Automatic Graph Generation for Document-Level Relation Extraction

Relation extraction, one of important natural language processing tasks, aims to find out the semantic relations between entities in text. It has been raised to the document level recently, which is a complex task that requires a logical inference to extract relations from multiple entities in intra-and inter-sentence. Existing methods mainly utilize co-references or syntactic trees to establish static document-level graphs. Different from these heuristic models which may not always yield the optimal structures, we propose a novel model that can automatically construct task-specific graphs. We regard the process of establishing a graph as a sequence problem and utilize a gated recurrent unit (GRU) recurrent neural networks (RNNs) to construct a document graph. Afterwards, we utilize the graph convolutional networks (GCNs) to calculate the node representation. Particularly, our model exhibits comparable performance to state-of-the-art models on the large-scale human annotated document-level relation extraction dataset (DocRED).