DialogConv: A Lightweight Fully Convolutional Network for Multi-view Response Selection

Anonymous

Abstract: Current end-to-end retrieval-based dialogue systems are primarily based on Recurrent Neural Networks or Transformers with attention mechanisms. Despite promising results have been achieved, these models usually suffer from slow inference speed or an enormous amount of parameters. In this paper, we propose a novel lightweight fully convolutional architecture called DialogConv for the response selection. DialogConv is built exclusively on convolutions for distilling the matching features of context and response. The dialogue is modeled in a 3D view, where DialogConv conducts convolution operations on embedding dimension, word dimension and utterance dimension iteratively to capture richer semantic information from a multi-view of context. On four benchmark datasets, DialogConv is approximately 4.0x smaller and up to 27x faster in inference compared with strong baselines. Moreover, DialogConv can achieve competitive performance results on four public datasets.