| Yusuf Evren Aykaç, Merve Özkan, Refik Samet Lexicon-Augmented Explainable Aspect-based Sentiment Analysis for Azerbaijani Finance and Business Text |
|---|
| Abstract. Aspect-based sentiment analysis (ABSA) for Azerbaijani has received limited attention, and finance / business text introduces domain terms whose polarity can be context-dependent. We study a Finance / Business Azerbaijani ABSA dataset with ∼27.2K aspect-level instances (≈18 tokens on average) and examine whether the SentiAzNet polarity lexicon can complement a multilingual transformer. Using XLMRoBERTa as the base encoder, we integrate lexicon cues in two lightweight ways: (1) concatenating token-level polarity features and (2) adding an attention bias towards lexicon-matched tokens. We also define SentiAzNetFin through a small set of finance-specific polarity overrides. Lexicon coverage is sparse (3.17% of tokens; ∼38% of instances have ≥1 hit), so the lexicon acts as an optional cue rather than a primary signal. Across 5-fold cross-validation, the attention-bias model with SentiAzNet-Fin reaches 0.769 accuracy and 0.738 macro-F1, improving over vanilla XLM-R by 3.6 macro-F1 points (McNemar p < 0.001). For analysis, we compare SHAP attributions, attention saliency, and lexicon hits; agreement between these signals tends to coincide with higher correctness. We further summarize frequent errors such as sarcasm, neutral–negative ambiguity, mixed sentiment, and code-switching. |
| Keywords: Aspect-based sentiment analysis, XLM-RoBERTa, financial sentiment analysis, sentiment lexicon, explainable AI, SHAP, attention |
 Download PDF |
| DOI: https://doi.org/10.54381/itta2026.3.02 |