Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

S$^2$GSL: Incorporating Segment to Syntactic Enhanced Graph Structure Learning for Aspect-based Sentiment Analysis

Bingfeng Chen, Qihan Ouyang, Yongqi Luo, Boyan Xu, Ruichu Cai, Zhifeng Hao

TL;DR

S^2GSL tackles ABSA by decomposing complex sentence structures into localized semantic segments via constituent trees and integrating a learnable syntax-based latent graph, all fused through a self-adaptive aggregation network. The dual-branch architecture—Segment-aware Semantic Graph Learning (SeSG) and Syntax-based Latent Graph Learning (SyLG)—mitigates semantic and syntactic noise while preserving relevant relations for each aspect. Empirical results on four benchmarks show S^2GSL achieves state-of-the-art performance on Laptop, Restaurant, and MAMS, with strong results on Twitter, and ablations confirm the critical contributions of both branches and the adaptive fusion mechanism. The work also analyzes dynamic local attention, dependency-label usefulness, and performance under ChatGPT prompts, highlighting practical implications for robust, fine-grained ABSA in real-world settings.

Abstract

Previous graph-based approaches in Aspect based Sentiment Analysis(ABSA) have demonstrated impressive performance by utilizing graph neural networks and attention mechanisms to learn structures of static dependency trees and dynamic latent trees. However, incorporating both semantic and syntactic information simultaneously within complex global structures can introduce irrelevant contexts and syntactic dependencies during the process of graph structure learning, potentially resulting in inaccurate predictions. In order to address the issues above, we propose S$^2$GSL, incorporating Segment to Syntactic enhanced Graph Structure Learning for ABSA. Specifically,S$^2$GSL is featured with a segment-aware semantic graph learning and a syntax-based latent graph learning enabling the removal of irrelevant contexts and dependencies, respectively. We further propose a self-adaptive aggregation network that facilitates the fusion of two graph learning branches, thereby achieving complementarity across diverse structures. Experimental results on four benchmarks demonstrate the effectiveness of our framework.

S$^2$GSL: Incorporating Segment to Syntactic Enhanced Graph Structure Learning for Aspect-based Sentiment Analysis

TL;DR

S^2GSL tackles ABSA by decomposing complex sentence structures into localized semantic segments via constituent trees and integrating a learnable syntax-based latent graph, all fused through a self-adaptive aggregation network. The dual-branch architecture—Segment-aware Semantic Graph Learning (SeSG) and Syntax-based Latent Graph Learning (SyLG)—mitigates semantic and syntactic noise while preserving relevant relations for each aspect. Empirical results on four benchmarks show S^2GSL achieves state-of-the-art performance on Laptop, Restaurant, and MAMS, with strong results on Twitter, and ablations confirm the critical contributions of both branches and the adaptive fusion mechanism. The work also analyzes dynamic local attention, dependency-label usefulness, and performance under ChatGPT prompts, highlighting practical implications for robust, fine-grained ABSA in real-world settings.

Abstract

Previous graph-based approaches in Aspect based Sentiment Analysis(ABSA) have demonstrated impressive performance by utilizing graph neural networks and attention mechanisms to learn structures of static dependency trees and dynamic latent trees. However, incorporating both semantic and syntactic information simultaneously within complex global structures can introduce irrelevant contexts and syntactic dependencies during the process of graph structure learning, potentially resulting in inaccurate predictions. In order to address the issues above, we propose SGSL, incorporating Segment to Syntactic enhanced Graph Structure Learning for ABSA. Specifically,SGSL is featured with a segment-aware semantic graph learning and a syntax-based latent graph learning enabling the removal of irrelevant contexts and dependencies, respectively. We further propose a self-adaptive aggregation network that facilitates the fusion of two graph learning branches, thereby achieving complementarity across diverse structures. Experimental results on four benchmarks demonstrate the effectiveness of our framework.
Paper Structure (30 sections, 26 equations, 12 figures, 10 tables)

This paper contains 30 sections, 26 equations, 12 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Proposed SGSL
  3. Context Encoding Module
  4. Segment-aware Semantic Graph Learning
  5. Syntax-based Latent Graph Learning
  6. Self-adaptive Aggregation Module
  7. Training
  8. Experiments
  9. Datasets
  10. Implementation Details
  11. Baselines
  12. Overall Performance
  13. Ablation Study
  14. Discuss and Analysis
  15. Effect of Dynamic Local Attention
  16. ...and 15 more sections

Figures (12)

  • Figure 1: The attention weight of the aspect word "atmosphere" in relation to other words in the sentence."×" refers to noise information for "atmosphere".
  • Figure 2: Dependency tree of an example sentence.
  • Figure 3: A segment matrix at the third layer of the sample sentence constituent tree divides the sentence into three semantically complete paragraphs.
  • Figure 4: The overall architecture of S$^2$GSL, which is composed primarily of SeSG, SyLG, and Self-adaptive Aggregation Modules.
  • Figure 5: The overall architecture of Self-adaptive Aggregation Module.
  • ...and 7 more figures