Model-Based Simulation for Optimising Smart Reply

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 12030–12043 July 9-14, 2023 ©2023 Association for Computational Linguistics Model-Based Simulation for Optimising Smart Reply Benjamin Towle1, Ke Zhou1,2 1University of Nottingham 2Nokia Bell Labs {benjamin.towle, ke.zhou}@nottingham.ac.uk Abstract Smart Reply (SR) systems present a user with a set of replies, of which one can be selected in place of having to type out a response. To perform well at this task, a system should be able to effectively present the user with a di- verse set of options, to maximise the chance that at least one of them conveys the user’s de- sired response. This is a significant challenge, due to the lack of datasets containing sets of responses to learn from. Resultantly, previous work has focused largely on post-hoc diversi- fication, rather than explicitly learning to pre- dict sets of responses. Motivated by this prob- lem, we present a novel method SIMSR, that employs model-based simulation to discover high-value response sets, through simulating possible user responses with a learned world model. Unlike previous approaches, this allows our method to directly optimise the end-goal of SR–maximising the relevance of at least one of the predicted replies. Empirically on two pub- lic datasets, when compared to SoTA baselines, our method achieves up to 21% and 18% im- provement in ROUGE score and Self-ROUGE score respectively. 1 Introduction Automated response suggestion, or Smart Reply (SR), is rapidly becoming a staple feature of many email and chat systems such as Gmail, Skype, Out- look, Microsoft Teams, LinkedIn and Facebook Messenger. Given a message, SR systems present the user with a selection of possible responses, e.g. How are you? → {I’m good; I’m ok; Not great}, which they can click in place of having to type out a reply. With the growth of communication over smaller devices that are poorly suited for manual typing (Varcholik et al., 2012; Palin et al., 2019), such as smartphones and smart watches, SR is becoming an increasingly more im- portant feature. While early methods in SR incorporated sequence-to-sequence models (Kannan et al., 2016), the current mainstream approach favours Matching models which separately encode the mes- sage and reply into a shared latent space and re- trieve the nearest neighbour response (Deb et al., 2019; Zhang et al., 2021; Deb et al., 2021). This has advantages in a production context, as it en- ables the model to retrieve replies from a fixed response set, maintaining greater controllability of model outputs; further, the latent representations for the response set can be pre-computed prior to inference, enabling faster latency. However, the naive approach of simply retriev- ing top-K highest-scoring candidates from the Matching model often fails to produce a sufficiently diverse set of reply options. For instance, in re- sponse to the message How are you?, if the first predicted response is I’m good, predicting I’m doing well as the second response pro- vides limited incremental value, as it carries equiva- lent semantic meaning. …

Content truncated to 500 words