MEMD: A Diversity-Promoting Learning Framework for Short-Text Conversation

Proceedings of the 27th International Conference on Computational Linguistics, pages 1281–1291 Santa Fe, New Mexico, USA, August 20-26, 2018. 1281 MEMD: A Diversity-Promoting Learning Framework for Short-Text Conversation Meng Zou Xihan Li Haokun Liu Zhihong Deng∗ Key Laboratory of Machine Perception (Ministry of Education) School of Electronics Engineering and Computer Science {mengzou,Peking University, Beijing 100871, China xihanli, 1300012769, zhdeng}@pku.edu.cn Abstract Neural encoder-decoder models have been widely applied to conversational response generation, which is a research hot spot in recent years. However, conventional neural encoder-decoder mod- els tend to generate commonplace responses like "I don't know" regardless of what the input is. In this paper, we analyze this problem from a new perspective: latent vectors. Based on it, we pro- pose an easy-to-extend learning framework named MEMD (Multi-Encoder to Multi-Decoder), in which an auxiliary encoder and an auxiliary decoder are introduced to provide necessary training guidance without resorting to extra data or complicating network's inner structure. Experimental results demonstrate that our method effectively improve the quality of generated responses ac- cording to automatic metrics and human evaluations, yielding more diverse and smooth replies. 1 Introduction Human-computer conversation is attracting particular attention recently. Research in this field falls into two categories: the retrieval-based method (Ji et al., 2014; Yan et al., 2017; Wu et al., 2017) and the generative method (Shang et al., 2015; Serban et al., 2016). While the retrieval-based method can guarantee completeness of output sentences, it fails to customized for particular posts from users. By contrast, the generative method may produce sentences with grammatical errors, however, it shows great promise in flexibility, which gives rise to a research hot spot. Figure 1: Conventional encoder-decoder model. Blue part represents the encoder, while red part repre- sents the decoder. Of the generative method, neural encoder-decoder based model (Sutskever et al., 2014) has become the mainstream (Figure 1). (Shang et al., 2015) firstly formalized the generation of response as a decoding process based on the latent representation of the input text, and both encoding and decoding are realized with recurrent neural networks (RNN). Such formalization is widely adopted by later work. However, the conventional encoder-decoder model's performance is far from satisfactory, for it tends to generate meaningless and generic responses like "I don't know". To tackle the issue of response diversity, lots of models have been proposed, and they can be broadly divided into three categories: (1) introducing external priori knowledge into the procedure of encod- ing/decoding (Mou et al., 2016; Xing et al., 2017), which usually requires pretreatment on extra large *Zhihong Deng is the corresponding author. This work is licenced under a Creative Commons Attribution 4.0 International Licence. Licence details: http:// creativecommons.org/licenses/by/4.0/ 1282 data. (2) complicating network's structure to enhance model's capacity for encoding/decoding (Zhou et al., 2017; Serban et al., 2017a; Serban et al., 2017b; Zhao et al., 2017; Clark and Cao, 2017; Shen et al., 2017). (3) directly modifying the objective function to penalize generation probability of trivial re- sponses (Li et al., 2016). However, these models merely …

Content truncated to 500 words