Search

International audience; The current rise of Open Science and Reproducibility in the Life Sciences requires the creation of rich, machine-actionable metadata in order to better share and reuse biological digital resources such as datasets, bioinformatics tools, training materials, etc. For this purpose, FAIR principles have been defined for both data and metadata and adopted by large communities, leading to the definition of specific metrics. However, automatic FAIRness assessment is still difficult because computational evaluations frequently require technical expertise and can be time-consuming. As a first step to address these issues, we propose FAIR-Checker , a web-based tool to assess the FAIRness of metadata presented by digital resources. FAIR-Checker offers two main facets: a “Check” module providing a thorough metadata evaluation and recommendations, and an “Inspect” module which assists users in improving metadata quality and therefore the FAIRness of their resource. FAIR-Checker leverages Semantic Web standards and technologies such as SPARQL queries and SHACL constraints to automatically assess FAIR metrics. Users are notified of missing, necessary, or recommended metadata for various resource categories. We evaluate FAIR-Checker in the context of improving the FAIRification of individual resources, through better metadata, as well as analyzing the FAIRness of more than 25 thousand bioinformatics software descriptions.

International audience; Discrete-time discrete-state finite Markov chains are versatile mathematical models for a wide range of real-life stochastic processes. One of most common tasks in studies of Markov chains is computation of the stationary distribution. Without loss of generality, and drawing our motivation from applications to large networks, we interpret this problem as one of computing the stationary distribution of a random walk on a graph. We propose a new controlled, easily distributed algorithm for this task, briefly summarized as follows: at the beginning, each node receives a fixed amount of cash (positive or negative), and at each iteration, some nodes receive 'green light' to distribute their wealth or debt proportionally to the transition probabilities of the Markov chain; the stationary probability of a node is computed as a ratio of the cash distributed by this node to the total cash distributed by all nodes together. Our method includes as special cases a wide range of known, very different, and previously disconnected methods including power iterations, Gauss-Southwell, and online distributed algorithms. We prove exponential convergence of our method, demonstrate its high efficiency, and derive scheduling strategies for the green-light, that achieve convergence rate faster than state-of-the-art algorithms.

International audience; This paper presents the construction of a Knowledge Graph (KG) of Educational Resources (ER), where RDF reification is essential. The ERs are described based on the subjects they cover considering their relevance. RDF reification is used to incorporate this subject’s relevance. Multiple reification models with distinct syntax and performance implications for storage and query processing exist. This study aims to experimentally compare four statement-based reification models with four triplestores to determine the most pertinent choice for our KG. We built four versions of the KG. Each version has a distinct reification model, namely standard reification, singleton properties, named graphs, and RDF-star, which were obtained using RML mappings. Each of the four triplestores (Virtuoso, Jena, Oxigraph, and GraphDB) was setup four times (except for Virtuoso, which does not support RDF-star), and seven different SPARQL queries were experimentally evaluated. This study shows that standard reification and named graphs lead to good performance. It also shows that, in the particular context of the used KG, Virtuoso outperforms Jena, GraphDB, and Oxigraph in most queries. The recent specification of RDF-star and SPARQL-star sheds light on statement-level annotations. The empirical study reported in this paper contributes to the efforts towards the efficient usage of RDF reification. In addition, this paper shares the pipeline of the KG con- struction using standard semantic web technologies.

International audience; We consider the problem of automatically inferring the (LaTeX) document class used to write a scientific article from its PDF representation. Applications include improving the performance of information extraction techniques that rely on the style used in each document class, or determining the publisher of a given scientific article. We introduce two approaches: a simple classifier based on hand-coded document style features, as well as a CNN-based classifier taking as input the bitmap representation of the first page of the PDF article. We experiment on a dataset of around 100k articles from arXiv, where labels come from the source LaTeX document associated to each article. Results show the CNN approach significantly outperforms that based on simple document style features, reaching over 90% average F1-score on a task to distinguish among several dozens of the most common document classes.

International audience; BACKGROUND: Medical diagnostic support systems are important tools in the field of radiology. However, the precision obtained, during the exploitation of high homogeneity image datasets, needs to be improved. OBJECTIVE: To develop a new learning system dedicated to public health practitioners. This study presents an upgraded version dedicated to radiology experts for better clinical decision-making when diagnosing and treating the patient (CAD approach). METHODS: Our system is a hybrid approach based on a matching of semantic and visual attributes of images. It is a combination of two complementary subsystems to form the intermodal system. The first one named α based on semantic attributes. Indexing and image retrieval based on specific key words. The second system named β based on low-level attributes. Vectors characterizing the digital content of the image (color, texture and shape) represent images. Our image database consists of 930 X-ray images including 320 mammograms acquired from the mini-MIAS database of mammograms and 610 X-rays acquired from the Public Hospital Establishment (EPH-Rouiba Algeria). The combination of two subsystems gives rise to the intermodal system: α-subsystem offers an overall result (based on visual descriptors), then β-subsystem (low level descriptors) refines the result and increases relevance. RESULTS: Our system can perform a specific image search (in a database of images with very high homogeneity) with an accuracy of around 90% for a recall of 25% . The average (overall) accuracy of the system exceeds 70% . CONCLUSION: The results obtained are very encouraging, and demonstrate efficiency of our approach, particularly for the intermodal system.

International audience; Lock-in and path-dependency are well-known concepts in economics dealing with unbalanced development of alternative options. Lock-in was studied in various sectors, considering production or consumption sides. Lock-in in academic research went little addressed. Yet, science develops through knowledge accumulation and cross-fertilisation of research topics, that could lead to similar phenomena when some topics do not sufficiently benefit from accumulation mechanisms, reducing innovation opportunities from the concerned field consequently. We introduce an original method to explore these phenomena by comparing topic trajectories in research fields according to strong or weak accumulative processes over time. We combine the concepts of ‘niche’ and ‘mainstream’ from transition studies with scientometric tools to revisit Callon’s strategic diagram with a diachronic perspective of topic clusters over time. Considering the trajectories of semantic clusters, derived from titles and authors’ keywords extracted from scholarly publications in the Web of Science, we applied our method to two competing research fields in food sciences and technology related to pulses and soya over the last 60 years worldwide. These highly interesting species for the sustainability of agrifood systems experienced unbalanced development and thus is under-debated. Our analysis confirms that food research for soya was more dynamic than for pulses: soya topic clusters revealed a stronger accumulative research path by cumulating mainstream positions while pulses research did not meet the same success. This attempt to unpack research lock-in for evaluating the competition dynamics of scientific fields over time calls for future works, by strengthening the method and testing it on other research fields.

