A Multi-Seed Comparative Study of Active Learning Query Strategies for Bangla Product Review Sentiment Analysis

This study provides a controlled multi-seed comparison of four active-learning query strategies for annotation-efficient Bangla product-review sentiment analysis using a fixed human-adjudicated gold test set. It evaluates not only final Macro F1 but also full learning trajectories, AULC, cross-seed stability, and class-wise behavior, showing that more complex acquisition strategies do not necessarily outperform Random sampling and highlighting a persistent Neutral-class bottleneck.