Search

International audience; We introduce a new approach to improve and adapt transformers for multi-label text classification. Dependencies between labels are an important factor in the multi-label context. Our proposed strategies take advantage of co-occurrences between labels. Our first approach consists in updating the final activation of each label by a weighted sum of all activations by these occurrence probabilities. The second proposed method consists in including the activations of all labels in the prediction. This is done using an approach similar to the 'self-attention' mechanism. As the most known multi-label datasets tend to have a small cardinality, we propose a new dataset, called 'arXiv-ACM', comprised of scientific abstracts from arXiv, tagged with their ACM keywords. We show that our approaches contribute to a performance gain, establishing a new state of the art for the studied datasets.; Nous présentons une nouvelle approche pour améliorer et adapter les transformeurs pour la classification multi-labels de textes. Les dépendances entre les labels sont un facteur important dans le contexte multi-labels. Les stratégies que nous proposons tirent profit des co-occurrences entre labels. Notre première approche consiste à mettre à jour l'activation de chaque label par une somme pondérée de toutes les activations par les probabilités d'occurrence. La deuxième méthode proposée consiste à inclure les activations de tous les labels dans la prédiction, en utilisant une approche similaire au mécanisme de 'selfattention'. Les jeux de données multi-labels les plus connus ont tendance à avoir une faible cardinalité, nous proposons un nouveau jeu de données, appelé 'arXiv-ACM', composé de résumés scientifiques d'arXiv, étiquetés avec leurs mots-clés ACM. Nous montrons que nos approches contribuent à un gain de performance, établissant un nouvel état de l'art pour les jeux de données étudiés.

International audience; Pre-trained language models have proven to be effective in multi-class text classification. Our goal is to study and improve this approach for multi-label text classification, a task that has been surprisingly little explored in the last few years despite its many real world applications. In this paper, our originality is to propose architectures for the classification layers that are used on top of transformers to improve their performance for multi-label classification. Our contribution involves the evaluation of thresholding methods on several transformers, either by computing an individual threshold for each label (IT) or a global one (GCT). We also propose two approaches for multi-label text classification. The first consists in adding a parameter for learning the number of labels present for a given example (NHA). The second approach consists in adding a layer to the classification layers in order to learn the features for selecting the relevant labels while avoiding the use of thresholds (TL). We evaluate these approaches on two English corpora of newspaper articles and scientific papers and then on a new multi-label dataset of French scientific article abstracts publicly available. The evaluations show that the performance of our proposals exceeds that of state-of-the-art multi-label text classification methods for the evaluated datasets, and are transposable to any multi-label classification problem.

In several real-world applications, extracting spatial semantics from text is critical. Spatial Role Labeling (SpRL) introduces a language-independent annotation scheme used in these applications, particularly for reasoning purposes. This paper proposes, first of all, a transfer learning method with a word embeddings-based approach for SpRL. Then, we enhance the word vectors with POS tags and CNN-based character-level representations. Finally, we propose a Residual BiLSTM CRF deep learning model to identify the spatial roles. The experimental results on two datasets: SemEval-2012 and SemEval-2013 Task 3, show that the proposed model outperforms other machine learning approaches.

International audience; The analysis of user-generated content on the Internet has become increasingly popular for a wide variety of applications. One particular type of content is represented by the user reviews for programs, multimedia, products, and so on. Investigating the opinion contained by reviews may help in following the evolution of the reviewed items and thus in improving their quality. Detecting contradictory opinions in reviews is crucial when evaluating the quality of the respective resource. This article aims to estimate the contradiction intensity (strength) in the context of online courses (MOOC). This estimation was based on review ratings and on sentiment polarity in the comments, with respect to specific aspects, such as “lecturer”, “presentation”, etc. Between course sessions, users stop reviewing, and also, the course contents may evolve. Thus, the reviews are time dependent, and this is why they should be considered grouped by the course sessions. Having this in mind, the contribution of this paper is threefold: (a) defining the notion of subjective contradiction around specific aspects and then estimating its intensity based on sentiment polarity, review ratings, and temporality; (b) developing a dataset to evaluate the contradiction intensity measure, which was annotated based on a user study; (c) comparing our unsupervised method with supervised methods with automatic feature selection, over the dataset. The dataset collected from coursera.org is in English. It includes 2244 courses and 73,873 user-generated reviews of those courses.The results proved that the standard deviation of the ratings, the standard deviation of the polarities, and the number of reviews are suitable features for predicting the contradiction intensity classes. Among the supervised methods, the J48 decision trees algorithm yielded the best performance, compared to the naive Bayes model and the SVM model.

