{"id":984,"date":"2021-12-15T08:02:18","date_gmt":"2021-12-15T07:02:18","guid":{"rendered":"https:\/\/pageperso.lis-lab.fr\/bernard.espinasse\/nouveausite\/?page_id=984"},"modified":"2024-01-08T15:10:59","modified_gmt":"2024-01-08T14:10:59","slug":"these-alaeddine-moussa","status":"publish","type":"page","link":"https:\/\/pageperso.lis-lab.fr\/bernard.espinasse\/index.php\/these-alaeddine-moussa\/","title":{"rendered":"These Alaeddine Moussa"},"content":{"rendered":"

\"\"<\/strong>Alaeddine Moussa<\/strong>, \u00ab Etiquetage de r\u00f4les spatiaux par apprentissage profond bas\u00e9 sur une repr\u00e9sentation vectorielle enrichie \u00bb. Th\u00e8se en cotutelle entre Aix-Marseille Universit\u00e9 et l’Universit\u00e9 de la Manouba (ENSI), Tunisie, soutenue le 5 d\u00e9cembre 2023 \u00e0 Tunis.<\/p>\n

Manuscrit<\/a><\/span><\/p>\n

R\u00e9sum\u00e9<\/strong><\/span><\/p>\n

L’une des fonctions essentielles du langage naturel concerne l\u2019\u00e9vocation de relations spatiales entre objets. Des constructions linguistiques peuvent notamment exprimer des relations spatiales entre objets ainsi que des mod\u00e8les de mouvement de ces objets dans l’espace. La compr\u00e9hension de ces \u00e9nonc\u00e9s spatiaux est un probl\u00e8me majeur dans de nombreux domaines, comme la robotique, la navigation, la gestion du trafic et les syst\u00e8mes de r\u00e9ponse aux requ\u00eates. L\u2019\u00e9tiquetage des r\u00f4les spatiaux (Spatial Role Labelling \u2013 SpRL en anglais), propose des sch\u00e9mas d’annotation ind\u00e9pendant de la langue consistant en un ensemble de r\u00f4les spatiaux dans le but de couvrir tous les aspects des concepts spatiaux notamment les relations spatiales statiques et dynamiques. La plupart des syst\u00e8mes automatiques de SpRL permettant d\u2019extraire automatiquement les r\u00f4les spatiaux d\u2019un texte sont des syst\u00e8mes bas\u00e9s sur des m\u00e9thodes d\u2019apprentissage traditionnelles, principalement statistiques. Dans le cadre de cette th\u00e8se nous \u00e0 l\u2019extraction automatiquement de ces r\u00f4les spatiaux par l\u2019apprentissage profond (Deep Learning). Apr\u00e8s un \u00e9tat de l\u2019art sur d’\u00e9tiquetage automatique de r\u00f4les spatiaux (SpRL), les syst\u00e8mes automatique SpRL existants, et les corpus de r\u00e9f\u00e9rences permettant de les \u00e9valuer, nous proposons plusieurs syst\u00e8mes SpRL \u00e0 base d\u2019apprentissage profond que nous \u00e9valuons sur ces corpus de r\u00e9f\u00e9rences et comparons leurs performances avec celles d\u2019autres syst\u00e8mes existants. Le premier syst\u00e8me propos\u00e9 s\u2019appuie sur une repr\u00e9sentation vectorielle du texte \u00e0 analyser en utilisant des vecteurs de mots avec des balises POS et des repr\u00e9sentations au niveau des caract\u00e8res bas\u00e9es sur CNN et enfin un mod\u00e8le d’apprentissage profond BiLSTM-CRF pour identifier les r\u00f4les spatiaux. Le deuxi\u00e8me syst\u00e8me propos\u00e9 utilise une repr\u00e9sentation vectorielle du texte obtenu par un plongement de mots sp\u00e9cifique, alternatif au mod\u00e8le sac de mots appliqu\u00e9 aux plongements de mots classique et permettant de prendre en compte la syntaxe du texte dans la repr\u00e9sentation vectorielle. Le dernier syst\u00e8me adopte une approche bas\u00e9e sur les \u00ab Transformer \u00bb mettant en \u0153uvre un m\u00e9canisme d\u2019attention permettant de tenir compte de la combinaison de tous les mots du contexte, en pond\u00e9rant chacun d\u2019entre eux. Les mod\u00e8les de plongement de mots obtenus permettent de cr\u00e9er des mod\u00e8les \u00ab contextualis\u00e9s \u00bb notamment avec BERT (Bidirectional Encoder Representations from Transformers) capables de produire des repr\u00e9sentations de mots qui d\u00e9pendent du contexte.<\/span><\/p>\n

