Local-to-global GCN with knowledge-aware representation for distantly supervised relation extraction

Distantly supervised relation extraction is a prevalent technique for identifying semantic relations between two entities. Most prior models cannot distinguish the local and global information in the long-range dependency among words effectively. The latent semantic information presented in existing knowledge graphs is completely neglected, as the knowledge graph information is simply used as a label to specify the class of relations instead of being treated as a graph. Moreover, previous studies only utilized a selective attention mechanism over sentences to alleviate the impact of noise; they did not consider the implicit interaction between sentences in a sentence bag. In this paper, (1) we propose a knowledge-aware framework to enhance word representations, which can highlight the importance of key words and relation clues; (2) we adopt a piecewise self-attention mechanism to model long-range dependency among words; (3) we design a heterogeneous graph structure for each sentence bag and introduce a heterogeneous graph convolutional network with knowledge attention to aggregate the implicit interaction among sentences from a local-to-global perspective. Experimental results on two popular relation extraction datasets demonstrate that our model obtains more discriminative relation representations and outperforms most of the state-of-the-art models. (C) 2021 Elsevier B.V. All rights reserved.