynamic becomes smoother. We report experimental results obtained on three tasks: unconditional text generation, question generation, and abstractive summarization. For the first time, to the best of our knowledge, the proposed language GANs compare favorably to MLE, and obtain improvements over the state-of-the-art on the considered tasks.

Abstract Ensuring reproducibility is key to all scientific domains. As Information Retrieval (IR) experiments are often composed of several steps that can be shared between tested models, and rely on various resources, it is difficult to keep track of all the experimental settings and to ensure experiments can be reproduced easily. In this demo paper, we present two managers, Experimaestro and Datamaestro, and their add-ons for IR, designed to help to define and run experimental plans.

International audience

Unsupervised relation extraction aims at extracting relations between entities in text.Previous unsupervised approaches are either generative or discriminative.In a supervised setting, discriminative approaches, such as deep neural network classifiers, have demonstrated substantial improvement.However, these models are hard to train without supervision, and the currently proposed solutions are unstable.To overcome this limitation, we introduce a skewness loss which encourages the classifier to predict a relation with confidence given a sentence, and a distribution distance loss enforcing that all relations are predicted in average.These losses improve the performance of discriminative based models, and enable us to train deep neural networks satisfactorily, surpassing current state of the art on three different datasets.

Patrick Bordes, Eloi Zablocki, Laure Soulier, Benjamin Piwowarski, Patrick Gallinari. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Thomas Scialom, Sylvain Lamprier, Benjamin Piwowarski, Jacopo Staiano. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Zero-Shot Learning (ZSL) aims at classifying unlabeled objects by leveraging auxiliary knowledge, such as semantic representations. A limitation of previous approaches is that only intrinsic properties of objects, e.g. their visual appearance, are taken into account while their context, e.g. the surrounding objects in the image, is ignored. Following the intuitive principle that objects tend to be found in certain contexts but not others, we propose a new and challenging approach, context-aware ZSL, that leverages semantic representations in a new way to model the conditional likelihood of an object to appear in a given context. Finally, through extensive experiments conducted on Visual Genome, we show that contextual information can substantially improve the standard ZSL approach and is robust to unbalanced classes.

Representing the semantics of words is a long-standing problem for the natural language processing community. Most methods compute word semantics given their textual context in large corpora. More recently, researchers attempted to integrate perceptual and visual features. Most of these works consider the visual appearance of objects to enhance word representations but they ignore the visual environment and context in which objects appear. We propose to unify text-based techniques with vision-based techniques by simultaneously leveraging textual and visual context to learn multimodal word embeddings. We explore various choices for what can serve as a visual context and present an end-to-end method to integrate visual context elements in a multimodal skip-gram model. We provide experiments and extensive analysis of the obtained results.

We address the task of node classification in heterogeneous networks, where the nodes are of different types, each type having its own set of labels, and the relations between nodes may also be of different types. A typical example is provided by social networks where node types may for example be users, content, or films, and relations friendship , like , authorship . Learning and performing inference on such heterogeneous networks is a recent task requiring new models and algorithms. We propose a model, Labeling Heterogeneous Network (LaHNet) , a transductive approach to classification that learns to project the different types of nodes into a common latent space. This embedding is learned so as to reflect different characteristics of the problem such as the correlation between node labels, as well as the graph topology. The application focus is on social graphs, but the algorithm is general and can be used for other domains. The model is evaluated on five datasets representative of different instances of social data.

Most collaborative filtering systems, such as matrix factorization, use vector representations for items and users. Those representations are deterministic, and do not allow modeling the uncertainty of the learned representation, which can be useful when a user has a small number of rated items (cold start), or when there is conflicting information about the behavior of a user or the ratings of an item. In this paper, we leverage recent works in learning Gaussian embeddings for the recommendation task. We show that this model performs well on three representative collections (Yahoo, Yelp and MovieLens) and analyze learned representations.

User social networks are a useful information for many information access related tasks, such as recommendation or information retrieval. In such tasks, recent papers have exploited the polarity of these links (friend/enemy) by capturing more precisely social patterns. This negative information being relatively scarce, a recent work proposed to infer it in social networks that contain none. However, this work relies on the direct interaction between users. In this paper, we pursue this approach under the assumption that we do not have access to this kind of data neither, thus allowing to cope with most social networks, where users can rate items and have friendship relationships. We exploit the user ratings polarity, i.e the fact that a rating can be positive (like) or negative (dislike), to infer negative ties. Experiments on the Epinions dataset show the potential of our approach.

