Cross-lingual Aspect-based Sentiment Analysis Based on Semi-supervised Knowledge Distillation
Bolun Zhang, Yahui Zhao, Guozhe Jin, Rongyi Cui · 2025
Aspect-based sentiment analysis aims to predict the sentiment polarity of a text concerning specific aspects. Although substantial advances have been made in high-resource languages, ABSA still faces significant difficulties in low-resource languages, primarily the limited availability of labeled data, which restricts the model's capacity to effectively learn and represent sentiment information. To address this problem, our paper proposes a cross-lingual aspect-based sentiment analysis approach based on semi-supervised knowledge distillation, leveraging labeled data from high-resource languages to transfer sentiment analysis capabilities to unlabeled data in low-resource languages. Firstly, we reduce the disparity in linguistic expressions across languages by employing an aspect code-switching method and applying synonym substitution for data augmentation, enriching the source language data and enhancing the model's generalization ability. Secondly, we use a multilingual pre-trained model as the foundation of the student model and train it using a knowledge distillation approach, which effectively improves sentiment classification performance in the target low-resource language. To further enhance the model's performance, we introduce a voting-based pseudo-hard labels generation strategy to improve the reliability of pseudo-hard labels, thereby reinforcing the model's acquisition of task-specific and language-specific knowledge. Experiments conducted on the SemEval-2016 benchmark dataset, covering five languages, demonstrate that our method achieves significant performance improvements in ABSA tasks across multiple languages, showcasing its robust adaptability in cross-lingual settings.