Twitter Topic Classification

Proceedings of the 29th International Conference on Computational Linguistics, pages 367–377 October 12–17, 2022. 367 A Personalized Dialogue Generator with Implicit User Persona Detection Itsugun Cho1 Dongyang Wang1∗ Ryota Takahashi1∗ Hiroaki Saito1 Keio University, Japan1 {choitsugun, wangdongyang, ryota.0226.tokky}@keio.jp hxs@ics.keio.ac.jp Abstract Current works in the generation of personalized dialogue primarily contribute to the agent pre- senting a consistent personality and driving a more informative response. However, we found that the generated responses from most previ- ous models tend to be self-centered, with little care for the user in the dialogue. Moreover, we consider that human-like conversation is essen- tially built based on inferring information about the persona of the other party. Motivated by this, we propose a novel personalized dialogue generator by detecting an implicit user persona. Because it is hard to collect a large number of detailed personas for each user, we attempted to model the user’s potential persona and its representation from dialogue history, with no external knowledge. The perception and fader variables were conceived using conditional vari- ational inference. The two latent variables sim- ulate the process of people being aware of each other’s persona and producing a corresponding expression in conversation. Finally, posterior- discriminated regularization was presented to enhance the training procedure. Empirical stud- ies demonstrate that, compared to state-of-the- art methods, our approach is more concerned with the user’s persona and achieves a consid- erable boost across both automatic metrics and human evaluations. 1 Introduction Personalized dialogue modeling is an attractive re- search topic in deep learning, where studies have explored the possibility of incorporating personal facts into the end-to-end generative framework. The established practice of assigning agents a pre- defined character improves the engagingness and consistency of open-domain dialogue. However, such models cannot generate distinguishable re- sponses while interacting with different users be- cause they do not take into consideration who the ∗Equal contribution. Figure 1: An example of dialogue generation with the implicit persona detection. The incorporated persona and the corresponding user’s real persona are in bold. other party is. As Shum et al. (2018) pointed out, a good chit-chat bot not only generates interest- ing responses but also resonates with interlocutors. However, there has been little research conducted into how to make the agent effectively mine a user’s persona to generate customized responses. To this end, this research studied personalized dialogue generation in which we aimed to have the agent recognize the other party’s potential per- sona by exploiting the dialogue itself and output personalized responses conditioned on the differ- ent target users. A simple illustration depicting this process is provided in Figure 1. Inspired by the impressive effectiveness of conditional varia- tional autoencoders (CVAEs) (Sohn et al., 2015; Zhao et al., 2017) with diverse response model- ing, we propose a personalized dialogue generator that detects an implicit user persona using condi- tional variational inference. Specifically, our model fits the profile descriptions of the other party to a multivariate isotropic Gaussian distribution using a latent variable (perception variable) during training. Because responses from …

Content truncated to 500 words