Multidocument summarization (MDS) aims for each given query to extract compressed and relevant information with respect to the different query‐related themes present in a set of documents. Many approaches operate in two steps. Themes are first identified from the set, and then a summary is formed by extracting salient sentences within the different documents of each of the identified themes. Among these approaches, latent semantic analysis (LSA) based approaches rely on spectral decomposition techniques to identify the themes. In this article, we propose a major extension of these techniques that relies on the quantum information access (QIA) framework. The latter is a framework developed for modeling information access based on the probabilistic formalism of quantum physics. The QIA framework not only points out the limitations of the current LSA‐based approaches, but motivates a new principled criterium to tackle multidocument summarization that addresses these limitations. As a byproduct, it also provides a way to enhance the LSA‐based approaches. Extensive experiments on the DUC 2005, 2006 and 2007 datasets show that the proposed approach consistently improves over both the LSA‐based approaches and the systems that competed in the yearly DUC competitions. This demonstrates the potential impact of quantum‐inspired approaches to information access in general, and of the QIA framework in particular.

Collaborative filtering (CF) aims to recommend items based on prior user interaction. Despite their success, CF techniques do not handle data sparsity well, especially in the case of the cold start problem where there is no past rating for an item. In this paper, we provide a framework, which is able to tackle such issues by considering item-related emotions and semantic data. In order to predict the rating of an item for a given user, this framework relies on an extension of Latent Dirichlet Allocation, and on gradient boosted trees for the final prediction. We apply this framework to movie recommendation and consider two emotion spaces extracted from the movie plot summary and the reviews, and three semantic spaces: actor, director, and genre. Experiments with the 100K and 1M MovieLens datasets show that including emotion and semantic information significantly improves the accuracy of prediction and improves upon the state-of-the-art CF techniques. We also analyse the importance of each feature space and describe some uncovered latent groups.

The probabilistic formalism of quantum physics is said to provide a sound basis for building a principled information retrieval framework. Such a framework can be based on the notion of information need vector spaces where events, such as document relevance or observed user interactions, correspond to subspaces. As in quantum theory, a probability distribution over these subspaces is defined through weighted sets of state vectors (density operators), and used to represent the current view of the retrieval system on the user information need. Tensor spaces can be used to capture different aspects of information needs. Our evaluation shows that the framework can lead to acceptable performance in an ad-hoc retrieval task. Going beyond this, we discuss the potential of the framework for three active challenges in information retrieval, namely, interaction, novelty and diversity.

We explore a set of hypothesis on user behavior that are potentially at the origin of the (Mean) Average Precision (AP) metric. This allows us to propose a more realistic version of AP where users click non-deterministically on relevant documents and where the number of relevant documents in the collection needs not be known in advance. We then depart from the assumption that a document is either relevant or irrelevant and we use instead relevance judgment similar to editorial labels used for Discounted Cumulated Gain (DCG). We assume that clicked documents provide users with a certain level of "utility" and that a user ends a search when she gathered enough utility. Based on the query logs of a commercial search engine we show how to evaluate the utility associated with a label from the record of past user interactions with the search engine and we show how the two different user models can be evaluated based on their ability to predict accurately future clicks. Finally, based on these user models, we propose a measure that captures the relative quality of two rankings.

Mining user web search activity potentially has a broad range of applications including web result pre-fetching, automatic search query reformulation, click spam detection, estimation of document relevance and prediction of user satisfaction. This analysis is difficult because the data recorded by search engines while users interact with them, although abundant, is very noisy. In this work, we explore the utility of mining search behavior of users, represented by observed variables including the time the user spends on the page, and whether the user reformulated his or her query. As a case study, we examine the contribution this data makes to predicting the relevance of a document in the absence of document content models. To this end, we first propose a method for grouping the interactions of a particular user according to the different tasks he or she undertakes. With each task corresponding to a distinct information need, we then propose a Bayesian Network to holistically model these interactions. The aim is to identify distinct patterns of search behaviors. Finally, we join these patterns to a list of custom features and we use gradient boosted decision trees to predict the relevance of a set of query document pairs for which we have relevance assessments. The experimental results confirm the potential of our model, with significant improvements in precision for predicting the relevance of documents based on a model of the user's search and click behavior, over a baseline model using only click and query features, with no Bayesian Network input.

