PhD 2024S: James Finch
Leveraging Diverse Data Generation for Domain-Adaptable Dialogue State Tracking
James Finch
Date: 2024-03-29 / 3:30 ~ 4:30 PM
Location: MSC W303
Abstract
This work investigates improving domain adaptability in Dialogue State Tracking (DST), a crucial task for integrating conversational AI to real-world software applications. DST produces structured state representations that track important information in dialogue, which can be used as an interface to external software components and for controlling dialogue model behavior. However, obtaining DST models that can robustly adapt to new application domains is an ongoing research challenge. The proposed work aims to improve the utility of DST by making the domain adaptation of DST models more effective and cost-efficient. To achieve this, a new task is proposed called Dialogue State Generation (DSG). The goal of DSG is to infer both the schema and values of dialogue state in unseen dialogue domains, and experimental results demonstrate the effectiveness of the presented DSG approach for tackling the challenge of domain generalizability. The DSG approach is then extended for Slot Schema Induction, which is shown to be the first practical method for discovering a consistent set of new slot types from unlabeled data. Finally, the novel DSG and Schema Induction approaches are leveraged to generate a synthetic DST dataset with silver dialogue state labels that covers 1,000 different domains, an order of magnitude more than any existing dataset. An evaluation of few- and zero-shot DST models trained on the domain-diverse synthetic data demonstrates a substantial positive impact on DST domain adaptation. These contributions improve the feasibility of integrating conversational AI in real-world applications, taking steps towards the global improvement of software applications’ efficacy and ease of use.