Research map / Large language models / Language, safety and interpretability
Uni-MIS: United Multiple Intent Spoken Language Understanding via Multi-View Intent-Slot Interaction
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 Uni-MIS, a joint multi-intent SLU model that fuses utterance-, chunk-, and token-level intent-slot interaction over a RoBERTa encoder with chunk-level intent detection and adaptive intent-slot graph slot filling, and evaluates ChatGPT on the task.
Problem
Multi-intent spoken language understanding requires joint intent detection and slot filling, but prior token-level intent-slot interaction underuses intent information at other granularities and suffers from error propagation.
Approach
Models intent-slot interaction at three views (utterance [CLS] intent, chunk-level sliding-window intent, token-level intent), fuses them with BiLSTM before adaptive intent-slot graph interaction for slot filling, and jointly trains with chunk-level intent detection.
Key results
- On MixATIS, Uni-MIS reaches 52.5% overall accuracy vs. 49.4% for CLID (Roberta), and 88.3 slot F1.
- On MixSNIPS, Uni-MIS reaches 83.4% overall accuracy vs. 82.2% for CLID (Roberta), and 96.4 slot F1.
- ChatGPT (gpt-3.5-turbo) with 3-shot prompting scores only 0.36% overall accuracy on MixATIS and 0.09% on MixSNIPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Uni-MIS instead of CLID (RoBERTa) | Multi-intent SLU · MixATIS | Overall accuracy: 3.1% | Helped |
| Use Uni-MIS instead of CLID (RoBERTa) | Multi-intent SLU · MixATIS | Slot F1: 2.4% | Helped |
| Use Uni-MIS instead of CLID (RoBERTa) | Multi-intent SLU · MixSNIPS | Overall accuracy: 1.2% | Helped |
| Use Uni-MIS instead of CLID (RoBERTa) | Multi-intent SLU · MixSNIPS | Slot F1: 0.2% | Helped |
| Use Uni-MIS instead of CLID (RoBERTa) | Multi-intent SLU · MixSNIPS | Intent accuracy: 0.4% | Helped |
| add utterance intent-slot interaction (ablation) | Uni-MIS · MixATIS | Overall accuracy: decrease of 1.4% | Helped |
| add utterance intent-slot interaction (ablation) | Uni-MIS · MixSNIPS | Overall accuracy: decrease of 1.3% | Helped |
| add chunk intent-slot interaction (ablation) | Uni-MIS · MixATIS | Overall accuracy: reduction of 0.6% | Helped |
| add chunk intent-slot interaction (ablation) | Uni-MIS · MixSNIPS | Overall accuracy: reduction of 1.3% | Helped |
| add token intent-slot interaction (ablation) | Uni-MIS · MixATIS | Overall accuracy: decline of 1.0% | Helped |
| add token intent-slot interaction (ablation) | Uni-MIS · MixSNIPS | Overall accuracy: decline of 1.3% | Helped |
Surprising findings
- ChatGPT achieves much higher intent accuracy on MixSNIPS (71.49%) than on MixATIS (21.62%), though its overall and slot scores are near zero.
Limitations
- Gains over CLID (Roberta) on MixSNIPS are small (1.2 points overall accuracy).
- Slot F1 decreases as the number of intents grows, and multi-intent ambiguity hurts slot filling.
- The ChatGPT benchmark is limited to a single prompt design with 3-shot examples.
Experimental setup
Datasets: MixATIS MixSNIPS
Benchmarks: MixATIS MixSNIPS
Metrics: slot F1 intent accuracy overall accuracy
Baselines: Bi-Model Slot-Gated SF-ID Stack-Propagation AGIF GL-GIN SDJN CLID CLID (Roberta)
Builds on: CLID AGIF GL-GIN Stack-Propagation
Closest papers by meaning
- Towards Multi-Intent Spoken Language Understanding via Hierarchical Attention and Optimal Transport (AAAI 2024 · similarity 0.93)
- Aligner²: Enhancing Joint Multiple Intent Detection and Slot Filling via Adjustive and Forced Cross-Task Alignment (AAAI 2024 · similarity 0.93)
- 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 0.87)
- 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 0.85)
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.