Abstract: Named entity recognition and relation extraction are two key steps to construct a knowledge graph. To solve the problem that the performance of the existing named entity recognition and relation extraction models would sharply decline due to the lack of a large amount of annotated data in the epilepsy domain, the named entity recognition and relation extraction models were improved according to the data characteristics of the papers in the epilepsy domain. A zero-resource named entity recognition and relation extraction model in the epilepsy domain is proposed based on medical data and a pre-training model from similar domains. The performance of the existing unsupervised and semi-supervised models on the epilepsy paper data set was evaluated, and then a domain adversarial network and a relation discriminator were introduced based on the characteristics of the data set to effectively improve the construction effect of the epilepsy domain knowledge graph. EEG features of epilepsy patients were embedded into the knowledge graph in a visual modality. While improving the interpretability of EEG analysis, it builds a more intuitive multi-modal knowledge graph.

Key words: Knowledge graph, Named entity recognition, Relation extraction, Epilepsy