Controlling Dialogue Generation with Semantic Exemplars

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3018–3029 June 6–11, 2021. ©2021 Association for Computational Linguistics 3018 Controlling Dialogue Generation with Semantic Exemplars Prakhar Gupta1, Jeffrey P. Bigham1,2, Yulia Tsvetkov1, Amy Pavel2 1Language Technologies Institute, Carnegie Mellon University, 2Human-Computer Interaction Institute, Carnegie Mellon University, {prakharg,jbigham,ytsvetko,apavel}@cs.cmu.edu Abstract Dialogue systems pretrained with large lan- guage models generate locally coherent re- sponses, but lack the fine-grained control over responses necessary to achieve specific goals. A promising method for controlling generated responses is exemplar-based generation, in which models edit exemplar responses that are retrieved from training data, or hand-written to strategically address discourse-level goals, to fit new dialogue contexts. We present an Exemplar-based Dialogue GEneration model, EDGE, that uses the semantic frames present in exemplar responses to guide response gener- ation. We show that controlling dialogue gen- eration based on the semantic frames of ex- emplars improves the coherence of generated responses, while preserving semantic meaning and conversation goals present in exemplar re- 1 sponses. 1 Introduction Large pre-trained language models (Radford et al., 2019; Devlin et al., 2019) currently used to power dialogue generation systems produce increasingly fluent and appropriate responses for novel dialogue contexts (Wolf et al., 2019; Zhang et al., 2020; Budzianowski and Vulic´, 2019). However, the gen- erated responses are often uninformative or incon- sistent with high-level constraints of a dialogue system and the tasks it supports. Prior work added high-level control for specific intents such as po- liteness (Niu and Bansal, 2018), emotions (Zhong et al., 2019) and persona (Song et al., 2019) through a fixed set of coarse labels, but these methods re- quire manually labelling data for each new intent. One approach for adding control over response intents is to use response exemplars that are hand- written or strategically curated to promote high- level goals without explicit labels. By condition- Context My friends and I have started eating vegan food since yesterday. Exemplar Eggs are very beneficial for your body. Frames FOOD USEFULNESS BODY-PARTS Responses Vegan food can be good for your health. Vegetables can do wonders for your body Vegan food is very healthy. Exemplar I want to drink milk as well. Frames DESIRING INGESTION FOOD Responses You want to eat some vegan food? We eat a lot of vegetables. It’s delicious. We like to eat organic food. Table 1: EDGE generates responses to dialogue con- texts by conditioning the response generation on the se- mantic frames of existing response exemplars to create coherent and controlled replies. ing on response exemplars, we can generate coher- ent responses that follow the intents of the exem- plars without manually labeling vast amounts of data. Current exemplar-based methods (Cai et al., 2019b,a; Wu et al., 2019) have two key drawbacks: (1) the models often overfit to the training data, then produce incoherent responses by copying ir- relevant tokens from exemplar responses into the generated responses, and (2) the models often learn to ignore the exemplars, then …

Content truncated to 500 words