MRF-Chat: Improving Dialogue with Markov Random Fields

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4925–4936 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 4925 MRF-Chat: Improving Dialogue with Markov Random Fields Ishaan Grover∗, Matthew Huggins∗, Cynthia Breazeal and Hae Won Park Massachusetts Institute of Technology {igrover, hugginsm, breazeal, haewon}@mit.edu Abstract Recent state-of-the-art approaches in open- domain dialogue include training end-to-end deep-learning models to learn various conver- sational features like emotional content of re- sponse, symbolic transitions of dialogue con- texts in a knowledge graph and persona of the agent and the user, among others. While neural models have shown reasonable results, modelling the cognitive processes that humans use when conversing with each other may im- prove the agent’s quality of responses. A key element of natural conversation is to tai- lor one’s response such that it accounts for concepts that the speaker and listener may or may not know and the contextual rele- vance of all prior concepts used in conversa- tion. We show that a rich representation and explicit modeling of these psychological pro- cesses can improve predictions made by ex- isting neural network models. In this work, we propose a novel probabilistic approach us- ing Markov Random Fields (MRF) to aug- ment existing deep-learning methods for im- proved next utterance prediction. Using hu- man and automatic evaluations, we show that our augmentation approach significantly im- proves the performance of existing state-of- the-art retrieval models for open-domain con- versational agents. 1 Introduction With advances in deep learning, the natural lan- guage understanding community has seen a re- cent proliferation of open-domain dialogue systems such as See et al. (2019); Kulikov et al. (2019); Roller et al. (2021) as well as competitions such as the Amazon Alexa Prize (Khatri et al., 2018) and ConvAI (Burtsev et al., 2018; Dinan et al., 2019). Existing approaches can be classified into two main categories: generative models and retrieval models. While the former produce responses from a genera- tive language model (Serban et al., 2016), retrieval ∗Equal Contribution models aim to select the best response from a set of candidate responses given a conversation history. This paper focuses on retrieval-based models. Most of the prior work in retrieval-based mod- els has focused on training end-to-end models us- ing different architectures (eg. Key-Value Memory Networks (Miller et al., 2016), gated self-attention (Zhang et al., 2018b), poly-encoder (Humeau et al., 2019) on specific datasets to statistically learn vari- ous conversational features like emotional content of response (Rashkin et al., 2019), symbolic tran- sitions of dialogue contexts in a knowledge graph (Moon et al., 2019) and even the persona of the agent and the user (Zhang et al., 2018a)). In natural conversations, while humans often view each other as cognitive agents, we observe that prior work has not focused on the cognitive processes that humans use when conversing with each other. We posit that explicitly modeling these cognitive processes and using these models alongside existing statistical approaches can improve state-of-the-art. In conversational inference, the cognitive theory of mutual …

Content truncated to 500 words