Towards Nested and Fine-Grained Open Information Extraction

Open Information Extraction is a crucial task in natural language processing with wide applications. Existing efforts only work on extracting simple flat triplets that are not minimized, which neglect triplets of other kinds and their nested combinations. As a result, they cannot provide comprehensive extraction results for its downstream tasks. In this paper, we define three more fine-grained types of triplets, and also pay attention to the nested combination of these triplets. Particular, we propose a novel end-to-end joint extraction model, which identifies the basic semantic elements, comprehensive types of triplets, as well as their nested combinations from plain texts jointly. In this way, information is shared more thoroughly in the whole parsing process, which also lets the model achieve more fine-grained knowledge extraction without relying on external NLP tools or resources. Our empirical study on datasets of two domains, Building Codes and Biomedicine, demonstrates the effectiveness of our model comparing to state-of-the-art approaches.