Jointly Optimizing Diversity and Relevance in Neural Response Generation

Proceedings of NAACL-HLT 2019, pages 1229–1238 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1229 Jointly Optimizing Diversity and Relevance in Neural Response Generation Xiang Gao Sungjin Lee Yizhe Zhang Chris Brockett Michel Galley Jianfeng Gao Bill Dolan Microsoft Research, Redmond, WA, USA {xiag,sule,yizzhang,chrisbkt,mgalley,jfgao,billdol}@microsoft.com Abstract Although recent neural conversation models have shown great potential, they often generate bland and generic responses. While various approaches have been explored to diversify the output of the conversation model, the improve- ment often comes at the cost of decreased rel- evance (Zhang et al., 2018). In this paper, we propose a SPACEFUSION model to jointly optimize diversity and relevance that essen- tially fuses the latent space of a sequence- to-sequence model and that of an autoen- coder model by leveraging novel regulariza- tion terms. As a result, our approach induces a latent space in which the distance and direction from the predicted response vector roughly match the relevance and diversity, respectively. This property also lends itself well to an intu- itive visualization of the latent space. Both au- tomatic and human evaluation results demon- strate that the proposed approach brings signif- icant improvement compared to strong base- lines in both diversity and relevance. 1 1 Introduction The field of neural response generation is advanc- ing rapidly both in terms of research and commer- cial applications (Gao et al., 2019; Zhou et al., 2018; Yoshino et al., 2019; Zhang et al., 2019). Nevertheless, vanilla sequence-to-sequence (S2S) models often generate bland and generic responses (Li et al., 2016a). Li et al. (2016a) encourage di- versity by re-ranking the beam search results ac- cording to their mutual information with the con- versation context. However, as beam search itself often produces lists of nearly identical sequences, this method can require a large beam width (e.g. 200). As a result, re-ranking can be extremely 1An implementation of our model is available at https: //github.com/golsun/SpaceFusion 2For simplicity, we omitted the response at the center: ”I would love to play this game”. See Table 2 for more details. No I don’t I’m not interested in the game When will you? When? I’d love to play it Yes I do. [Context] Anyone want to start this game? Figure 1: Illustration of one context and its multiple responses in the latent space induced by our model. Distance and direction from the predicted response vector given the context roughly match the relevance and diversity, respectively. Based on the example in Table 2.2 time-consuming, raising difficulties for real-time applications. This highlights the need to improve the diversity of candidates before re-ranking, and the need to optimize for diversity during training rather than just at the decoding stage. While various approaches have been explored to diversify the output of conversation models, the improvement often comes at the cost of decreased response relevance along other dimensions. For instance, Zhao et al. (2017) present an approach to enhancing diversity by mapping diverse re- sponses to a probability distribution using a con- ditional variational …

Content truncated to 500 words