[4Xin1] Poster session 2

[4Xin1-08] Generating Readme with Heuristics-Augmented Large Language Models

〇Yuta Koreeda¹, Terufumi Morishita¹, Osamu Imaichi¹, Yasuhiro Sogawa¹ (1.Hitachi, Ltd.)

Keywords:Natural language processing, Large language model, Software engineering, Natural language generation

Writing a readme is a critical part of software development but automatically creating one remains a challenge as it requires generating abstract description from thousands of lines of code. In this paper, we show that large language models are capable of generating a coherent and factually correct readme if we can identify a code fragment that is representative of the repository. Based on this finding, we developed representative code identification based on heuristics and weak supervision. We show the efficacy of the proposed system through human and automated evaluations.