Multi-granularity Neural Networks for Document-Level Relation Extraction

Document-level relation extraction aims to extract relations from multiple sentences in a document. However, it remains challenging to obtain rich semantic information across multiple sentences for relation prediction. In this paper, a multi-granularity relation extraction (MGRE) neural network is proposed, which integrates multiple granularity semantic features (i.e., entity level, sentence level and document level), to capture the semantic interactions among entities and sentences in a document. For entities, the shortest dependency path is utilized to obtain head-to-tail entity representations, which is further used for acquiring the relation between entity pairs through a translation strategy. Then, at the sentence level, a convolutional neural network is used to extract semantic features for each sentence. While for the documents, an attention mechanism is adopted to fuse multiple sentence-level feature vectors into document-level semantic features. Finally, entity representation and document representation are combined into a holistic representation for relation prediction. Extensive experiments are conducted on the DocRED dataset against state-of-the-art methods, and the comparative results demonstrate the superiority of MGRE on document-level relation extraction.