Research map / Detection and segmentation / CNN features
Towards Explainable Joint Models via Information Theory for Multiple Intent Detection and Slot Filling
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 a theoretical information-gain framework (cross-task information gain) for multi-intent SLU and InfoJoint, a multi-stage iterative joint model trained with multi-level contrastive learning and margin constraints, which outperforms prior joint models on MixATIS and MixSNIPS.
Problem
Joint models for multi-intent detection and slot filling are designed heuristically and lack theoretical measurement of interaction quality and explainability of design and optimization.
Approach
Defines cross-task information gain (CIG) as an information-theoretic measure of joint-interaction quality, models the dual-task interaction as a multi-stage Markov process with proven constraints (intrinsic, directional, upper-bound) and convergence, and instantiates it as InfoJoint, which maximizes CIG via multi-level cross-task InfoNCE contrastive learning with margin penalties and entropy-base
Key results
- InfoJoint (Tmax=10) reaches 52.5% overall accuracy on MixATIS vs 51.3% for Co-guiding Net, and 78.9% on MixSNIPS vs 77.5% for Co-guiding Net and SSRAN.
- With RoBERTa encoder, InfoJoint reaches 58.6 overall on MixATIS and 86.1 on MixSNIPS, above Co-guiding Net+RoBERTa (57.5, 85.3).
- ChatGPT achieves only 34.2% overall accuracy on MixATIS in the authors' evaluation, far below InfoJoint.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use InfoJoint (Tmax=10) instead of Co-guiding Net | Multi-intent SLU · MixATIS | Overall Acc: 51.3 -> 52.5 | Helped |
| use InfoJoint (Tmax=10) instead of SSRAN | Multi-intent SLU · MixSNIPS | Overall Acc: 77.5 -> 78.9 | Helped |
| increase Tmax from 5 to 10 (ablation) | InfoJoint · MixATIS | Overall Acc: 51.9 -> 52.5 | Helped |
| increase Tmax from 5 to 10 (ablation) | InfoJoint · MixSNIPS | Overall Acc: 78.2 -> 78.9 | Helped |
| add multi-level contrastive loss, directional constraints, time coding, and entropy weighting (ablation) | InfoJoint · MixATIS | Overall Acc: 43.2 -> 52.5 | Helped |
| add multi-level contrastive loss, directional constraints, time coding, and entropy weighting (ablation) | InfoJoint · MixSNIPS | Overall Acc: 74.5 -> 78.9 | Helped |
| add utterance-level contrastive loss to token-level contrastive loss (ablation) | InfoJoint · MixATIS | Overall Acc: 46.2 -> 52.5 | Helped |
| add utterance-level contrastive loss to token-level contrastive loss (ablation) | InfoJoint · MixSNIPS | Overall Acc: 76.4 -> 78.9 | Helped |
| add token-level contrastive loss to utterance-level contrastive loss (ablation) | InfoJoint · MixATIS | Overall Acc: 49.7 -> 52.5 | Helped |
| add token-level contrastive loss to utterance-level contrastive loss (ablation) | InfoJoint · MixSNIPS | Overall Acc: 77.5 -> 78.9 | Helped |
| add both directional margin penalties (ablation) | InfoJoint · MixATIS | Overall Acc: 47.6 -> 52.5 | Helped |
| add both directional margin penalties (ablation) | InfoJoint · MixSNIPS | Overall Acc: 76.9 -> 78.9 | Helped |
Surprising findings
- ChatGPT zero-shot is competitive on intent accuracy (66.1 on MixATIS) but very poor on slot F1 (43.7) and overall accuracy (34.2).
Limitations
- Convergence guarantee relies on theoretical propositions; the energy function is only estimated via Monte Carlo and cannot be computed exactly.
- Gains over strongest baselines are modest on MixSNIPS slot F1 (0.3%).
Experimental setup
Datasets: MixATIS MixSNIPS
Benchmarks: MixATIS MixSNIPS
Metrics: intent accuracy slot F1 overall accuracy
Baselines: Bi-Model Stack-Propagation Joint Multiple ID-SF AGIF GL-GIN GISCO Co-guiding Net SSRAN ChatGPT
Builds on: AGIF Co-guiding Net SSRAN InfoNCE Stack-Propagation
Closest papers by meaning
- Aligner²: Enhancing Joint Multiple Intent Detection and Slot Filling via Adjustive and Forced Cross-Task Alignment (AAAI 2024 · similarity 0.84)
- Decoupling Representation and Knowledge for Few-Shot Intent Classification and Slot Filling (AAAI 2024 · similarity 0.83)
- Uni-MIS: United Multiple Intent Spoken Language Understanding via Multi-View Intent-Slot Interaction (AAAI 2024 · similarity 0.83)
- Towards Multi-Intent Spoken Language Understanding via Hierarchical Attention and Optimal Transport (AAAI 2024 · similarity 0.82)
- Divide-Solve-Combine: An Interpretable and Accurate Prompting Framework for Zero-shot Multi-Intent Detection (AAAI 2025 · similarity 0.82)
- Generalized Intention Modeling in Multi-Agent Reinforcement Learning (NeurIPS 2026 · similarity 0.82)
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.