Scholarlink

<p>Aspect Sentiment Triplet Extraction (ASTE) is a sophisticated, fine-grained sentiment analysis task that aims to identify aspects, corresponding opinion expressions, and their associated sentiment polarities within a given sentence, thereby enabling comprehensive sentiment analysis. Recent years have witnessed notable advancements and widespread research attention on span-based end-to-end models for ASTE. However, such models often overlook the rich semantic and syntactic information embedded in textual inputs, leading to suboptimal performance in capturing complex contextual semantics. To address this critical limitation, we propose a Semantic-Syntactic Enhanced ASTE model (SSE-ASTE), which leverages span-level semantic and syntactic features to achieve precise extraction of aspect sentiment triplets and accurate sentiment prediction. Specifically, we employ Graph Neural Networks (GNNs) to encode semantic and syntactic information, thereby enhancing span representations for more effective integration of aspect and opinion spans. We further incorporate a Top-K pruning strategy to filter and retain high-quality aspect and opinion spans. Additionally, a bidirectional decoding structure is adopted to comprehensively extract aspect-opinion span pairs, ensuring more complete and accurate triplet identification. Extensive experiments conducted on SemEval datasets demonstrate the exceptional performance of the SSE-ASTE model. It outperforms most state-of-the-art baselines, with consistent improvements across Precision, Recall, and F1&#xa0;&#xa0;-score metrics. Notably, in direct comparative experiments on the ASTE task, our proposed model even surpasses the well-known large language model GPT-4&#xa0;&#xa0;. These comprehensive experimental results fully validate the effectiveness and superior performance of the SSE-ASTE model for the ASTE task.</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

A span-level aspect sentiment triplet extraction model by semantic-syntactic enhancement

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • Zhiyuan Zhang,
  • Shanshan Xie,
  • Yuanzhe Hu,
  • Mengyao Sun,
  • Dong Chen

摘要

Aspect Sentiment Triplet Extraction (ASTE) is a sophisticated, fine-grained sentiment analysis task that aims to identify aspects, corresponding opinion expressions, and their associated sentiment polarities within a given sentence, thereby enabling comprehensive sentiment analysis. Recent years have witnessed notable advancements and widespread research attention on span-based end-to-end models for ASTE. However, such models often overlook the rich semantic and syntactic information embedded in textual inputs, leading to suboptimal performance in capturing complex contextual semantics. To address this critical limitation, we propose a Semantic-Syntactic Enhanced ASTE model (SSE-ASTE), which leverages span-level semantic and syntactic features to achieve precise extraction of aspect sentiment triplets and accurate sentiment prediction. Specifically, we employ Graph Neural Networks (GNNs) to encode semantic and syntactic information, thereby enhancing span representations for more effective integration of aspect and opinion spans. We further incorporate a Top-K pruning strategy to filter and retain high-quality aspect and opinion spans. Additionally, a bidirectional decoding structure is adopted to comprehensively extract aspect-opinion span pairs, ensuring more complete and accurate triplet identification. Extensive experiments conducted on SemEval datasets demonstrate the exceptional performance of the SSE-ASTE model. It outperforms most state-of-the-art baselines, with consistent improvements across Precision, Recall, and F1  -score metrics. Notably, in direct comparative experiments on the ASTE task, our proposed model even surpasses the well-known large language model GPT-4  . These comprehensive experimental results fully validate the effectiveness and superior performance of the SSE-ASTE model for the ASTE task.