Yan Xiang, Anlan Zhang, Jinlei Shi, and Yuxin Huang, Journal of Information Processing Systems Vol. 21, No. 2, pp. 180-192, Apr. 2025  

https://doi.org/10.3745/JIPS.04.0343
Keywords: Aspect Category Detection, Aspect-Level Sentiment Analysis, Label Filtering, Topic Model, Weakly Supervised Method
Fulltext: PDF Full Text PubReader

Abstract

Aspect category detection is crucial for aspect-level sentiment analysis. Traditionally, supervised methods require a large number of labeled samples to achieve effectiveness, whereas unsupervised methods rely heavily on human judgment, which can compromise identification accuracy. To address these limitations, we propose a weakly supervised aspect category detection method that uses a hierarchical label filtering mechanism. Our approach begins by assigning preliminary pseudo-labels to comments based on topic similarity. Subsequently, unreliable pseudo-labeled samples are filtered out using semantic similarity. Finally, high-confidence training samples are selected using both high and low thresholds. Using the hierarchical label filtering mechanism employed in these three stages, we constructed a high-confidence training dataset, which was used to train a topic representation-enhanced classifier for aspect category detection. The proposed method, which was evaluated on three publicly available datasets, outperformed existing baseline models, thereby reducing the need for human intervention.

Statistics
Cite this article