An Automated Toxicity Classification on Social Media Using LSTM and Word Embedding.

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Automated toxicity identification is vital for analyzing unfiltered online content.
• Training datasets can introduce bias, leading to inaccurate classification of toxic language.
• Existing methods struggle with nuanced toxicity detection in diverse texts.

Purpose of the Study:

• To enhance text classification accuracy for identifying toxicity.
• To mitigate bias in toxicity detection models.
• To improve the F1-score in classifying toxic comments.

Main Methods:

• Assessed unsupervised methods using state-of-the-art models and external embeddings.
• Implemented a long short-term memory (LSTM) deep learning model with Glove word embeddings.
• Utilized LSTM with word embeddings generated by Bidirectional Encoder Representations from Transformers (BERT).
• Trained and tested models on large, qualitative datasets of classified comments.

Main Results:

• Achieved 94% accuracy and an F1-score of 0.89 using LSTM with BERT word embeddings for binary classification (toxic/nontoxic).
• The LSTM and BERT combination outperformed LSTM alone and LSTM with Glove embeddings.
• Demonstrated improved performance by incorporating high-quality word embeddings from larger text corpora.

Conclusions:

• LSTM combined with BERT word embeddings offers a superior approach for accurate toxicity detection in online comments.
• Leveraging advanced embeddings like BERT can significantly relieve bias and enhance classification metrics.
• Future work should focus on further refining models with extensive, high-quality textual data for robust toxicity analysis.