Research map / Label-efficient and robust learning / Recognition and long-tail
Variational Hybrid-Attention Framework for Multi-Label Few-Shot Aspect Category Detection
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 VHAF, a variational hybrid-attention framework that replaces prototype point estimates with aspect-specific Gaussian distributions and an adaptive threshold for multi-label few-shot aspect category detection.
Problem
Multi-label few-shot aspect category detection (FS-ACD) where prototypical networks use point estimates that are sensitive to noise from irrelevant aspects and biased by scarce support data.
Approach
Combines hybrid attention (aspect-wise and cross-instance attention) to get aspect-specific embeddings, then models each aspect as a variational Gaussian distribution with a class-conditioned prior, scores queries by negative KL divergence, and learns a per-aspect threshold with a focal loss.
Key results
- On FewAsp, VHAF reaches 98.17% AUC and 89.22% macro-F1 in 10-way 5-shot, versus 97.51% and 81.73% for the LDF baseline (LPN-level comparison is closer).
- On FewAsp (multi) 10-way 5-shot, VHAF achieves 96.01 AUC and 75.92 macro-F1, improving over LPN by 1.50 AUC and 8.64 macro-F1 points.
- Ablation on FewAsp shows removing the variational distribution inference drops macro-F1 from 89.22 to 83.06 in 10-way 5-shot.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use VHAF instead of LPN | FS-ACD · FewAsp (single), 10-way 5-shot | AUC: 99.14 -> 99.16 | Helped |
| Use VHAF instead of LPN | FS-ACD · FewAsp (single), 10-way 5-shot | macro-F1: 89.40 -> 89.60 | Helped |
| Use VHAF instead of LPN | FS-ACD · FewAsp (multi), 10-way 5-shot | AUC: 94.51 -> 96.01 | Helped |
| Use VHAF instead of LPN | FS-ACD · FewAsp (multi), 10-way 5-shot | macro-F1: 67.28 -> 75.92 | Helped |
| Use VHAF instead of LPN | FS-ACD · FewAsp, 10-way 5-shot | AUC: 95.36 -> 97.02 | Helped |
| Use VHAF instead of LPN | FS-ACD · FewAsp, 10-way 5-shot | macro-F1: 71.42 -> 79.72 | Helped |
| Add the full hybrid-attention feature extraction (AWA and CIA) (ablation) | VHAF · FewAsp, 10-way 5-shot | macro-F1: 83.06 -> 89.22 | Helped |
| Add the full hybrid-attention feature extraction (AWA and CIA) (ablation) | VHAF · FewAsp (multi), 10-way 5-shot | macro-F1: 64.46 -> 75.92 | Helped |
| Add aspect-wise attention (ablation) | VHAF · FewAsp, 10-way 5-shot | macro-F1: 87.15 -> 88.38 | Helped |
| Add aspect-wise attention (ablation) | VHAF · FewAsp (multi), 10-way 5-shot | macro-F1: 69.53 -> 74.99 | Helped |
| Add cross-instance attention (ablation) | VHAF · FewAsp, 10-way 5-shot | macro-F1: 85.82 -> 88.38 | Helped |
| Add cross-instance attention (ablation) | VHAF · FewAsp (multi), 10-way 5-shot | macro-F1: 72.65 -> 74.99 | Helped |
Limitations
- Evaluated only on YelpAspect-derived FewAsp datasets with BERT-base encoder
- Performance on multi-aspect sentences remains lower than on single-aspect sentences
Experimental setup
Datasets: FewAsp (single) FewAsp (multi) FewAsp
Benchmarks: FewAsp (single) FewAsp (multi) FewAsp
Metrics: AUC macro-F1
Baselines: Matching Network Prototypical Network Relation Network Graph Network Proto-HATT Proto-AWATT LDF LPN
Builds on: Proto-AWATT prototypical network variational few-shot learning focal loss BERT
Closest papers by meaning
- VERO: Verification and Zero-Shot Feedback Acquisition for Few-Shot Multimodal Aspect-Level Sentiment Classification (AAAI 2025 · similarity 0.83)
- A Novel Energy Based Model Mechanism for Multi-Modal Aspect-Based Sentiment Analysis (AAAI 2024 · similarity 0.83)
- You Only Read Once: Constituency-Oriented Relational Graph Convolutional Network for Multi-Aspect Multi-Sentiment Classification (AAAI 2024 · similarity 0.82)
- Seek Common Ground While Reserving Differences: Semi-Supervised Image-Text Sentiment Recognition (CVPR 2025 · similarity 0.82)
- Multi-Label Few-Shot Image Classification via Pairwise Feature Augmentation and Flexible Prompt Learning (AAAI 2025 · similarity 0.82)
- Few-Shot Learning from Augmented Label-Uncertain Queries in Bongard-HOI (AAAI 2024 · similarity 0.81)
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.