International audience; K-anonymity is the most widely used approach to privacy preserving microdata which is mainly based on generalization. Although generalization-based k-anonymity approaches can achieve the privacy protection objective, they suffer from information loss. Clustering-based approaches have been successfully adapted for k-anonymization as they enhance the data quality, however, the computational complexity of finding an optimal solution has shown as NP-hard. Nature-inspired optimization algorithms are effective in finding solutions to complex problems. We propose, in this paper, a novel algorithm based on a simple nature-inspired metaheuristic called Black Hole Algorithm (BHA), to address such limitations. Experiments on real data set show that data utility has been improved by our approach compared to k-anonymity, BHA-based k-anonymity and clustering-based k-anonymity approaches.

A major drawback of nonlinear dimensionality reduction (DR) techniques is their inability to preserve some authentic information from the source domain, leading to projections that are often hard to interpret when it comes to observing anything other than the topological structure of the data. In this paper, we propose a nonlinear DR approach enforcing projection constraints resulting from an a priori knowledge about the structure of the data in multivariate periodical time series. We then propose several ways of exploiting this constrained projection to extract user-relevant information, such as the nominal behavior of a periodical dynamical system or the deviant behaviors which may occur at different time scales. The techniques are demonstrated on both a synthetic dataset composed of simulated multivariate data exhibiting a periodical behavior, and a real dataset corresponding to six months of sensor data acquisitions and recordings inside experimental buildings. 1 We would like to thank the Institut National de l'Energie Solaire (INES) and the CEA, LITEN, Laboratoire Energetique du Bâtiment for providing the data.

The work presented in this "Habilitation à Diriger des Recherches" (Accreditation to Supervise Research) focuses on the analysis and indexing of scientific texts and lies at the intersection of two research themes: Natural Language Processing (NLP), which involves the analysis, understanding, and generation of natural language, and Information Retrieval (IR), which studies ways to retrieve information from a collection of documents. We are interested in the question of scholarly document retrieval, which involves searching for documents in the scientific literature (e.g., articles, books, theses) related to a specific subject of study. More specifically, our research aims to enhance the metadata associated with documents to improve their accessibility and dissemination.Our work focuses on the development of automated methods for keyword generation, which are characterized by the unique utilization of graph-based techniques and node ranking algorithms. We delve into the issue of indirectly evaluating automatically generated keywords through application-specific tasks and their utilization in search engines and academic recommendation systems. We present our efforts into constructing linguistic resources, developing software tools, and their dissemination within the scientific community. Finally, we conclude with some prospective insights into keyword indexing and, more broadly, the emerging research at the intersection of NLP and IR themes.; Les travaux présentés dans cette habilitation à diriger des recherches (HDR) ont pour objet l'analyse et l'indexation des textes scientifiques, et se situent à la croisée de deux thématiques de recherche : celle du Traitement Automatique des Langues (TAL) qui concerne l'analyse, la compréhension et la production de langage naturel, et celle de la Recherche d'Information (RI) qui étudie la manière de retrouver des informations dans une collection de documents. Nous nous intéressons à la problématique de la recherche bibliographique, c'est-à-dire la recherche de documents dans la littérature scientifique (e.g. articles, ouvrages, thèses) en rapport avec un sujet d'étude, et plus particulièrement à l'enrichissement des métadonnées associées aux documents pour en améliorer l'accessibilité et la diffusion.Nos travaux concernent le développement de méthodes automatisées de génération de mots-clés dont la singularité réside dans l'utilisation de méthodes de graphes et d'algorithmes d'ordonnancement de sommets. Nous nous penchons sur la problématique de l'évaluation indirecte des mots-clés générés au travers de tâches applicatives et de leur exploitation dans les moteurs de recherche et de recommandation académique. Nous présentons les travaux que nous avons menés dans la construction de ressources langagières, le développement d'outils logiciels et leur valorisation dans la communauté scientifique. Nous terminons par quelques réflexions prospectives sur l'indexation par mots-clés et plus généralement sur les travaux de recherche émergeant de l'intersection des thématiques du TAL et de la RI.

International audience; L'analyse historique des dépôts de HAL montre depuis la forte inflexion de 2016 une tendance accrue à l'autoarchivage de ses publications corrélée à une forte baisse qualitative des métadonnées. Par effet direct des principes d'indexation si de plus en plus de notices sont déposées celles-ci sont de moins en moins lisibles. SoVisu offre aux chercheurs d'un établissement la possibilité d’évaluer la qualité descriptive de chacune de ses productions à l’aide d’un score sur les métadonnées et réalise une cartographie lexicale globale extraite des productions et des descriptions accessibles. En cartographiant la lisibilité de chacun (par quels mots les travaux peuvent être retrouvés), SoVisu dépasse la simple incitation au dépôt car il permet aux chercheurs d’entrer « du bon pied » en SO en gérant à priori leurs identifiants et leur indexation. Une seconde fonctionnalité de SoVisu est la recherche d'informations. Conçu pour répondre à un besoin d’identification de l’expertise d’un établissement, SoVisu permet d'assembler la production scientifique structurée à différents degrés de granularité par mimétisme de l'organisation hiérarchique de la recherche (individu, groupes, labos, établissement). Au plan de l’écosystème de la SO, SoVisu se positionne au plus près des « clients » pour leur faciliter l’accès aux différents outils et services en dispositif de médiatisation. Réciproquement, le résultat de qualification et de vérification des données est une donnée utile pour la représentation scientifique des entités utile à de nombreuses briques d'un système d'information recherche.

National audience

International audience; Le grand public a tendance à éviter les sources fiables telles que la littérature scientifique en raison de leur langage complexe et du manque de connaissances nécessaires. Au lieu de cela, il s’appuie sur des sources superficielles, trouvées sur internet ou dans les médias sociaux et qui sont pourtant souvent publiées pour des raisons commerciales ou politiques, plutôt que pour leur valeur informative. La simplification des textes peut-elle contribuer à supprimer certains de ces obstacles à l’accès ? Cet article présente l’action « CLEF 2023 SimpleText » qui aborde les défis techniques et d’évaluation de l’accès à l’information scientifique pour le grand public. Nous fournissons des données réutilisables et des critères de référence pour la simplification des textes scientifiques et encourageons les recherches visant à faciliter à la compréhension des textes complexes.

International audience; This article presents MORFITT, the first multi-label corpus in French annotated in specialties in the medical field. MORFITT is composed of 3,624 abstracts of scientific articles from PubMed, annotated in 12 specialties for a total of 5,116 annotations. We detail the corpus, the experiments and the preliminary results obtained using a classifier based on the pre-trained language model CamemBERT. These preliminary results demonstrate the difficulty of the task, with a weighted average F1-score of 61.78 %.; Cet article présente MORFITT, le premier corpus multi-labels en français annoté en spécialités dans le domaine médical. MORFITT est composé de 3 624 résumés d'articles scientifiques issus de PubMed, annotés en 12 spécialités pour un total de 5 116 annotations. Nous détaillons le corpus, les expérimentations et les résultats préliminaires obtenus à l'aide d'un classifieur fondé sur le modèle de langage pré-entraîné CamemBERT. Ces résultats préliminaires démontrent la difficulté de la tâche, avec un score F1 moyen pondéré de 61,78 %.

