A Diverse Bengali-English Code-Mixed Dataset for Sentiment Analysis

Abstract

Code-mixed data, blending two or more languages within sentences, offers valuable insights for low-resource languages like Bengali, which have limited annotated cor pora. While sentiment analysis has been widely explored in various languages, code mixed Bengali remains underrepresented, lacking a comprehensive benchmark dataset. To address this gap, we introduce BnSentMix, a sentiment analysis dataset for code mixed Bengali-English, consisting of 20,000 samples annotated with four sentiment labels: positive, negative, neutral, and mixed. The data was sourced from e-commerce websites, YouTube, and Facebook, ensuring linguistic diversity and reflecting real world, code-mixed scenarios. Our dataset captures a wide variety of user-generated content, providing robust coverage of both informal and formal language styles. We utilized a novel automated text filtering pipeline that employs fine-tuned pre-trained language models to detect and extract code-mixed samples, ensuring high-quality data. To evaluate the dataset, we applied 11 baseline approaches, ranging from tra ditional machine learning models to advanced transformer-based architectures. Our best model achieved an accuracy of 69.5% and an F1 score of 68.8%.