Research map / Label-efficient and robust learning / Recognition and long-tail
Winnie: Task-Oriented Dialog System with Structure-Aware Contrastive Learning and Enhanced Policy Planning
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 Winnie, an end-to-end task-oriented dialog system built on T5-base that adds a dialogue-level Transformer over utterance representations, supervised contrastive learning, and a system act classification loss, fine-tuned without extra pre-training.
Problem
End-to-end task-oriented dialog systems underuse dialog structure (interactions among user utterances, belief states, DB results, system acts, responses) and dialog policy, and existing system-act improvements rely on extra pre-training corpora.
Approach
Model dialog structure as a residual via a dialogue-level Transformer over max-pooled per-turn T5 representations, train it with supervised contrastive loss grouping same-category turn components, and add a system act classification auxiliary loss during fine-tuning, with no additional pre-training corpora.
Key results
- On MultiWOZ2.2 Winnie reaches a combined score of 103.4, 3.2 above the prior best MTTOD (100.2) and above GALAXY, which uses extra pre-training corpora.
- On CamRest676 Winnie reaches a combined score of 118.3 (vs. GALAXY 117.3) with a match rate of 100.0.
- On In-Car Winnie reaches a combined score of 109.5 (vs. GALAXY 107.5); ablation on MultiWOZ shows SCL alone +1.2 and SAC alone +1.0 combined over the T5 baseline (99.9), and both together 103.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add supervised contrastive learning (ablation) | T5-base TOD system · MultiWOZ2.2 | Combined: 99.9 -> 100.9 | Helped |
| add system act classification (ablation) | T5-base TOD system · MultiWOZ2.2 | Combined: 99.9 -> 101.1 | Helped |
| add both system act classification and supervised contrastive learning (ablation) | T5-base TOD system · MultiWOZ2.2 | Combined: 99.9 -> 103.4 | Helped |
| add Winnie’s contrastive learning and system act classification | T5-base TOD system · MultiWOZ2.2 | Combined: 100.2 -> 103.4 | Helped |
| use Winnie instead of GALAXY | End-to-end TOD system · CamRest676 | Combined: 117.3 -> 118.3 | Helped |
| use Winnie instead of GALAXY | End-to-end TOD system · In-Car | Combined: 107.5 -> 109.5 | Helped |
| increase the SCL temperature to 15.0 (ablation) | Winnie · MultiWOZ2.2 | Combined | Helped |
Surprising findings
- System act classification gave a larger gain than supervised contrastive learning in the ablation, which the authors attribute to system act quality strongly affecting response quality.
Limitations
- Temperature hyperparameter for SCL is sensitive; values below 15.0 degrade performance.
- CamRest676 results rely on warm-starting from a MultiWOZ2.2-trained model because the dataset is small.
Experimental setup
Datasets: MultiWOZ2.2 CamRest676 Stanford In-Car Assistant
Benchmarks: MultiWOZ2.2 CamRest676 Stanford In-Car Assistant
Metrics: Inform Success SuccessF1 BLEU Combined score Match rate
Baselines: MTTOD GALAXY PPTOD BORT MinTL SOLOIST UBAR LABES Sequicity T5-base fine-tuning baseline
Builds on: T5 Supervised Contrastive Learning MTTOD GALAXY
Closest papers by meaning
- GALAXY: A Generative Pre-trained Model for Task-Oriented Dialog with Semi-supervised Learning and Explicit Policy Injection (AAAI 2022 · similarity 0.87)
- CINS: Comprehensive Instruction for Few-Shot Learning in Task-Oriented Dialog Systems (AAAI 2022 · similarity 0.86)
- TA&AT: Enhancing Task-Oriented Dialog with Turn-Level Auxiliary Tasks and Action-Tree Based Scheduled Sampling (AAAI 2024 · similarity 0.86)
- An Efficient Self-Learning Framework For Interactive Spoken Dialog Systems (ICML 2024 · similarity 0.84)
- Learning to Clarify: Multi-turn Conversations with Action-Based Contrastive Self-Training (ICLR 2025 · similarity 0.84)
- Structure-Aware Multimodal Sequential Learning for Visual Dialog (AAAI 2024 · similarity 0.83)
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.