Members of the OntoPortal Alliance; There is an explosion in the number of ontologies and semantic artefacts being produced in science. This paper discusses the need for common platforms to receive, host, serve, align, and enable their reuse. Ontology repositories and semantic artefact catalogues are necessary to address this need and to make ontologies FAIR (Findable, Accessible, Interoperable, and Reusable). The OntoPortal Alliance (https://ontoportal.org) is a consortium of research and infrastructure teams dedicated to promoting the development of such repositories based on the open, collaboratively developed OntoPortal software. We present the OntoPortal technology as a generic resource to build ontology repositories and semantic artefact catalogues that can support resources ranging from SKOS thesauri to OBO, RDF-S, and OWL ontologies. The paper reviews the features of OntoPortal and presents the current and forthcoming public and open repositories built with the technology maintained by the Alliance.

International audience; L’université de Toulon a développé et mis en œuvre un outil libre, Sovisu, permettant à chaque chercheur de gérer son identité et son profil numérique par la synthèse de la production individuelle et l’accès direct aux notices HAL. SoVisu se pose en brique élémentaire d’un Système d’Information Recherche, actuellement manquant au plan national. En cartographiant la lisibilité de chacun (par quels mots les travaux peuvent être retrouvés), SoVisu dépasse la simple incitation au dépôt car il permet aux chercheurs d’entrer « du bon pied » en SO en gérant à priori leurs identifiants. Conçu pour répondre à un besoin d’identification de l’expertise d’un établissement. SoVisu permet d'assembler la production scientifique structurée à différents degrés de granularité par mimétisme de l'organisation hiérarchique de la recherche (individu, groupes, labos, établissement). Au plan de l’écosystème de la SO, SoVisu se positionne au plus près des « clients » pour leur faciliter l’accès aux différents outils et services en dispositif de médiatisation, et de gestion qualité des données pour permettant de projeter l'intégration de briques intelligentes pour accompagner la recherche à de multiples niveaux.

International audience; ObjectivesTo describe psychosocial constraints and mental health of hospital workers, and to identify the psychosocial constraints significantly associated with mental health difficulties, especially in two groups: caregivers and other hospital workers.MethodData about working conditions and health status collected by the Evrest National observatory in 2018–2019 during occupational health consultation were used. Psychosocial constraints and mental health among caregivers, other hospital workers and non-hospital workers were described.ResultsThere were 1251 hospital workers (843 caregivers, 408 other hospital workers) and 25 129 other workers. Intensity and working time (time pressure, extra working time, missing or shortening a meal), and ethical dilemmas (not having the means to ensure high-quality work, too rapidly handling a procedure that would require more painstaking care) were significantly more reported by the caregivers than by the other hospital workers (50.8% vs 44.2%, 43.4% vs 32.5%, 47.2% vs 17.2%, 21.4% vs 16.4% and 41.5% vs 29.0% respectively). Prevalence of psychological distress was not significantly higher for caregivers (12.3%) than for other hospital workers (12.4%) but was significantly higher than for other workers (7.3%). For caregivers, factors significantly associated with psychological distress were time pressure (Odds Ratio adjusted on sociodemographic factors (OR) = 2.33 CI95% [1.35–4.04]), “difficulties to reconcile private life and work life” (OR = 2.95 [1.54–5.69]), “work not recognized in the professional setting” (OR = 1.89 [1.08–3.31]) and “fear of losing one's job” (OR = 2.98 [1.53–5.8]). For other hospital workers, they were “difficulties to reconcile private life and work life (OR = 2.76 [1.04–7.30]), “insufficient possibilities of mutual aid” (OR = 2.85 [1.24–6.53] and “not having the means to ensure high-quality work” (OR = 3.42 [1.62–7.21]).ConclusionFactors significantly associated with psychological distress were not the same for caregivers and other hospital workers, nor were they the most frequently reported. Detailed description of the constraints according to group of workers could help to develop a high-priority preventive program regarding psychosocial risk factors.; ObjectifsDécrire les contraintes psychosociales et la santé mentale des travailleurs hospitaliers et identifier les contraintes significativement associées à des troubles neuropsychiques, en particulier chez les soignants et les autres salariés hospitaliers.MéthodeLes données collectées par l'observatoire national Evrest durant les consultations de santé au travail en 2018-2019 ont été utilisées. Les contraintes psychosociales et la santé mentale ont été décrites parmi les personnels hospitaliers soignants, les autres salariés hospitaliers et les travailleurs non hospitaliers.RésultatsLa population d'étude comprenait 1 251 travailleurs hospitaliers (843 soignants, 408 non soignants) et 25 129 autres travailleurs. L'intensité et le temps de travail (forte pression temporelle, dépasser les horaires normaux, sauter ou écourter un repas), les conflits de valeur (ne pas avoir les moyens de faire un travail de qualité, traiter trop vite une opération qui demanderait davantage de soin) étaient significativement plus rapportés par les soignants que par les non soignants (respectivement 50,8% vs 44,2%, 43,4% vs 32,5%, 47,2% vs 17,2%, 21,4% vs 16,4% and 41,5% vs 29,0%). La prévalence des troubles neuropsychiques n'était pas significativement différente entre les soignants (12,3%) et les non soignants (12,4%), mais était significativement plus importante que chez les autres travailleurs (7,3%). Chez les soignants, les facteurs significativement associés à la présence de troubles neuropsychiques étaient la forte pression temporelle (OR = 2,33 IC95% [1,35–4,04]), les difficultés pour concilier vie privée et vie professionnelle (OR = 2,95 [1,54–5,69]), l'absence de reconnaissance par l'entourage professionnel (OR = 1.89 [1.08–3.31]) et la peur de perdre son travail (OR = 2,98 [1,53–5,8]). Pour les autres travailleurs hospitaliers, il s'agissait des difficultés pour concilier vies privée et professionnelle (OR = 2,76 [1,04–7,30]), de manquer d'entraide (OR = 2,85 [1,24–6,53] et manquer de moyens pour faire un travail de qualité (OR = 3,42 [1,62–7,21]).ConclusionsLes facteurs associés aux troubles neuropsychiques ne sont pas les mêmes chez les soignants et les autres travailleurs hospitaliers. Il ne s'agit pas non plus des facteurs les plus fréquemment rapportés. La description détaillée des facteurs selon le groupe de travailleurs peut être utilisée pour cibler les actions de prévention prioritaires en termes de risque psychosociaux.

