Evaluation of active generation of interlocutor profiling sentences from utterances and their implicit context
Year: 2024
Type: conference-paper
Abstract: Recently, it has been reported that the quality of chats can be improved by utilizing information about what kind of persona the chat interlocutor and what kind of personality the system itself mimics in order to generate responses using LLMs. In these studies, personality data is stored in a form of statements describing interlocutor’s profile. In order to control LLM using profile describing sentences in actual chats, it is not enough to prepare them in advance, but it is necessary to actively add them through the chats. However, not enough research has been conducted on generating the profiling sentences from the utterances in the chat dialogs. In particular, there are no studies on the generation of profiling sentences that can only be generated with context the context. Therefore, in this study, we propose a task to generate profiling sentences from the target utterances based on context in a chat dialog in Japanese, and create a dataset for this task. Experiments on the created dataset and analysis show that LLM can generate profiling sentences while taking the context into account.
Authors: Muraji, Shinji, Rzepka, Rafal, Itoh, Toshihiko
Institution: Hokkaido University
Language: English
Topic: Speech and dialogue systems
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Award: G2097858196
OpenAlex ID: W7144615855