BGSGA: Combining Bi-GRU and Syntactic Graph Attention for Improving Distant Supervision Relation Extraction

Distant supervision Relation Extraction(RE) aligns entities in a Knowledge Base (KB) with text to automatically construct large-labeled corpus, which alleviates the need for manual annotation in traditional RE. However, most existing models can't take full advantage of the syntactic structure information of each word in the dependency tree. In this paper, we propose BGSGA, a novel distant supervision RE model, to capture both semantic information and syntactic structure in the bag. BGSGA constructs the syntactic graph by combining the dependency trees of the sentences in the bag and then employs a syntactic structure attention mechanism to update the word embedding obtained from Bi-GRU. The syntactic structure attention mechanism captures the cross-sentence information with different weight by implementing the special attention operation between the target word and its neighbors of first-order and second-order in the graph. The experiments show that BGSGA outperforms our baseline models on benchmark datasets.