Scholarlink

<p>Aspect-based sentiment analysis (ABSA) focuses on identifying the sentiment polarity expressed toward particular aspects or attributes in a text. Existing methods often rely on large-scale annotated datasets, leading to unsatisfactory results when training data are limited. To address this challenge, we introduce BiF-ABSA, a few-shot framework that leverages a bidirectional learning strategy. The proposed framework combines the tasks of “Text-to-Label” and “Label-to-Text”, forming a generative closed-loop learning paradigm that enhances the model’s generalization ability while improving data quality through sample diversity augmentation. To further improve generalization, we introduce a unified infill module, which redefines ABSA as a text completion task, thus bridging the gap with pre-trained models and strengthening task adaptability. In addition, a focus-based text span localization module is incorporated to enhance the accuracy of aspect term extraction. Experimental results demonstrate that BiF-ABSA significantly improves performance across multiple benchmark datasets, achieving gains of 1.27 and 2.49 percentage points under the 2%-shot and 5%-shot settings, respectively. These results confirm the framework’s effectiveness and stability in few-shot settings. Ablation experiments further demonstrate that each component plays a vital role in boosting accuracy and generalization. In conclusion, the framework offers a practical and extensible solution for few-shot sentiment analysis, particularly under low-resource conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Generative few-shot aspect-based sentiment analysis based on bidirectional learning

  • 期刊论文

  • 发表日期：

    2025年12月22日

查看全文
  • Hongwei Chen,
  • Chuanyu Xie,
  • Wei Wang,
  • Changyong Lin

摘要

Aspect-based sentiment analysis (ABSA) focuses on identifying the sentiment polarity expressed toward particular aspects or attributes in a text. Existing methods often rely on large-scale annotated datasets, leading to unsatisfactory results when training data are limited. To address this challenge, we introduce BiF-ABSA, a few-shot framework that leverages a bidirectional learning strategy. The proposed framework combines the tasks of “Text-to-Label” and “Label-to-Text”, forming a generative closed-loop learning paradigm that enhances the model’s generalization ability while improving data quality through sample diversity augmentation. To further improve generalization, we introduce a unified infill module, which redefines ABSA as a text completion task, thus bridging the gap with pre-trained models and strengthening task adaptability. In addition, a focus-based text span localization module is incorporated to enhance the accuracy of aspect term extraction. Experimental results demonstrate that BiF-ABSA significantly improves performance across multiple benchmark datasets, achieving gains of 1.27 and 2.49 percentage points under the 2%-shot and 5%-shot settings, respectively. These results confirm the framework’s effectiveness and stability in few-shot settings. Ablation experiments further demonstrate that each component plays a vital role in boosting accuracy and generalization. In conclusion, the framework offers a practical and extensible solution for few-shot sentiment analysis, particularly under low-resource conditions.