{"id":7728,"date":"2026-11-19T08:00:26","date_gmt":"2026-11-19T07:00:26","guid":{"rendered":"https:\/\/samovar.telecom-sudparis.eu\/?p=7728"},"modified":"2026-11-19T08:00:26","modified_gmt":"2026-11-19T07:00:26","slug":"avis-de-soutenance-de-madame-khouloud-mnassri","status":"publish","type":"post","link":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/2026\/11\/19\/avis-de-soutenance-de-madame-khouloud-mnassri\/","title":{"rendered":"AVIS DE SOUTENANCE de Madame Khouloud MNASSRI"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">L&rsquo;Ecole doctorale : Ecole Doctorale de l&rsquo;Institut Polytechnique de Paris<br><br>et le Laboratoire de recherche SAMOVAR &#8211; Services r\u00e9partis, Architectures, Mod\u00e9lisation, Validation, Administration des R\u00e9seaux<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">pr\u00e9sentent<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">l\u2019AVIS DE SOUTENANCE de Madame Khouloud MNASSRI<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Autoris\u00e9e \u00e0 pr\u00e9senter ses travaux en vue de l\u2019obtention du Doctorat de l&rsquo;Institut Polytechnique de Paris, pr\u00e9par\u00e9 \u00e0 l&rsquo;Institut Polytechnique de Paris T\u00e9l\u00e9com SudParis en :<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Math\u00e9matiques et Informatique<\/h2>\n\n\n\n<h1 class=\"wp-block-heading\">\u00ab Compr\u00e9hension du langage subjectif en traitement automatique des langues : de l\u2019acquisition des connaissances \u00e0 leur r\u00e9utilisation gouvern\u00e9e \u00bb<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">le VENDREDI 27 NOVEMBRE 2026 \u00e0 9h30<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u00e0<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Salle E206<br>T\u00e9l\u00e9com SudParis, 9 rue Charles Fourier, 91011 \u00c9vry-Courcouronnes<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Membres du jury :<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>M. Noel&nbsp;CRESPI<\/strong>, Professeur, Institut Polytechnique de Paris T\u00e9l\u00e9com SudParis, FRANCE &#8211; Directeur de th\u00e8se<br><strong>M. Gareth&nbsp;TYSON<\/strong>, Professor, University of Oxford, ROYAUME-UNI &#8211; Rapporteur<br><strong>M. Rub\u00e9n&nbsp;CUEVAS RUMIN<\/strong>, Associate Professor, Universidad Carlos III de Madrid, ESPAGNE &#8211; Rapporteur<br><strong>Mme Elena&nbsp;CABRIO<\/strong>, Professeure des universit\u00e9s, Universit\u00e9 C\u00f4te d&rsquo;Azur, Inria, CNRS, I3S, FRANCE &#8211; Examinateur<br><strong>Mme Silvia&nbsp;GIORDANO<\/strong>, Full professor, SUPSI &#8211; University of Applied Sciences and Arts of Southern Switzerland, SUISSE &#8211; Examinateur<br><strong>M. Reza&nbsp;FARAHBAKHSH<\/strong>, Associate Professor, Institut Polytechnique de Paris T\u00e9l\u00e9com SudParis,, FRANCE &#8211; Co-encadrant de th\u00e8se<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u00ab Compr\u00e9hension du langage subjectif en traitement automatique des langues : de l\u2019acquisition des connaissances \u00e0 leur r\u00e9utilisation gouvern\u00e9e \u00bb<\/h2>\n\n\n\n<h2 class=\"wp-block-heading\">pr\u00e9sent\u00e9 par Madame Khouloud MNASSRI<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>R\u00e9sum\u00e9 :<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">La compr\u00e9hension du langage subjectif constitue un d\u00e9fi majeur en traitement automatique des langues. Dans l\u2019analyse des sentiments et la d\u00e9tection des discours haineux, l\u2019interpr\u00e9tation d\u00e9pend du contexte linguistique et culturel, suscitant parfois des d\u00e9saccords m\u00eame entre annotateurs. Ces difficult\u00e9s s\u2019accentuent dans les environnements multilingues, o\u00f9 les donn\u00e9es annot\u00e9es sont in\u00e9galement disponibles. L\u2019entra\u00eenement des mod\u00e8les et leur sollicitation r\u00e9p\u00e9t\u00e9e soul\u00e8vent \u00e9galement des questions de passage \u00e0 l\u2019\u00e9chelle. Cette th\u00e8se examine comment acqu\u00e9rir, transf\u00e9rer et r\u00e9utiliser les connaissances, tout en \u00e9tudiant les conditions dans lesquelles les d\u00e9cisions ant\u00e9rieures restent fiables. La premi\u00e8re partie explore des alternatives \u00e0 l\u2019annotation intensive et \u00e0 l\u2019entra\u00eenement sp\u00e9cifique aux t\u00e2ches. Un cadre adversarial g\u00e9n\u00e9ratif semi-supervis\u00e9 exploite des donn\u00e9es non annot\u00e9es pour la d\u00e9tection multilingue des discours haineux, atteignant des performances comparables \u00e0 celles d\u2019une m\u00e9thode de r\u00e9f\u00e9rence avec environ six fois moins de donn\u00e9es annot\u00e9es dans la configuration \u00e9tudi\u00e9e. Si cette approche r\u00e9duit les besoins en annotation, elle n\u00e9cessite toujours un entra\u00eenement sp\u00e9cifique. Cela