Speaker Identification and Dataset Construction Using LLMs: A Case Study on Japanese Narratives

AbstractSpeaker identification in narrative analysis is a challenging task due to complex dialogues, diverse utterance patterns, and ambiguous character references. Cosly and time-intensive manual annotation limits the scalability of high-quality dataset creation.This study demonstrates a cost-efficient approach of constructing speaker identification datasets by combining small-scale manual annotation with LLM-based labeling. A subset of data is manually annotated and is used to guide LLM predictions with a few-shot approach followed by refinement through minimal human corrections. Our results show that LLMs achieve approximately 90% accuracy on challenging narratives, such as the “Three Kingdoms” dataset, underscoring the importance of targeted human corrections. This approach proves effective for constructing scalable and cost-efficient datasets for Japanese and complex narratives.