Document-Level Iterative Entity and Relation Extraction for Materials Scientific Literature

The existing document-level entity and relation extraction methods mainly concentrate on generic semantics. However, for scientific literature, especially in the materials domain, there are a large number of entities that are iteratively generated by other entities under certain conditions, which are different from traditional overlapping entities and overlapping relations that overlap at least one entity. To address these challenges, this paper proposes a new entity type in the field of information extraction: Iterative Entity as above mentioned. Meanwhile, we propose a document-level Iterative Entity and Relation Extraction method for Materials scientific literature, namely MatIERE, which contains two modules: 1) named entity extraction model for materials scientific literature with hybrid rule and semantic block labeling, which is used to extract entity such as materials, processes, and material pronouns, etc., while a materials process knowledge base is introduced to refine the extraction results; 2) rule-based iterative entity and relation recognition algorithm: firstly, we use material pronouns as relation trigger words to search for the corresponding entity or iterative entity within the context and establish the relation; then we propose to add the entity-relation triples extracted from current iteration as iterative entity to the entity set, which is input for the next iteration to extract the document-level iterative entity and relation. In the experiments, we first construct a dataset with a total of 48,714 entities and 22,885 document-level relations containing iterative entities from materials scientific literature. The comparison results show that our approach significantly outperforms other baseline models.