Noisy Label Regularisation for Textual Regression

Proceedings of the 29th International Conference on Computational Linguistics, pages 461–470 October 12–17, 2022. 461 CGIM: A Cycle Guided Interactive Learning Model for Consistency Identification in Task-oriented Dialogue Libo Qin1, Qiguang Chen1, Tianbao Xie1, Qian Liu2, Shijue Huang1, Wanxiang Che1∗, Zhou Yu3 1Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, China 2 Beihang University, Beijing, China 3Columbia University {lbqin,tianbaoxie,car}@ir.hit.edu.cn; qian.liu@buaa.edu.cn; zy2461@columbia.edu Abstract Consistency identification in task-oriented di- alog (CI-ToD) usually consists of three sub- tasks, aiming to identify inconsistency be- tween current system response and current user response, dialog history and the corre- sponding knowledge base. This work aims to solve CI-ToD task by introducing an ex- plicit interaction paradigm, Cycle Guided Interactive learning Model (CGIM), which achieves to make information exchange ex- plicitly from all the three tasks. Specifically, CGIM relies on two core insights, referred to as guided multi-head attention module and cycle interactive mechanism, that collaborate from each other. On the one hand, each two tasks are linked with the guided multi-head attention module, aiming to explicitly model the interaction across two related tasks. On the other hand, we further introduce cycle in- teractive mechanism that focuses on facilitat- ing model to exchange information among the three correlated sub-tasks via a cycle interac- tion manner. Experimental results on CI-ToD benchmark show that our model achieves the state-of-the-art performance, pushing the over- all score to 56.3% (5.0% point absolute im- provement). In addition, we find that CGIM is robust to the initial task flow order. 1 Introduction Consistency identification task in dialogue has the potential benefits of preventing inconsistent re- sponse generation (Welleck et al., 2019), which has attracted increasing attention. Recent years have witnessed two promising research directions in consistency identification. The first focuses on consistency identification in open-domain di- alogue (Zhang et al., 2018; Zheng et al., 2019). The second direction consider consistency iden- tification in task-oriented dialogue system (CI- ToD) (Qin et al., 2021b). In this work, we focus on the latter. Recently, Qin et al. (2021b) intro- ∗Email corresponding. (a) Traditional Multi-task Learning Model. (b) Cycle Guided Interative Learning Model (CGIM). KBCIT QCIT HCIT Encoder ✓ ✓ ✓ Encoder QCIT KBCIT HCIT ✘ ✘ Figure 1: (a) Traditional multi-task learning models learn mutual information across tasks via an implicit interaction manner vs. (b) Our proposed cycle inter- active learning model explicitly consider cross-impact across three tasks via an explicit interaction manner. duces a benchmark (CI-ToD) for consistency iden- tification in task-oriented dialogue to facilitate the relevant research. CI-ToD introduces three sub- tasks including: (1) dialogue history consistency identification task (HCIT) to judge whether gen- erated response is inconsistent with dialogue his- tory; (2) user query consistency identification task (QCIT) to detect the consistency status between query and system response, and (3) knowledge base consistency identification task (KBCIT) to determine system response is contradicted with the corresponding knowledge base. Intuitively, the three tasks are closely related, indicating information of one task can be utilized in other related tasks. For example, if …

Content truncated to 500 words