Associative knowledge feature vector inferred on external knowledge base for dialog state tracking

The dialog state tracker is one of the most important modules on task-oriented dialog systems, its accuracy strongly affects the quality of the system response. The architecture of the tracker has been changed from pipeline processing to an end-to-end approach that directly estimates a user's intention from each current utterance and a dialog history because of the growth in the use of the neural-network-based classifier. However, tracking appropriate slot-value pairs of dialog states that are not explicitly mentioned in user utterances is still a difficult problem. In this research, we propose creating feature vectors by using inference results on an external knowledge base. This inference process predicts associative entities in the knowledge base, which contribute to the dialog state tracker for unseen entities of utterances. We extracted a part of a graph structure from an external knowledge base (Wikidata). Label propagation was used for inferring associative nodes (entities) on the graph structure to produce feature vectors. We used the vectors for the input of a fully connected neural network (FCNN) based tracker. We also introduce a convolutional neural network (CNN) tracker as a state-of-the-art tracker and ensemble models of FCNN and CNN trackers. We used a common test bed, Dialog State Tracking Challenge 4 for experiments. We confirmed the effectiveness of the associative knowledge feature vector, and one ensemble model outperformed other models.