Web page segmentation (WPS) aims to break a web page into different segments with coherent intra- and inter-semantics. By evidencing the morpho-dispositional semantics of a web page, WPS has traditionally been used to demarcate informative from non-informative content, but it has also evidenced its key role within the context of non-linear access to web information for visually impaired people. For that purpose, a great deal of ad hoc solutions have been proposed that rely on visual, logical, and/or text cues. However, such methodologies highly depend on manually tuned heuristics and are parameter-dependent. To overcome these drawbacks, principled frameworks have been proposed that provide the theoretical bases to achieve optimal solutions. However, existing methodologies only combine few discriminant features and do not define strategies to automatically select the optimal number of segments. In this article, we present a multi-objective clustering technique called MCS that relies on \( K \) -means, in which (1) visual, logical, and text cues are all combined in a early fusion manner and (2) an evolutionary process automatically discovers the optimal number of clusters (segments) as well as the correct positioning of seeds. As such, our proposal is parameter-free, combines many different modalities, does not depend on manually tuned heuristics, and can be run on any web page without any constraint. An exhaustive evaluation over two different tasks, where (1) the number of segments must be discovered or (2) the number of clusters is fixed with respect to the task at hand, shows that MCS drastically improves over most competitive and up-to-date algorithms for a wide variety of external and internal validation indices. In particular, results clearly evidence the impact of the visual and logical modalities towards segmentation performance.

The Web and social media have become the main source of information for citizens, with the risk that users rely on shallow information in sources prioritizing commercial or political incentives rather than the correctness and informational value. Non-experts tend to avoid scientific literature due to its complex language or their lack of prior background knowledge. Text simplification promises to remove some of these barriers. The CLEF 2022 SimpleText track addresses the challenges of text simplification approaches in the context of promoting scientific information access, by providing appropriate data and benchmarks, and creating a community of NLP and IR researchers working together to resolve one of the greatest challenges of today. The track will use a corpus of scientific literature abstracts and popular science requests. It features three tasks. First, content selection (what is in, or out?) challenges systems to select passages to include in a simplified summary in response to a query. Second, complexity spotting (what is unclear?) given a passage and a query, aims to rank terms/concepts that are required to be explained for understanding this passage (definitions, context, applications). Third, text simplification (rewrite this!) given a query, asks to simplify passages from scientific abstracts while preserving the main content.

International audience; A reliable representation of the user's knowledge state during a learning search session is crucial to understand their real information needs. When a search system is aware of such a state, it can adapt the search results and provide greater support for the user's learning objectives. A common practice to track the user's knowledge state is to consider the content of the documents they read during their search session(s). However, most current work ignores entity mentions in the documents, which, when linked to knowledge graphs, can be a source of valuable information regarding the user's knowledge. To fill this gap, we extend RULK-Representing User Knowledge in Search-as-Learning-with entity linking capabilities. The extended framework RULK NE represents and tracks user knowledge as a collection of such entities. It eventually estimates the user knowledge gain-learning outcome-by measuring the similarity between the represented knowledge and the learning objective. We show that our methods allow for up to 10% improvements when estimating user knowledge gains.

International audience; In this paper we propose several new quantum computation algorithms as an original contribution to the domain of PageRank algorithm theory, Spectral Graph Theory and Quantum Signal Processing. We first propose an application to PageRank of the HHL quantum algorithm for linear equation systems. We then introduce one of the first Quantum-Based Algorithms to perform a directed Graph Fourier Transform with a low gate complexity. After proposing a generalized PageRank formulation, based on ideas stemming from Spectral Graph Theory, we show how our quantum directed graph Fourier Transform can be applied to compute our generalized version of the PageRank.

International audience; In recent years, a large number of RDF datasets have been built and published on the Web in fields as diverse as linguistics or life sciences, as well as general datasets such as DBpedia or Wikidata. The joint exploitation of these datasets requires specific knowledge about their content, access points, and commonalities. However, not all datasets contain a self-description, and not all access points can handle the complex queries used to generate such a description. In this article, we provide a standard-based approach to generate the description of a dataset. The generated descriptions as well as the process of their computation are expressed using standard vocabularies and languages. We implemented our approach into a framework, called IndeGx, where each indexing feature and its computation is collaboratively and declaratively defined in a GitHub repository. We have experimented IndeGx on a set of 339 RDF datasets with endpoints listed in public catalogs, over 8 months. The results show that we can collect, as much as possible, important characteristics of the datasets depending on their availability and capacities. The resulting index captures the commonalities, variety and disparity in the offered content and services and it provides an important support to any application designed to query RDF datasets.

We present a new pre-training method, Multimodal Inverse Cloze Task, for Knowledge-based Visual Question Answering about named Entities (KVQAE). KVQAE is a recently introduced task that consists in answering questions about named entities grounded in a visual context using a Knowledge Base. Therefore, the interaction between the modalities is paramount to retrieve information and must be captured with complex fusion models. As these models require a lot of training data, we design this pre-training task from existing work in textual Question Answering. It consists in considering a sentence as a pseudo-question and its context as a pseudo-relevant passage and is extended by considering images near texts in multimodal documents. Our method is applicable to different neural network architectures and leads to a 9% relative-MRR and 15% relative-F1 gain for retrieval and reading comprehension, respectively, over a no-pre-training baseline., Accepted at ECIR 2023

soumission à Episciences, In the world of e-learning, learning systems have sought to adapt the user's profile and the content offered to them. However, from the point of view of collaboration between learners based on adaptation to the learner profile, this adaptation has not been sufficiently explored as an important aspect of the e-learning process. Adaptation will allow users with similar or very similar profiles to be grouped together to learn in harmony while maintaining motivation and commitment to learning. This should increase the success rate of learners. This will also allow us to reuse learning paths with good success rates for future recommendations to users with the same profile. In this paper, we focus on this aspect and propose a learning system that controls learning paths adapted to the users' profile and that allows collaborative learning of users in a synchronous way. After an overview of the existing work in the field of adaptive e-learning, we propose an architecture for the piloting of this type of collaborative adaptive learning based on ontologies and orchestrated by a multi-agent system. The latter is responsible for the piloting of learning paths, the recommendation of paths in collaborative or non-collaborative mode through communication between the different agents involved, and the management of events captured by the system., Dans le monde de la formation en ligne, les systèmes d'apprentissage ont cherché à proposer un contenu adapté à tout utilisateur selon son profil.Cette adaptation a comme objectif général de permettre à l'apprenant de tirer le meilleur parti du contenu exposé des ressources d'apprentissage.Cependant, du point de vue de la collaboration entre apprenants selon l'adaptation au profil des apprenants, cette adaptation n'a pas été suffisamment explorée comme un aspect important du processus d'apprentissage en ligne.L'adaptation va permettre de grouper des utilisateurs de profil similaire ou très proche pour apprendre en harmonie tout en gardant la motivation et l'engagement nécessaire afin d'augmenter le taux de réussite des apprenants. Cet état de fait doit permettre aussi de réutiliser certains parcours d'apprentissage avec de bons taux de réussite pour de futures recommandations aux utilisateurs ayant le même profil. Dans cet article, nous mettons l'accent sur cet aspect et proposons un système d'apprentissage qui recommande des parcours d'apprentissage adaptés au profil des utilisateurs et qui permet un apprentissage collaboratif des utilisateurs de façon synchrone. Après un tour d'horizon de l'existant dans le domaine de l'apprentissage adaptatif en ligne, nous proposerons une architecture pour le pilotage de ce type d'apprentissage adaptatif collaboratif. Cette solution est fondée sur des ontologies et orchestrée par un système multi-agents. Ce dernier est responsable du pilotage des parcours d'apprentissage, de la recommandation de parcours en mode collaboratif ou non par le biais d'une communication entre les différents agents intervenants et de la gestion des événements captés par le système.

