The fifth Dialog State Tracking Challenge (DSTC5) introduces a newcross-language dialog state tracking scenario, where the participants are askedto build their trackers based on the English training corpus, while evaluatingthem with the unlabeled Chinese corpus. Although the computer-generatedtranslations for both English and Chinese corpus are provided in the dataset,these translations contain errors and careless use of them can easily hurt theperformance of the built trackers. To address this problem, we propose amultichannel Convolutional Neural Networks (CNN) architecture, in which wetreat English and Chinese language as different input channels of one singleCNN model. In the evaluation of DSTC5, we found that such multichannelarchitecture can effectively improve the robustness against translation errors.Additionally, our method for DSTC5 is purely machine learning based andrequires no prior knowledge about the target language. We consider this adesirable property for building a tracker in the cross-language context, as notevery developer will be familiar with both languages.
展开▼
