In task-oriented dialogue systems, generating consistent dialogue responses is crucial for ensuring the reliability of applications. However, ensuring that the system provides non-contradictory information throughout the entire dialogue process has been a long-standing challenge. A pioneering study conducted by the research team led by Libo Qin has introduced a modularized pre-training framework (MPFToD), aimed at significantly enhancing the consistency identification capability (Consistency Identification in Task-oriented Dialogue, CI-ToD) in task-oriented dialogues by using a large amount of knowledge-free dialogue data for pre-training. This advancement is not only of profound significance to academia but also holds substantial practical application value for deployed dialogue systems. The research findings have been published in Frontiers of Computer Science on 15 October 2025 co-published by Higher Education Press and Springer Nature.
Consistency identification in task-oriented dialogue (CI-ToD) is an emerging and increasingly important field of research, capable of preventing the generation of inconsistent dialogue responses. The core of CI-ToD lies in determining whether the system's response contradicts the knowledge base (KB), dialogue history, and user queries. By introducing MPFToD, the research team has, for the first time, explored a pre-training paradigm for CI-ToD. MPFToD, through its modular design, is capable of utilizing a large amount of knowledge-free dialogue data for pre-training, thereby solving the problem of data scarcity in traditional methods. Specifically, MPFToD breaks down CI-ToD into three sub-modules and designs pre-training tasks for each, including query response matching pre-training, dialogue history consistency identification pre-training, and knowledge base mask language modeling. This modular pre-training approach not only improves the model's consistency identification capability but also significantly enhances the overall performance of the dialogue system.

In the CI-ToD benchmark tests, MPFToD has demonstrated its outstanding performance, increasing the state-of-the-art technology level from 56.3% to 61.0%, achieving an absolute improvement of 4.7%. Furthermore, the research team has also proven the transferability of MPFToD in other downstream tasks (such as dialogue act recognition, sentiment classification, and table fact checking), showing its broad application prospects. This achievement not only promotes the research of task-oriented dialogue systems but also provides new ideas and methods for the development of dialogue systems in practical applications.
DOI: 10.1007/s11704-024-3778-9