Deep learning for religious and continent-based toxic content detection and classification.

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning (ML)
• Computational Linguistics

Background:

• Online platforms facilitate expression, leading to increased toxic language like racism and sexual harassment.
• Accurate toxic language identification is crucial for online safety and ethical AI.
• Existing machine learning models sometimes misclassify non-toxic comments containing identity terms (e.g., religious or racial groups) as toxic.

Purpose of the Study:

• To analyze and compare modern deep learning algorithms for multilabel toxic comment classification.
• To evaluate performance in classifying religious and race/ethnicity-based toxic comments.
• To investigate the impact of various word embeddings (GloVe, Word2vec, FastText) versus no embeddings on classification accuracy.

Main Methods:

• Utilized deep learning models for multilabel classification of toxic comments.
• Conducted experiments on two scenarios: religious toxicity and race/ethnicity toxicity.
• Compared performance using different word embedding techniques (GloVe, Word2vec, FastText) and an ordinary embedding layer.

Main Results:

• The Convolutional Neural Network (CNN) model achieved the best performance in both classification scenarios.
• Performance was evaluated using standard multilabel classification metrics.
• The study demonstrated CNN's effectiveness in handling nuanced toxic language detection.

Conclusions:

• CNN models are highly effective for multilabel toxic comment classification, particularly for religious and race/ethnicity-based toxicity.
• The choice of word embeddings can influence model performance, but CNNs show robust results.
• Further research can build upon these findings for more accurate and equitable AI moderation systems.