International audience

International audience; Les modèles de langue pré-entraînés ont prouvé leur efficacité dans la classification de texte multiclasses. Notre objectif est d'étudier et d'améliorer ce type d'approches pour la classification multilabels de texte, une tâche étonnamment peu explorée au cours de ces toutes dernières années. Cette tâche a pourtant des applications industrielles importantes telles que la recommandation de contenu, l'extraction de méta-données pour l'enrichissement des bases de données ou le routage automatique multicritères des emails. Dans cet article, notre originalité est de proposer des méthodes d'exploitation des activations des couches de sortie des transformeurs pour améliorer la performance de ces modèles pour la classification multilabels. Notre contribution concerne l'évaluation de l'utilité des méthodes de seuillage sur plusieurs modèles d'apprentissage profond, en calculant un seuil de classification global pour optimiser l'ensemble des classes (SGO), ou un seuil individuel propre à chaque classe étudiée (SI). Elle concerne aussi la proposition de deux approches pour la classification multilabels de texte. La première approche (NPA) consiste à ajouter un paramètre pour l'apprentissage du nombre de classes et/ou labels N présentes pour un exemple donné, pour considérer les classes qui correspondent aux N activations les plus élevées comme étant des labels valides. La deuxième approche (TL) consiste à ajouter une couche au transformeur pour l'apprentissage des critères utiles pour la sélection des labels pertinents. Nous évaluons ces approches sur des corpus d'articles de journaux et d'articles scientifiques. Nous avons aussi constitué et mis à disposition un jeu de données de résumés d'articles scientifiques en français que nous avons conçu à partir du dépôt d'archives ouvertes 'HAL'. Ces évaluations montrent que la performance de nos propositions dépasse celles des méthodes de l'état de l'art de classification multilabels de texte pour les jeux de données étudiés, et sont transposables à tout problème de classification multilabels utilisant les réseaux de neurones.

International audience; Les agents virtuels peuvent être utilisés pour promouvoir des comportements pro-sociaux. Dans cet article, nous explorons le potentiel des agents virtuels et de la réalité virtuelle pour lutter contre les discriminations sociales à travers l'induction d'empathie chez l'utilisateur. Nous présentons différents systèmes développés dans cette perspective et discutons les facteurs d'influence, les outils de mesures des performances de ces systèmes, ainsi que les limites actuelles de ces travaux.

The digititalization of recruitment processes did not smooth out all frictions related to this activity. Automatically searching for an ideal candidate is still hindered by one major issue: adequately modeling the needs expressed in natural language in a job offer. This problem is even more obvious when one considers the use case of recruiting expert profiles. Indeed, these profiles represent a minority of the recruitments. To assist a recruitment process concerning these profiles, an automated system requires an accurate knowledge of their activity sector. In this article we introduce the architecture of an expert profile recommender system applied to the process industry sector., La dématérialisation des processus de recrutement n'a pas fait disparaître toutes les frictions inhérentes à cette activité. La recherche automatisée d'un candidat idéal se heurte toujours à la difficulté à modéliser correctement les besoins exprimés en langage naturel dans une offre d'emploi. Le recrutement d'experts, notamment, est particulièrement difficile. En effet, ces profils concernent une proportion réduite des recrutements et leur prise en charge informatisée nécessite une connaissance précise du secteur d'activité concerné. Dans cet article, nous proposons l'architecture d'un système de recommandation de profils experts dans l'industrie des procédés afin d'assister ce type de recrutements.

Extractive multi-document summarization (MDS) is the process of automatically summarizing a collection of documents by ranking sentences according to their importance and informativeness. Text representation is a fundamental process that affects the effectiveness of many text summarization methods. Word embedding representations have shown to be effective for several Natural Language Processing (NLP) tasks including Automatic Text Summarization (ATS). However, most of these representations do not consider the order and the semantic relationships between words in a sentence. This does not fully allow grasping the sentence semantics and the syntactic relationships between sentences constituents. In this paper, to overcome this problem, we propose an unsupervised method for generic extractive multi-document summarization based on the sentence embedding representations and the centroid approach. The proposed method selects relevant sentences according to the final score obtained by combining three scores: sentence content relevance, sentence novelty, and sentence position scores. The sentence content relevance score is computed as the cosine similarity between the centroid embedding vector of the cluster of documents and the sentence embedding vectors. The sentence novelty metric is explicitly adopted to deal with redundancy. The sentence position metric assumes that the first sentences of a document are more relevant to the summary, and it assigns high scores to these sentences. Moreover, this paper provides a comparative analysis of nine sentence embedding models used to represent sentences as dense vectors in a low dimensional vector space in the context of extractive multi-document summarization. Experiments are performed on the standard DUC’2002–2004 benchmark datasets and the Multi-News dataset. The overall obtained results have shown that our method outperforms several state-of-the-art methods and achieves promising results compared to the best performing methods including supervised deep learning based methods.

International audience; Text representation is a fundamental cornerstone that impacts the effectiveness of several text summarization methods. Transfer learning using pre-trained word embedding models has shown promising results. However, most of these representations do not consider the order and the semantic relationships between words in a sentence, and thus they do not carry the meaning of a full sentence. To overcome this issue, the current study proposes an unsupervised method for extractive multi-document summarization based on transfer learning from BERT sentence embedding model. Moreover, to improve sentence representation learning, we fine-tune BERT model on supervised intermediate tasks from GLUE benchmark datasets using single-task and multi-task fine-tuning methods. Experiments are performed on the standard DUC’2002–2004 datasets. The obtained results show that our method has significantly outperformed several baseline methods and achieves a comparable and sometimes better performance than the recent state-of-the-art deep learning–based methods. Furthermore, the results show that fine-tuning BERT using multi-task learning has considerably improved the performance.

Extractive query-focused multi-document summarization (QF-MDS) is the process of automatically generating an informative summary from a collection of documents that answers a pre-given query. Sentence and query representation is a fundamental cornerstone that affects the effectiveness of several QF-MDS methods. Transfer learning using pre-trained word embedding models has shown promising performance in many applications. However, most of these representations do not consider the order and the semantic relationships between words in a sentence, and thus they do not carry the meaning of a full sentence. In this paper, to deal with this issue, we propose to leverage transfer learning from pre-trained sentence embedding models to represent documents’ sentences and users’ queries using embedding vectors that capture the semantic and the syntactic relationships between their constituents (words, phrases). Furthermore, BM25 and semantic similarity function are linearly combined to retrieve a subset of sentences based on their relevance to the query. Finally, the maximal marginal relevance criterion is applied to re-rank the selected sentences by maintaining query relevance and minimizing redundancy. The proposed method is unsupervised, simple, efficient, and requires no labeled text summarization training data. Experiments are conducted using three standard datasets from the DUC evaluation campaign (DUC’2005–2007). The overall obtained results show that our method outperforms several state-of-the-art systems and achieves comparable results to the best performing systems, including supervised deep learning-based methods.

