Data Augmentation for Copy-Mechanism in Dialogue State Tracking

Abstract

Traditional dialogue state tracking (DST) approaches need a predefined ontology to provide candidate values for each slot. To handle unseen slot values, the copy-mechanism has been widely used in DST models recently, which copies slot values from user utterance directly. Even though the state-of-the-art approaches have shown a promising performance on several benchmarks, there is still a significant gap between seen slot values (values that occur in both training set and test set) and unseen ones (values that only occur in the test set). In this paper, we aim to find out the factors that influence the generalization capability of the copy-mechanism based DST model. Our key observations include two points: 1) performance on unseen values is positively related to the diversity of slot values in the training set; 2) randomly generated strings can enhance the diversity of slot values as well as real values. Based on these observations, an interactive data augmentation algorithm is proposed to train copy-mechanism models, which augments the input dataset by duplicating user utterances and replacing the real slot values with randomly generated strings. Experimental results on three widely used datasets: WoZ 2.0, DSTC2 and Multi-WoZ demonstrate the effectiveness of our approach.