Steering Output Style and Topic in Neural Response Generation

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2140–2150 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Steering Output Style and Topic in Neural Response Generation Di Wang1,2, Nebojsa Jojic1, Chris Brockett1, and Eric Nyberg2 1Microsoft Research, Redmond, WA, USA 2School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, USA diwang@cs.cmu.edu, jojic@microsoft.com Abstract We propose simple and flexible training and decoding methods for influencing out- put style and topic in neural encoder- decoder based language generation. This capability is desirable in a variety of appli- cations, including conversational systems, where successful agents need to produce language in a specific style and generate responses steered by a human puppeteer or external knowledge. We decompose the neural generation process into empirically easier sub-problems: a faithfulness model and a decoding method based on selective- sampling. We also describe training and sampling algorithms that bias the genera- tion process with a specific language style restriction, or a topic restriction. Human evaluation results show that our proposed methods are able to to restrict style and topic without degrading output quality in conversational tasks. 1 Introduction Neural encoder-decoder models have demon- strated great promise in many sequence gener- ation tasks, including neural machine transla- tion (Sutskever et al., 2014; Cho et al., 2014; Bahdanau et al., 2014; Luong et al., 2015; Wu et al., 2016), image captioning (Xu et al., 2015), summarization (Rush et al., 2015; Gu et al., 2016; Kikuchi et al., 2016), and conversation gen- eration (Vinyals and Le, 2015; Sordoni et al., 2015; Shang et al., 2015; Shao et al., 2017; Li et al., 2015). These encouraging early successes have motivated research interest in training more natural-sounding conversational systems based on large volumes of open-domain human-to-human interactions. In order to create more human-like chat context Where are you? Twitter bot i’m on my way to london now. Star Wars bot we’re heading to the dark side of jabba’s palace. Hillary bot i’m here in philadelphia, pennsylvania Trump bot i went to the white house. we were in brooklyn. Kennedy bot i am on the question of the united states. chat context What are you doing? Twitter bot i’m just chilling with my dad, i’m just sitting in the kitchen and listening to music Star Wars bot i’m doing my magic solo. Hillary bot i’m running for president, i’m going to be talking about some of these things Trump bot i’m building a wall, right now Kennedy bot today, i am asking the congress for a new program to make a new effort to increase the tax privileges and to stimulate Table 1: Example responses from our Star Wars, Hillary, Trump, and Kennedy bots with scented conversation models. patterns of conversation, the agents need to have recognizable (and tunable) style, just as individ- ual humans do, and also need to accept guidance from separate information processing modules in order to increase quality of responses. In an ex- treme case, an agent may be micro-managed by a human …

Content truncated to 500 words