Research map / Vision-language and multimodal / Human and embodied modeling
Towards Multi-Intent Spoken Language Understanding via Hierarchical Attention and Optimal Transport
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 HAOT, a single-stage multi-intent SLU model that uses hierarchical attention to split utterances into intent scopes and optimal transport to get bidirectional slot-intent guidance.
Problem
Multi-intent spoken language understanding suffers from scope barrier (irrelevant context hurting each intent's prediction) and unidirectional guidance (intent-to-slot only, not slot-to-intent).
Approach
Replaces standard encoder attention with hierarchical attention driven by a learned, progressively increasing neighboring-affinity attention mask that merges tokens into intent-specific clusters, and models slot-intent mutual guidance as a relaxed optimal transport cost added to the training loss.
Key results
- HAOT reaches 55.6% overall accuracy on MixATIS, 3.4 points above the prior best (Co-guiding Net, 51.3 / 52.2 reported in table), with slot F1 92.8 and intent accuracy 82.2.
- On MixSNIPS HAOT achieves 79.7% overall accuracy, 2.2 points above the previous best, with slot F1 96.4 and intent accuracy 98.9.
- Replacing the relaxed transport with exact IPOT gives similar accuracy (55.8 vs 55.6 on MixATIS) but about twice the per-batch training time (7.3 vs 3.7).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add hierarchical attention instead of conventional attention (ablation) | HAOT · MixATIS | overall accuracy: 51.8 -> 55.6 | Helped |
| Add hierarchical attention instead of conventional attention (ablation) | HAOT · MixSNIPS | overall accuracy: 76.9 -> 79.7 | Helped |
| Replace conventional attention with hierarchical attention and increase encoder depth to six layers (ablation) | HAOT · MixATIS | overall accuracy: 52.1 -> 55.6 | Helped |
| Replace conventional attention with hierarchical attention and increase encoder depth to six layers (ablation) | HAOT · MixSNIPS | overall accuracy: 77.1 -> 79.7 | Helped |
| Add the optimal transport loss (ablation) | HAOT · MixATIS | overall accuracy: 51.4 -> 55.6 | Helped |
| Add the optimal transport loss (ablation) | HAOT · MixSNIPS | overall accuracy: 76.6 -> 79.7 | Helped |
| Replace relaxed moving distance with Sinkhorn (ablation) | HAOT · MixATIS | overall accuracy: 55.6 -> 52.8 | Hurt |
| Replace relaxed moving distance with Sinkhorn (ablation) | HAOT · MixSNIPS | overall accuracy: 79.7 -> 78.3 | Hurt |
| Replace relaxed moving distance with IPOT (ablation) | HAOT · MixATIS | overall accuracy: 55.6 -> 55.8 | No clear effect |
| Replace relaxed moving distance with IPOT (ablation) | HAOT · MixSNIPS | overall accuracy: 79.7 -> 79.8 | No clear effect |
| Use relaxed moving distance instead of Sinkhorn (ablation) | HAOT · MixATIS | training time per batch: 8.7 -> 3.7 | Helped |
| Use relaxed moving distance instead of Sinkhorn (ablation) | HAOT · MixSNIPS | training time per batch: 9.9 -> 4.8 | Helped |
Surprising findings
- ChatGPT performs far below fine-tuned models on this task (34.2% overall accuracy on MixATIS, 43.7 slot F1), indicating a large gap for LLM prompting on SLU.
Limitations
- The gain over strong baselines on MixSNIPS intent accuracy is small (0.5 points), and the authors' own claim of 3.4 points on MixATIS is relative to the prior best row.
- Relaxed optimal transport drops the second marginal constraint, so it only gives a lower bound on the exact transport cost.
Experimental setup
Datasets: MixATIS MixSNIPS
Benchmarks: MixATIS MixSNIPS
Metrics: overall accuracy slot F1 intent accuracy
Baselines: Bi-Model SF-ID Stack-Propagation Joint Multiple ID-SF AGIF GL-GIN LR-Transformer SDJN GISCo Co-guiding Net ReLa-Net DARER2
Builds on: AGIF GL-GIN Co-guiding Net Stack-Propagation Tree Transformer
Closest papers by meaning
- Uni-MIS: United Multiple Intent Spoken Language Understanding via Multi-View Intent-Slot Interaction (AAAI 2024 · similarity 0.93)
- Aligner²: Enhancing Joint Multiple Intent Detection and Slot Filling via Adjustive and Forced Cross-Task Alignment (AAAI 2024 · similarity 0.92)
- Introducing Visual Scenes and Reasoning: A More Realistic Benchmark for Spoken Language Understanding (AAAI 2026 · similarity 0.88)
- MIntRec2.0: A Large-scale Benchmark Dataset for Multimodal Intent Recognition and Out-of-scope Detection in Conversations (ICLR 2024 · similarity 0.88)
- Divide-Solve-Combine: An Interpretable and Accurate Prompting Framework for Zero-shot Multi-Intent Detection (AAAI 2025 · similarity 0.87)
- MMSU: A Massive Multi-task Spoken Language Understanding and Reasoning Benchmark (ICLR 2026 · similarity 0.86)
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.