Neural Response Generation via GAN with an Approximate Embedding Layer

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 617–626 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Neural Response Generation via GAN with an Approximate Embedding Layer∗ Zhen Xu1, Bingquan Liu1, Baoxun Wang2, Chengjie Sun1, Xiaolong Wang1, Zhuoran Wang2 and Chao Qi2 1School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China 2Tricorn (Beijing) Technology Co., Ltd, Beijing, China 1{zxu,bqliu,cjsun,wangxl}@insun.hit.edu.cn 2{wangbaoxun, wangzhuoran, qichao}@trio.ai Abstract This paper presents a Generative Adver- sarial Network (GAN) to model single- turn short-text conversations, which trains a sequence-to-sequence (Seq2Seq) net- work for response generation simulta- neously with a discriminative classifier that measures the differences between human-produced responses and machine- generated ones. In addition, the proposed method introduces an approximate embed- ding layer to solve the non-differentiable problem caused by the sampling-based output decoding procedure in the Seq2Seq generative model. The GAN setup pro- vides an effective way to avoid non- informative responses (a.k.a “safe re- sponses”), which are frequently observed in traditional neural response generators. The experimental results show that the proposed approach significantly outper- forms existing neural response generation models in diversity metrics, with slight increases in relevance scores as well, when evaluated on both a Mandarin corpus and an English corpus. 1 Introduction After achieving remarkable successes in Machine Translation (Sutskever et al., 2014; Cho et al., 2014), neural networks with the encoder-decoder architectures (a.k.a sequence-to-sequence models, Seq2Seq) have been proven to be a functioning method to model short-text conversations (Vinyals and Le, 2015; Shang et al., 2015), where the corresponding task is often called Neural Re- sponse Generation. The advantage of applying ∗The work was done when the first author was an intern at Tricorn (Beijing) Technology Co., Ltd. Seq2Seq models to conversation generation is that the training procedure can be performed end-to-end in an unsupervised manner, based on human-generated conversational utterances (typ- ically query-response pairs mined from social networks). One of the potential applications of such neural response generators is to improve the capability of existing conversational interfaces (informally also known as chatbots) by enabling them to go beyond predefined tasks and chat with human users in an open domain. However, previous research has indicated that na¨ıve implementations of Seq2Seq based conver- sation models tend to suffer from the so-called “safe response” problem (Li et al., 2016a), i.e. such models tend to generate non-informative responses that can be associated to most queries, e.g. “I don’t know”, “I think so”, etc. This is due to the fundamental nature of statistical models, which fit sufficiently observed examples better than insufficiently observed ones. Concretely, the space of open-domain conversations is so large that in any sub-sample of it (i.e. a training set), the distribution of most pieces of information are relatively much sparser when compared to safe response patterns. Furthermore, since a safe response can be of relevance to a large amount of diverse queries, a statistical learner will tend to minimize its empirical risk in the response generation process by capturing those safe responses …

Content truncated to 500 words