Research map / Label-efficient and robust learning / Recognition and long-tail
Modeling Adaptive Inter-Task Feature Interactions via Sentiment-Aware Contrastive Learning for Joint Aspect-Sentiment Prediction
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 AIFI, a contrastive-learning alignment method for joint aspect-sentiment prediction whose temperature is adaptively set by a learned inter-task correlation network, reporting state-of-the-art F1 on three ALSA datasets.
Problem
Joint aspect prediction and sentiment prediction suffer from negative transfer when inter-task feature interactions are fixed by rigid regularization alignment, and the correlation strength between aspect and sentiment features varies across sentences.
Approach
Align AP-specific and SP-specific representations of the same sentence as positives with InfoNCE over in-batch negatives, and predict per-sentence correlation strength with a feature correlation network whose output sets the InfoNCE temperature (smaller for strongly correlated sentences), trained on a count of non-neutral aspect sentiments.
Key results
- AIFI reaches F1 of 0.7991 on DR, 0.6963 on DL and 0.6130 on DT, outperforming HI-ASA by +1.22, +0.92 and +1.26 F1 points.
- On the SP sub-task AIFI improves accuracy over HI-ASA by +0.38, +1.11 and +1.67 points on DR, DL and DT.
- Removing contrastive learning lowers joint F1 notably; ChatGPT few-shot reaches only 0.6762 F1 on DR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AIFI instead of HI-ASA | joint aspect-sentiment prediction · DR | F1: 0.7879 -> 0.7991 | Helped |
| use AIFI instead of HI-ASA | joint aspect-sentiment prediction · DL | F1: 0.6871 -> 0.6963 | Helped |
| use AIFI instead of HI-ASA | joint aspect-sentiment prediction · DT | F1: 0.6004 -> 0.6130 | Helped |
| add contrastive learning (ablation) | AIFI · DR | F1: 0.7931 -> 0.7991 | Helped |
| add contrastive learning (ablation) | AIFI · DL | F1: 0.6602 -> 0.6963 | Helped |
| add contrastive learning (ablation) | AIFI · DT | F1: 0.5887 -> 0.6130 | Helped |
| add adaptive temperature to contrastive learning (ablation) | AIFI · DR | F1: 0.7955 -> 0.7991 | Helped |
| add adaptive temperature to contrastive learning (ablation) | AIFI · DL | F1: 0.6847 -> 0.6963 | Helped |
| add adaptive temperature to contrastive learning (ablation) | AIFI · DT | F1: 0.6086 -> 0.6130 | Helped |
Surprising findings
- ChatGPT with few-shot prompting lags far behind fine-tuned BERT-based models on all three datasets.
- On the DT Twitter dataset, AIFI's AP F1 is lower than HI-ASA (0.7523 vs 0.7621), attributed to short tweets limiting context.
Limitations
- Performance on short Twitter sentences is weaker for aspect prediction, since limited context makes correlation strength hard to estimate.
- Error analysis shows difficulty with multiple opinion terms per aspect and implicit sentiment expressions.
- Correlation ground truth relies on a heuristic count of non-neutral aspect sentiments.
Experimental setup
Datasets: DR (SemEval 2014/2015/2016 restaurant) DL (SemEval 2014 laptop) DT (Twitter)
Benchmarks: DR DL DT
Metrics: precision recall F1 accuracy
Baselines: SPAN-BERT DOER S-AESC SPRM HI-ASA ChatGPT
Builds on: InfoNCE HI-ASA multi-gate mixture-of-experts BERT
Closest papers by meaning
- You Only Read Once: Constituency-Oriented Relational Graph Convolutional Network for Multi-Aspect Multi-Sentiment Classification (AAAI 2024 · similarity 0.86)
- A Novel Energy Based Model Mechanism for Multi-Modal Aspect-Based Sentiment Analysis (AAAI 2024 · similarity 0.85)
- Exploring Model Editing for LLM-based Aspect-Based Sentiment Classification (AAAI 2025 · similarity 0.84)
- Aspect-Based Sentiment Analysis with Explicit Sentiment Augmentations (AAAI 2024 · similarity 0.84)
- SimRP: Syntactic and Semantic Similarity Retrieval Prompting Enhances Aspect Sentiment Quad Prediction (AAAI 2025 · similarity 0.83)
- 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.