BABSA: A large scale bangla aspect based sentiment analysis dataset.

Area of Science:

• Natural Language Processing
• Computational Linguistics

Background:

• Bangla Aspect-Based Sentiment Analysis (ABSA) research is hindered by a lack of comprehensive, high-quality datasets.
• The widespread use of Bangla in digital communications necessitates specialized resources for sentiment analysis.

Purpose of the Study:

• Introduce BABSA, a novel dataset for Bangla ABSA.
• Facilitate aspect extraction and aspect-specific sentiment classification in Bangla.
• Address the gap in fine-grained sentiment analysis resources for the Bangla language.

Main Methods:

• Compiled 15,860 instances from five diverse sources, including web-scraped news data.
• Implemented a rigorous three-pass manual annotation protocol with clear guidelines for aspect terms and sentiment.
• Ensured high annotation quality with an inter-annotator agreement score of 0.84 (Cohen's kappa).

Main Results:

• Developed BABSA, a large-scale dataset covering 21 domains with detailed annotations for aspect terms and sentiment.
• The dataset includes metadata for advanced downstream analysis and is split into train-test sets.
• Achieved high inter-annotator agreement, ensuring the reliability of the annotations.

Conclusions:

• BABSA provides a crucial resource for advancing Bangla ABSA research and development.
• The dataset's scale, diversity, and fine-grained annotations are suitable for training and evaluating various NLP models.
• Public availability of BABSA promotes reproducibility and further research in Bangla language understanding.