Conversational information retrieval is challenging since it requires the consideration of the conversation history which potentially gives rise to topic shifts and coreference resolution across previous turns. To address these challenges, previous work mainly rely on traditional fine-tuning of ad-hoc retrievers on conversational datasets or extrapolates their generalizability through multi-tasking. However, this mainstream approach is costly—since it requires model re-training—and exhibits catastrophic forgetting, where the model loses its foundational ad-hoc retrieval performance. In this paper, we fill this gap by introducing model merging as a training-free strategy enabling the design of a single retrieval model that operates across both ad-hoc and conversational settings with no additional fine-tuning. We conduct experiments using linear and non-linear parameter-wise merging strategies—namely Model Soup and Slerp—on standard ad-hoc search and conversational retrieval datasets. Our results demonstrate that model merging significantly enhances the ad-hoc search capabilities of conversational retrievers while improving generalizability across task-specific datasets, achieving up to 15% higher NDCG@3 under zero-shot conditions.

Electronic Health Records (EHRs) provide crucial information for clinical decision-making. However, their high-dimensionality, heterogeneity, and sparsity make clinical prediction challenging. Large Language Models (LLMs) allowed progress towards addressing this challenge by leveraging parametric medical knowledge to enhance EHR data for clinical prediction tasks. Despite the significant achievements made so far, most of the existing approaches are fundamentally task-agnostic in the sense that they deploy LLMs as EHR encoders or EHR completion modules without fully integrating signals from the prediction tasks. This naturally hinders task performance accuracy. In this work, we propose Rewrite-To-Predict (ReToP), an LLM-based framework that addresses this limitation through an end-to-end training of an EHR rewriter and a clinical predictor. To cope with the lack of EHR rewrite training data, we generate synthetic pseudo-labels using clinical-driven feature selection strategies to create diverse patient rewrites for fine-tuning the EHR rewriter. ReToP aligns the rewriter with prediction objectives using a novel Classifier Supervised Contribution (CSC) score that enables the EHR rewriter to generate clinically relevant rewrites that directly enhance prediction. Our ReToP framework surpasses strong baseline models across three clinical tasks on MIMIC-IV. Moreover, the analysis of ReToP shows its generalizability to unseen datasets and tasks with minimal fine-tuning while preserving faithful rewrites and emphasizing task-relevant predictive features.

Whether to retrieve, answer, translate, or reason, multimodality opens up new challenges and perspectives. In this context, we are interested in answering questions about named entities grounded in a visual context using a Knowledge Base (KB). To benchmark this task, called KVQAE (Knowledge-based Visual Question Answering about named Entities), we provide ViQuAE, a dataset of 3.7K questions paired with images. This is the first KVQAE dataset to cover a wide range of entity types (e.g. persons, landmarks, and products). The dataset is annotated using a semi-automatic method. We also propose a KB composed of 1.5M Wikipedia articles paired with images. To set a baseline on the benchmark, we address KVQAE as a two-stage problem: Information Retrieval and Reading Comprehension, with both zero- and few-shot learning methods. The experiments empirically demonstrate the difficulty of the task, especially when questions are not about persons. This work paves the way for better multimodal entity representations and question answering. The dataset, KB, code, and semi-automatic annotation pipeline are freely available at https://github.com/PaulLerner/ViQuAE.

In this demo paper, we present a new open-source python module for building information retrieval pipelines with transformers namely CHERCHE. Our aim is to propose an easy to plug tool capable to execute, simple but strong, state-of-the-art information retrieval models. To do so, we have integrated classical models based on lexical matching but also recent models based on semantic matching. Indeed, a large number of models available on public hubs can be now tested on information retrieval tasks with only a few lines. CHERCHE is oriented to newcomers into the neural information retrieval field that want to use transformer-based models in small collections without struggling with heavy tools. The code and documentation of CHERCHE is public available at https://github.com/raphaelsty/cherche

While great progress is made in the area of information access, there are still open issues that involve designing intelligent systems supporting task-based search. Despite the importance of task-based search, the information retrieval and information science communities still feel the lack of open-ended and annotated datasets that enable the evaluation of a number of related facets of search tasks in downstream applications. Existing datasets are either sampled from large-scale logs but provide poor annotations, or sampled from lower-scale user studies but focus on ranked list evaluation. In this work, we present CoST: a novel richly annotated dataset for evaluating complex search tasks, collaboratively designed by researchers from the computer science and cognitive psychology domains, and intended to answer a wide range of research questions dealing with task-based search. CoST includes 5667 queries recorded in 630 task-based sessions that result from a user study involving 70 french native participants who are expert in one among 3 different domains of expertise (computer science, medicine, psychology). Each participant completed 15 tasks with 5 different types of cognitive complexity (fact-finding, exploratory learning, decision-making, problem-solving, multicriteria-inferential). In addition to search data (e.g., queries and clicks), CoST provides task and session-related data, task annotations and query annotations. We illustrate possible usages of CoST through the evaluation of query classification models and the understanding of the effect of task complexity and domain on user's search behavior.

