RELATION EXTRACTION OF ADVERSE DRUG REACTION IN SOCIAL MEDIA BASED ON GRAPH INFERENCE NETWORK

In order to solve the problem of long distance dependence and semantic complexity caused by cross-sentence entities in Chinese medical social media users' consultation records, we propose a relation extraction model of adverse drug reactions based on graph inference network. A heterogeneous mention graph was constructed for the text, and document node was added. All mention nodes were connected to document node through document edges to better model the long-distance dependency in the document. The document-aware representation of each mention node was obtained by using graph convolutional network for the mention graph, and the entity-level graph was constructed by merging the mention nodes that pointed to the same entity in the mention graph. Link prediction mechanism was proposed to extract the relation between unrelated entity pairs in entity graph, and the relation between target entity pairs could be further inferred by predicting the relation between related entities. Experimental results on the data obtained from the "Good Doctor Online" website show that the proposed model can effectively improve the performance of document-level relation extraction.