BanglaSentiEmo: Cross-Task Affective Structure from a Shared BanglaBERT Encoder for Joint Sentiment-Emotion Classification in Bengali

Abstract—Sentiment analysis and emotion recognition are predominantly treated as independent tasks in Bengali NLP, despite both being grounded in the same underlying affective signal. Existing multi-task learning (MTL) studies either pair sentiment with structurally unrelated tasks, or rely on mul-tilingual encoders trained on two disjoint corpora without a genuinely shared encoder. In this work, we propose a shared-encoder BanglaBERT framework for joint sentiment (5-class) and emotion (7-class) classification of Bengali social media text. We first quantify the sentiment-emotion dependence on the full dataset (χ2 = 11090.04, df = 24, p < 0.001; bias-corrected Cramer’s V = 0.684; normalized mutual information = 0.556; information gain = 1.373 bits), confirming a strong association that motivates joint modeling. Under a matched protocol repeated over five random seeds, the shared encoder shows no statistically significant change in sentiment macro-F1 (STL 65.25 ±1.55 vs. MTL 65.39 ±1.45; paired t-test p = 0.893) and a statistically significant regression in emotion macro-F1 (STL 69.21 ±0.87 vs. MTL 67.39 ±1.02; p = 0.023), so raw per-task F1 does
not reliably improve under the joint objective. Despite this, error analysis on a representative run shows correlated task performance: joint (both-correct) accuracy of 55.91% exceeds the 47.13% expected under independence (+18.6% relative), with a Pearson correlation of 0.4084 (p= 3.334 ×10−25) between per-post sentiment and emotion-correctness, indicating the shared representation captures genuine cross-task affective structure even where this does not consistently translate into higher macro-F1.Index Terms—Multi-task learning, BanglaBERT, sentiment analysis, emotion recognition, Bengali natural language processing, shared encoder, low-resource languages.