Deep learning based sentiment analysis and offensive language identification on multilingual code-mixed data.

Area of Science:

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

Background:

• Existing sentiment analysis and offensive language identification research primarily focuses on monolingual datasets.
• Code-mixed data, combining multiple languages, presents unique challenges due to inherent noise.
• Advancements are concentrated on high-resource languages, leaving low-resource languages underserved.

Purpose of the Study:

• To develop a system for sentiment analysis and offensive language identification in low-resource, code-mixed Tamil and English data.
• To address the challenges of noise and semantic information extraction in code-mixed text.
• To compare the effectiveness of machine learning, deep learning, and pre-trained models for these tasks.

Main Methods:

• Utilized a dataset from the DravidianLangTech@ACL2022 shared task on Multitask Learning.
• Employed machine learning, deep learning, and pre-trained models including BERT, RoBERTa, and adapter-BERT.
• Applied word embedding techniques for extracting semantically meaningful information from code-mixed data.

Main Results:

• The adapter-BERT model demonstrated superior performance compared to other trained models.
• Achieved an accuracy of 65% for sentiment analysis.
• Achieved an accuracy of 79% for offensive language identification.

Conclusions:

• The proposed system effectively performs sentiment analysis and offensive language identification on low-resource, code-mixed data.
• Adapter-BERT shows significant promise for handling the complexities of code-mixed text in NLP tasks.
• The findings contribute to advancing NLP capabilities for under-resourced language communities.