International audience; Given the high data volumes in time series applications, or simply the need for fast response times, it is usually necessary to rely on alternative, shorter representations of time series, usually with information loss. This incurs approximate comparisons of time series where precision is a major issue. We propose a new representation approach called ASAX, coming with two techniques ASAX EN and ASAX SAE, for segmenting time series before their transformation into symbolic representations. Our solution can reduce significantly the error incurred by possible splittings at different steps of the representation calculation, by taking into account the entropy of the representations (ASAX EN) or the sum of absolute errors (ASAX SAE), particularly for datasets with unbalanced (non-uniform) distributions. This is particularly useful for time series similarity search, which is the core of many data analytics tasks. We provide theoretical guarantees on the lower bound of similarity measures, and our experiments illustrate that our approach can improve significantly the time series representation quality.

International audience; Manufacturing industry needs access to the data in order to realise its activities but also to generate new value-added knowledge. Nevertheless, it is confronted with a large and growing volume of heterogeneous data which limits its ability to exploit them optimally. Moreover, the data are distributed within different heterogeneous information systems, which limits the relationship exploration under the information retrieval process. Usually, the challenge is addressed by trying to manage and normalize the data structure in order to faster searching and exploiting them in a manufacturing context. For their part, the authors present i-Dataquest, an information retrieval system supported by (i) a graph-oriented model built from the structured and unstructured data of the company and (ii) a query system answering ‘what’ and ‘about what’ and (iii) generating three different results: a list of items, a list of property values and a list of sentences. The i-Dataquest prototype is built using Neo4J for the graph system generation, ConceptNet for lexical resource management and StandfordNLP for natural language processing. An evaluation of the prototype’s performance is conducted through a data set representing a drone manufacturer. The results show that the transformation of specific content such as tables in the graph and the semantic expansion of queries significantly improves the recall and precision measures. The results also suggest improving filtering less relevant results by considering particularly queries looking for a specific value.

International audience; In this paper, we describe the plans for the first LongEval CLEF 2023 shared task dedicated to evaluating the temporal persistence of Information Retrieval (IR) systems and Text Classifiers. The task is motivated by recent research showing that the performance of these models drops as the test data becomes more distant, with respect to time, from the training data. LongEval differs from traditional shared IR and classification tasks by giving special consideration to evaluating models aiming to mitigate performance drop over time. We envisage that this task will draw attention from the IR community and NLP researchers to the problem of temporal persistence of models, what enables or prevents it, potential solutions and their limitations.

International audience; In the context of Epidemic Intelligence, many Event-Based Surveillance (EBS) systems have been proposed in the literature to promote the early identification and characterization of potential health threats from online sources of any nature. Each EBS system has its own surveillance definitions and priorities, therefore this makes the task of selecting the most appropriate EBS system for a given situation a challenge for end-users. In this work, we propose a new evaluation framework to address this issue. It first transforms the raw input epidemiological event data into a set of normalized events with multi-granularity, then conducts a descriptive retrospective analysis based on four evaluation objectives: spatial, temporal, thematic and source analysis. We illustrate its relevance by applying it to an Avian Influenza dataset collected by a selection of EBS systems, and show how our framework allows identifying their strengths and drawbacks in terms of epidemic surveillance.

Open data are gold mines because they can be used to create services that develop a smart city while improving users' living conditions. Several research works go in this direction, presenting open data impact in the smart city for some, while others have focused on data processing methods. We have therefore deemed it necessary to make a state of the art on these different issues. The particularity of our study is that it shows the link between open data and smart city in all its aspects, describing what kind of open data is suitable for the smart city, how it is important for its development, and how these open data are processed to create services. Thus, in this article, we first present a review of existing surveys since 2015. Then, we present different smart city dimensions based on open data as well as some applications, and we detail how to process these data. We end with a list of open data sources as well as some challenges and solutions related to smart city services.

International audience

Factorization machines (FMs) are a powerful tool for regression and classification in the context of sparse observations, that has been successfully applied to collaborative filtering, especially when side information over users or items is available. Bayesian formulations of FMs have been proposed to provide confidence intervals over the predictions made by the model, however they usually involve Markov-chain Monte Carlo methods that require many samples to provide accurate predictions, resulting in slow training in the context of large-scale data. In this paper, we propose a variational formulation of factorization machines that allows us to derive a simple objective that can be easily optimized using standard mini-batch stochastic gradient descent, making it amenable to large-scale data. Our algorithm learns an approximate posterior distribution over the user and item parameters, which leads to confidence intervals over the predictions. We show, using several datasets, that it has comparable or better performance than existing methods in terms of prediction accuracy, and provide some applications in active learning strategies, e.g., preference elicitation techniques., 8 pages, 4 figures, 4 tables. Proceedings of the IEEE BigData 2022 conference

Keyphrase generation is the task consisting in generating a set of words or phrases that highlight the main topics of a document. There are few datasets for keyphrase generation in the biomedical domain and they do not meet the expectations in terms of size for training generative models. In this paper, we introduce kp-biomed, the first large-scale biomedical keyphrase generation dataset with more than 5M documents collected from PubMed abstracts. We train and release several generative models and conduct a series of experiments showing that using large scale datasets improves significantly the performances for present and absent keyphrase generation. The dataset is available under CC-BY-NC v4.0 license at https://huggingface.co/ datasets/taln-ls2n/kpbiomed., Accepted at the 13th International Workshop on Health Text Mining and Information Analysis (LOUHI 2022)

International audience; Abstract:Clustering is one of the most important data mining techniques, its objective is to regroup similar objects into groups, with the aim of maximizing the intra-cluster similarit and minimizing the inter-cluster similarity in unsupervised way. Recently, support-based clustering methods attracted a lot of attention, especially Support Vector Clustering (SVC) due to its capability to overcome the main hardships of classical clustering methods. SVC can easily handle complex shape clusters and identify their number without initialization. SVC undergoes on two main steps, training and labeling, the first one consist of solving a quadratic programming problem (QPP) to obtain a decision mathematics function, which is used in the next step to label all objects with their appropriate clusters. However, training an SVC model (solving a QPP) and labelling objects using huge data sets can lead to a high computation burden, in order to surmount this main issue and trying to improve the SVC performance, many methods and techniques was proposed in literature. In this paper, we aim to highlight and classify some of the most insightful works proposed by researchers according to their targeted SVC step.

