DLCEncDec: A Fully Character-level Encoder-Decoder Model for Neural Responding Conversation

摘要

Recent years have witnessed a surge of interest in building conversation systems such as smart agents or chatbots. The most existing generation-based neural responding conversation systems are implemented by RNN Encoder-Decoder framework relying on word-level modeling with explicit segmentation. A word-level model typically maintains a fixed vocabulary, which correspondingly encounters the unknown words and segmentation issues. In this paper, we proposed a fully character-level Encoder-Decoder model DLCEncDec without explicit segmentation for neural responding conversation. DLCEncDec utilizes both of fine-grained character embedding features and coarse-grained n-gram features. Coarse-grained n-gram features are captured by constructing a convolutional layer and a four-layer highway network on the top of the character embeddings. The appearance of out-of-vocabulary words (i.e. unknown words) can be addressed due to the fully character-level operating. We evaluate the DLCEncDec on a Chinese corpus consisting of 4.44 million message-response pairs from Sina Weibo. Experimental results show that our fully character-level model DLCEncDec significantly outperforms baseline models in terms of BLEU and ROUGE.