Social media text analytics of Malayalam-English code-mixed using deep learning

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Social Media Analysis

Background:

• Social media text, often code-mixed, presents significant challenges for automated analysis due to informal language and vocabulary.
• Existing methods struggle with the nuances of Malayalam-English code-mixed data, necessitating specialized approaches for tasks like offensive language identification and sentiment analysis.

Purpose of the Study:

• To develop and evaluate a framework for offensive language identification and sentiment analysis on Malayalam-English code-mixed text.
• To investigate the impact of embedding methods, deep learning architectures, and translation techniques on model performance.

Main Methods:

• Utilized Word2Vec and FastText for feature engineering, exploring dependencies among embeddings.
• Compared various deep learning models, including uni-directional, bi-directional, hybrid, and transformer approaches.
• Incorporated selective translation and transliteration, alongside hyper-parameter optimization.

Main Results:

• Achieved high F1-Scores: 0.76 for the Forum for Information Retrieval Evaluation (FIRE) 2020 dataset and 0.99 for the European Chapter of the Association for Computational Linguistics (EACL) 2021 dataset.
• The proposed strategy outperformed benchmarked models for Malayalam-English code-mixed messages.
• Detailed error analysis provided insights into model limitations and areas for improvement.

Conclusions:

• The developed framework demonstrates significant effectiveness in analyzing Malayalam-English code-mixed text for critical NLP tasks.
• This work represents a substantial advancement in processing informal, multilingual social media data, contributing to societal well-being.
• The findings highlight the importance of tailored methods for handling code-mixed language in digital communication.