DravidianCodeMix: sentiment analysis and offensive language identification dataset for Dravidian languages in...

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Social Media Analysis

Background:

• Under-resourced Dravidian languages present significant challenges for NLP tasks.
• Social media platforms are rich sources of multilingual user-generated content, often exhibiting code-mixing.
• Existing datasets often lack sufficient coverage for low-resource languages and specific annotation tasks.

Purpose of the Study:

• To develop and release a novel, manually annotated dataset for sentiment analysis and offensive language identification.
• To support research in natural language processing for Tamil, Kannada, and Malayalam.
• To capture code-mixing phenomena prevalent in multilingual social media data.

Main Methods:

• Manual annotation of over 60,000 YouTube comments by volunteer annotators.
• Annotation focused on sentiment analysis and offensive language identification.
• Dataset creation involved collecting and processing comments in Tamil-English, Kannada-English, and Malayalam-English.

Main Results:

• A comprehensive multilingual dataset comprising 44,000 Tamil-English, 7,000 Kannada-English, and 20,000 Malayalam-English comments.
• High inter-annotator agreement (Krippendorff's alpha) demonstrating annotation quality.
• Baseline machine learning and deep learning experiments established performance benchmarks.

Conclusions:

• The released dataset provides a valuable resource for advancing NLP research in under-resourced Dravidian languages.
• The dataset's inclusion of code-mixing phenomena offers unique opportunities for studying language interaction.
• Availability on GitHub and Zenodo promotes accessibility and further research.