MeExtRel: a dataset of inter-method extend-use relationships based on citation sentence classification

Abstract The exponential growth in the number of documents has promoted fine-grained research on scientific and technological documents. These include techniques for entity identification and relationship extraction, abstract generation, knowledge extraction, and other forms of fine-grained research. Nevertheless, the aforementioned research is only capable of extracting fine-grained knowledge from a single document, and is thus unable to correlate such knowledge across different documents. The research methods is the most critical content of a paper. Therefore, clarifying the inheritance, improvement, or extension of research methods (defined in this paper as inter-method extend-use relationships) is particularly important for understanding the continuity and innovation of scientific research. In order to address this issue, this paper puts forward a new research task, namely the identification of inter-method extend-use relationships, and proposes the construction of a corresponding dataset. The dataset is a hierarchical dataset containing two levels of structure. The upper layer is a citation relationship classification dataset, while the lower layer is a extend-use relationship dataset. The former classifies citation sentences in the literature into three categories: background citation, method citation and result citation. The latter further classifies method citation sentences into two subcategories on the basis of the former: extend-use relationships and non-extend-use relationships. To verify the feasibility of our task and the validity of our dataset, a fine-tuned SciBERT model was employed to classify citation and extend-use relationships. Additionally, the challenges of extracting inter-method extend-use relationships are pointed out in conjunction with error analyses.