Rethinking Text Segmentation and Semantic Filtering for Aspect-Based Sentiment Analysis in Financial News

Abstract

Aspect-based sentiment analysis (ABSA) has been widely applied to identify aspect terms and classify their associated sentiment polarity. However, in financial news, where sentences are often long, information-dense, and structurally complex, it remains unclear whether text segmentation and semantic filtering improve the effectiveness of pipeline-based ABSA. This study investigates the effects of sentence complexity, text segmentation, and semantic filtering on ABSA performance in financial news. Using the PyABSA framework with the FAST-LCF architecture, we evaluate two core tasks, namely Aspect Term Extraction (ATE) and Aspect Polarity Classification (APC), on the SentFin and FinNews datasets. The experiments compare the baseline setting with two text segmentation strategies, Sentencizer and Senter, as well as semantic filtering based on Semantic Role Labeling (SRL). Performance is assessed using Precision, Recall, F1-score, and Accuracy. The results show that increasing sentence complexity is associated with substantial performance degradation, particularly for APC, where accuracy on SentFin declines from 92.91% to 78.76%. Moreover, neither text segmentation nor SRL-based filtering consistently improves ATE or APC performance; in several settings, both strategies reduce F1-score. These findings indicate that, for financial news ABSA, preserving discourse context is more beneficial than aggressive structural simplification, and they call into question the assumption that lighter or more selective inputs necessarily improve pipeline effectiveness.