Research map / Vision-language and multimodal / Cross-modal learning
Aligner²: Enhancing Joint Multiple Intent Detection and Slot Filling via Adjustive and Forced Cross-Task Alignment
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 Aligner² for joint multi-intent detection and slot filling, adding an adjustive cross-task aligner (label-embedding projection before task interaction) and a forced cross-task aligner (REINFORCE with SLU-metric rewards after interaction), and reports gains when plugged into several existing models on MixATIS and MixSNIPS.
Problem
Multi-intent spoken language understanding suffers from misaligned intent scopes over utterance fragments and from independent decoding of intent and slot predictions, which lowers utterance-level overall accuracy.
Approach
Before interaction, task-specific hidden states are projected into a joint label embedding space and re-expressed via dot products with label embeddings (ACA) to align scopes; after interaction, REINFORCE with a self-critical reference reward optimizes overall SLU accuracy as the reward (FCA) to align the two subtasks' predictions.
Key results
- On MixATIS, plugging Aligner² into ReLa-Net raises overall accuracy from 52.2% to 52.8%, and into SSRAN from 48.7% to 49.4%.
- On MixSNIPS, Aligner² with ReLa-Net reaches 76.5% overall accuracy (vs 76.1%) and with SSRAN 77.8% (vs 77.4%).
- With RoBERTa encoder, Aligner² reaches 55.6% overall accuracy on MixATIS, above ReLa-Net + RoBERTa at 54.4%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Aligner2 to ReLa-Net | ReLa-Net · MixATIS | Overall accuracy: 52.2 -> 52.8 | Helped |
| Add Aligner2 to ReLa-Net | ReLa-Net · MixSNIPS | Overall accuracy: 76.1 -> 76.5 | Helped |
| Add Aligner2 to SSRAN | SSRAN · MixATIS | Overall accuracy: 48.7 -> 49.4 | Helped |
| Add Aligner2 to SSRAN | SSRAN · MixSNIPS | Overall accuracy: 77.4 -> 77.8 | Helped |
| Add FCA (ablation) | ReLa-Net with Aligner2 · MixATIS | Overall accuracy: 52.4 -> 52.8 | Helped |
| Add FCA (ablation) | ReLa-Net with Aligner2 · MixSNIPS | Overall accuracy: 76.2 -> 76.5 | Helped |
| Add ACA (ablation) | ReLa-Net with Aligner2 · MixATIS | Overall accuracy: 52.6 -> 52.8 | Helped |
| Add ACA (ablation) | ReLa-Net with Aligner2 · MixSNIPS | Overall accuracy: 76.3 -> 76.5 | Helped |
| Add FCA (ablation) | SSRAN with Aligner2 · MixATIS | Overall accuracy: 48.9 -> 49.4 | Helped |
| Add FCA (ablation) | SSRAN with Aligner2 · MixSNIPS | Overall accuracy: 77.5 -> 77.8 | Helped |
| Add ACA (ablation) | SSRAN with Aligner2 · MixATIS | Overall accuracy: 49.1 -> 49.4 | Helped |
| Add ACA (ablation) | SSRAN with Aligner2 · MixSNIPS | Overall accuracy: 77.7 -> 77.8 | Helped |
Surprising findings
- Optimizing the reward with only slot F1 or only intent accuracy as the reference worsens the other subtask, exacerbating misalignment.
- Projection strategy (cosine, Euclidean, correlation) has minimal influence on ACA performance.
Limitations
- Gains over baselines are modest (roughly 0.4% to 0.8% absolute overall accuracy).
- Evaluated only on two multi-intent datasets and some single-intent datasets; no LLM-based evaluation.
Experimental setup
Datasets: MixATIS MixSNIPS ATIS SNIPS
Benchmarks: MixATIS MixSNIPS ATIS SNIPS
Metrics: slot F1 intent accuracy overall accuracy
Baselines: AGIF GL-GIN SDJN GISC Co-guiding Net ReLa-Net SSRAN
Builds on: AGIF GL-GIN ReLa-Net SSRAN self-critical sequence training
Closest papers by meaning
- Uni-MIS: United Multiple Intent Spoken Language Understanding via Multi-View Intent-Slot Interaction (AAAI 2024 · similarity 0.93)
- Towards Multi-Intent Spoken Language Understanding via Hierarchical Attention and Optimal Transport (AAAI 2024 · similarity 0.92)
- MIntRec2.0: A Large-scale Benchmark Dataset for Multimodal Intent Recognition and Out-of-scope Detection in Conversations (ICLR 2024 · similarity 0.88)
- Introducing Visual Scenes and Reasoning: A More Realistic Benchmark for Spoken Language Understanding (AAAI 2026 · similarity 0.88)
- Enhancing Speech Large Language Models through Reinforced Behavior Alignment (NeurIPS 2026 · similarity 0.86)
- Divide-Solve-Combine: An Interpretable and Accurate Prompting Framework for Zero-shot Multi-Intent Detection (AAAI 2025 · 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.