Search engine click logs provide an invaluable source of relevance information but this information is biased because we ignore which documents from the result list the users have actually seen before and after they clicked. Otherwise, we could estimate document relevance by simple counting. In this paper, we propose a set of assumptions on user browsing behavior that allows the estimation of the probability that a document is seen, thereby providing an unbiased estimate of document relevance. To train, test and compare our model to the best alternatives described in the Literature, we gather a large set of real data and proceed to an extensive cross-validation experiment. Our solution outperforms very significantly all previous models. As a side effect, we gain insight into the browsing behavior of users and we can compare it to the conclusions of an eye-tracking experiments by Joachims et al. [12]. In particular, our findings confirm that a user almost always see the document directly after a clicked document. They also explain why documents situated just after a very relevant document are clicked more often.

In information retrieval research, comparing retrieval approaches requires test collections consisting of documents, user requests and relevance assessments. Obtaining relevance assessments that are as sound and complete as possible is crucial for the comparison of retrieval approaches. In XML retrieval, the problem of obtaining sound and complete relevance assessments is further complicated by the structural relationships between retrieval results. A major difference between XML retrieval and flat document retrieval is that the relevance of elements (the retrievable units) is not independent of that of related elements. This has major consequences for the gathering of relevance assessments. This article describes investigations into the creation of sound and complete relevance assessments for the evaluation of content-oriented XML retrieval as carried out at INEX, the evaluation campaign for XML retrieval. The campaign, now in its seventh year, has had three substantially different approaches to gather assessments and has finally settled on a highlighting method for marking relevant passages within documents—even though the objective is to collect assessments at element level. The different methods of gathering assessments at INEX are discussed and contrasted. The highlighting method is shown to be the most reliable of the methods.

Web search engines consistently collect information about users interaction with the system: they record the query they issued, the URL of presented and selected documents along with their ranking. This information is very valuable: It is a poll over millions of users on the most various topics and it has been used in many ways to mine users interests and preferences. Query logs have the potential to partially alleviate the search engines from thousand of searches by providing a way to predict answers for a subset of queries and users without knowing the content of a document. Even if the predicted result is at rank one, this analysis might be of interest: If there is enough confidence on a user's click, we might redirect the user directly to the page whose link would be clicked. In this paper, we present three different models for predicting user clicks, ranging from most specific ones (using only past user history for the query) to very general ones (aggregating data over all users for a given query). The former model has a very high precision at low recall values, while the latter can achieve high recalls. We show that it is possible to combine the different models to predict with high accuracy (over 90%) a high subset of query sessions (24% of all the sessions).

Standard Information Retrieval (IR) metrics are not well suited for new paradigms like XML or Web IR in which retrievable information units are document elements and/or sets of related documents. Part of the problem stems from the classical hypotheses on the user models: They do not take into account the structural or logical context of document elements or the possibility of navigation between units. This article proposes an explicit and formal user model that encompasses a large variety of user behaviors. Based on this model, we extend the probabilistic precision-recall metric to deal with the new IR paradigms.

Standard Information Retrieval (IR) metrics assume a simple model where documents are understood as independent units. Such an assumption is not adapted to new paradigms like XML or Web IR where retrievable informations are parts of documents or sets of related documents. Moreover, classical hypotheses assumes that the user ignores the structural or logical context of document elements and hence the possibility of navigation between units. EPRUM is a generalisation of Precision-Recall (PR) that aims at allowing the user to navigate or browse in the corpus structure. Like the Cumulated Gain metrics, it is able to handle continuous valued relevance. We apply and compare EPRUM in the context of XML Retrieval -- a very active field for evaluation metrics. We also explain how EPRUM can be used in other IR paradigms.

Comparing retrieval approaches requires test collections, which consist of documents, queries and relevance assessments. Obtaining consistent and exhaustive relevance assessments is crucial for the appropriate comparison of retrieval approaches. Whereas the evaluation methodology for flat text retrieval approaches is well established, the evaluation of XML retrieval approaches is a research issue. This is because XML documents are composed of nested components that cannot be considered independent in terms of relevance. This paper describes the methodology adopted in INEX (the INitiative for the Evaluation of XML Retrieval) to ensure consistent and exhaustive relevance assessments.