Abstract:
Fine-grained entity typing results can serve as important information for entities while constructing knowledge bases. It is a challenging task due to the use of large tag sets and the requirement of understanding the context.
We find that, in some cases, existing neural fine-grained entity typing models may ignore the semantic information in the context that is important for typing.
To address this problem, we propose to exploit semantic relations extracted from the sentence to improve the use of context. The used semantic relations are mainly those that are between the mention and the other words or phrases in the sentence.
We investigate the use of two types of semantic relations: hypernym relation, and verb-argument relation. Our approach combine the predictions made based on different semantic relations and the predictions of a base neural model to produce the final results. We conduct experiments on two commonly used datasets: FIGER (GOLD) and BBN. Our approach achieves at least 2\% absolute strict accuracy improvement on both datasets compared with a strong BERT based model.
