Retrieval-Enhanced Adversarial Training for Neural Response Generation

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3763–3773 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 3763 Retrieval-Enhanced Adversarial Training for Neural Response Generation Qingfu Zhu], Lei Cui[, Weinan Zhang]\, Furu Wei[, Ting Liu]\⇤ ]Harbin Institute of Technology, Harbin, China [Microsoft Research Asia, Beijing, China \Peng Cheng Laboratory, Shenzhen, China {qfzhu, wnzhang, tliu}@ir.hit.edu.cn {lecu, fuwei}@microsoft.com Abstract Dialogue systems are usually built on ei- ther generation-based or retrieval-based ap- proaches, yet they do not benefit from the advantages of different models. In this pa- per, we propose a Retrieval-Enhanced Ad- versarial Training (REAT) method for neural response generation. Distinct from existing approaches, the REAT method leverages an encoder-decoder framework in terms of an ad- versarial training paradigm, while taking ad- vantage of N-best response candidates from a retrieval-based system to construct the dis- criminator. An empirical study on a large scale public available benchmark dataset shows that the REAT method significantly outperforms the vanilla Seq2Seq model as well as the con- ventional adversarial training approach. 1 Introduction Dialogue systems intend to converse with humans with a coherent structure. They have been widely used in real-world applications, including cus- tomer service systems, personal assistants, and chatbots. Early dialogue systems are often built using the rule-based method (Weizenbaum, 1966) or template-based method (Litman et al., 2000; Schatzmann et al., 2006; Williams and Young, 2007), which are usually labor-intensive and dif- ficult to scale up. Recently, with the rise of social networking, conversational data have ac- cumulated to a considerable scale. This pro- moted the development of data-driven methods, including retrieval-based methods (Shang et al., 2015; Sordoni et al., 2015; Vinyals and Le, 2015; Wen et al., 2017) and generation-based meth- ods (Leuski et al., 2006; Ji et al., 2014; Yan et al., 2016). Retrieval-based methods reply to users by searching and re-ranking response candidates ⇤Corresponding author. MSG I made strawberry shortcake. GT Where did you learn that, it is sweet and cheery. RSP How did you make it? It looks delicious. C#1 Could you tell me how this thing is cooked? C#2 Tiramisu is my favorite dessert. It’s so delicious. Table 1: An example of a message (MSG), a ground- truth response (GT), a generated response (RSP) and N-best response candidates (C#1 and C#2) during the training process. Similar contents in the response and candidates are in boldface. from a pre-constructed response set. Written mainly by humans, these responses are always diverse and informative, but may be inappropri- ate to input messages due to their being pre- pared in advance and thus incapable of being customized (Shang et al., 2015). In contrast, generation-based methods can produce responses tailored to the messages. The most common method of this category in recent years is the se- quence to sequence (Seq2Seq) model (Sutskever et al., 2014; Shang et al., 2015; Vinyals and Le, 2015). In practice, it usually suffers from the prob- lem of generating generic responses, such as “I don’t know” and “Me, too” …

Content truncated to 500 words