Titre :	A Deep Learning Approach to Detecting Offensive Tweets
Type de document :	texte imprimé
Auteurs :	feriel Lafi, Auteur ; Assala Bennour, Auteur ; Lakhfif, Abdelaziz, Directeur de thèse
Année de publication :	2023
Importance :	1 vol (63 f .)
Format :	29cm
Langues :	Français (fre)
Catégories :	Thèses & Mémoires:Informatique
Mots-clés :	detecting offensive language Natural Language Processing
Index. décimale :	004 Informatique
Résumé :	Social media platforms, such as Twitter, have become integral to our daily lives, providing a wide range of services. However, these platforms also expose users to the risk of cyberbullying, which can result in mental and emotional abuse. Pre-trained transformers have become the standard models in natural language processing due to their outstanding performance across various tasks and languages. However, the majority of these models have been trained in languages that already possess abundant text resources, such as English, French, and others. As a result, there is a need for increased focus on low-resource languages that have limited text data, such as Arabic, particularly the Algerian dialect. In this thesis, our primary focus is on developping an automated solution for detecting offensive language in Arabic tweets. We used deep learning algorithms, including MarBERT, DziriBert, and Arabert transformers. To train and evaluate the models, we utilized the OSACT 2022 and DziriOFN datasets. Through experiments, we explore the impact of emojis on model performance and the need for additional training on the Algerian dialect. Our findings demonstrate that emojis provide valuable contextual information, improving the models understanding and classification capabilities. TheMarBERT v2 model achieves an impressive F1 score of 85.1%, while DziriBert excels in detecting offensive language in the Algerian dialect with an F1 score of 87%. These results emphasize the importance of dialect-specific training.
Côte titre :	MAI/0706
En ligne :	https://drive.google.com/file/d/1y5pgvD9jTrDs-bysqsBvcmKQScebxK74/view?usp=drive [...]
Format de la ressource électronique :	pdf

A Deep Learning Approach to Detecting Offensive Tweets [texte imprimé] / feriel Lafi, Auteur ; Assala Bennour, Auteur ; Lakhfif, Abdelaziz, Directeur de thèse . - 2023 . - 1 vol (63 f .) ; 29cm.
Langues : Français (fre)

Catégories :	Thèses & Mémoires:Informatique
Mots-clés :	detecting offensive language Natural Language Processing
Index. décimale :	004 Informatique
Résumé :	Social media platforms, such as Twitter, have become integral to our daily lives, providing a wide range of services. However, these platforms also expose users to the risk of cyberbullying, which can result in mental and emotional abuse. Pre-trained transformers have become the standard models in natural language processing due to their outstanding performance across various tasks and languages. However, the majority of these models have been trained in languages that already possess abundant text resources, such as English, French, and others. As a result, there is a need for increased focus on low-resource languages that have limited text data, such as Arabic, particularly the Algerian dialect. In this thesis, our primary focus is on developping an automated solution for detecting offensive language in Arabic tweets. We used deep learning algorithms, including MarBERT, DziriBert, and Arabert transformers. To train and evaluate the models, we utilized the OSACT 2022 and DziriOFN datasets. Through experiments, we explore the impact of emojis on model performance and the need for additional training on the Algerian dialect. Our findings demonstrate that emojis provide valuable contextual information, improving the models understanding and classification capabilities. TheMarBERT v2 model achieves an impressive F1 score of 85.1%, while DziriBert excels in detecting offensive language in the Algerian dialect with an F1 score of 87%. These results emphasize the importance of dialect-specific training.
Côte titre :	MAI/0706
En ligne :	https://drive.google.com/file/d/1y5pgvD9jTrDs-bysqsBvcmKQScebxK74/view?usp=drive [...]
Format de la ressource électronique :	pdf