| Biomedical literature has become a treasure house of knowledge that urgently needs to be explored and is an essential data resource in the biomedical domain.The chemical-protein relation extraction is to automatically extract chemicals and proteins and their relations from unstructured biomedical literature.It is of great significance for pharmacological and clinical research,playing a key role in drug discovery,understanding the molecular mechanism of adverse drug reactions,describing drug metabolism,or drawing regulatory networks of importance for systems pharmacology.Therefore,it has important research significance and social value to extract chemical-protein relations from biomedical literature.This work includes three key tasks,i.e.,named entity recognition,relation extraction,and joint extraction of entities and relations.For named entity recognition,the existing methods cannot effectively use the BIO labeling scheme to recognize target entities.To address this issue,the proposed method uses BERT in the machine reading comprehension framework to recognize chemical and protein entities.Experimental results show that the proposed model has certain advantages in chemical and protein entity recognition performance compared with the existing models.Compared with using BERT in sequence labeling,using BERT in the machine reading comprehension framework can enhance the model’s ability to recognize chemical and protein entities.For relation extraction,the existing methods cannot effectively distinguish the importance of tokens at different positions in the sequence.To address this issue,the proposed method introduces Gaussian probability distribution to increase the weights of the target word and its adjacent words.Meanwhile,it uses external knowledge to guide the model to extract chemicalprotein relations.Experimental results show that the proposed method is reliable and effective,and significantly improves model performance on chemical-protein interaction extraction.The introduction of Gaussian probability distribution to enhance the weights of the target word and its adjacent words is beneficial to improving model performance.Using external knowledge to guide the model also contributes to chemical-protein interaction extraction.There is a complementary relationship between the two,and using these two aspects of information can further improve model performance.For joint extraction of entities and relations,the existing joint extraction methods are difficult to effectively extract overlapping triplets.To solve this issue,a novel method for the joint extraction of entities and relations is proposed.The proposed method formulates the joint extraction as extracting entity spans in the machine reading comprehension framework and combines the extracted entities and relations into triplets using a proposed EntityCombine algorithm.Experimental results show that,compared with the existing joint extraction methods,the proposed method can effectively extract triplets and is especially good at extracting overlapping triplets in the CHEMPROT dataset. |