International audience; Data is the central element of a geographic information system (GIS) and its cost is often high because of the substantial investment that allows its production. However, these data are often restricted to a service or a category of users. This has highlighted the need to propose and optimize the means of enriching spatial information relevant to a larger number of users. In this chapter, a data enrichment approach that integrates recent advances in machine learning; more precisely, the use of deep learning to optimize the enrichment of GDBs is proposed, specifically, during the topic identification phase. The evaluation of the approach was completed showing its performance.

International audience

In this paper, we introduce FreSaDa, a French Satire Data Set, which is composed of 11,570 articles from the news domain. In order to avoid reporting unreasonably high accuracy rates due to the learning of characteristics specific to publication sources, we divided our samples into training, validation and test, such that the training publication sources are distinct from the validation and test publication sources. This gives rise to a cross-domain (cross-source) satire detection task. We employ two classification methods as baselines for our new data set, one based on low-level features (character n-grams) and one based on high-level features (average of CamemBERT word embeddings). As an additional contribution, we present an unsupervised domain adaptation method based on regarding the pairwise similarities (given by the dot product) between the training samples and the validation samples as features. By including these domain-specific features, we attain significant improvements for both character n-grams and CamemBERT embeddings., Comment: Accepted at IJCNN 2021

International audience; We propose in this paper a statistical model in the perspective of predicting listener's feedbacks in a conversation. The first contribution of the paper is a study of the prediction of all feedbacks, including those in overlap with the speaker with a good accuracy. Existing model are good at predicting feedbacks during a pause, but reach a very low success level for all feedbacks. We give in this paper a first step towards this complex problem. The second contribution is a model predicting precisely the type of the feedback (generic vs. specific) as well as other specific features (valence expectation) useful in particular for generating feedbacks in dialogue systems. This work relies on an original corpus.

All over the world, people receive daily news on many subjects through web-based information sharing platforms such as social networks. However, some of such news are false (fake) with the potential to deceive them. Thus, the automatic detection of false news is a major issue and is gaining careful attention from the scientific community. In this paper, we present experimental analysis using both supervised and unsupervised learning on the Fake.Br corpus, a fake news dataset in Brazilian Portuguese. We propose a classification method for fake news detection based on distinct types of features, and deep learning supervised algorithms. Our best classification model achieved F1 scores up to 96% and was compared with other non-deep learning classifiers. Furthermore, we provide a complementary analysis of the same dataset by performing topic modeling based on both uni-grams and bi-grams.

The majority of the documents produced and exchanged through medias and social networks are unstructured. Due to the amount of these unstructured documents on the Web, their exploitation represents a tedious or even impossible task for human beings without assistance by dedicated algorithms and specialized computer systems in document classification or information extraction. To be efficient and relevant, such systems have to understand the content of these unstructured documents. The context (or topic) of a document is one of the basic information essential for the understanding of its content, and the more precise the context of a document, the more relevant its understanding will be. This paper presents a precise context identification approach that is evaluated quantitatively and qualitatively on several reference corpora and compared to other context identification systems. The contexts identified by our model are much more precise than those identified by these others systems.

International audience; Book search is a challenging task due to discrepancies between the content and description of books, on one side, and the ways in which people query for books, on the other. However, online reviewers provide an opinionated description of the book, with alternative features that describe the emotional and experiential aspects of the book. Therefore, locating emotional sentences within reviews, could provide a rich alternative source of evidence to help improve book recommendations. Specifically, sentiment analysis (SA) could be employed to identify salient emotional terms, which could then be used for query expansion? This paper explores the employment ofSA based query expansion, in the book search domain. We introduce a sentiment-oriented method for the selection of sentences from the reviews of top rated book. From these sentences, we extract the terms to be employed in the query formulation. The sentence selection process is based on a semi-supervised SA method, which makes use of adapted word embeddings and lexicon seed-words.Using the CLEF 2016 Social Book Search (SBS) Suggestion TrackCollection, an exploratory comparison between standard pseudo-relevance feedback and the proposed sentiment-based approach is performed. The experiments show that the proposed approach obtains 24%-57% improvement over the baselines, whilst the classic technique actually degrades the performance by 14%-51%.

This research is a part of a work which addresses the issue of reconfiguration of production systems that aims to extend the assembly line life and to offer solutions to the introduction of a new product or the ability to change the existent one or the capacity regulation. Through this article, we propose a new reconfiguration process, which is defined by structured reconfiguration strategies divided to several operations. Each strategy follows specific steps in response to a reconfiguration issue. The process solves the issue by taking into account the impact limitation, which is to respond as locally as possible before considering cooperative or global solutions. The definition of these reconfiguration strategies allows us to implement a concatenation of steps and actions that characterize each point of decision-making. We propose then a model that defines each operation and the relation-ship between them and the production system entities, taking into account the various parameters of the system. We use the example of the automotive assembly lines to cover the analysis process.

