Automated Speech Act Annotation in a Russian Spoken Corpus Using Large Language Models: A Comparative Study

P. 912–920.

Sherstinova T., Viktoria Firsanova

The research focuses on the automatic annotation of a linguistic corpus using large language models (LLMs). Annotating a corpus is a crucial step in its creation, as it determines the practical scope and applications of the resource being developed. This study explores the annotation of oral speech transcripts at the pragmatic level using speech acts that reflect the speaker's intent and purpose. Typically, this task is performed manually by experts, which greatly limits the volume of annotated data that can be produced. In this work, an attempt was made to automatically annotate speech acts using five LLMs commonly used for processing Russian texts – ChatGPT, GigaCHAT, YandexGPT, Mistral, and Gemini. A comparative analysis of the automatic annotation results was conducted, highlighting the strengths and weaknesses of each model. . The findings suggest that employing LLMs for corpus annotation is a promising approach, with ChatGPT and Gemini demonstrating particular effectiveness in speech act categorization. However, for Russian, language-specific models like GigaCHAT and YandexGPT are preferred when language-specific information is needed.

Keywords: LLM

Publication based on the results of: Text as Big Data: methods and models for big text data analysis (2024)

In book

PROCEEDING OF THE 36TH CONFERENCE OF FRUCT ASSOCIATION

[б.и.], 2024.