Research On Biomedical Entity Relation Extraction Algorithm And Application

Biomedical research is closely related to human's health,thus has achieved much attention.In recent years,the number of biomedical literatures has shown a rapid growth.On the one hand,massive biomedical literature is a valuable resource for biomedical experts.On the other hand,manual extraction of useful information from unstructured text is time-consuming and labor-intensive,which reduces the efficiency of biomedical research to some extent.Therefore,biomedical text mining technology has been widely used to address the problem.Biomedical relation extraction is one of the most important tasks in text mining.Currently,mainstream relation extraction method is single-task learning based on deep neural networks.Single-task method cannot effectively utilize the correlation among related tasks in the same domain,which limited the performance and generalization ability of the model.Therefore,we constructed multi-task models including a fully-shared model,a shared-private model and a main-auxiliary model based on the attention mechanism.The private space in the multi-task model extracts the private features of each task.Meanwhile,the shared space can extract shared features among multiple tasks to supplement and enhance the private features.Especially,the proposed attention-based main-auxiliary model uses the attention layer to assign different weights for the different auxiliary tasks according to the effect of each auxiliary task on the main task,which not only makes good use of the positive effects,but also avoids the noise of the auxiliary tasks,thus improving the performance of the model.In addition,we extracted useful information from unstructured biomedical literatures using an automatic relation extraction system based on the deep learning method and constructed a structured disease-specific knowledge graph.Since it is difficult to conduct computation and reasoning on a knowledge graph,we researched on the knowledge graph embedding method including translation-based models and bilinear models,which represent entities and relations in a knowledge graph as vectors in the continuous low-dimensional vector space.Knowledge graph embedding method is a vital step for the downstream tasks,such as question answering and knowledge graph completion.In biomedical domain,research on human malignant neoplasms is of great significance to human life.Protein-protein interactions(PPIs)associated with malignant neoplasms can reveal the molecular mechanisms behind the diseases,which have great value for healthcare professionals.Although there are some cancer-related PPIs databases,they are manually curated and constructed by human experts.As the number of biomedical literatures is rapidly increasing,manual extraction of PPIs is time consuming and inefficient.To this end,we used an automatic relation extraction tool to extract the cancer-related PPIs from large quantities of biomedical literatures,and constructed a protein-protein interactions database associated with human malignant neoplasms.The structured database allows biomedical researchers to directly obtain structured PPIs without digging into a large amount of biomedical literatures,significantly improving the efficiency of research on PPIs related to human malignant neoplasms.