Transfer Learning and Deep Learning for Multilingual Algerian Dialect Hate Speech Detection

Abstract

Social network have a great impact not only on information availability but also on its generation by enabling users to add their own contents to web without having any specific programming skills. These new communication services make human communication easier and people closer to each other. Despite to their uncountable benefits, social media have their negative implications due to their misuse. Abusive behaviors like Hate speech and all other shapes of toxicity have increasingly taken place in these communication means making of online environments more and more toxic. The aim of this project is dedicated, first to hate speech detection using Deep learning and Transfer learning models applied on Algerian dialect written in multilingual language, second to create a new dataset related to this problematic. In this study, we test several models, Models of machine learning With TF-IDF and With Word2Vec embedding (Randomforestclassifier / MultinomialNB / Logistic Regression / SGD Classifier/ SVC classifier), Deep learning models Without / With FastText (LSTM/ GRU/ Bi-LSTM/ Bi-GRU/ CNN), Transfer learning models (Distilbert-base-uncased/ Bert-base-multilingual-uncased/ Bert-base-arabic), Bert layer with GRU and LSTM (BERT embedding +LSTM/ BERT embedding +GRU/ BERT embedding+GRU and LSTM) and Ensemble learning to improve an automated detection of Algerian dialect written in multilingual language. The best models is transfer learning (Bert layer with GRU )F1-Score 86 % accuracy 87% afetr Deep learning models ( LSTM )F1-Score 87% accuracy 85%.