New Bangla Emotion Dataset Tackles Multi-Label Detection in World's 7th Most Spoken Language

A research team has introduced MONOVAB, a new annotated dataset designed to advance emotion recognition in Bangla, the seventh most widely spoken language globally. The work tackles a largely unexplored challenge: identifying multiple emotional states within a single text passage, which differs from existing approaches that focus on simple sentiment polarity or single-emotion classification.

The Challenge#

Bangla's complex linguistic structure has historically made accurate emotion and sentiment extraction difficult. While researchers have made progress on detecting broad positive/negative sentiment or assigning individual emotion categories, the field of multi-label emotion detection—recognizing that a text can express several feelings simultaneously—remains largely uncharted in Bangla research. This gap has hindered the development of more nuanced natural language processing tools for the language.

Building the Corpus#

To address this limitation, the researchers compiled and manually annotated a new corpus drawn from Facebook data. Rather than applying generic labeling schemes, they employed a context-based annotation methodology to ensure that emotional labels reflected the actual meaning conveyed by each text sample. This careful approach aims to produce higher-quality training data for machine learning models.

Model Performance#

The team evaluated multiple machine learning architectures on the new dataset. BERT (Bidirectional Encoder Representations from Transformers), a prominent transformer-based model, outperformed other approaches tested in the study. The researchers subsequently deployed the best-performing BERT model in a web application, creating a live demonstration of the system's multi-label emotion detection capabilities in Bangla.

Looking Forward#

This work opens new possibilities for emotion-aware natural language processing in Bangla and establishes a foundation for future research in multi-label emotion recognition for the language community.