CoDial: Interpretable Task-Oriented Dialogue Systems Through Dialogue Flow Alignment

ACL ARR 2025 February Submission4799 Authors

Abstract: It is often challenging to integrate specialized, unseen tasks to dialogue systems due to the high cost of expert knowledge, training data, and high technical difficulty. To facilitate cross-disciplinary collaboration, it is then essential to build frameworks with a low cost of specification to enable adoption by experts from other domains. In this work, we introduce CoDial (**Co**de for **Dial**ogue), a novel framework that converts structured expert knowledge, represented as a heterogeneous graph, into conversation logic. We create an implementation of our framework with Colang, a guardrailing language that aligns language models without additional training. Empirically, our framework achieves new state-of-the-art performance on the STAR dataset for inference-based models and is competitive with similar baselines on the well-known MultiWOZ dataset. We also demonstrate how to further improve experimental results with user feedback. CoDial is an interpretable and modifiable framework for task-oriented dialogue that can specify conversation logic in a strict zero-shot setting

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: task-oriented,grounded dialog,dialogue state tracking

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 4799