motive le recours \u00e0 la r\u00e9cup\u00e9ration de connaissances externes au moment de l\u2019inf\u00e9rence. Une base de connaissances multilingue partag\u00e9e combine des donn\u00e9es structur\u00e9es et non structur\u00e9es pour permettre cette approche. D\u2019abord appliqu\u00e9e \u00e0 la d\u00e9tection des discours haineux, elle est ensuite \u00e9tendue \u00e0 l\u2019analyse des sentiments dans douze langues, y compris en configuration zero-shot interlingue, o\u00f9 elle permet des gains de performance substantiels sans entra\u00eenement sp\u00e9cifique \u00e0 la t\u00e2che. Face \u00e0 la r\u00e9currence d\u2019opinions similaires et de contenus haineux, solliciter syst\u00e9matiquement les mod\u00e8les pour les analyser augmente les co\u00fbts d\u2019inf\u00e9rence. La r\u00e9utilisation des d\u00e9cisions ant\u00e9rieures offre une possibilit\u00e9 de r\u00e9duire ces co\u00fbts, mais soul\u00e8ve \u00e9galement une question de fiabilit\u00e9. Les exp\u00e9riences r\u00e9v\u00e8lent que cette r\u00e9utilisation am\u00e9liore les performances dans certaines langues, mais les d\u00e9grade dans d\u2019autres. Dans le langage subjectif, des messages similaires peuvent v\u00e9hiculer des significations diff\u00e9rentes, ce qui rend la seule similarit\u00e9 s\u00e9mantique insuffisante pour d\u00e9cider de leur r\u00e9utilisation. Ce constat motive MAMA, une architecture de m\u00e9moire multi-agents qui d\u00e9termine quand r\u00e9utiliser une d\u00e9cision ant\u00e9rieure et quelles nouvelles connaissances conserver. \u00c9valu\u00e9e sur cinq jeux de donn\u00e9es, MAMA maintient des performances proches de celles du mod\u00e8le de r\u00e9f\u00e9rence tout en r\u00e9duisant jusqu\u2019\u00e0 93 % le co\u00fbt des appels \u00e0 celui-ci. Ensemble, ces contributions font progresser la compr\u00e9hension du langage subjectif en r\u00e9duisant la d\u00e9pendance aux donn\u00e9es annot\u00e9es et en facilitant le transfert des connaissances entre les langues. Elles soulignent \u00e9galement l\u2019importance de gouverner la r\u00e9utilisation des d\u00e9cisions ant\u00e9rieures, ouvrant la voie \u00e0 des syst\u00e8mes de traitement automatique des langues plus accessibles, efficaces et fiables.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Abstract :<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Understanding subjective language is a major challenge in Natural Language Processing. In sentiment analysis and hate speech detection, interpretations depend on linguistic and cultural context, sometimes leading even human annotators to disagree. These difficulties grow in multilingual settings with uneven access to annotated data. Training and repeatedly querying models also raise scalability concerns. This thesis examines how knowledge can be acquired, transferred and reused while assessing when previous decisions remain reliable. The first part explores alternatives to extensive annotation and task-specific training. A semi-supervised generative adversarial framework uses unlabelled data for multilingual hate speech detection, achieving performance comparable to a baseline with approximately one-sixth of the labelled data in the setting studied. Although annotation needs are reduced, task-specific training remains necessary. This motivates a shift towards retrieval, allowing models to access external knowledge at inference time. A shared multilingual knowledge base combines structured and unstructured data to support this process. First applied to hate speech detection, the approach is then extended to sentiment analysis across twelve languages, including a zero-shot cross-lingual setting, where it achieves substantial performance gains without task-specific training. As similar opinions and harmful content recur, repeatedly asking models to assess them increases inference costs. Reusing previous decisions offers an opportunity to reduce this burden, but also raises a question of reliability. Experiments reveal that direct reuse improves performance in some languages while degrading it in others. In subjective language, similar messages may convey different meanings, making semantic similarity an insufficient basis for reuse. This finding motivates MAMA, a multi-agent memory framework that governs both when past decisions can be reused and what new knowledge should be retained. Across five benchmarks, MAMA maintains performance close to the reference model while reducing the cost of invoking it by up to 93%. Together, these contributions advance subjective language understanding by reducing dependence on annotated data and enabling knowledge transfer across languages. They also establish the importance of governing how previous decisions are reused, supporting more accessible, efficient and reliable language processing systems.