Hierarchical Dialog State Tracking with Unknown Slot Values

A Robust Data-Driven Approach for Dialogue State Tracking of Unseen Slot Values

“…The DB model relies on a semantic lexicon to replace slot names in the input to generic delexicalised token while all other approaches in Table 2 (including our approach) are data-driven. The approaches of neural belief tracker (NBT) , GLAD (Zhong et al, 2018) and StateNet PSI (Ren et al, 2018) are focused on modelling the seen slot-values, while HDSTM (Yang et al, 2018), multi-domain DST (Rastogi et al, 2017) and Pointer Net (Xu and Hu, 2018) addresses the problem of unseen slot values in some aspect. We can see from the results that our approach significantly improves on the NBT and Pointer Net on both DSTC2 and WoZ2.0 datasets proving its ability in performing state-of-the-art in predicting the seen slot-values.…”

“…The most relevant works addressing the problem of unseen slot-values are (Xu and Hu, 2018;Yang et al, 2018;Jang et al, 2016;Yoshida et al, 2018). (Xu and Hu, 2018) proposed a joint model for DST based on pointer networks which outputs a distribution over the tokens in user utterance.…”

“…This approach assumes no candidate set for the values and since it can only output from the input vocabulary, a semantic lexicon is required for post-processing of output to normalize for the value. (Yang et al, 2018) proposed an hierarchical dialog state tracker (HDSTM) that consists of a detection module to identify if an unseen slot value occurs in the user utterance. Based on the detection module output, the belief state is updated using an update mechanism.…”

“…Joint Goal DSTC2 WoZ2.0 Delexicalisation-Based (DB) Model 69.1 70.8 HDSTM (Yang et al, 2018) 68.4 84.5 Scalable multi-domain DST (Rastogi et al, 2017) 70.3 -Pointer Net (Xu and Hu, 2018) 72.1 -Neural Belief Tracker (NBT) -DNN 72.6 84.4 Neural Belief Tracker (NBT) -CNN 73.4 84.2 GLAD (Zhong et al, 2018) 74.5 88.1 StateNet PSI (Ren et al, 2018) 75.5 88.9 Our approach 73.8 87.5…”

“…Any update for the slot values requires new training data for that value and model retraining. While approaches such as (Yang et al, 2018;Yoshida et al, 2018;Xu and Hu, 2018) has tried to address this using data-driven approaches, they often make compromise in the performance of seen slot values to accommodate for the unseen slot values. Their evaluation shows that a large performance gap still persists in predicting for seen and unseen slot values.…”