Musical structure, defined as a simplified representation of the organization of a song, is an important musicological concept, but hard to automatically estimate. This thesis presents new methods to automatically estimate the structural segmentation of a song, focusing the study of music at the barscale. By developing a new segmentation algorithm (called ``CBM'') and by comparing several unsupervised compression schemes (from linear and multilinear algebra to neural networks), paradigms introduced in this thesis result in segmentation performance outperforming those of the unsupervised State-of-the-Art methods and almost similar with those of the global State-of-the-Art, obtained with supervised machine learning algorithms. In particular, as the methods described in this thesis are unsupervised, the estimation do not rely on annotated data, lowering the bias in the estimates related to ambiguity and subjectivity (inherent to musical structure) while limiting the loss in performance compared to the best supervised methods. In addition, some of the methods studied in this thesis (in particular Nonnegative Tucker Decomposition) allow to extract automatically interpretable parts of a song which may be used for other task than the estimation of structure, and participate in the development of interpretable machine and deep learning algorithms, which is a major field of research nowadays.; La structure musicale, définie comme la représentation simplifiée de l'organisation d'un morceau de musique, est un concept musicologique important mais néanmoins complexe à estimer automatiquement. Cette thèse présente de nouvelles méthodes pour estimer automatiquement la structure musicale, se focalisant sur l'étude à l'échelle de la mesure musicale. Par le développement d'un nouvel algorithme de segmentation (appelé ``CBM'') et par l'étude et la comparaison de différentes méthodes de compression non supervisées (allant de l'algèbre linéaire et multilinéaire aux réseaux de neurones), les paradigmes introduits dans cette thèse permettent d'obtenir des résultats quantitatifs dépassant l'Etat-de-l'Art non supervisé actuel et se rapprochant de l'Etat-de-l'Art global, issu de méthodes d'apprentissage avec supervision. En particulier, les méthodes décrites dans cette thèse étant non supervisées, l'estimation ne repose pas sur des bases de données annotées, permettant ainsi de mitiger les biais liés à l'ambiguïté et à la subjectivité (inhérents à la structure musicale), tout en limitant le perte en performance par rapport aux meilleures méthodes supervisées. Enfin, certaines méthodes étudiées dans cette thèse (en particulier la décomposition nonnégative en Tucker) permettent d'extraire automatiquement des parties interprétables de la chanson qui pourraient être utilisées pour d'autres tâches que l'estimation de structure, et s'intégrer dans le développement d'algorithmes interprétables d'apprentissage automatique profond, sujet de recherche majeur aujourd'hui.

Application to the 2022 IUF Chair Programme; Research misconduct and frauds pollute the scientific literature. Honest errors and malevolent data fabrication, image manipulation, journal hijacking, and plagiarism passed peer review unnoticed. Problematic papers deceive readers, authors citing them, and AI-powered literature-based discovery. Flagship publishers accepted hundreds flawed papers despite claiming to enforce peer review. This application ambitions to decontaminate the scientific literature using curative and preventive actions.; Méconduites et fraudes polluent la littérature scientifique. L’évaluation par les pairs peine à détecter les erreurs, fabrications de données, manipulations d’images, piratages de revues et plagiats. La littérature erronée trompe les lecteurs, les auteurs la citant et les IA qui génèrent des synthèses. Des éditeurs réputés publient des centaines d’articles erronés qui déjouent l’évaluation. Cette candidature ambitionne de dépolluer la littérature scientifique de façon curative et préventive.

International audience

This paper describes FuzViz, a tool to explore interactively massive relational data. FuzViz relies on a method building automatically linguistic summaries, that provide concise and intelligible insights in the data content. It offers an interactive view of these summaries, dynamically recomputed on demand. To ensure a fluid exploration of the data, FuzViz exploits the proposition of a highly efficient method for estimating the cardinality ofthe summary properties, estimated from statistics about the data distribution stored in the relational data base, consolidated by a sampling-based approach. The proposed workflow also involves a vocabulary inference mechanism from these statistics., Cet article présente un outil d'exploration interactive de données massives stockées dans un système de gestion de base de données (SGBD), nommé FuzViz. Il repose sur une méthode de construction automatique de résumés linguistiques, qui fournissent une synthèse concise et intelligible du contenu des données. Il offre une vue interactive de ces résumés recalculée dynamiquement selon les demandes de l'utilisateur. Pour assurer une exploration fluide des propriétés décrites par ces résumés, FuzViz s'appuie sur la proposition d'une méthode efficace d'estimations de leurs cardinalités, produites à partir des statistiques sur la distribution des données gérées par le SGBD et consolidées par une approche basée sur un échantillonnage. L'outil propose de plus un mécanisme d'inférence de vocabulaire flou à partir de ces statistiques.

The paper addresses the tasks of anomaly detection and explanation simultaneously, in the human-in-the-loop paradigm integrating the end-user expertise : it first proposes to exploit two complementary data representations to identify anomalies, namely the description induced by the raw features and the description induced by a userdefined vocabulary. These representations respectively lead to identify so-called data-driven and knowledge driven anomalies. The paper then proposes to confront these two sets of instances so as to improve the detection step and to dispose of tools towards anomaly explanations. It distinguishes and discusses three cases, underlining how the two description spaces can benefit from oneanother, in terms of accuracy and interpretability., La personnalisation par prise en compte de l'expertise de l'utilisateur des processus de détection d'anomalies est le sujet principal de cet article. Les anomalies sont identifiées dans deux espaces de représentation des données, l'espace initial composé des valeurs brutes et l'espace induit par un vocabulaire flou défini par l'utilisateur sur ces attributs initiaux. L'application de toute méthode de détection d'anomalies sur ces deux espaces conduit à différencier des anomalies issues des données brutes de celles issues du vocabulaire. Afin à la fois d'améliorer la détection des exceptions et de disposer d'outils pour les expliquer, ces deux ensembles d'anomalies sont confrontés. Trois situations émanant de cette comparaison sont étudiées afin de montrer comment l'utilisation combinée des deux espaces améliore l'efficacité de la procédure de détection d'anomalies ainsi que l'interprétabilité des résultats qu'elle génère.

Silk is a particularly important material in European history. Its trade has facilitated the growth of new economic centres, stimulated the development of innovative weaving techniques, and allowed the creation of exceptional objects. While many institutions specialise in its conservation, this fragile Cultural Heritage is under threat. In this article, we explain the methodology used by the Horizon 2020 SILKNOW project to better promote these Cultural Heritage collections. We present the CIDOC CRM-based data model that we designed for the creation of a Knowledge Graph. We also present the methodology used to integrate, in the data model, the annotations generated by the analysis of texts and images. Finally, we present the SIlKNOW extension, designed to model the complex semantics describing the production process of silk fabrics.