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>L&rsquo;Ecole doctorale : Ecole Doctorale de l&rsquo;Institut Polytechnique de Paris et le Laboratoire de recherche SAMOVAR &#8211; Services r\u00e9partis, Architectures, Mod\u00e9lisation, Validation, Administration des R\u00e9seaux pr\u00e9sentent l\u2019AVIS DE SOUTENANCE de Madame Khouloud MNASSRI Autoris\u00e9e \u00e0 pr\u00e9senter ses travaux en vue de l\u2019obtention du Doctorat de l&rsquo;Institut Polytechnique de Paris, pr\u00e9par\u00e9 \u00e0 l&rsquo;Institut Polytechnique de Paris [&hellip;]<\/p>\n","protected":false},"author":4,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ocean_front_end_style_editor":"no","ocean_post_layout":"","ocean_both_sidebars_style":"","ocean_both_sidebars_content_width":0,"ocean_both_sidebars_sidebars_width":0,"ocean_sidebar":"","ocean_second_sidebar":"","ocean_disable_margins":"enable","ocean_add_body_class":"","ocean_shortcode_before_top_bar":"","ocean_shortcode_after_top_bar":"","ocean_shortcode_before_header":"","ocean_shortcode_after_header":"","ocean_has_shortcode":"","ocean_shortcode_after_title":"","ocean_shortcode_before_footer_widgets":"","ocean_shortcode_after_footer_widgets":"","ocean_shortcode_before_footer_bottom":"","ocean_shortcode_after_footer_bottom":"","ocean_display_top_bar":"default","ocean_display_header":"default","ocean_header_style":"","ocean_center_header_left_menu":"","ocean_custom_header_template":"","ocean_custom_logo":0,"ocean_custom_retina_logo":0,"ocean_custom_logo_max_width":0,"ocean_custom_logo_tablet_max_width":0,"ocean_custom_logo_mobile_max_width":0,"ocean_custom_logo_max_height":0,"ocean_custom_logo_tablet_max_height":0,"ocean_custom_logo_mobile_max_height":0,"ocean_header_custom_menu":"","ocean_menu_typo_font_family":"","ocean_menu_typo_font_subset":"","ocean_menu_typo_font_size":0,"ocean_menu_typo_font_size_tablet":0,"ocean_menu_typo_font_size_mobile":0,"ocean_menu_typo_font_size_unit":"px","ocean_menu_typo_font_weight":"","ocean_menu_typo_font_weight_tablet":"","ocean_menu_typo_font_weight_mobile":"","ocean_menu_typo_transform":"","ocean_menu_typo_transform_tablet":"","ocean_menu_typo_transform_mobile":"","ocean_menu_typo_line_height":0,"ocean_menu_typo_line_height_tablet":0,"ocean_menu_typo_line_height_mobile":0,"ocean_menu_typo_line_height_unit":"","ocean_menu_typo_spacing":0,"ocean_menu_typo_spacing_tablet":0,"ocean_menu_typo_spacing_mobile":0,"ocean_menu_typo_spacing_unit":"","ocean_menu_link_color":"","ocean_menu_link_color_hover":"","ocean_menu_link_color_active":"","ocean_menu_link_background":"","ocean_menu_link_hover_background":"","ocean_menu_link_active_background":"","ocean_menu_social_links_bg":"","ocean_menu_social_hover_links_bg":"","ocean_menu_social_links_color":"","ocean_menu_social_hover_links_color":"","ocean_disable_title":"default","ocean_disable_heading":"default","ocean_post_title":"","ocean_post_subheading":"","ocean_post_title_style":"","ocean_post_title_background_color":"","ocean_post_title_background":0,"ocean_post_title_bg_image_position":"","ocean_post_title_bg_image_attachment":"","ocean_post_title_bg_image_repeat":"","ocean_post_title_bg_image_size":"","ocean_post_title_height":0,"ocean_post_title_bg_overlay":0.5,"ocean_post_title_bg_overlay_color":"","ocean_disable_breadcrumbs":"default","ocean_breadcrumbs_color":"","ocean_breadcrumbs_separator_color":"","ocean_breadcrumbs_links_color":"","ocean_breadcrumbs_links_hover_color":"","ocean_display_footer_widgets":"default","ocean_display_footer_bottom":"default","ocean_custom_footer_template":"","ocean_post_oembed":"","ocean_post_self_hosted_media":"","ocean_post_video_embed":"","ocean_link_format":"","ocean_link_format_target":"self","ocean_quote_format":"","ocean_quote_format_link":"post","ocean_gallery_link_images":"on","ocean_gallery_id":[],"footnotes":""},"categories":[286,402],"tags":[],"class_list":["post-7728","post","type-post","status-publish","format-standard","hentry","category-fractualites-ennews-fr","category-seminaires-ness-2013-fr","entry"],"_links":{"self":[{"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts\/7728","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/users\/4"}],"replies":[{"embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/comments?post=7728"}],"version-history":[{"count":1,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts\/7728\/revisions"}],"predecessor-version":[{"id":7729,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts\/7728\/revisions\/7729"}],"wp:attachment":[{"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/media?parent=7728"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/categories?post=7728"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/tags?post=7728"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}