Meet Your Favorite Character: Open-domain Chatbot Mimicking Fictional Characters with only a Few Utterances

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5114 - 5132 July 10-15, 2022 ©2022 Association for Computational Linguistics Meet Your Favorite Character: Open-domain Chatbot Mimicking Fictional Characters with only a Few Utterances Seungju Han† Beomsu Kim† Jin Yong Yoo† Seokjun Seo Sangbum Kim Enkhbayar Erdenee Buru Chang∗ Hyperconnect {seungju.han,beomsu.kim,jeffrey,seokjun.seo,airdish,enkhbayar.erdenee, buru.chang}@hpcnt.com Abstract In this paper, we consider mimicking fictional characters as a promising direction for building engaging conversation models. To this end, we present a new practical task where only a few ut- terances of each fictional character are available to generate responses mimicking them. Fur- thermore, we propose a new method named Pseudo Dialog Prompting (PDP) that gener- ates responses by leveraging the power of large- scale language models with prompts containing the target character’s utterances. To better re- flect the style of the character, PDP builds the prompts in the form of dialog that includes the character’s utterances as dialog history. Since only utterances of the characters are available in the proposed task, PDP matches each utter- ance with an appropriate pseudo-context from a predefined set of context candidates using a retrieval model. Through human and auto- matic evaluation, we show that PDP generates responses that better reflect the style of fictional characters than baseline methods. 1 Introduction How would you feel if you could talk to your fa- vorite character? In recent years, open-domain conversation mod- els (Adiwardana et al., 2020; Roller et al., 2021) have achieved remarkable progress with the devel- opment of large-scale language models (Radford et al., 2019; Brown et al., 2020). Meanwhile, recent studies have suggested several directions reflecting desirable traits of real-life conversation to make open-domain conversation models more engaging beyond plain chit-chat. Style-controlling conversa- tion models generate responses in the target styles such as emotion (Zhou et al., 2018; Demszky et al., 2020) and empathy (Rashkin et al., 2019). Persona- grounded conversation models (Zhang et al., 2018a; Kim et al., 2020; Majumder et al., 2020) produce †Equal contribution ∗Corresponding author Figure 1: Illustration of PDP. The retriever matches pseudo-context for utterances from the character, and utilizes them in a prompt while generating the response. responses that preserve consistent personalities by leveraging personal descriptions (e.g., "I have two dogs"). In this paper, we consider mimicking fic- tional characters as a promising direction for build- ing engaging conversation models. When it comes to building conversation models that mimic fictional characters, two major chal- lenges prevent us from directly applying previous models designed for conditional response gener- ation: (1) It is difficult to define fictional charac- ters with only a few descriptions, as in persona- grounded conversation models. Furthermore, it is not expressive enough to represent characters’ styles with discrete labels (e.g., angry, happy), as style-controlling conversation models do. (2) There lacks sufficient dialog data of fictional characters for training conversation models. It is inefficient to manually create dialog datasets of characters for training, especially considering that additional data is needed for …

Content truncated to 500 words