Scholarlink

<p>Aspect Sentiment Triplet Extraction (ASTE) is pivotal for nuanced sentiment analysis, demanding precise identification of aspect terms, opinion terms, and their associated sentiment polarities. Existing methods often rely on additional networks, such as graph neural networks, to exploit semantic and syntactic knowledge from reviews. However, these approaches tend to increase model complexity while neglecting the intrinsic relationship between target terms and their part-of-speech attributes. To address these limitations, we propose a novel <b>P</b>art-of-speech-<b>D</b>riven <b>P</b>rompt and <b>A</b>ttention (PDPA) framework for ASTE. Our approach integrates part-of-speech information into both a lightweight soft prompt mechanism and an attention network, significantly enhancing model efficiency and performance. Specifically, the soft prompt embedding, initialized with rich term features, guides a pre-trained language model to generate task-specific knowledge from a part-of-speech perspective, effectively bridging the gap between task requirements and feature representations. Furthermore, the attention network strengthens term representations and their interactions within a 2D table by incorporating part-of-speech information during the update process. Additionally, we introduce a boundary tag mapping strategy for triplet decoding, which transforms asymmetric regions into symmetric regions based on diagonals, thereby fully leveraging bidirectional interactions between entities that are often overlooked by conventional region-based detection methods. Extensive experiments on public benchmark datasets demonstrate that our PDPA framework achieves state-of-the-art performance, showcasing its effectiveness and efficiency in ASTE tasks.</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

PDPA: a part-of-speech-driven prompt and attention method for aspect sentiment triplet extraction

  • 期刊论文

  • 发表日期：

    2025年4月22日

查看全文
  • Guangchao Wang,
  • Zhaojun Li,
  • Cong Liu,
  • Rongyan Wang,
  • Changyou Guo

摘要

Aspect Sentiment Triplet Extraction (ASTE) is pivotal for nuanced sentiment analysis, demanding precise identification of aspect terms, opinion terms, and their associated sentiment polarities. Existing methods often rely on additional networks, such as graph neural networks, to exploit semantic and syntactic knowledge from reviews. However, these approaches tend to increase model complexity while neglecting the intrinsic relationship between target terms and their part-of-speech attributes. To address these limitations, we propose a novel Part-of-speech-Driven Prompt and Attention (PDPA) framework for ASTE. Our approach integrates part-of-speech information into both a lightweight soft prompt mechanism and an attention network, significantly enhancing model efficiency and performance. Specifically, the soft prompt embedding, initialized with rich term features, guides a pre-trained language model to generate task-specific knowledge from a part-of-speech perspective, effectively bridging the gap between task requirements and feature representations. Furthermore, the attention network strengthens term representations and their interactions within a 2D table by incorporating part-of-speech information during the update process. Additionally, we introduce a boundary tag mapping strategy for triplet decoding, which transforms asymmetric regions into symmetric regions based on diagonals, thereby fully leveraging bidirectional interactions between entities that are often overlooked by conventional region-based detection methods. Extensive experiments on public benchmark datasets demonstrate that our PDPA framework achieves state-of-the-art performance, showcasing its effectiveness and efficiency in ASTE tasks.