Scholarlink

Aspect sentiment quad prediction (ASQP) is a key aspect-based sentiment analysis task, focusing on predicting the quad sentiment elements in a given sentence. Recent research has achieved commendable results using generative models to achieve this task. However, these studies frequently fail to leverage the connections between sentiment components, overlooking the interplay within sentiment tuples and the impact of varied linguistic expressions on outcomes. In addition, previous works often ignore the potential built-in negative samples of the model, resulting in errors in quadruple prediction. To address these limitations, we propose multi-template sequence generation to achieve different input and target expressions, and we introduce unlikelihood learning to motivate the model to clearly distinguish semantically close or similar words, suppressing the generation of negative samples. Furthermore, we introduce a prediction correction strategy to optimize the generated output. Our method exhibits superiority over existing approaches, as evidenced by experimental results on 4 public datasets, outperforming various baseline methods and achieving better performance on benchmark test sets.

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

Leveraging Template Sequence Generation and Unlikelihood Learning for Aspect Sentiment Quadruplet Prediction

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yongmei Zhou,
  • Jiahao Chen,
  • Aimin Yang,
  • Jianghao Lin

摘要

Aspect sentiment quad prediction (ASQP) is a key aspect-based sentiment analysis task, focusing on predicting the quad sentiment elements in a given sentence. Recent research has achieved commendable results using generative models to achieve this task. However, these studies frequently fail to leverage the connections between sentiment components, overlooking the interplay within sentiment tuples and the impact of varied linguistic expressions on outcomes. In addition, previous works often ignore the potential built-in negative samples of the model, resulting in errors in quadruple prediction. To address these limitations, we propose multi-template sequence generation to achieve different input and target expressions, and we introduce unlikelihood learning to motivate the model to clearly distinguish semantically close or similar words, suppressing the generation of negative samples. Furthermore, we introduce a prediction correction strategy to optimize the generated output. Our method exhibits superiority over existing approaches, as evidenced by experimental results on 4 public datasets, outperforming various baseline methods and achieving better performance on benchmark test sets.