International audience; Les agents artificiels (robots humanoïdes et personnages virtuels), peuvent être exploités pour un monde meilleur : pour augmenter l'empathie entre les gens, améliorer leur bien-être, promouvoir les coopérations, encourager les comportements prosociaux, réduire les inégalités sociales par l'apprentissage, etc. S’intégrant dans le récent courant de recherche appelé l’Informatique Prosociale, un certain nombre de travaux dans le domaine des agents artificiels sont aujourd’hui menés avec cet objectif. Outre ces recherches dédiées, quelque que soit le domaine applicatif de nos travaux de recherche, il est impératif de veiller à la conception de nos agents artificiels. En particulier, l’apparence et le comportement de ces derniers peuvent avoir une influence inattendue et non souhaitable sur le comportement de l’utilisateur et ses croyances, non seulement au sein de l’environnement virtuel mais se poursuivant aussi ensuite dans le monde réel. Dans cet article, ces problématiques sont discutées à la lumière des résultats de recherche récents autour de l’influence des agents artificiels. De solutions sont discutés pour veiller à ce que nos agents artificiels puissent être utilisés pour tendre vers un monde meilleur.

International audience; Nous étudions l'impact de l'ajout de plusieurs modalités à la transcription de la parole pour la prédiction d'émotions. Ce travail a pour objectif de contribuer aux thérapies pour l'éducation émotionnelle des adolescents TSA (Trouble du Spectre Autistique) à partir d'une recommandation automatique de passages de séries TV.

International audience; Natural Language Processing (NLP) of textual data is usually broken down into a sequence of several subtasks, where the output of one the subtasks becomes the input to the following one, which constitutes an NLP pipeline. Many third-party NLP tools are currently available, each performing distinct NLP subtasks. However, it is difficult to integrate several NLP toolkits into a pipeline due to many problems, including different input/output representations or formats, distinct programming languages, and tokenization issues. This paper presents DeepNLPF, a framework that enables easy integration of third-party NLP tools, allowing the user to preprocess natural language texts at lexical, syntactic, and semantic levels. The proposed framework also provides an API for complete pipeline customization including the definition of input/output formats, integration plugin management, transparent multiprocessing execution strategies, corpus-level statistics, and database persistence. Furthermore, the DeepNLPF user-friendly GUI allows its use even by a non-expert NLP user. We conducted runtime performance analysis showing that DeepNLPF not only easily integrates existent NLP toolkits but also reduces significant runtime processing compared to executing the same NLP pipeline in a sequential manner.

International audience; Experience mining is considered a substantial extension of opinion mining. Experience mining covers the description of all events that are related to the user's perception in the interaction with the object. There is information about the user`s experience that cannot be obtained with polarity analysis or sentiment analysis. The information obtained from those analyses is the positive or negative evaluation of a product/service. But in opinion analysis there is more useful information for other users. In this work we propose a method to obtain expressions of user experiences that are difficult to obtain by some of the current methods. The contributions of this work are: 1) Classification of information of user experiences according to the intention of communication: the categories are: "Evaluation", "Sensation" "Recommendation" and "Target Users". 2) A rule-based method: rules take into account the lexical classification of words to classify sentences based on communicative intention. The results obtained show that the method based on classification rules can provide useful information about the user experience to other users

National audience; La recherche d’information et la recommandation correspondent à des processus cognitifs complexes. Pour être efficaces, les approches algorithmiques doivent s’adapter dynamiquement et disposer d’informations nombreuses, notamment sur les contextes d’utilisation et sur les préférences des utilisateurs. C’est ainsi que se sont développées des approches personnalisées d’apprentissage automatique, exploitant de très nombreuses données de façon peu transparente. Nous présentons les approches statistiques les plus populaires, du modèle vectoriel de recherche d’information aux approches connexionnistes à bases de réseaux neuronaux. Nous soulignons les limites inhérentes aux procédures d’apprentissage dans le cadre de tâches subjectives, notamment du point de vue des risques d’uniformisation contre lesquelles une pluralité de modèles est nécessaire. Nous proposons enfin d’étudier le comportement des modèles et des systèmes automatiques sous différents angles, selon des approches pluridisciplinaires.

International audience; The emergence of social media allows users to get opinions, suggestions, or recommendations from other users about complex information needs. Tasks, as the CLEF Social Book Search 2016 Suggestion Track, propose to pursue this issue. The originality is to deal with verbose queries of book recommendation in order to support users in searching for books in catalogues of professional meta-data and complementary social media (i.e. tags, authors, similar products). In this context, a new technique for community-of-books discovery based on frequent social information (i.e. tags, authors) of similar books are proposed for book recommendation. Our method allows detecting frequent sub-graphs of similar books and using them to enrich the results returned by a traditional information retrieval system. This approach is tested on a collection containing Amazon/LibraryThing book descriptions and a set of queries, extracted from the LibraryThing discussion forums.

International audience; Search systems based on both professional meta-data (e.g., title, description, etc.) and social signals (e.g., like, comment , rating, etc.) from social networks is the trending topic in information retrieval (IR) field. This paper presents 2SRM (Social Signals Relevance Model), an approach of IR which takes into account social signals (users' actions) as an additional information to enhance a search. We hypothesize that these signals can play a role to estimate a priori social importance (relevance) of the resource (document). In this paper, we first study the impact of each such signal on retrieval performance. Next, some social properties such as popularity, reputation and freshness are quantified using several signals. The 2SRM combines the social relevance, estimated from these social signals and properties, with the conventional textual relevance. Finally, we investigate the effect of the social signals on the retrieval effectiveness using state-of-the-art learning approaches. In order to identify the most effective signals, we adopt feature selection algorithms and the correlation between the signals. We evaluated the effectiveness of our approach on both IMDb (Internet Movie Databese) and SBS (Social Book Search) datasets containing movies and books resources and their social characteristics collected from several social networks. Our experimental results are statistically significant, and reveal that incorporating social signals in retrieval model is a promising approach for improving the retrieval performance.

