Exploiting the Syntax-Model Consistency for Neural Relation Extraction