Improving Dialogue State Tracking by Discerning the Relevant Context

Proceedings of NAACL-HLT 2019, pages 576–581 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 576 Improving Dialogue State Tracking by Discerning the Relevant Context Sanuj Sharma, Prafulla Kumar Choubey, Ruihong Huang Department of Computer Science and Engineering Texas A&M University (sanuj, prafulla.choubey, huangrh)@tamu.edu Abstract A typical conversation comprises of multi- ple turns between participants where they go back-and-forth between different topics. At each user turn, dialogue state tracking (DST) aims to estimate user’s goal by processing the current utterance. However, in many turns, users implicitly refer to the previous goal, ne- cessitating the use of relevant dialogue history. Nonetheless, distinguishing relevant history is challenging and a popular method of using dia- logue recency for that is inefficient. We, there- fore, propose a novel framework for DST that identifies relevant historical context by refer- ring to the past utterances where a particular slot-value changes and uses that together with weighted system utterance to identify the rel- evant context. Specifically, we use the cur- rent user utterance and the most recent sys- tem utterance to determine the relevance of a system utterance. Empirical analyses show that our method improves joint goal accuracy by 2.75% and 2.36% on WoZ 2.0 and Mul- tiWoZ 2.0 restaurant domain datasets respec- tively over the previous state-of-the-art GLAD model. 1 Introduction Dialog state tracking (DST) is a vital component in the task-oriented dialog systems which is used to estimate user’s goals and requests in order to plan next action and respond accordingly. At each turn, DST aims to identify the set of goals that a user aims to achieve and requests that are repre- sented as slot-value pairs. Typically, this decision is made by considering user utterance in the cur- rent turn or system actions in the previous turn. However, in many cases, the considered user ut- terance or system actions do not present enough information and refers to a previous utterance. As shown through an example in Figure 1, while exploring different available options, user User: hello, i'm looking for a restaurant with fair prices System: There are 31 places with moderate price range. Can you please tell me what kind of food you would like? Sys Act: food Turn Label: price range = moderate User: well I want to eat in the North, what's up that way? System: I have two options that fit that description, Golden Wok chinese restaurant and Nirala which serves Indian food. Do you have a preference? Sys Act: food Turn Label: area = north User: Can I have the address and phone number for the Golden Wok chinese restaurant? System: The phone number is 01223 350688. Turn Label: request = address, phone; food = chinese User: thank you. what is the address? System: The address is 191 Histon Road Chesterton. Turn Label: request = address User: Okay, what about Nirala, what's the address and phone of that? System: 7 Milton Road Chesterton and the number is 01223 360966 Turn Label: request = address, phone; food …

Content truncated to 500 words