Does Topic-Based Aspect Labeling Improve Sentiment Analysis? A Comparative Study of LDA-Guided Segmentation with BERT and TF-IDF
Abstrak
Aspect-based sentiment analysis (ABSA) enables fine-grained opinion understanding by associating sentiments with specific text aspects.However, manual aspect annotation remains a bottleneck for large-scale applications.This study investigates whether automatic aspect labeling using Latent Dirichlet Allocation (LDA) topic modeling can effectively support downstream sentiment classification without sacrificing accuracy.We used a product review dataset containing 26,873 labeled instances.Aspect labels were assigned automatically by selecting the dominant topic for each review based on LDA topic probabilities.We then trained and evaluated two sentiment classifiers, bidirectional encoder representations from transformers (BERT) and term frequency-inverse document frequency (TF-IDF) with logistic regression, under two conditions.The first was a baseline setting without topic segmentation.The second involved training separate models on topic-specific subsets created from the automatic aspect labels.Automatic annotation achieved 66% agreement with manual labeling.This result suggests moderate reliability but also reveals substantial ambiguity in the aspect assignments.Contrary to expectations, topic-based segmentation did not improve classification performance.The baseline BERT model achieved the highest performance, with an F1-score of 0.83 and an accuracy of 0.94.It outperformed all topic-specific BERT models, which reached an average F1-score of 0.79, as well as all TF-IDF-based variants.Student's t-tests revealed no statistically significant differences between segmented and non-segmented models (p > 0.05 for all comparisons).Error analysis identified three main sources of misclassification: keyword ambiguity, mixed-sentiment expressions, and topic overlap.We conclude that LDA offers a viable path toward automatic aspect labeling.However, topic-based segmentation does not automatically translate into performance gains in sentiment classification.Future work should explore more precise annotation methods and domain-adaptive topic modeling techniques.
Identifier
- DOI: https://doi.org/10.18280/isi.310427
- OpenAlex: https://openalex.org/W7163323043