Retrieval-guided Dialogue Response Generation via a Matching-to-Generation Framework

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1866–1875, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 1866 Retrieval-guided Dialogue Response Generation via a Matching-to-Generation Framework Deng Cai1∗ Yan Wang2∗ Wei Bi2 Zhaopeng Tu2 Xiaojiang Liu2 Shuming Shi2 1The Chinese University of Hong Kong, 2Tencent AI Lab thisisjcykcd@gmail.com {brandenwang,victoriabi,zptu,kieranliu,shumingshi}@tencent.com Abstract End-to-end sequence generation is a popular technique for developing open domain dia- logue systems, though they suffer from the safe response problem. Researchers have at- tempted to tackle this problem by incorpo- rating generative models with the returns of retrieval systems. Recently, a skeleton-then- response framework has been shown promis- ing results for this task. Nevertheless, how to precisely extract a skeleton and how to effec- tively train a retrieval-guided response gener- ator are still challenging. This paper presents a novel framework in which the skeleton ex- traction is made by an interpretable matching model and the following skeleton-guided re- sponse generation is accomplished by a sep- arately trained generator. Extensive exper- iments demonstrate the effectiveness of our model designs. 1 Introduction Sequence-to-sequence (seq2seq) neural models (Shang et al., 2015; Vinyals and Le, 2015; Sor- doni et al., 2015; Serban et al., 2016; Li et al., 2016a) have been popular for single-turn dialogue response generation. However, many of the gen- erated responses (e.g., “I don’t know” and “I think so”) appear to be generic and dull (safe response problem) (Li et al., 2016a). This problem is avoided in traditional retrieval systems (Ji et al., 2014; Hu et al., 2014) by preceding the selection of informative and engaging responses. It is of interest to benefit from both the gener- alization capacity of the seq2seq models and the information richness of the retrieved responses. Following the standard encoder-decoder frame- work, early attempts have either used an extra en- coder for the retrieved response (Song et al., 2016; ∗ This work was mainly done when Deng Cai was an intern at Tencent AI Lab. Yan Wang is the corresponding author. Query: How is your day today? Response: Great, I get promotion today. Retrieval guided Vanilla seq2seq Bad, I hate the weather. collapse mismatch retrieve generate generate Figure 1: The common problem for training a retrieval- guided generation model in previous work. The model is forced to neglect the retrieved response even though it is a proper response, due to the mismatch between the retrieved response and the target response. Pandey et al., 2018; Wu et al., 2019) or a uni- fied encoder for the concatenation of the query and the retrieved response (Weston et al., 2018). To prevent the inflow of erroneous information, Cai et al. (2019) proposed a general framework that first extracts a skeleton from the retrieved re- sponse and then generates the response based on the extracted skeleton. Despite their differences, a common issue is that the generation model easily learns to ignore the retrieved response entirely and collapses to a vanilla …

Content truncated to 500 words