Research map / Label-efficient and robust learning / Recognition and long-tail
TA&AT: Enhancing Task-Oriented Dialog with Turn-Level Auxiliary Tasks and Action-Tree Based Scheduled Sampling
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 turn-level auxiliary classification tasks on the encoder and action-tree based scheduled sampling for the decoder, built on a T5 shared-encoder two-decoder task-oriented dialog model, reaching SOTA among non-continual-pretraining methods on MultiWOZ 2.0/2.1/2.2.
Problem
End-to-end task-oriented dialog systems underuse intermediate state annotations and suffer from error accumulation when generated policies/actions are wrong, causing responses attached to incorrect actions.
Approach
Adds four turn-level auxiliary heads (slot type, slot change, action type, response keywords) over encoder turn representations, and performs sequence-level scheduled sampling in which ground-truth action trees are replaced by similar negative action trees sampled according to ordered tree-edit-distance similarity.
Key results
- Achieves combined scores of 109.27 on MultiWOZ 2.0, 108.03 on 2.1, and 103.59 on 2.2, SOTA among methods without continual pre-training.
- Improves over its MTTOD base by 2.23 (2.0), 1.81 (2.1), and 3.39 (2.2) combined points.
- Ablation shows removing the action-tree scheduled sampling (w.o. AT) drops combined score to 108.17 and removing turn-level tasks (w.o. TA) to 107.86 on MultiWOZ 2.0.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add slot type prediction loss (ablation) | TA&AT · MultiWOZ 2.0 | Comb: 108.59 -> 109.27 | Helped |
| Add slot change prediction loss (ablation) | TA&AT · MultiWOZ 2.0 | Comb: 108.83 -> 109.27 | Helped |
| Add action prediction loss (ablation) | TA&AT · MultiWOZ 2.0 | Comb: 109.00 -> 109.27 | Helped |
| Add response keywords prediction loss (ablation) | TA&AT · MultiWOZ 2.0 | Comb: 109.34 -> 109.27 | Hurt |
| Use action-tree scheduled sampling instead of normal teacher forcing (ablation) | TA&AT · MultiWOZ 2.0 | Comb: 108.17 -> 109.27 | Helped |
| Add turn-level auxiliary tasks (ablation) | TA&AT · MultiWOZ 2.0 | Comb: 107.86 -> 109.27 | Helped |
| Use TA&AT instead of MTTOD | T5-base end-to-end dialog system · MultiWOZ 2.0 | Comb: 107.04 -> 109.27 | Helped |
| Use TA&AT instead of MTTOD | T5-base end-to-end dialog system · MultiWOZ 2.1 | Comb: 106.22 -> 108.03 | Helped |
| Use TA&AT instead of MTTOD | T5-base end-to-end dialog system · MultiWOZ 2.2 | Comb: 100.2 -> 103.59 | Helped |
Surprising findings
- Removing the response keywords prediction loss slightly increased the combined score (+0.07), which the authors attribute possibly to randomness.
Limitations
- Still trails continual pre-training methods such as SPACE and GALAXY on MultiWOZ 2.0/2.1.
- Loss weighting across auxiliary tasks is not balanced; the response keywords task showed no benefit in ablation.
Experimental setup
Datasets: MultiWOZ 2.0 MultiWOZ 2.1 MultiWOZ 2.2
Benchmarks: MultiWOZ 2.0 MultiWOZ 2.1 MultiWOZ 2.2
Metrics: Inform Success BLEU Combined score
Baselines: SimpleTOD DoTS SOLOIST MinTL UBAR GALAXY BORT Mars MTTOD PPTOD SPACE
Builds on: MTTOD DialoFlow SPACE GALAXY Scheduled sampling
Closest papers by meaning
- Task-Oriented Dialog Systems That Consider Multiple Appropriate Responses under the Same Context (AAAI 2020 · similarity 0.90)
- GALAXY: A Generative Pre-trained Model for Task-Oriented Dialog with Semi-supervised Learning and Explicit Policy Injection (AAAI 2022 · similarity 0.89)
- StateTree: Enhancing Long-term Dialogue Reasoning via Reinforcement Learning (NeurIPS 2026 · similarity 0.88)
- Learning to Clarify: Multi-turn Conversations with Action-Based Contrastive Self-Training (ICLR 2025 · similarity 0.88)
- An Efficient Self-Learning Framework For Interactive Spoken Dialog Systems (ICML 2024 · similarity 0.87)
- From Retrieval to Generation: A Simple and Unified Generative Model for End-to-End Task-Oriented Dialogue (AAAI 2024 · similarity 0.87)
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.