TA&AT: Enhancing Task-Oriented Dialog with Turn-Level Auxiliary Tasks and Action-Tree Based Scheduled Sampling

Longxiang Liu, Xiuxing Li, Yang Feng

Abstract

Task-oriented dialog (TOD) systems have witnessed substantial progress due to conversational pre-training techniques. Yet, two significant challenges persist. First, most systems primarily utilize the latest turn's state label for the generator, overlooking the comprehensive value of state labels in boosting the model's understanding for future generations. Second, an overreliance on generated policy often leads to error accumulation, resulting in suboptimal responses when adhering to incorrect actions. To combat these challenges, we propose turn-level multi-task objectives for the encoder (TA) and an action tree-based scheduled sampling technique (AT) for the decoder. Our approach achieved state-of-the-art (SOTA) performance on the MultiWOZ dataset series among methods without continual pre-training and remains competitive with pre-trained SOTA methods.

1. Introduction

Building robust Task-Oriented Dialog (TOD) systems involves a complex pipeline: Natural Language Understanding (NLU), Dialog State Tracking (DST), Policy Learning (POL), and Natural Language Generation (NLG). While end-to-end models have simplified this pipeline, they often suffer from two bottlenecks: underutilized intermediate labels and exposure bias. Most systems use labels like belief states or system actions only to supervise the generator, ignoring their potential to sharpen the encoder's representation. Furthermore, during inference, models often struggle with error accumulation; a single incorrect action choice can lead to a cascade of nonsensical responses. We introduce the TA&AT framework to address these issues through structured supervision and robust sampling.

2. Methodology

2.1 Turn-Level Auxiliary Tasks (TA)

To enhance the encoder's understanding, we introduce four auxiliary classification heads used exclusively during training. These tasks force the model to build richer turn-level representations:

  • Slot Type Prediction: Identifies which slots (e.g., price, area) are active in the current turn.
  • Slot Change Prediction (Transition): Tracks how the belief state evolves (e.g., whether a slot is kept, changed, deleted, or newly introduced).
  • Action Type Prediction: Predicts the high-level system intent (e.g., inform, request).
  • Response Keywords Prediction: Predicts the essential values that must appear in the final response.

2.2 Action-Tree Based Scheduled Sampling (AT)

To mitigate exposure bias, we model hierarchical policies as trees (Domain → Action → Slot). We introduce a scheduled sampling mechanism that utilizes "Tree Edit Distance" to identify similar but incorrect action sequences. During training, the model is occasionally fed these "noisy" negative samples instead of ground truth actions. This forces the decoder to learn invariant response generation under perturbations, effectively bridging the gap between perfect training data and imperfect inference-time predictions.

Overview of the TA&AT framework, illustrating the turn-level auxiliary tasks (TA) integrated into the Dialog Encoder and the action-tree based scheduled sampling (AT) mechanism for the decoder.

3. Experimental Evaluation and Results

3.1 Performance on MultiWOZ

We evaluated TA&AT on MultiWOZ 2.0, 2.1, and 2.2. The model achieved a new SOTA Combined Score (Inform + Success + 0.5 * BLEU) for models without continual pre-training, reaching 109.27 on MultiWOZ 2.0. Notably, it outperformed several models that utilize expensive pre-training, particularly in the BLEU metric, indicating superior fluency and relevance.

3.2 Efficiency and Low-Resource Settings

Because the auxiliary tasks and scheduled sampling are training-time interventions, the inference speed remains identical to standard base models. In low-resource experiments (using only 10% of training data), TA&AT demonstrated significant robustness, outperforming the previous best model (Mars) by over 3 points in the combined score.

3.3 Qualitative Analysis

Qualitative comparisons reveal that TA&AT is more effective at capturing all required information slots requested by the user. While baseline models might miss critical details like addresses or phone numbers when the generated action is slightly off, TA&AT's robustness training allows it to recover and provide more complete information.

Qualitative comparison of dialogue responses: TA&AT successfully provides all requested information (address and phone), whereas the Mars baseline misses these details.

4. Conclusion

The TA&AT framework demonstrates that better utilization of existing labels through auxiliary tasks and structured noise injection can significantly improve TOD performance. By focusing on turn-level representations and action-tree robustness, we achieve SOTA results on standard benchmarks without the need for massive continual pre-training.

Summary updated

Abstract

Task-oriented dialog systems have witnessed substantial progress due to conversational pre-training techniques. Yet, two significant challenges persist. First, most systems primarily utilize the latest turn's state label for the generator. This practice overlooks the comprehensive value of state labels in boosting the model's understanding for future generations. Second, an overreliance on generated policy often leads to error accumulation, resulting in suboptimal responses when adhering to incorrect actions. To combat these challenges, we propose turn-level multi-task objectives for the encoder. With the guidance of essential information from labeled intermediate states, we establish a more robust representation for both understanding and generation. For the decoder, we introduce an action tree-based scheduled sampling technique. Specifically, we model the hierarchical policy as trees and utilize the similarity between trees to sample negative policy based on scheduled sampling, hoping the model to generate invariant responses under perturbations. This method simulates potential pitfalls by sampling similar negative policy, bridging the gap between task-oriented dialog training and inference. Among methods without continual pre-training, our approach achieved state-of-the-art (SOTA) performance on the MultiWOZ dataset series and was also competitive with pre-trained SOTA methods.
BibTeX
@inproceedings{liu2024ta,
  author = {Longxiang Liu and Xiuxing Li and Yang Feng},
  title = {{TA\&AT: Enhancing Task-Oriented Dialog with Turn-Level Auxiliary Tasks and Action-Tree Based Scheduled Sampling}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=KitRgxAH9M},
}
Figure