Generating High-Quality and Informative Conversation Responses with Sequence-to-Sequence Models

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2210–2219 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Generating High-Quality and Informative Conversation Responses with Sequence-to-Sequence Models Louis Shao1∗, Stephan Gouws3∗, Denny Britz2†, Anna Goldie2, Brian Strope1, Ray Kurzweil1 {overmind,sgouws,dennybritz,agoldie,bps,raykurzweil}@google.com 1Google Research and 2Google Brain Mountain View, CA, USA and 3Google Brain London, UK Abstract Sequence-to-sequence models have been applied to the conversation response gen- eration problem where the source se- quence is the conversation history and the target sequence is the response. Un- like translation, conversation responding is inherently creative. The generation of long, informative, coherent, and diverse responses remains a hard task. In this work, we focus on the single turn set- ting. We add self-attention to the de- coder to maintain coherence in longer responses, and we propose a practical approach, called the glimpse-model, for scaling to large datasets. We introduce a stochastic beam-search algorithm with segment-by-segment reranking which lets us inject diversity earlier in the generation process. We trained on a combined data set of over 2.3B conversation messages mined from the web. In human evalua- tion studies, our method produces longer responses overall, with a higher proportion rated as acceptable and excellent as length increases, compared to baseline sequence- to-sequence models with explicit length- promotion. A back-off strategy produces better responses overall, in the full spec- trum of lengths. 1 Introduction Building computer systems capable of general- purpose conversation is a challenging problem. However, it is a necessary step toward building in- telligent agents that can interact with humans via ∗Both authors contributed equally to this work. †Work done as a member of the Google Brain Residency program (g.co/brainresidency). natural language, and for eventually passing the Turing test. The sequence-to-sequence (seq2seq) model has proven very popular as a purely data- driven approach in domains that can be cast as learning to map to and from variable-length se- quences, with state-of-the art results in many do- mains, including machine translation (Cho et al., 2014; Sutskever et al., 2014; Wu et al., 2016). Neural conversation models are the latest devel- opment in the domain of conversation modeling, with the promise of training computers to converse in an end-to-end fashion (Vinyals and Le, 2015; Shang et al., 2015; Sordoni et al., 2015; Wen et al., 2016). Despite promising results, there are still many challenges with this approach. In particu- lar, these models produce short, generic responses that lack diversity (Sordoni et al., 2015; Li et al., 2015). Even when longer responses are explicitly encouraged (e.g. via length normalization), they tend to be incoherent (“The sun is in the center of the sun.”), redundant (“i like cake and cake”), or contradictory (“I don’t own a gun, but I do own a gun.”). In this paper, we provide two methods to ad- dress these issues with minimal modifications to the standard seq2seq model. First, we present a glimpse model that only trains on fixed-length segments of the target-side at a time, allowing us to …

Content truncated to 500 words