BARD10: A Bangla Authorship Attribution Dataset for Analyzing the Stylistic Role of Stop-Words

Abstract

Stop-word removal is a common preprocessing step in natural language processing, but in authorship attribution, frequent function words may preserve important stylistic information. For Bangla, the effect of stop-word removal remains insufficiently examined, particularly beyond literary corpora and under consistent comparisons between classical and neural models. This study introduces the Bangla Authorship Recognition Dataset of 10 Authors (BARD10), a curated dataset of 2337 postlevel texts from 10 contemporary Bangla authors, and evaluates it alongside the literary benchmark Bangla Authorship Attribution Dataset (BAAD16). We compare four model families: support vector machine (SVM), eXtreme Gradient Boosting (XGBoost), an embedding-based multilayer perceptron (MLP), and BanglaBERT, a Bangla-adapted bidirectional encoder representations from transformers (BERT) model, under stop-word–retained and stop-word–removed conditions. For the classical models, we further examine word-level, character-level, and subword-level term frequency–inverse document frequency (TF–IDF) representations. Across both datasets, sparse TF–IDF features combined with linear SVM achieve the strongest overall performance, reaching a macro-F1 of 0.9956 on BAAD16 and 0.9232 ± 0.0086 on BARD10. BanglaBERT remains competitive but does not surpass the best sparse-feature configurations. The results show that the effect of stop-word removal is not uniform across models, tokenization settings, or datasets. In particular, stop-words often preserve useful authorial information in BARD10, indicating their importance in contemporary online Bangla writing. A word-level Δ-recall analysis further shows that individual stop-word effects are heterogeneous and author-dependent. These findings suggest that Bangla stop-words should not be treated as uniformly disposable preprocessing noise and that interpretable sparse-feature baselines remain essential reference points for Bangla authorship attribution.