Abstract<\/strong><\/em><\/span><\/p>\n

One of the essential functions of natural language concerns the evocation of spatial relations between objects. In particular, linguistic constructs can express spatial relationships between objects as well as patterns of movement of these objects in space. Understanding these spatial statements is a major problem in many domains, such as robotics, navigation, traffic management, and query response systems. Spatial Role Labeling (SpRL) provides language-independent annotation schemes consisting of a set of spatial roles in order to cover all aspects of spatial concepts including static and dynamic spatial relationships. Most automatic SpRL systems that automatically extract spatial roles from a text are systems based on traditional, mainly statistical, learning methods. In this thesis, we focus on the automatic extraction of these spatial roles using Deep Learning. After a state of the art on automatic Spatial Role Labeling (SpRL), existing automatic SpRL systems, and reference corpora allowing to evaluate them, we propose several deep learning based SpRL systems that we evaluate on these reference corpora and compare their performances with those of other existing systems. The first proposed system relies on a vector representation of the text to be analyzed using word vectors with POS tags and character-level representations based on CNN and finally a BiLSTM-CRF deep learning model to identify spatial roles. The second proposed system uses a vector representation of the text obtained by a specific word embedding, as an alternative to the bag-of-words model applied to classical word embeddings and allowing to take into account the syntax of the text in the vector representation. The last system adopts a \u00ab\u00a0Transformer\u00a0\u00bb based approach implementing an attention mechanism allowing to take into account the combination of all the words in the context, by weighting each of them. The resulting word embedding models allow the creation of \u00ab\u00a0contextualized\u00a0\u00bb models, in particular with BERT (Bidirectional Encoder Representations from Transformers) capable of producing context-dependent word representations.<\/em><\/span><\/p>\n","protected":false},"excerpt":{"rendered":"

Alaeddine Moussa, \u00ab Etiquetage de r\u00f4les spatiaux par apprentissage profond bas\u00e9 sur une repr\u00e9sentation vectorielle enrichie \u00bb. Th\u00e8se en cotutelle entre Aix-Marseille Universit\u00e9 et l’Universit\u00e9 de la Manouba (ENSI), Tunisie, soutenue le 5 d\u00e9cembre 2023 \u00e0 Tunis. Manuscrit R\u00e9sum\u00e9 L’une des fonctions essentielles du langage naturel concerne l\u2019\u00e9vocation de relations spatiales entre objets. Des constructions … <\/p>\n

Continuer la lecture \u00ab\u00a0These Alaeddine Moussa\u00a0\u00bb<\/span><\/a><\/p>\n","protected":false},"author":2,"featured_media":0,"parent":0,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","meta":{"_crdt_document":"","footnotes":""},"class_list":["post-984","page","type-page","status-publish","hentry","entry"],"_links":{"self":[{"href":"https:\/\/pageperso.lis-lab.fr\/bernard.espinasse\/index.php\/wp-json\/wp\/v2\/pages\/984","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/pageperso.lis-lab.fr\/bernard.espinasse\/index.php\/wp-json\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/pageperso.lis-lab.fr\/bernard.espinasse\/index.php\/wp-json\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/pageperso.lis-lab.fr\/bernard.espinasse\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/pageperso.lis-lab.fr\/bernard.espinasse\/index.php\/wp-json\/wp\/v2\/comments?post=984"}],"version-history":[{"count":10,"href":"https:\/\/pageperso.lis-lab.fr\/bernard.espinasse\/index.php\/wp-json\/wp\/v2\/pages\/984\/revisions"}],"predecessor-version":[{"id":1708,"href":"https:\/\/pageperso.lis-lab.fr\/bernard.espinasse\/index.php\/wp-json\/wp\/v2\/pages\/984\/revisions\/1708"}],"wp:attachment":[{"href":"https:\/\/pageperso.lis-lab.fr\/bernard.espinasse\/index.php\/wp-json\/wp\/v2\/media?parent=984"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}