Dialogue State Tracking with Incremental Reasoning

Dialogue State Tracking with Incremental Reasoning Lizi Liao, Le Hong Long, Yunshan Ma, Wenqiang Lei, Tat-Seng Chua School of Computing National University of Singapore {liaolizi.llz, yunshan.ma, wenqianglei}@gmail.com lehonglong@u.nus.edu chuats@comp.nus.edu.sg Abstract Tracking dialogue states to better interpret user goals and feed downstream policy learning is a bottleneck in dialogue management. Common practice has been to treat it as a problem of classifying dialogue content into a set of pre-defined slot-value pairs, or generating values for different slots given the dialogue history. Both have limitations on considering dependencies that occur on dialogues, and are lacking of reasoning capabilities. This paper proposes to track dialogue states gradually with reasoning over dialogue turns with the help of the back-end data. Empirical results demonstrate that our method outperforms the state-of-the- art methods in terms of joint belief accuracy for MultiWOZ 2.1, a large-scale human–human dialogue dataset across multiple domains. 1 Introduction Dialogue State Tracking (DST) usually works as a core component to monitor the user’s intentional states (or belief states) and is cru- cial for appropriate dialogue management. A state in DST typically consists of a set of dialogue acts and slot value pairs. Consider the task of restaurant reservation as shown in Figure 1. In each turn, the user may inform the agent of particular goals (e.g. single one as inform(food=Indian) or composed one as inform(area=center,food=Jamaican)). Such goals given during a turn are referred as turn belief. The joint belief is the set of accu- mulated turn goals updated until the current turn, which summarizes the information needed to successfully maintain and finish the dialogue. Traditionally, dialogue system is supported by a domain ontology, which defines a collection of slots and the values that each slot can take. The aim of DST is to identify good features or patterns, and map to entries such as specific slot- value pairs in the ontology. It is often treated as a classification problem. Therefore, most efforts center on (1) finding salient features: from hand- crafted features (Wang and Lemon, 2013; Sun et al., 2014a), semantic dictionaries (Henderson et al., 2014b; Rastogi et al., 2017), to neural network extracted features (Mrkˇsi´c et al., 2017); or (2) investigating effective mappings: from rule-based models (Sun et al., 2014b), generative models (Thomson and Young, 2010; Williams and Young, 2007), to discriminative ones (Lee and Eskenazi, 2013; Ren et al., 2018; Xie et al., 2018). On the other hand, some researchers attack these methods’ over-dependence on domain ontology. They perform DST in the absence of a comprehensive domain ontology and handle unknown slot values by generating words from dialogue history or knowledge source (Rastogi et al., 2017; Xu and Hu, 2018; Wu et al., 2019). However, the critical problem of modeling the dependencies and reasoning over dialogue history is not well researched. Many existing methods work on turn level only, which takes in the cur- rent turn utterance and outputs the corresponding turn belief (Henderson et al., 2014b; Zilka and Jurcicek, 2015; Rastogi et al., 2017; Xu and Hu, 2018). Compared to …

Content truncated to 500 words