Building End-to-End Task-oriented Dialogue Systems Via CNNs and Attention Mechanisms

Abstract

We put forward a general solution towards building task-oriented dialogue systems via convolutional neural networks (CNNs) and Attention Mechanisms. The dominant task-oriented dialogue models follow pipeline design and usually rely on complex recurrent neural networks(RNNs). In that case, its structure is complicated and the amount of parameters exceeds on million. In this paper, we propose a novel end-to-end task-oriented dialogue system with a clear structure, just based on CNNs and attention mechanisms, totally dispensing with recurrent neural networks (RNNs). Our model is based entirely on sequence-to-sequence (seq2seq) architecture and realizes the true meaning of end-to-end. Experimental results indicate that our model outperforms the state-of-the-art approaches on Cam-Rest676 and KVRET datasets in two evaluation indexes, namely, task completion and quality of language generation. Meanwhile, the training speed of our model is increased by 3–10 times and the amount of parameters is just 1/3.