FGCS: A Fine-Grained Scientific Information Extraction Dataset in Computer Science Domain

As scientific communities grow and evolve, more and more papers are published, especially in computer science field (CS). It is important to organize scientific information into structured knowledge bases extracted from a large corpus of CS papers, which usually requires Information Extraction (IE) about scientific entities and their relationships. In order to construct high-quality structured scientific knowledge bases by supervised learning way, as far as we know, in computer science field, there have been several handcrafted annotated entity-relation datasets like SciERC and SciREX, which are used to train supervised extracted algorithms. However, almost all these datasets ignore the annotation of following fine-grained named entities: nested entities, discontinuous entities and minimal independent semantics entities. To solve this problem, this paper will present a novel Fine-Grained entity-relation Extraction dataset in Computer Science field (FGCS), which contains rich fine-grained entities and their relationships. The proposed dataset includes 1,948 sentences of 6 entity types with up to 7 layers of nesting and 5 relation types. Extensive experiments show that the proposed dataset is a good benchmark for measuring an information extraction model’s ability of recognizing fine-grained entities and their relations. Our dataset is publicly available at https://github.com/broken-dream/FGCS.