A Unified Knowledge Extraction Method Based on BERT and Handshaking Tagging Scheme

In the actual knowledge extraction system, different applications have entity classes and relationship schema, so generalization migration ability of are very important. By training a model in source domain applying to an arbitrary target directly, open technology becomes crucial mitigate issues. Traditional models cannot be directly transferred new domains also extract undefined relation types. order deal with above issues, this paper, we proposed end-to-end Chinese open-domain model, TPORE (Extract Open-domain Relations through Token Pair linking), which combined BERT handshaking tagging scheme. can alleviate nested entities relations Additionally, loss function that conducts pairwise comparison category score non-target automatically balance weight was adopted, experiment results indicate bring speed performance improvements. The extensive experiments demonstrate method significantly surpass strong baselines. Specifically, our approach achieve state-of-the-art Relation Extraction (ORE) benchmarks (COER SAOKE). COER dataset, F1 increased from 66.36% 79.63%, SpanSAOKE 46.0% 54.91%. medical domain, obtain close compared SOTA CMeIE CMeEE datasets.