Spans Detection of Toxic Phrases in Arabic Tweets
Azzam Radman, Mohammed Atros, Rehab Duwairi · 2022
In this paper, we investigate and develop different techniques and deep learning models to detect the spans of characters within an Arabic content that drive a model to classify it as being toxic. Incorporating a model capable of providing such a detailed output into an automated toxicity detection system would significantly reduce the amount of time required to investigate measures taken by automated systems against contents classified as toxic. The dataset used in this study contains 1800 tweets and was originally used for sentiment analysis, however it was re-annotated on the character level to match the requirements of this work. The proposed approach has achieved 0.8289 on the modified F1-score metric and is based on both word2vec word embeddings and BERT-base pooled embeddings. To our knowledge, this is the first effort aiming at approaching this task in Arabic contents.