BEmoC: A Corpus for Identifying Emotion in Bengali Texts

Area of Science:

• Natural Language Processing (NLP)
• Computational Linguistics
• Affective Computing

Background:

• Emotion classification in text is gaining importance for Web 2.0 applications.
• Bengali, a low-resource language, faces challenges in automatic emotion classification due to limited resources.
• A benchmark corpus is essential for developing effective Bengali emotion classifiers.

Purpose of the Study:

• To describe the development of a benchmark emotional corpus for Bengali texts, named BEmoC.
• To facilitate the classification of six basic emotions (anger, fear, surprise, sadness, joy, disgust) in Bengali.
• To provide a foundational resource for advancing NLP research in low-resource languages.

Main Methods:

• Corpus development involved four key steps: data crawling, pre-processing, labelling, and verification.
• A dataset of 7000 Bengali texts was manually labelled into six distinct emotion categories.
• Inter-annotator agreement was assessed using Cohen's kappa statistic.

Main Results:

• The developed Bengali Emotion Corpus (BEmoC) contains 7000 labelled texts across six emotion categories.
• A Cohen's kappa score of 0.969 indicated high agreement between corpus annotators and experts.
• Analysis revealed that Bengali emotion word distribution follows Zipf's law, with insights into coding reliability and emotion density.

Conclusions:

• BEmoC is a valuable benchmark corpus for Bengali emotion classification, addressing the scarcity of resources for low-resource languages.
• The high annotator agreement validates the quality and reliability of the corpus.
• This resource supports further research and development of NLP applications for Bengali, particularly in areas like sentiment analysis and user behavior understanding.