Named entity processing over historical texts is more and more being used due to the massive documents and archives being stored in digital libraries. However, due to the poor annotated resources of historical nature, information extraction performances fall behind those on contemporary texts. In this paper, we introduce the development of the NewsEye resource, a multilingual dataset for named entity recognition and linking enriched with stances towards named entities. The dataset is comprised of diachronic historical newspaper material published between 1850 and 1950 in French, German, Finnish, and Swedish. Such historical resource is essential in the context of developing and evaluating named entity processing systems. It evenly allows enhancing the performances of existing approaches on historical documents which enables adequate and efficient semantic indexing of historical documents on digital cultural heritage collections.

The Information Retrieval (IR) community has witnessed a flourishing development of deep neural networks, however, only a few managed to beat strong baselines. Among them, models like DRMM and DUET were able to achieve better results thanks to the proper handling of exact match signals. Nowadays, the application of pre-trained language models to IR tasks has achieved impressive results exceeding all previous work. In this paper, we assume that established IR cues like exact term-matching, proven to be valuable for deep neural models, can be used to augment the direct supervision from labeled data for training these pre-trained models. To study the effectiveness of this assumption, we propose MarkedBERT a modified version of one of the most popular pre-trained models via language modeling tasks, BERT. MarkedBERT integrates exact match signals using a marking technique that locates and highlights Exact Matched query-document terms using marker tokens. Experiments on MS MARCO Passage Ranking task show that our rather simple approach is actually effective. We find that augmenting the input with marker tokens allows the model to focus on valuable text sequences for IR.

To fulfill their information needs, users submit sets of related queries to available search engines. Query logs record users' activities along with timestamps and additional search-related information. The analysis of those chronological query logs enables the modeling of search tasks from user interactions. Previous research works rely on clicked URLs and surrounding queries to determine if adjacent queries are part of the same search tasks to segment the query logs properly. However, waiting for clicked URLs or future adjacent queries could render the use of these methods unfeasible in user supporting applications that require model results on the fly. Therefore, we propose a model for sequential search log segmentation. The proposed model uses only query pairs and their time span, generating results suited for on the fly user supporting applications, with improved accuracy over existing search segmentation approaches. We also show the advantages of fine-tuning the proposed model for adjusting the architecture to a small annotated collection.

Users convert their information needs to search queries, which are then run on available search engines. Query logs registered by search engines enable the automatic identification of the search tasks that users perform to fulfill their information needs. Search engine logs contain queries in multiple languages, but most existing methods for search task identification are not multilingual. Some methods rely on search context training of custom embeddings or external indexed collections that support a single language, making it challenging to support the multiple languages of queries run in search engines. Other methods depend on supervised components and user identifiers to model search tasks. The supervised components require labeled collections, which are difficult and costly to get in multiple languages. Also, the need for user identifiers renders these methods unfeasible in user agnostic scenarios. Hence, we propose an unsupervised multilingual approach for search task identification. The proposed approach is user agnostic, enabling its use in both user-independent and personalized scenarios. Furthermore, the multilingual query representation enables us to address the existing trade-off when mapping new queries to the identified search tasks.

Clustering is a central task in unsupervised learning. Recent advances that perform clustering into learned deep features (such as DEC[14], IDEC [6] or VaDe [10]) have shown improvements over classical algorithms, but most of them are based on the Euclidean distance. Moreover, symmetry-based distances have shown to be a powerful tool to distinguish symmetric shapes -- such as circles, ellipses, squares, etc. This paper presents an adaptation of symmetry-based distances into deep clustering algorithms, named SymDEC. Our results show that the proposed strategy outperforms significantly the existing Euclidean-based deep clustering as well as recent symmetry-based algorithms in several of the synthetic symmetric and UCI studied datasets.

B-CUBED metrics have recently been adopted in the evaluation of clustering results as well as in many other related tasks. However, this family of metrics is not well adapted when datasets are unbalanced. This issue is extremely frequent in Web results, where classes are distributed following a strong unbalanced pattern. In this paper, we present a modified version of B-CUBED metrics to overcome this situation. Results in toy and real datasets indicate that the proposed adaptation correctly considers the particularities of unbalanced cases.

Different important studies in Web search results clustering have recently shown increasing performances motivated by the use of external resources. Following this trend, we present a new algorithm called Dual C-Means, which provides a theoretical background for clustering in different representation spaces. Its originality relies on the fact that external resources can drive the clustering process as well as the labeling task in a single step. To validate our hypotheses, a series of experiments are conducted over different standard datasets and in particular over a new dataset built from the TREC Web Track 2012 to take into account query logs information. The comprehensive empirical evaluation of the proposed approach demonstrates its significant advantages over traditional clustering and labeling techniques.