Span-based Multi-dimensional Neural Network for Joint Entity Relation Extraction

This paper designs an effective joint modeling method to handle information in multiple dimensions. Traditional span-based models tend to focus on the semantics within the span but fail to capture the interaction between the span and information in other dimensions. This paper proposes a Span-based Multi-dimensional Neural Network for Joint Extraction of Entities and Relations (MD-SpERE). This network extracts the context span information and simultaneously extracts the information on the start and end positions of the span as well as the word length. During the modeling process, joint modeling is conducted on the start and end positions of the span and the span width, and joint modeling is also carried out on the word length and the word semantics. Experiments on three public datasets, namely SciERC, Conll04 and ADE, have verified the effectiveness of the proposed model.