An Interpretable Neuro-Symbolic Reasoning Framework for Task-Oriented Dialogue Generation

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4918 - 4935 May 22-27, 2022 c⃝2022 Association for Computational Linguistics An Interpretable Neuro-Symbolic Reasoning Framework for Task-Oriented Dialogue Generation Shiquan Yang1, Rui Zhang2, Sarah Erfani1, and Jey Han Lau1 1The University of Melbourne, 2www.ruizhang.info 1{shiquan@student., sarah.erfani@, laujh@}unimelb.edu.au, 2rayteam@yeah.net Abstract We study the interpretability issue of task- oriented dialogue systems in this paper. Pre- viously, most neural-based task-oriented dia- logue systems employ an implicit reasoning strategy that makes the model predictions un- interpretable to humans. To obtain a trans- parent reasoning process, we introduce neuro- symbolic to perform explicit reasoning that justifies model decisions by reasoning chains. Since deriving reasoning chains requires multi- hop reasoning for task-oriented dialogues, ex- isting neuro-symbolic approaches would in- duce error propagation due to the one-phase design. To overcome this, we propose a two- phase approach that consists of a hypothesis generator and a reasoner. We first obtain mul- tiple hypotheses, i.e., potential operations to perform the desired task, through the hypothe- sis generator. Each hypothesis is then verified by the reasoner, and the valid one is selected to conduct the final prediction. The whole sys- tem is trained by exploiting raw textual dia- logues without using any reasoning chain an- notations. Experimental studies on two public benchmark datasets demonstrate that the pro- posed approach not only achieves better results, but also introduces an interpretable decision process. Code and data: https://github. com/shiquanyang/NS-Dial. 1 Introduction Neural task-oriented dialogue systems have en- joyed a rapid progress recently (Peng et al., 2020; Hosseini-Asl et al., 2020; Wu et al., 2020), achiev- ing strong empirical results on various benchmark datasets such as SMD (Eric et al., 2017) and Multi- WOZ (Budzianowski et al., 2018). However, most existing approaches suffer from the lack of explain- ability due to the black-box nature of neural net- works (Doshi-Velez and Kim, 2017; Lipton, 2018; Bommasani et al., 2021), which may hurt the trust- worthiness between the users and the system. For External Knowledge Base (KB) Cityroom Price_range Moderate Chadstone Located_in Leichhardt Cityroom Next_to Palm_Lawn Gonville_Hotel Price_range Expensive Palm_Lawn Located_in Chadstone Gonville_Hotel Located_in Moorabbin User: Can you recommend me a hotel located in Leichhardt ? System: Cityroom is a nice one there. [Chadstone, Located_in, Leichhardt] [Cityroom, Located_in, Leichhardt] Generated hypotheses: ...... [Cityroom, Located_in, Leichhardt] Verification: [Cityroom, Next_to, Palm_Lawn], [Palm_Lawn, Located_in, Chadstone], ...... Figure 1: An example dialogue that incorporates exter- nal KB. The context entity (i.e., Leichhardt) and answer entity (i.e., Cityroom) are marked as Red and Yellow, respectively. The triple containing the context entity and answer entity is not directly stored in KB and should be derived by a reasoning chain formed by multiple KB triplets. instance, in Figure 1, a user is asking for a ho- tel recommendation at a given location. The sys- tem performs reasoning on a knowledge base (KB) and incorporates the correct entity in the response. However, when the system fails to provide the cor- rect entities, it would be difficult …

Content truncated to 500 words