Research map / Label-efficient and robust learning / Recognition and long-tail
Aspect-Based Sentiment Analysis with Explicit Sentiment Augmentations
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 ABSA-ESA, which post-trains T5 to generate explicit sentiment augmentations for aspect terms and appends them to input sentences to handle implicit sentiment in ABSA.
Problem
Aspect-based sentiment analysis struggles with implicit sentiment sentences that lack distinct opinion words, making polarity hard to determine.
Approach
Generate explicit, aspect-related sentiment sentences with a post-trained T5 (using syntax distance weighting, unlikelihood contrastive regularization, and constrained beam search) and concatenate them to the input before T5-encoder classification.
Key results
- ABSA-ESA achieves 88.29% overall accuracy on Rest.14 and 82.44% on Lap.14, outperforming baselines in the T5-base setting.
- Implicit sentiment accuracy improves over CLEAN by 1.12 points on Rest.14 and 1.71 points on Lap.14.
- Removing Syntax Distance Weighting drops implicit accuracy by 2.62 on Rest.14 and 2.42 on Lap.14.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add explicit sentiment augmentations | T5-base ABSA · Rest.14 | Implicit accuracy: 1.12↑ | Helped |
| Add explicit sentiment augmentations | T5-base ABSA · Lap.14 | Implicit accuracy: 1.71↑ | Helped |
| Add Syntax Distance Weighting (ablation) | ABSA-ESA · Rest.14 | Implicit accuracy: 2.62↓ | Helped |
| Add Syntax Distance Weighting (ablation) | ABSA-ESA · Lap.14 | Implicit accuracy: 2.42↓ | Helped |
| Add Unlikelihood Contrastive Regularization (ablation) | ABSA-ESA without SDW · Rest.14 | Implicit accuracy | No clear effect |
| Add Unlikelihood Contrastive Regularization (ablation) | ABSA-ESA without SDW · Lap.14 | Implicit accuracy: 77.90 -> 79.31 | Helped |
| Add Constrained Beam Search (ablation) | ABSA-ESA · Rest.14 | Implicit accuracy: 67.81 -> 70.78 | Helped |
| Add Constrained Beam Search (ablation) | ABSA-ESA · Lap.14 | Implicit accuracy: 78.65 -> 80.00 | Helped |
| Add Syntax Distance Weighting and Unlikelihood Contrastive Regularization (ablation) | ABSA-ESA · Rest.14 | Implicit accuracy: 68.39 -> 70.78 | Helped |
| Add Syntax Distance Weighting and Unlikelihood Contrastive Regularization (ablation) | ABSA-ESA · Lap.14 | Implicit accuracy: 77.90 -> 80.00 | Helped |
Surprising findings
- Unlikelihood Contrastive Regularization alone offers limited gains on implicit accuracy and may reduce performance, per the ablation.
- THOR's implicit F1 on Lap.14 (67.63) trails a basic Flan-T5 prompt baseline (76.40).
Limitations
- Augmentation generation relies on rule-based selection from the training dataset and GloVe similarity for aspect-term matching.
- Gains from UCR on implicit accuracy are limited according to ablation.
Experimental setup
Datasets: Rest.14 Lap.14
Benchmarks: Rest.14 Lap.14
Metrics: accuracy macro-F1 implicit sentiment accuracy explicit sentiment accuracy average prediction entropy
Baselines: BERT-base BT BERTAsp EDA PT ADA R-GAT CLEAN ASA-WD C3DA T5Asp THOR
Builds on: T5 GloVe BERTAsp THOR C3DA
Closest papers by meaning
- DanceHA: A Multi-Agent Framework for Document-Level Aspect-Based Sentiment Analysis (AAAI 2026 · similarity 0.89)
- TextGT: A Double-View Graph Transformer on Text for Aspect-Based Sentiment Analysis (AAAI 2024 · similarity 0.87)
- You Only Read Once: Constituency-Oriented Relational Graph Convolutional Network for Multi-Aspect Multi-Sentiment Classification (AAAI 2024 · similarity 0.85)
- Aspect Enhancement and Text Simplification in Multimodal Aspect-Based Sentiment Analysis for Multi-Aspect and Multi-Sentiment Scenarios (AAAI 2025 · similarity 0.85)
- SimRP: Syntactic and Semantic Similarity Retrieval Prompting Enhances Aspect Sentiment Quad Prediction (AAAI 2025 · similarity 0.85)
- Counterfactual-Enhanced Information Bottleneck for Aspect-Based Sentiment Analysis (AAAI 2024 · similarity 0.84)
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.