Scholarlink

Aspect-based Sentiment Analysis (ABSA) employing generative pre-trained language models (PLMs) has emerged as a prominent approach in sentiment analysis, particularly in addressing the challenging task of Aspect-Category-Opinion-Sentiment Quadruple Extraction (ACOSQE). The primary challenges of ACOSQE manifest in two key areas: the scarcity and heterogeneity of datasets, and the complexity of the aspect and opinion terminology embedded in texts. To address these limitations, we propose MvGEN-SCL, a novel framework incorporating two techniques for enhanced structured generation in ACOSQE. First, We introduce an efficient multi-view augmentation strategy that fixes the aspect term to reduce computational overhead while maintaining performance, while generating multiple labels from a single input by leveraging human-like problem-solving intuitions from different perspectives. Furthermore, we enhance the model’s training objective by incorporating supervised contrastive learning as an auxiliary task, enabling the model to effectively capture both explicit and implicit relationships between target elements. Extensive experimental results demonstrate that the proposed MvGEN-SCL achieves strong performance when implemented on the T5-large model. Additionally, MvGEN-SCL achieved notable improvements on challenging implicit aspect and opinion tasks that have traditionally proven difficult for existing ACOSQE methods.( https://github.com/Rastimus/MvGEN-SCL-for-ACOSQE )

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

Improving Generative Aspect-Based Sentiment Analysis via Multi-view Data Augmentation and Supervised Contrastive Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yanhong Wang,
  • Lei Chen,
  • Xin Wang,
  • Lin Yao

摘要

Aspect-based Sentiment Analysis (ABSA) employing generative pre-trained language models (PLMs) has emerged as a prominent approach in sentiment analysis, particularly in addressing the challenging task of Aspect-Category-Opinion-Sentiment Quadruple Extraction (ACOSQE). The primary challenges of ACOSQE manifest in two key areas: the scarcity and heterogeneity of datasets, and the complexity of the aspect and opinion terminology embedded in texts. To address these limitations, we propose MvGEN-SCL, a novel framework incorporating two techniques for enhanced structured generation in ACOSQE. First, We introduce an efficient multi-view augmentation strategy that fixes the aspect term to reduce computational overhead while maintaining performance, while generating multiple labels from a single input by leveraging human-like problem-solving intuitions from different perspectives. Furthermore, we enhance the model’s training objective by incorporating supervised contrastive learning as an auxiliary task, enabling the model to effectively capture both explicit and implicit relationships between target elements. Extensive experimental results demonstrate that the proposed MvGEN-SCL achieves strong performance when implemented on the T5-large model. Additionally, MvGEN-SCL achieved notable improvements on challenging implicit aspect and opinion tasks that have traditionally proven difficult for existing ACOSQE methods.( https://github.com/Rastimus/MvGEN-SCL-for-ACOSQE )