Publication 05 · Aspect-level sentiment

Graph Convolution over the Semantic-Syntactic Hybrid Graph.

Syntax tells you who modifies what; semantics and affect tell you what it means. SSGCN refuses to choose.

Published · IJCNN 2022 Hybrid graph Gate fusion SenticNet

Junjie Xu, Shuwen Yang, Luwei Xiao, Zhichao Fu, Xingjiao Wu, Tianlong Ma, Liang He · East China Normal University

Graph convolution over dependency trees became the default for aspect-level sentiment classification — yet it reads only syntax. Fixes that added semantic relations or affective dependency each addressed one limitation at a time. SSGCN builds a semantic latent graph from self-attention, enhances the dependency graph with SenticNet affective knowledge, and lets a gate mechanism dynamically combine both — reaching state-of-the-art on three benchmarks.

3benchmark datasets with state-of-the-art results
2graph sources: semantic latent + enhanced dependency
1gate mechanism fusing both dynamically
2022IJCNN · the earliest first-author architecture
01

The question

In "Despite the awful food, the service is excellent", the dependency tree correctly links "food" to "awful" and "service" to "excellent". But syntax alone is brittle in short, noisy contexts — and it knows nothing of sentiment beyond what the sentence surface offers.

Prior work added either semantic latent relations or affective knowledge. The question SSGCN answers: what happens when both information sources are modeled jointly, and fused adaptively?

02

The approach

SSGCN maintains two parallel graphs. A semantic latent graph, parsed from self-attention networks, captures relations the dependency tree never draws. A syntactic dependency graph enhanced with SenticNet carries affective knowledge along grammatical structure.

A gate mechanism then decides — per input — how much each graph contributes to the aspect-specific representation. Fusion is dynamic, not a fixed weighted average.

SSGCN overall architecture: word embedding, semantic latent graph, enhanced syntactic graph, hybrid graph convolution with gates
Fig. 1Overall SSGCN architecture — word embedding feeds both the semantic latent graph and the SenticNet-enhanced syntactic graph; gates fuse them for aspect-specific prediction.
03

What it shows

On three benchmark datasets, SSGCN outperforms syntax-only and single-enhancement baselines, achieving state-of-the-art performance — evidence that semantic and affective relations are complementary, not redundant.

The gate analysis shows the fusion weighting shifts across inputs: short noisy sentences lean more on the semantic latent graph, grammatically clean sentences lean more on the enhanced dependency graph.

SSGCN ablation: accuracy and F1 versus the number of GCN layers
Fig. 2Ablation over GCN depth — accuracy and F1 respond to the number of graph-convolution layers, supporting the hybrid-graph design.
04

Claim boundary

✓ Established

  • Joint modeling of semantic latent and affective-enhanced syntactic relations
  • Dynamic gate fusion rather than fixed combination
  • State-of-the-art across three benchmarks at publication
  • The first first-author architecture — the starting point of the affective-computing line

✗ Explicitly not claimed

  • English restaurant/laptop/twitter benchmarks only
  • SenticNet coverage bounds the affective enhancement
  • No generative or cross-lingual claim
  • Superseded in raw numbers by later work — kept for the design lesson
05

Citation

@inproceedings{xu2022ssgcn,
  title     = {Graph Convolution over the Semantic-Syntactic Hybrid Graph
               Enhanced by Affective Knowledge for Aspect-level
               Sentiment Classification},
  author    = {Xu, Junjie and Yang, Shuwen and Xiao, Luwei and Fu, Zhichao
               and Wu, Xingjiao and Ma, Tianlong and He, Liang},
  booktitle = {2022 International Joint Conference on Neural Networks (IJCNN)},
  year      = {2022},
  doi       = {10.1109/IJCNN55064.2022.9892027}
}