https:\/\/teams.microsoft.com\/l\/meetup-join\/19%3ameeting_MDA4MWE2MzctMDBkMS00Njc3LWE3YTEtZWQ4Yzk0OTU1ODA3%40thread.v2\/0?context=%7b%22Tid%22%3a%222b1be215-5cf1-43de-8bc7-946bdf57f9cd%22%2c%22Oid%22%3a%222d076f5a-9749-47e3-a73e-33e405eb993a%22%7d<\/a><\/p>\n\n\n\nMembres du jury :<\/strong><\/p>\n\n\n\nM. Noel CRESPI<\/strong>, Professeur, T\u00e9l\u00e9com SudParis, FRANCE – Directeur de these
M. Reza FARAHBAKHSH<\/strong>, Ma\u00eetre de conf\u00e9rences, T\u00e9l\u00e9com SudParis, FRANCE – Co-encadrant de these
M. Christophe CERISARA<\/strong>, Charg\u00e9 de recherche, CNRS – LORIA laboratory in Nancy, FRANCE – Examinateur
Mme Ioana MANOLESCU<\/strong>, Directrice de recherche, Inria , FRANCE – Rapporteur
Mme Tiziana MARGARIA<\/strong>, Professeure, University of Limerick, IRLANDE – Rapporteur
Mme Gabriella PASI<\/strong>, Professeure, University of Milano-Bicocca, ITALIE – Examinateur<\/p>\n\n\n\n\u00ab Am\u00e9lioration de la d\u00e9tection des fausses informations : des approches hybrides et multilingues \u00e0 l\u2019augmentation de donn\u00e9es bas\u00e9e sur les LLM et l\u2019analyse des biais \u00bb<\/h2>\n\n\n\npr\u00e9sent\u00e9 par Madame Razieh CHALEHCHALEH<\/h2>\n\n\n\n
R\u00e9sum\u00e9 :<\/strong><\/p>\n\n\n\nLa prolif\u00e9ration des fausses informations sur les plateformes en ligne repr\u00e9sente une menace critique pour les individus et les soci\u00e9t\u00e9s du monde entier. Cette th\u00e8se fait progresser la d\u00e9tection automatique des fausses informations gr\u00e2ce \u00e0 quatre contributions compl\u00e9mentaires qui r\u00e9pondent aux limites fondamentales des approches actuelles : la d\u00e9pendance aux seules caract\u00e9ristiques de contenu, la recherche centr\u00e9e sur l’anglais, la raret\u00e9 des donn\u00e9es et les biais inh\u00e9rents aux grands mod\u00e8les de langage (LLM). Premi\u00e8rement, nous introduisons BRaG, un nouveau cadre hybride qui int\u00e8gre des repr\u00e9sentations textuelles bas\u00e9es sur BERT, une mod\u00e9lisation par r\u00e9seaux de neurones r\u00e9currents des s\u00e9quences d’engagement des utilisateurs, et un encodage par r\u00e9seaux de neurones sur graphes des graphes de propagation. En exploitant conjointement le contenu et le contexte social, BRaG surpasse les mod\u00e8les de r\u00e9f\u00e9rence unimodaux. Deuxi\u00e8mement, nous \u00e9tudions la d\u00e9tection multilingue des fausses informations en \u00e9valuant des mod\u00e8les multilingues (mBERT, XLM-RoBERTa, LASER-LR) dans des sc\u00e9narios d’entra\u00eenement monolingues, multilingues et interlingues (zero-shot cross-lingual), et en les comparant \u00e0 des strat\u00e9gies bas\u00e9es sur la traduction. Les r\u00e9sultats soulignent l’efficacit\u00e9 des mod\u00e8les et des donn\u00e9es multilingues, offrant des perspectives pour les contextes de langues peu dot\u00e9es. Troisi\u00e8mement, nous explorons l’augmentation de donn\u00e9es bas\u00e9e sur les LLM en utilisant Llama 3. Nous examinons le prompting zero-shot et few-shot, des taux d’augmentation vari\u00e9s, l’augmentation sp\u00e9cifique aux classes, ainsi que le sous-\u00e9chantillonnage al\u00e9atoire par rapport \u00e0 celui bas\u00e9 sur la similarit\u00e9, afin de g\u00e9n\u00e9rer des donn\u00e9es synth\u00e9tiques pour des classifieurs bas\u00e9s sur BERT. Avec des configurations appropri\u00e9es, les donn\u00e9es augment\u00e9es permettent d’obtenir des am\u00e9liorations par rapport aux r\u00e9f\u00e9rences multilingues. Enfin, nous menons la premi\u00e8re \u00e9tude syst\u00e9matique des biais de genre dans l’annotation de fausses informations par des LLM. En augmentant le jeu de donn\u00e9es LIAR avec des variantes genr\u00e9es des intitul\u00e9s de poste des locuteurs et en \u00e9valuant six LLM de l’\u00e9tat de l’art \u00e0 travers plusieurs m\u00e9triques d’\u00e9quit\u00e9, nous constatons un comportement sensible au genre constant. Cela se manifeste par deux formes de biais : l’instabilit\u00e9 (jugements incoh\u00e9rents) et la directionnalit\u00e9 (diff\u00e9rences syst\u00e9matiques entre les genres). Ces r\u00e9sultats montrent que les jugements de v\u00e9racit\u00e9 rendus par les LLM sont influenc\u00e9s par la pr\u00e9sentation du genre, soulignant la n\u00e9cessit\u00e9 de strat\u00e9gies d’att\u00e9nuation conscientes des biais. \u00c0 travers des exp\u00e9riences approfondies sur des jeux de donn\u00e9es r\u00e9els, cette th\u00e8se fournit de nouvelles m\u00e9thodologies et des perspectives qui am\u00e9liorent les syst\u00e8mes de d\u00e9tection automatique de fausses informations.<\/p>\n\n\n\n