International audience; In this paper, we present a tool allowing dynamic prediction and visualization of an individual's local brain activity during a conversation. The prediction module of this tool is based on classifiers trained using a corpus of human-human and human-robot conversations including fMRI recordings. More precisely, the module takes as input behavioral features computed from raw data, mainly the participant and the interlocutor speech but also the participant's visual input and eye movements. The visualisation module shows in real-time the dynamics of brain active areas synchronised with the behavioral raw data. In addition, it shows which integrated behavioral features are used to predict the activity in individual brain areas.

International audience

International audience; Most of the queries submitted to search engines are composed of keywords but it is not enough for users to express their needs. Through verbose natural language queries, users can express complex or highly specific information needs. However, it is difficult for search engine to deal with this type of queries. Moreover, the emergence of social medias allows users to get opinions, suggestions or recommendations from other users about complex information needs. In order to increase the understandability of user needs, tasks as the CLEF Social Book Search Suggestion Track have been proposed from 2011 to 2016. The aim is to investigate techniques to support users in searching for books in catalogs of professional metadata and complementary social media. In this context, we introduce in the current paper a statical approach to deal with long verbose queries in Social Information Retrieval (SIR) by taking Social Book Search(SBS) as a study case. firstly, a morphosyntactic analysis was introduced to reduce verbose queries, the second step is based on expanding the reduced queries using association rules mining combined with Pseudo relevance feedback. Experiments on SBS 2014 and 2016 collections show significant improvement in the retrieval performance.

International audience; This paper proposes an approach on Facebook search in Arabic and English, which exploits several users' traces (e.g. comment, share, reactions) left on Facebook posts to estimate their social importance. Our goal is to show how these social traces (signals) can play a vital role in improving Arabic and English Facebook search. Firstly, we identify polarities (positive or negative) carried by the textual signals (e.g. comments) and non-textual ones (e.g. the reactions love and sad) for a given Facebook posts. Therefore, the polarity of each comment expressed in Arabic or in English on a given Facebook post, is estimated on the basis of a neural sentiment model. Secondly , we group signals according to their complementarity using attributes (features) selection algorithms. Thirdly, we apply learning to rank (LTR) algorithms to re-rank Facebook search results based on the selected groups of signals. Finally, experiments are carried out on 13,500 Facebook posts, collected from 45 topics, for each of the two languages. Experiments results reveal that Random Forests was the most effective LTR approach for this task, and for the both languages. However, the best appropriate features selection algorithms are ReliefFAttributeEval and InfoGainAttributeEval for Arabic and English Facebook search task, respectively.

International audience; This paper proposes an approach on Facebook search in Arabic, which exploits several users' traces (e.g. comment, share, reactions) left on Facebook posts to estimate their social importance. Our goal is to show how these social traces (signals) can play a vital role in improving Arabic Facebook search. Firstly, we identify polarities (positive or negative) carried by the textual signals (e.g. comments) and non-textual ones (e.g. the reactions love and sad) for a given Facebook post. Therefore, the polarity of each comment expressed on a given Facebook post, is estimated on the basis of a neural sentiment model in Arabic language. Secondly, we group signals according to their complementarity using features selection algorithms. Thirdly, we apply learning to rank (LTR) algorithms to re-rank Facebook search results based on the selected groups of signals. Finally, experiments are carried out on 13,500 Facebook posts, collected from 45 topics in Arabic language. Experiments results reveal that Random Forests combined with ReliefFAttributeEval (RLF) was the most effective LTR approach for this task.

International audience; Extractive summarization consists of generating a summary by ranking sentences from the original texts according to their importance and salience. Text representation is a fundamental process that affects the effectiveness of many text summarization methods. Distributed word vector representations have been shown to improve Natural Language Processing (NLP) tasks, especially Automatic Text Summariza-tion (ATS). However, most of them do not consider the order and the context of the words in a sentence. This does not fully allow grasping the sentence semantics and the syntactic relationships between sentences constituents. In this paper, to overcome this problem, we propose a deep neural network model based-method for extractive single document sum-marization using the state-of-the-art sentence embedding models. Experiments are performed on the standard DUC2002 dataset using three sentence embedding models. The obtained results show the effectiveness of the used sentence embedding models for ATS. The overall comparison results show that our method outperforms eight well-known ATS baselines and achieves comparable results to the state-of-the-art deep learning based methods.

International audience; A key challenge when studying human-agent interaction is the evaluation of user's experience. In virtual reality, this question is addressed by studying the sense of "presence" and"co-presence", generally assessed thanks to well-grounded subjective post-experience questionnaires. In this article, we aim at exploring behavioral measures of presence and co-presence by analyzing multimodal cues produced during an interaction both by the user and the virtual agent. In our study, we started from a corpus of human-agent interaction collected in a task-oriented context: a virtual environment aiming at training doctors to break bad news to a patient (played by a virtual agent). Based on this corpus, we have used machine learning algorithms to explore the possibility of predicting user's sense of presence and co-presence. In particular, we have applied and compared two techniques, Random forest and SVM, both showing very good results in predicting the level of presence and co-presence.

