Scholarlink

<p>Aspect-based sentiment analysis aims to infer the sentiment polarity associated with a given aspect term at a fine-grained level. However, many prior methods underutilize external knowledge and lack a principled mechanism for dynamically fusing semantic and syntactic representations. This work proposes KEDGE (Knowledge-Enhanced Dual-Graph Encoder), which jointly leverages Wiktionary definitions and SenticNet polarity to refine semantics and augment graph structure, while learning to fuse multi-view signals in a sample-conditional manner. Concretely, the model comprises three complementary components: (i) a contextual semantic branch that performs cross-knowledge interaction between sentence encodings and aspect-related definitional text, followed by self-attention and pooling to form a compact semantic representation; (ii) a position-aware syntactic branch that enriches the dependency graph with an aspect prior and SenticNet scores, then propagates features via multi-layer graph convolution with distance-aware masking, and a semantics-guided attention readout to form a structural representation; and (iii) a conditional dual-path fusion module that integrates a gated nonlinear interaction path with a lightweight mixture-of-experts path, using a top-level gate to adaptively arbitrate their contributions. We further calibrate affective priors by constructing <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\hbox {SenticNet}^{*}\)</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mtext>SenticNet</mtext> <mrow /> <mrow> <mrow /> <mo>∗</mo> </mrow> </mmultiscripts> </math></EquationSource> </InlineEquation> via paraphrase aggregation over Wiktionary and SenticNet. Experiments on five public ABSA benchmarks (Laptop14, Restaurant14, Twitter, Restaurant15, and Restaurant16) show that the proposed approach achieves strong performance.</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

KEDGE: knowledge-enhanced dual-graph encoder for aspect-based sentiment analysis

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Kexin Jiang,
  • Yue Qin,
  • Xiaoqin Xiao,
  • Yahui Zhao

摘要

Aspect-based sentiment analysis aims to infer the sentiment polarity associated with a given aspect term at a fine-grained level. However, many prior methods underutilize external knowledge and lack a principled mechanism for dynamically fusing semantic and syntactic representations. This work proposes KEDGE (Knowledge-Enhanced Dual-Graph Encoder), which jointly leverages Wiktionary definitions and SenticNet polarity to refine semantics and augment graph structure, while learning to fuse multi-view signals in a sample-conditional manner. Concretely, the model comprises three complementary components: (i) a contextual semantic branch that performs cross-knowledge interaction between sentence encodings and aspect-related definitional text, followed by self-attention and pooling to form a compact semantic representation; (ii) a position-aware syntactic branch that enriches the dependency graph with an aspect prior and SenticNet scores, then propagates features via multi-layer graph convolution with distance-aware masking, and a semantics-guided attention readout to form a structural representation; and (iii) a conditional dual-path fusion module that integrates a gated nonlinear interaction path with a lightweight mixture-of-experts path, using a top-level gate to adaptively arbitrate their contributions. We further calibrate affective priors by constructing \(\hbox {SenticNet}^{*}\) SenticNet ∗ via paraphrase aggregation over Wiktionary and SenticNet. Experiments on five public ABSA benchmarks (Laptop14, Restaurant14, Twitter, Restaurant15, and Restaurant16) show that the proposed approach achieves strong performance.