Research And Design Of Chemical-protein Relation Extraction Based On Deep Learning

Along with the burst increase of biomedical literature volume,how to extract valuable information and knowledge with effecacy has turned into an urgent problem.Chemical Protein Relation Extraction(CPRE)refers to the automatic extraction of interactions between chemicals and proteins from biomedical literature,including activation,inhibition,antagonism,and catalysis,etc.It has important significance for the construction of biomedical knowledge graphs,precision medicine and new drug development.This study on extraction of chemical-protein relation mainly includes the following points:(1)We propose the extraction of chemical-protein relation based on the shortest dependency path and ensemble learning.Herein,a bidirectional LSTM model based on the shortest dependency path and attention mechanism is constructed and applied to chemical protein relation extraction.Features,including part of speech,position and dependency type on the shortest dependent path,are taken into full consideration.Experiments on the BioCreative VI CHEMPROT task demonstrate that the proposed method achieves better F1-value performance.Moreover,the ensemble learning method has further improved the performance of chemical protein relation extraction.(2)We compare the performance scores of chemical protein relationship extraction on different pre-trained language models.In view of the progress of the pre-trained models represented by BERT in natural language processing,we employ the current pre-trained models such as BERT,BioBERT,XLNet,etc.in chemical protein relation extraction,followed by performance comparison and analysis among those models.The results demonstrate that the BioBERT model based on biomedical corpora achieves the best performance in chemical protein relation extraction.(3)We design and implement an information extraction platform BioPIE(Platform for Biomedical Information Extraction)for biomedical domain.With the characteristics of versatility and flexibility,the platform can support corpora with different annotation levels such as instance level,sentence level,abstract level and full text level,implementing various fundamental tasks including named entity recognition and relation extraction,as well as other complicated information extraction tasks via class inheritance and rewriting.