International audience; A key challenge when studying human-agent interaction is the evaluation of user's experience. In virtual reality, this question is addressed by studying the sense of "presence" and"co-presence", generally assessed thanks to well-grounded subjective post-experience questionnaires. In this article, we aim at exploring behavioral measures of presence and co-presence by analyzing multimodal cues produced during an interaction both by the user and the virtual agent. In our study, we started from a corpus of human-agent interaction collected in a task-oriented context: a virtual environment aiming at training doctors to break bad news to a patient (played by a virtual agent). Based on this corpus, we have used machine learning algorithms to explore the possibility of predicting user's sense of presence and co-presence. In particular, we have applied and compared two techniques, Random forest and SVM, both showing very good results in predicting the level of presence and co-presence.

International audience

The first emojis were created in 1999. Since then, their popularity constantly raised in communication systems. Being images representing either an idea, a concept, or an emotion, emojis are available to the users in multiple software contexts: instant messaging, emails, forums, and other types of social medias. Their usage grew constantly and, associated to the constant addition of new emojis, there are now more than 2,789 standard emojis since winter 2018.To access a specific emoji, scrolling through huge emoji librairies or using a emoji search engines is not enough to maximize their usage and their diversity. An emoji recommendation system is required. To answer this need, we present our research work focused on the emoji recommendation topic. The objectives are to create an emoji recommender system adapted to a private and informal conversational context. This system must enhance the user experience, the communication quality, and take into account possible new emerging emojis.Our first contribution is to show the limits of a emoji prediction for the real usage case, and to demonstrate the need of a more global recommandation. We also verify the correlation between the real usage of emojis representing facial expressions and a related theory on facial expressions. We also tackle the evaluation part of this system, with the metrics' limits and the importance of a dedicated user interface.The approach is based on supervised and unsupervised machine learning, associated to language models such as word embeddings. Several parts of this work were published in national and international conferences, including the best software award and best poster award for the social media track.; Depuis leur apparition en 1999, les emojis ont une popularité grandissante dans les systèmes de communication. Ces petites images pouvant représenter une idée, un concept ou une émotion, se retrouvent disponibles aux utilisateurs dans de nombreux contextes logiciels : messagerie instantanée, courriel, forums et autres réseaux sociaux en tout genre. Leur usage, en hausse constante, a entraîné l'apparition récurrente de nouveaux emojis, allant jusqu'à 2 789 emojis standardisés en fin d'année 2018.Le parcours de bibliothèques d'emojis ou l'utilisation de moteur de recherche intégré n'est plus suffisant pour permettre à l'utilisateur de maximiser leur utilisation ; une recommandation d'emojis adaptée est nécessaire. Pour cela, nous présentons nos travaux de recherche axés sur le thème de la recommandation d'emojis. Ces travaux ont pour objectif de créer un système de recommandation automatique d'emojis adapté à un contexte conversationnel informel et privé. Ce système doit améliorer l'expérience utilisateur et la qualité de la communication, en plus de pouvoir prendre en compte d'éventuels nouveaux emojis. Dans le cadre de cette thèse, nous contribuons tout d'abord en montrant les limites d'usage réel d'une prédiction d'emojis ainsi que la nécessité de prédire des notions plus générales. Nous vérifions également si l'usage réel des emojis représentant une expression faciale d'émotion correspond à l'existant théorique. Enfin, nous abordons les pistes d'évaluation d'un tel système par l'insuffisance des métriques, et l'importance d'une interface utilisateur dédiée.Pour ce faire, nous utilisons une approche orientée apprentissage automatique à la fois supervisée et non supervisée, ainsi que la conception de modèles de langues ou, plus précisément, de plongements lexicaux. Plusieurs composantes de ce travail ont donné lieu à des publications nationales et internationales, dont le prix du meilleur logiciel reproductible à la conférence CICLing 2018, ainsi que celui du meilleur poster pour la catégorie des médias sociaux SONAMA à la conférence SAC 2018.

International audience; We argue that verbose natural language queries used for software retrieval contain many terms that follow specific discourse rules, yet hinder retrieval. Through verbose queries, users can express complex or highly specific information needs. However, it is difficult for search engine to deal with this type of queries. Moreover, the emergence of social me-dias allows users to get opinions, suggestions, or recommendations from other users about complex information needs. In order to increase the understanding of user needs, a tasks, as the CLEF Social Book Search Classification Track, the aims is to investigates how systems can automatically identify book search requests in online forums. In this respect, we introduce in the present paper a new approach to automatically detect the type of each thread. Our proposal aims to identify book search queries by syntactic patterns, association rules between terms and tex-tual sequences mining.

International audience; Deep Learning is more and more used in NLP tasks, such as in relation classification of texts. This paper assesses the impact of syntactic dependencies in this task at two levels. The first level concerns the generic Word Embedding (WE) as input of the classification model, the second level concerns the corpus whose relations have to be classified. In this paper, two classification models are studied, the first one is based on a CNN using a generic WE and does not take into account the dependencies of the corpus to be treated, and the second one is based on a compositional WE combining a generic WE with syntactical annotations of this corpus to classify. The impact of dependencies in relation classification is estimated using two different WE. The first one is essentially lexical and trained on the Wikipedia corpus in English, while the second one is also syntactical, trained on the same previously annotated corpus with syntactical dependencies. The two classification models are evaluated on the SemEval 2010 reference corpus using these two generic WE. The experiments show the importance of taking dependencies into account at different levels in the relation classification.

