A Transformer-based Threshold-Free Framework for Multi-Intent NLU

Proceedings of the 29th International Conference on Computational Linguistics, pages 745–755 October 12–17, 2022. 745 TopKG: Target-oriented Dialog via Global Planning on Knowledge Graph Zhitong Yang1,3,2, Bo Wang2,1∗, Jinfeng Zhou2,1, Yue Tan2,1, Dongming Zhao4, Kun Huang4, Ruifang He2,1, Yuexian Hou2 1State Key Laboratory of Communication Content Cognition, People’s Daily Online, Beijing, China 2College of Intelligence and Computing, Tianjin University, Tianjin, China 3School of New Media and Communication, Tianjin University, Tianjin, China 4AI Lab, China Mobile Communication Group Tianjin Co., Ltd. {yyyyyyzt, bo_wang, jfzhou, tanyue_098}@tju.edu.cn Abstract Target-oriented dialog aims to reach a global target through multi-turn conversation. The key to the task is the global planning towards the target, which flexibly guides the dialog con- cerning the context. However, existing target- oriented dialog works take a local and greedy strategy for response generation, where global planning is absent. In this work, we propose global planning for target-oriented dialog on a commonsense knowledge graph (KG). We design a global reinforcement learning with the planned paths to flexibly adjust the local re- sponse generation model towards the global tar- get. We also propose a KG-based method to col- lect target-oriented samples automatically from the chit-chat corpus for model training. Ex- periments show that our method can reach the target with a higher success rate, fewer turns, and more coherent responses. 1 Introduction Human-like dialog agents have three types of approaches: open-domain (Zhang et al., 2019a; Huang et al., 2020), task-oriented (Budzianowski et al., 2018; Rastogi et al., 2020; Yang et al., 2020), and target-oriented dialog (Tang et al., 2019; Qin et al., 2020; Zhong et al., 2021). The open-domain dialog only requires the dialog generation to be flu- ent and context coherent. In contrast, typical task- oriented dialog further completes a specific task by understanding users’ intention and collecting the required information of predefined sub-tasks of the intention. However, as a more challenging task, target-oriented dialog aims to achieve a global target that often can not be clearly defined as sub- tasks. The dialog agents are required to lead the conversation to the target flexibly, and the process is excepted to be coherent, effective, and successful. Due to its purpose and flexibility, target-oriented dialog agents have a broad-based demand, e.g., con- versational recommendation (Li et al., 2018; Kang ∗Corresponding author. et al., 2019), psychotherapy (Sharma et al., 2020), and education (Clarizia et al., 2018). In these fields, a typical expectation of target-oriented dialog is to actively lead the conversation by smoothly chang- ing the dialog topic to a designated one, e.g., a prod- uct, a stimulus of mind, and a knowledge point. To reach a target topic effectively and coherently in dialog, existing approaches primarily represent the topic as keywords and adopt a two-stage archi- tecture, i.e., predicting a next-turn keyword and keyword-augmented response retrieval (Tang et al., 2019). In this direction, Xu et al. (2020b) further introduces reinforcement learning with “target sim- ilarity” rewards to target-oriented dialog learning. However, the target-oriented dialog is a typical knowledge-rich task. Although dialog context can support …

Content truncated to 500 words