International audience

International audience; Nowadays, job recommender systems are more useful in the fight against unemployment due to their strong presence in e-recruitment platforms that are becoming very popular. Most of the recommender systems based on machine learning that recommend jobs to profiles use a vector representation of job offers based on keywords. However, these keywords are results of vectorization which is applied on a collection of documents where each one is a job offer. In this case, each keyword discriminates one job offer from another, whereas it can be preferable that each keyword discriminates one class from another. The aim of this paper is to improve the recommendation of job offers to users, by proposing to apply vectorization on a class-oriented collection of documents in order to obtain more useful keywords for the representation of job offers. In the case of job recommendation, each class-oriented document corresponds to a user profile. Experiments are done on two datasets (Nigam and Minajobs), using TF-IDF and Doc2Vec as vectorization techniques, Naive Bayes and Decision Trees as supervised machine learning models for top-N recommendation, and Precision, MRR and MAP as evaluation metrics. Our results show that, whatever the case, the best performance is always reach by a job recommender system resulting from our contribution. Compared to classic job recommender systems, the improvement rates can go up to 13% and 24% for systems based on Naive Bayes and go up to 55% and 46% for those based on Decision tree, respectively in the Nigam and Minajobs datasets.; Aujourd'hui, les systèmes de recommandation d'emploi sont plus utiles dans la lutte contre le chômage en raison de leur forte présence dans les plateformes de recrutement en ligne qui deviennent très populaires. La plupart des systèmes de recommandation basés sur l'apprentissage automatique qui recommandent des emplois à des profils utilisent une représentation vectorielle des offres d'emploi basée sur des mots-clés. Cependant, ces mots-clés sont les résultats d'une vectorisation appliquée à une collection de documents dont chacun est une offre d'emploi. Dans ce cas, chaque mot-clé discrimine une offre d'emploi par rapport à une autre, alors qu'il peut être préférable que chaque mot-clé discrimine une classe par rapport à une autre. L'objectif de cet article est d'améliorer la recommandation d'offres d'emploi aux utilisateurs, en proposant d'appliquer la vectorisation sur une collection de documents orientée classe afin d'obtenir des mots-clés plus utiles pour la représentation des offres d'emploi. Dans le cas de la recommandation d'emploi, chaque document orienté classe correspond à un profil d'utilisateur. Les expériences sont réalisées sur deux jeux de données (Nigam et Minajobs), en utilisant TF-IDF et Doc2Vec comme techniques de vectorisation, Naive Bayes et Decision Trees comme modèles d'apprentissage automatique supervisés pour la recommandation top-N, et Precision, MRR et MAP comme métriques d'évaluation. Nos résultats montrent que, dans tous les cas, la meilleure performance est toujours atteinte par un système de recommandation d'emploi issu de notre contribution. Par rapport aux systèmes classiques de recommandation d'emploi, les taux d'amélioration peuvent aller jusqu'à 13% et 24% pour les systèmes basés sur Naive Bayes et jusqu'à 55% et 46% pour ceux basés sur Decision tree, respectivement dans les jeux de données Nigam et Minajobs.

International audience; Top-N recommendation systems are important because they influence user choices on a daily basis on streaming and e-commerce platforms. Recommendation graphs are one of the commonly used approaches for computing top-N recommendations, as they are intuitive and interpretable. The idea is to build a graph from explicit data of users' actions on products, and then apply a random walk algorithm like PageRank to it to propose recommendations. However, implicit data that are not directly perceptible in the history of users' actions, such as latent data resulting from matrix factorization, are not exploited in recommendation graphs. Indeed, at the end of a matrix factorization, users and products are represented in a new space defined by the latent factors, in which a user and an item are close if the item is of the user's taste. In this paper, we propose to exploit the latent information resulting from matrix factorization to enrich recommendation graphs. Experiments are performed on six datasets, using three metrics for evaluating top-N recommendations: Precision, MAP and Hit-ratio. Our results show that the graphs enriched with latent information are better in 97% of the cases compared to the initial graph. The performance improvement can go up to at least 33%, depending on the dataset.; Les systèmes de recommandation top-N sont importants car ils influencent quotidiennement les choix des utilisateurs des plateformes notamment celles de streaming et de commerce électronique. Les graphes de recommandation sont l'une des approches couramment utilisées pour le calcul des recommandations top-N, car ces derniers sont intuitifs et interprétables. Le principe est de construire un graphe à partir des données explicites des actions des utilisateurs sur les produits, puis d'y appliquer un algorithme de marche aléatoire comme le PageRank pour proposer des recommandations. Cependant, les données implicites qui ne sont pas directement perceptibles dans l'historique des actions des utilisateurs à l'exemple des données latentes résultantes de la factorisation matricielle ne sont pas exploitées dans les graphes de recommandation. En effet, à l'issue d'une factorisation matricielle les utilisateurs et les produits sont représentés dans un nouvel espace défini par les facteurs latents, dans lequel un utilisateur et un item sont proches si l'item est du goût de l'utilisateur. Dans cet article, nous proposons d'exploiter les informations latentes résultantes de la factorisation matricielle pour enrichir les graphes de recommandation. Des expérimentations sont effectuées sur six jeux de données, en utilisant trois métriques d'évaluation des recommandations top-N : Précision, MAP et Hit-ratio. Nos résultats montrent que les graphes enrichis par les informations latentes sont meilleurs dans 97% des cas comparé au graphe initial. L'amélioration des performances peut aller jusqu'à 33% au moins, selon le jeu de données.

International audience; In the era of Big Data and Open Data, a massive and heterogeneous collections of documents (from text to multimedia) are created, managed and stored electronically. to make these documents more usable, a manual and/or automatic indexing process allows to create a representation of documents by a set of metadata, descriptors and social tags. These representations then make it easier to find information in a massive and scalable collection of documents from different sources (social networks, open data, …) to respond to user information needs (user requests). Numerous research studies have been carried out to propose indexing approaches depending on the type of indexed documents. Also, the evolution of indexing Methods, documents representation, electronic content, Big Data and Open Data. This paper presents a state of the art of approaches and methodologies ranging from manual and automatic indexing to algorithmic methods in the era of Big Data and Open Data.

International audience; This chapter is about an intelligent system to support ideas management as a result of a multi-agent system used in a distributed system with heterogeneous information as ideas and knowledge, after the results about an ontology to describe meaning of these ideas. The intelligent system assists participants of the creativity workshop to manage their ideas and consequently proposing an ontology dedicated to ideas. During the creative workshop many creative activities and collaborative creative methods are used by roles immersed in this creativity workshop event where they share knowledge. The collaboration of these roles is physically distant, their interactions might be synchrony or asynchrony, and the information of the ideas are heterogeneous, so we can say that the process is distributed. Those ideas are writing in natural language by participants which have a role and the ideas are heterogeneous since some of them are described by schema, text or scenario of use. This paper presents first, our MAS and second our Ontology design.