International audience; We argue that verbose natural language queries used for software retrieval contain many terms that follow specific discourse rules, yet hinder retrieval. Through verbose queries, users can express complex or highly specific information needs. However, it is difficult for search engine to deal with this type of queries. Moreover, the emergence of social me-dias allows users to get opinions, suggestions, or recommendations from other users about complex information needs. In order to increase the understanding of user needs, a tasks, as the CLEF Social Book Search Classification Track, the aims is to investigates how systems can automatically identify book search requests in online forums. In this respect, we introduce in the present paper a new approach to automatically detect the type of each thread. Our proposal aims to identify book search queries by syntactic patterns, association rules between terms and tex-tual sequences mining.

International audience; Handling long queries can involve either reducing its size by retaining only useful sentences, or decomposing the long query into several short queries based on their content. A proper sentence classification improves the utility of these procedures. Can Sentiment Analysis has a role in sentence classification? This paper analysis the correlation between sentiment analysis and sentence classification in long book-search queries. Also, it studies the similarity in writing style between book reviews and sentences in book-search queries. To accomplish this study, a semi-supervised method for sentiment intensity prediction, and a language model based on book reviews are presented. In addition to graphical illustrations reflecting the feedback of this study, followed by interpretations and conclusions.

International audience; This paper proposes an approach on Facebook Search in Arabic language, which exploits all traces of users (e.g. polarity, share, like, haha) left on Facebook publications to estimate their social importance. Our goal is to show how these signals can play a vital role in improving Arabic Facebook search. First, the identification of the polarities (positive or negative) carried by the textual signals (e.g. comments) and non-textual ones (e.g. the reactions love and sad) for a given Facebook publication. Therefore, the polarity of each comment expressed on a given Facebook post, is estimated on the basis of a neural model of sentiment in Arabic language. Secondly, grouping signals according to their complementarity using selection algorithms. Third, the application of learning to rank algorithms to re-rank Facebook search results based on the selected groups of signals. Finally, experiments are carried out on 13500 Face-book posts, collected from 45 topics in Arabic language. Experiments reveal promising results.; Cet article propose une approche de recherche d'information (RI) en langue arabe sur Facebook, qui exploite toutes les traces des utilisateurs (ex. polarité, partage, j'aime, haha) laissées sur des publications Facebook pour estimer leur importance sociale. Notre objectif est de montrer comment ces signaux peuvent jouer un rôle vital dans l'amélioration de la recherche en langue arabe sur Facebook. Premièrement, des polarités (positive ou négative) portée par les signaux textuels (ex. commentaires) et non textuels (ex. les réactions j'adore et triste) ont été identifiées pour chaque publication Facebook. Par conséquent, la polarité de chaque com-mentaire exprimé sur une publication donnée, est estimée sur la base d'un modèle neuronal de sentiment en langue arabe. Deuxièmement, des signaux en fonction de leur complémentarité ont été regroupés en utilisant des algorithmes de sélection. Troisièmement, des algorithmes de learning to rank ont été appliqués pour re-ordonner les résultats de recherche de Facebook en fonction des groupes de signaux sélectionnés. Enfin, des expérimentations sont réalisées sur 13500 publications Facebook, collectées à partir de 45 requêtes en langue arabe. Les expéri-mentations révèlent des résultats prometteurs pour la RI en langue arabe sur Facebook.

We present a novel functional magnetic resonance imaging paradigm for second-person neuroscience. The paradigm compares a human social interaction (human–human interaction, HHI) to an interaction with a conversational robot (human–robot interaction, HRI). The social interaction consists of 1 min blocks of live bidirectional discussion between the scanned participant and the human or robot agent. A final sample of 21 participants is included in the corpus comprising physiological (blood oxygen level-dependent, respiration and peripheral blood flow) and behavioural (recorded speech from all interlocutors, eye tracking from the scanned participant, face recording of the human and robot agents) data. Here, we present the first analysis of this corpus, contrasting neural activity between HHI and HRI. We hypothesized that independently of differences in behaviour between interactions with the human and robot agent, neural markers of mentalizing (temporoparietal junction (TPJ) and medial prefrontal cortex) and social motivation (hypothalamus and amygdala) would only be active in HHI. Results confirmed significantly increased response associated with HHI in the TPJ, hypothalamus and amygdala, but not in the medial prefrontal cortex. Future analysis of this corpus will include fine-grained characterization of verbal and non-verbal behaviours recorded during the interaction to investigate their neural correlates. This article is part of the theme issue ‘From social brains to social robots: applying neurocognitive insights to human–robot interaction'.

International audience; This is a report on the ninth edition of the Conference and Labs of the Evaluation Forum (CLEF 2018), held in early September 2018, in Avignon, France. CLEF was a four day event combining a Conference and an Evaluation Forum. The Conference featured keynotes by Nicholas Belkin, Julio Gonzalo, and Gabriella Pasi, and presentation of 29 peer reviewed research papers covering a wide range of topics in addition to many posters. The Evaluation Forum consisted to ten Labs: CENTRE, CheckThat, DynSe, eRisk, eHealth, ImageCLEF, LifeCLEF, Cultural Microblog Contextualization, PAN, and PIR-CLEF, addressing a wide range of tasks, media, languages, and ways to go beyond standard test collections.

