FFS: Few-Shot Language Feedback for Domain Adaptation in End-to-End Dialogue State Tracking

ACL ARR 2025 February Submission4611 Authors

Abstract: End-to-end task-oriented dialogue (TOD) systems have become increasingly feasible due to advancements in language modeling. However, tasks such as dialogue state tracking (DST) remain challenging, particularly in domain adaptation, where models must generalize to unseen domains without additional supervision. While Large Language Models (LLMs) exhibit strong fine-tuning performance and even generalization, they still make mistakes and it can be difficult to correct those errors through fine-tuning. In this work, we propose a method that enables improvement of a fine-tuned LLM by incorporating few-shot language feedback. Our approach follows a two-step process: first, we bootstrap a draft model using data augmentation techniques to improve schema robustness. This model is then applied to a validation set, where incorrect predictions are identified. In the second step, expert annotators provide targeted natural language feedback on a subset of these errors, explicitly guiding the model on how to improve its performance on the task. The model is then fine-tuned again on both the original data and the feedback-augmented examples. Experiments on MultiWOZ and SpokenWOZ demonstrate that integrating language feedback in this manner improves DST performance by up to 5.8\% in unseen domains.

Paper Type: Short

Research Area: Language Modeling

Research Area Keywords: Machine Learning for NLP, Generation, Efficient/Low-Resource Methods for NLP

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 4611