Research map / Graph learning / GNN architectures
TextGT: A Double-View Graph Transformer on Text for Aspect-Based Sentiment Analysis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TextGT, a double-view graph Transformer that alternates dense GIN-style graph convolution (TextGINConv) with Transformer layers for aspect-based sentiment analysis, reporting state-of-the-art accuracy on SemEval and Twitter benchmarks.
Problem
Aspect-based sentiment analysis (predicting aspect-level sentiment) with GNNs suffers from over-smoothing, and graph Transformers had not been explored for ABSA or tightly combined with sequence-view Transformers.
Approach
Alternate graph convolution layers (dense TextGINConv over dependency graphs with edge features) and Transformer layers in series within each block, with a middle LayerNorm+ReLU layer and skip connections, so the graph and sequence views are tightly coupled and over-smoothing is mitigated.
Key results
- TextGT+BERT reaches 87.31% accuracy on Restaurant and 82.27% micro-F1 on Restaurant, outperforming DualGCN+BERT (87.13 / 81.16).
- TextGT (BiLSTM) achieves 85.17 accuracy on Restaurant, 78.64 on Laptop and 76.51 on Twitter, best among non-BERT methods.
- On Rest16, TextGT+BERT attains 92.21 accuracy and 81.48 micro-F1, versus 91.97 / 79.56 for Sentic GCN+BERT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TextGT+BERT instead of DualGCN+BERT | ABSA models · Restaurant | Accuracy: 87.13 -> 87.31 | Helped |
| use TextGT+BERT instead of DualGCN+BERT | ABSA models · Laptop | Micro-F1: 78.10 -> 78.71 | Helped |
| use TextGT+BERT instead of DGEDT+BERT | ABSA models · Twitter | Accuracy: 77.90 -> 77.70 | Hurt |
| use TextGT+BERT instead of SSEGCN+BERT | ABSA models · Twitter | Micro-F1: 76.36 -> 76.45 | Helped |
| use TextGT+BERT instead of Sentic GCN+BERT | ABSA models · Rest16 | Accuracy: 91.97 -> 92.21 | Helped |
| replace Transformer layers with TextGINConv (ablation) | TextGT · Restaurant, Laptop, Twitter | Accuracy: 85.17 -> 80.07; 78.64 -> 74.68; 76.51 -> 74.89 | Hurt |
| replace TextGINConv layers with Transformer layers (ablation) | TextGT · Restaurant, Laptop, Twitter | Accuracy: 85.17 -> 81.77; 78.64 -> 75.16; 76.51 -> 75.48 | Hurt |
| remove LayerNorm + ReLU middle layer (ablation) | TextGT · Restaurant, Laptop, Twitter | Accuracy: 85.17 -> 83.82; 78.64 -> 75.63; 76.51 -> 76.22 | Hurt |
| replace TextGINConv with TextGCNConv (ablation) | TextGT · Restaurant, Laptop, Twitter | Accuracy: 85.17 -> 83.56; 78.64 -> 77.37; 76.51 -> 75.33 | Hurt |
| replace TextGINConv with TextGATConv (ablation) | TextGT · Restaurant, Laptop, Twitter | Accuracy: 85.17 -> 82.13; 78.64 -> 75.63; 76.51 -> 75.78 | Hurt |
Surprising findings
- TextGT+BERT does not take first place on Laptop and Twitter, though results are comparable to the best.
Limitations
- Gains on Laptop and Twitter with BERT are not the best, only comparable.
- Edge feature updating is not used; edge features remain constant across layers.
- Evaluated only on ABSA; applicability to other NLP tasks is untested.
Experimental setup
Datasets: SemEval 2014 Restaurant SemEval 2014 Laptop SemEval 2016 Rest16 Twitter
Benchmarks: SemEval 2014 Restaurant SemEval 2014 Laptop Twitter SemEval 2016 Rest16
Metrics: accuracy micro-F1
Baselines: IAN RAM MGAN TNet ASGCN CDT BiGCN kumaGCN InterGCN R-GAT DualGCN DGEDT
Builds on: GINE GIN GraphGPS R-GAT DualGCN
Closest papers by meaning
- You Only Read Once: Constituency-Oriented Relational Graph Convolutional Network for Multi-Aspect Multi-Sentiment Classification (AAAI 2024 · similarity 0.88)
- Aspect-Based Sentiment Analysis with Explicit Sentiment Augmentations (AAAI 2024 · similarity 0.87)
- A Novel Energy Based Model Mechanism for Multi-Modal Aspect-Based Sentiment Analysis (AAAI 2024 · similarity 0.84)
- Aspect Enhancement and Text Simplification in Multimodal Aspect-Based Sentiment Analysis for Multi-Aspect and Multi-Sentiment Scenarios (AAAI 2025 · similarity 0.84)
- DEQA: Descriptions Enhanced Question-Answering Framework for Multimodal Aspect-Based Sentiment Analysis (AAAI 2025 · similarity 0.84)
- DanceHA: A Multi-Agent Framework for Document-Level Aspect-Based Sentiment Analysis (AAAI 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.