Mathematics learning has become one of the most researched fields in education. Particularly, word or story problem solving skills have been gaining an enormous amount of attention from researchers and practitioners. Within this context, several studies have been done in order to analyze the impact that serious games have on learning processes and, in particular, on the development of word problem solving skills. However, little is known regarding how games may influence student acquisition of the process skills of problem solving. In a first attempt, this theoretical paper deals with word problem solving skill enhancement in second-grade school children by means of a practical educational serious game that addresses general and specific abilities involved in problem solving, focusing on how different parts of a solution effort relate to each other. The serious game is based on Polya’s problem solving model. The emphasis of using the specific model was on dividing the problem solving procedure into stages and the concentration on the essential details of a problem solving process and the relationships between the various parts of the solution.

Relation Extraction (RE) consists in detecting and classifying semantic relations between entities in a sentence. The vast majority of the state-of-the-art RE systems relies on morphosyntactic features and supervised machine learning algorithms. This paper tries to answer important questions concerning both the impact of semantic based features, and the integration of external linguistic knowledge resources on RE performance. For that, a RE system based on a logical and relational learning algorithm was used and evaluated on three reference datasets from two distinct domains. The yielded results confirm that the classifiers induced using the proposed richer feature set outperformed the classifiers built with morphosyntactic features in average 4% (F1-measure).

​Consumers are used to consulting posted reviews on the Internet before buying a product. But it’s difficult to know the global opinion considering the important number of those reviews. Sentiment analysis afford detecting polarity (positive, negative, neutral) in a expressed opinion and therefore classifying those reviews. Our purpose is to determine the influence of emotions on the polarity of books reviews. We define “bag-of-words” representation models of reviews which use a lexicon containing emotional (anticipation, sadness, fear, anger, joy, surprise, trust, disgust) and sentimental (positive, negative) words. This lexicon afford measuring felt emotions types by readers. The implemented supervised learning used is a Random Forest type. The application concerns Amazon platform’s reviews., Les consommateurs ont l'habitude de consulter les critiques postées sur internet avant d'acheter un produit. Mais, il est difficile pour le consommateur de connaître l'opinion globale du produit vu le nombre important de ces critiques. L'analyse des sentiments permet de détecter la polarité (positive, négative ou neutre) sur une opinion exprimée et donc de classer ces critiques. Notre but est de déterminer l'influence de l'expression des émotions sur l'analyse de la polarité des critiques de livres. Nous définissons des modèles de représentation "sac de mots" de critiques qui s'appuient sur un lexique contenant des mots porteurs de sentiments (positif, négatif) et d'émotions (anticipation, tristesse, peur, colère, joie, surprise, confiance, dégoût). Ce lexique permet de mesurer les types d'émotions ressenties par les lecteurs. L'apprentissage supervisé mis en oeuvre est de type forêt aléatoire (Random Forest). L'application concerne des critiques de la plateforme Amazon.

International audience; This paper proposes an approach on Facebook Search in Arabic language, which exploits all traces of users (e.g. polarity, share, like, haha) left on Facebook publications to estimate their social importance. Our goal is to show how these signals can play a vital role in improving Arabic Facebook search. First, the identification of the polarities (positive or negative) carried by the textual signals (e.g. comments) and non-textual ones (e.g. the reactions love and sad) for a given Facebook publication. Therefore, the polarity of each comment expressed on a given Facebook post, is estimated on the basis of a neural model of sentiment in Arabic language. Secondly, grouping signals according to their complementarity using selection algorithms. Third, the application of learning to rank algorithms to re-rank Facebook search results based on the selected groups of signals. Finally, experiments are carried out on 13500 Face-book posts, collected from 45 topics in Arabic language. Experiments reveal promising results.; Cet article propose une approche de recherche d'information (RI) en langue arabe sur Facebook, qui exploite toutes les traces des utilisateurs (ex. polarité, partage, j'aime, haha) laissées sur des publications Facebook pour estimer leur importance sociale. Notre objectif est de montrer comment ces signaux peuvent jouer un rôle vital dans l'amélioration de la recherche en langue arabe sur Facebook. Premièrement, des polarités (positive ou négative) portée par les signaux textuels (ex. commentaires) et non textuels (ex. les réactions j'adore et triste) ont été identifiées pour chaque publication Facebook. Par conséquent, la polarité de chaque com-mentaire exprimé sur une publication donnée, est estimée sur la base d'un modèle neuronal de sentiment en langue arabe. Deuxièmement, des signaux en fonction de leur complémentarité ont été regroupés en utilisant des algorithmes de sélection. Troisièmement, des algorithmes de learning to rank ont été appliqués pour re-ordonner les résultats de recherche de Facebook en fonction des groupes de signaux sélectionnés. Enfin, des expérimentations sont réalisées sur 13500 publications Facebook, collectées à partir de 45 requêtes en langue arabe. Les expéri-mentations révèlent des résultats prometteurs pour la RI en langue arabe sur Facebook.

International audience

International audience; In human-agent interaction, one key challenge is the evaluation of the user's experience. In the virtual reality domain, the sense of presence and co-presence, reflecting the psychological immersion of the user, is generally assessed through well-grounded subjective post-experience questionnaires. In this article, we aim at presenting a new way to automatically predict the sense of presence and co-presence of a user at the end of an interaction based on specific verbal and non-verbal behavioral cues automatically computed. A random forest algorithm has been applied on a human-agent interaction corpus collected in the specific context of a virtual environment developed to train doctors to break bad news to a virtual patient. The performance of the models demonstrate the capacity to automatically and accurately predict the level of presence and co-presence, but also show the relevancy of the verbal and non-verbal behavioral cues as objective measures of presence.

International audience