Abstract :<\/strong><\/p>\n\n\n\nThe proliferation of fake news on online platforms poses critical threats to individuals and societies worldwide. This thesis advances automated fake news detection through four complementary contributions that address key limitations in current approaches: reliance on content-only features, English-centric research, data scarcity, and inherent bias in Large Language Models (LLMs). First, we introduce BRaG, a novel hybrid framework that integrates BERT-based text representations, recurrent neural network modeling of user engagement sequences, and graph neural network encoding of propagation graphs. By jointly leveraging content and social context features, BRaG outperforms single-modality baselines. Second, we investigate multilingual fake news detection, evaluating multilingual models (mBERT, XLM-RoBERTa, LASER-LR) under monolingual, multilingual, and zero-shot cross-lingual training scenarios, and comparing them with translation-based strategies. The results highlight the effectiveness of multilingual models and training data, offering insights for low-resource language settings. Third, we explore LLM-based data augmentation using Llama 3, examining zero-shot and few-shot prompting, varying augmentation rates, class-specific augmentation, and random vs. similarity-based subsampling to generate synthetic training data for BERT-based classifiers. With appropriate configurations, augmented data yields improvements over the multilingual baselines. Finally, we conduct the first systematic investigation of gender bias in LLM-based fake news annotation. By augmenting the LIAR dataset with gender-variant versions of speaker job titles and evaluating six state-of-the-art LLMs across multiple fairness metrics, we find consistent gender-sensitive behavior resulting in two bias manifestations\u2014instability (inconsistent judgments) and directionality (systematic differences across genders). These results show that LLM-driven veracity judgments are influenced by gender presentation, underscoring the need for bias-aware mitigation strategies. Through extensive experiments on real-world datasets, this thesis provides novel methodologies and insights that enhance automated fake news detection systems.<\/p>\n","protected":false},"excerpt":{"rendered":"
L’Ecole doctorale : Ecole Doctorale de l’Institut Polytechnique de Paris et le Laboratoire de recherche SAMOVAR – Services r\u00e9partis, Architectures, Mod\u00e9lisation, Validation, Administration des R\u00e9seaux pr\u00e9sentent l\u2019AVIS DE SOUTENANCE de Madame Razieh CHALEHCHALEH Autoris\u00e9e \u00e0 pr\u00e9senter ses travaux en vue de l\u2019obtention du Doctorat de l’Institut Polytechnique de Paris, pr\u00e9par\u00e9 \u00e0 T\u00e9l\u00e9com SudParis en : […]<\/p>\n","protected":false},"author":4,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"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-7039","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\/7039","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=7039"}],"version-history":[{"count":1,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts\/7039\/revisions"}],"predecessor-version":[{"id":7040,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts\/7039\/revisions\/7040"}],"wp:attachment":[{"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/media?parent=7039"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/categories?post=7039"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/tags?post=7039"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}