Quand elle s’interesse a l’apprentissage, la psychologie cognitive a pour objectif de comprendre comment les humains apprennent. L’ingenierie pedagogique a pour objectif de fournir des connaissances et des methodes utiles lors de la conception de situations d’enseignement. Les deux disciplines n’ont donc ni le meme objet (apprentissage vs. enseignement), ni le meme statut (science vs. ingenierie). L’objectif de cet article est de montrer comment on peut articuler connaissances en psychologie cognitive et ingenierie pedagogique. La theorie de la charge cognitive est utilisee comme exemple pour montrer comment on peut produire des connaissances pour l’ingenierie pedagogique a partir d’hypotheses fondees sur une description precise des structures et des fonctions cognitives impliquees dans les apprentissages. Cette theorie illustre que les connaissances pour l’ingenierie sont conditionnelles : leur mise en œuvre depend de la situation a concevoir et notamment des connaissances anterieures des eleves.

In this article, we study the polarity detection problem using linear supervised classifiers. We show the interest of penalizing the document frequencies in the regularization process to increase the accuracy. We propose a systematic comparison of different loss and regularization functions on this particular task using the Amazon dataset. Then, we evaluate our models according to three criteria: accuracy, sparsity and subjectivity. The subjectivity is measured by projecting our dictionary and optimized weight vector on the SentiWordNet lexicon. This original approach highlights a bias in the selection of the relevant terms during the regularization procedure: frequent terms are overweighted compared to their intrinsic subjectivities.We show that this bias appears whatever the chosen loss or regularization and on all datasets: it is closely link to the gradient descent technique. Penalizing the document frequency during the learning step enables us to improve significantly our performances. A lot of sentimental markers appear rarely and thus, are unappreciated by statistical learning algorithms. Explicitly boosting their influences leads to increasing the accuracy in the sentiment classification task.

International audience

This brief addresses understandability of modern machine learning networks with respect to the statistical properties of their convolution layers. It proposes a set of tools for categorizing a convolution layer in terms of kernel property (meanlet, differencelet, or distrotlet) or kernel sequence property (frame spectra and intralayer correlation matrix). These tools are expected to be relevant for determining the generalization capabilities of a convolutional neural network. In particular, this brief highlights that the less frequency penalizing network among AlexNet, GoogleNet, RESNET101, and VGG19 is the more relevant one in terms of solutions for low-level ice-sheet feature enhancement.

International audience; Unstructured data in electronic health records, represented by clinical texts, are a vast source of healthcare information because they describe a patient's journey, including clinical findings, procedures, and information about the continuity of care. The publication of several studies on temporal relation extraction from clinical texts during the last decade and the realization of multiple shared tasks highlight the importance of this research theme. Therefore, we propose a review of temporal relation extraction in clinical texts. We analyzed 105 articles and verified that relations between events and document creation time, a coarse temporality type, were addressed with traditional machine learning–based models with few recent initiatives to push the state-of-the-art with deep learning–based models. For temporal relations between entities (event and temporal expressions) in the document, factors such as dataset imbalance because of candidate pair generation and task complexity directly affect the system's performance. The state-of-the-art resides on attention-based models, with contextualized word representations being fine-tuned for temporal relation extraction. However, further experiments and advances in the research topic are required until real-time clinical domain applications are released. Furthermore, most of the publications mainly reside on the same dataset, hindering the need for new annotation projects that provide datasets for different medical specialties, clinical text types, and even languages.

International audience; With around 300 millions people worldwide suffering from depression, the detection of this disorder is crucial and a challenge for individual and public health. As with many diseases, early detection means better medical management; the use of social media messages as potential clues to depression is an opportunity to assist in this early detection by automatic means. This chapter is based on the participation of the CNRS IRIT laboratory in the early detection of depressive people (eRisk) task at the CLEF evaluation forum. Early depression detection differs from depression detection in that it considers temporality; the system must make its decision about a user’s possible depression with as little data as possible. In this chapter we re-evaluate the models we have developed for our participation at eRisk over the years on the different collections, to obtain a more robust comparison. We also add new models. We use well-established classification methods, such as Logistic regression, Random forest, and Support Vector Machine (SVM). The users’ data from which the system should detect if they are depressed, are represented as vectors composed of (a) various task-oriented features including depression related lexicons and (b) word and document embeddings, extracted from the users’ posts. We perform an ablation study to analyze the most important features for our models. We also use BERT deep learning architecture for comparison purposes, both for depression detection and early depression detection. According to our results, well-established machine learning models are still better than more modern models for -early- detection of depression.

International audience; Tracking news stories in documents is a way to deal with the large amount of information that surrounds us everyday, to reduce the noise and to detect emergent topics in news. Since the Covid-19 outbreak, the world has known a new problem: infodemic. News article titles are massively shared on social networks and the analysis of trends and growing topics is complex. Grouping documents in news stories lowers the number of topics to analyse and the information to ingest and/or evaluate. Our study proposes to analyse news tracking with little information provided by titles on social networks. In this paper, we take advantage of datasets of public news article titles to experiment news tracking algorithms on short messages. We evaluate the clustering performance with little amount of data per document. We deal with the document representation (sparse with TF-IDF and dense using Transformers [26]), its impact on the results and why it is key to this type of work. We used a supervised algorithm proposed by Miranda et al. [22] and K-Means to provide evaluations for different use cases. We found that TF-IDF vectors are not always the best ones to group documents, and that algorithms are sensitive to the type of representation. Knowing this, we recommend taking both aspects into account while tracking news stories in short messages. With this paper, we share all the source code and resources we handled.; Le suivi des actualités dans les documents est un moyen de traiter la grande quantité d'informations qui nous entoure chaque jour, de réduire le bruit et de détecter les sujets émergents dans les actualités. Depuis l'épidémie de Covid-19, le monde connaît un nouveau problème : l'infodémie. Les titres des articles d'actualité sont massivement partagés sur les réseaux sociaux et l'analyse des tendances et des sujets en expansion est complexe. Le regroupement de documents dans des articles d'actualité réduit le nombre de sujets à analyser et d'informations à ingérer et/ou à évaluer. Notre étude propose d'analyser le suivi des actualités avec peu d'informations fournies par les titres sur les réseaux sociaux. Dans cet article, nous tirons parti de jeux de données de titres d'articles d'actualité publics pour expérimenter des algorithmes de suivi des actualités sur des messages courts. Nous évaluons la performance du clustering avec une faible quantité de données par document. Nous traitons de la représentation du document (clairsemée avec TF-IDF et dense avec Transformers [26]), de son impact sur les résultats et de la raison pour laquelle elle est essentielle pour ce type de travail. Nous avons utilisé un algorithme supervisé proposé par Miranda et al. [22] et K-Means pour fournir des évaluations pour différents cas d'utilisation. Nous avons constaté que les vecteurs TF-IDF ne sont pas toujours les meilleurs pour regrouper les documents, et que les algorithmes sont sensibles au type de représentation. Sachant cela, nous recommandons de prendre en compte ces deux aspects lors du suivi des nouvelles dans les messages courts. Avec cet article, nous partageons tout le code source et les ressources que nous avons manipulés.