Abstract:Multi-domain dialogue state tracking (MDST) is a crucial component of task-oriented dialogue systems. In the context of multi-turn dialogues between the user and the system, MDST necessitates the continuous keeping track of dialogue states based on the information present in the current dialogue utterance and the dialogue states from the preceding turn. Recent work achieves the successful execution of multi-domain dialogue tasks by adopting an approach that treats each state as an individual label, while regrettably neglecting the potential benefits of incorporating domain-specific information associated with these states. Simultaneous, existing models exhibit a deficiency in effectively modeling the explicit correlations between dialogue contextual semantics and dialogue states. In this paper, we introduce the module of multi-domain gate and interactive dual attention as novel solutions to address the aforementioned concerns. For the efficient exploitation of domain-specific information within states, we leverage the multi-domain gate as indices to amplify the states pertinent to the current utterance domain while filtering out irrelevant states. Interactive dual attention comprises utterance attention and slot attention, effectively modeling the correlation between dialogue utterances and slots. Additionally, interactive dual attention ensures that each dialogue utterance interacts with the slots once to derive all state updates, thereby ensuring computational efficiency. Specifically, slot attention models the associations between slots by incorporating semantic features to forecast updates in slot values. Meanwhile, utterance attention captures the semantics of dialogue context and integrates it with slot name features to generate dialogue states. All the aforementioned modules are designed based on a slot-independent framework, enabling efficient scalability of slots and circumventing issues related to model input limitations. The experimental results on the multi-domain dialogues dataset MultiWOZ 2.4 demonstrate the superior performance of our model compared to the baselines. Additionally, we conduct a comprehensive analysis of the effectiveness of the multi-domain gate and interactive dual attention modules, elucidating their contribution to the performance of the model through visualization and case studies.

Chinese Dialogue Analysis Using Multi-Task Learning Framework

Dialog Act Annotation for Chinese Daily Conversation

Multi-Task Learning based Online Dialogic Instruction Detection with Pre-trained Language Models

Improving Multi-Party Dialogue Discourse Parsing via Domain Integration

Teacher-Student Framework Enhanced Multi-domain Dialogue Generation

Multijugate Dual Learning for Low-Resource Task-Oriented Dialogue System.

Multi-turn dialogue comprehension from a topic-aware perspective

Multi-task learning with graph attention networks for multi-domain task-oriented dialogue systems

Multi-domain gate and interactive dual attention for multi-domain dialogue state tracking

Different Strokes for Different Folks: Investigating Appropriate Further Pre-training Approaches for Diverse Dialogue Tasks

Towards Generalized Models for Task-oriented Dialogue Modeling on Spoken Conversations

DialogAct2Vec: Towards End-to-End Dialogue Agent by Multi-Task Representation Learning

LLM–Assisted Data Augmentation for Chinese Dialogue–Level Dependency Parsing

MDD-Eval: Self-Training on Augmented Data for Multi-Domain Dialogue Evaluation

Unsupervised Mutual Learning of Dialogue Discourse Parsing and Topic Segmentation

Few‐Shot Contrastive Learning‐Based Multi‐Round Dialogue Intent Classification Method

Improved Dynamic Memory Network for Dialogue Act Classification with Adversarial Training

OmniDialog: An Omnipotent Pre-training Model for Task-Oriented Dialogue System

Learning an Effective Context-Response Matching Model with Self-Supervised Tasks for Retrieval-based Dialogues

JDDC 2.1: A Multimodal Chinese Dialogue Dataset with Joint Tasks of Query Rewriting, Response Generation, Discourse Parsing, and Summarization

Sentiment Classification in Customer Service Dialogue with Topic-Aware Multi-Task Learning