RepoFromPaper: An Approach to Extract Software Code Implementations from Scientific Publications

Abstract

AbstractAn increasing amount of scientists link to their research software code implementations in their academic publications in order to support the reusability of their results. However, research papers usually contain many code links (e.g., from reused tools or existing competing efforts) making it challenging to automatically establish clear links between papers and their corresponding implementations. This paper presents RepoFromPaper, an approach for automatically extracting the main code implementation associated with a research paper, based on the context in which that link is mentioned. Our approach uses fine-tuned language models to retrieve the top candidate sentences where a code implementation may be found, and uses custom heuristics to link candidate sentences back to their corresponding URL (footnote, reference or full-text mention). We evaluated RepoFromPaper on 150 research papers, obtaining an F1 score of 0.94. We also run our approach on nearly 1800 papers from the CS.AI Arxiv category, discovering 604 paper-repository links and making them available to the community.