Disambiguation Model for Bio-Medical Named Entity Recognition

Abstract

Discovery of biomedical named entities is one of the preliminary steps for many biomedical texts mining task. In the biomedical domain, typical entities are present, including disease, chemical, gene, and protein. To find these entities, currently, a deep learning-based approach applied into the Biomedical Named Entity Recognition (Bio_NER) which gives prominent results. Although deep learning-based approach gives a satisfactory result, still a tremendous amount of data is required for training because a lack of data can be one of the barriers in the performance of Bio_NER. There is one more obstacle in the path of Bio_NER is polysemy or misclassification of the entity in bio-entity. Which means one biomedical entity might have a different meaning in different places, i.e., a gene named entity may be labeled as disease name. When Conditional Random Field combined with deep learning-based approach i.e. Bidirectional Long Short Term Memory (Bi-LSTM), It mistakenly labeled a gene entity “BRCA1” as a disease entity which is “BRCA1 abnormality” or “Braca1-deficient” present in the training dataset. Similarly, “VHL (Von Hippel-Lindau disease),” which is one of the genes named labeled as a disease by Bi-LSTM CRF Model. One more problem is addressed in this chapter, as bio-med domain, entities are long and complex like cell whose name is “A375M (B-Raf (V600E)) is a human melanoma cell line”, in this biomedical entity, multiple words are present, but still it is difficult to find the context information of this particular bio-entity. For lack of data and entity misclassification problem, this chapter embeds multiple Bio_NER models. In the proposed model, the model trained with different datasets is connected so that the targeted model obtained the information by combining another model, which reduce the false-positives rate. Recurrent Neural Network (RNN) which is dependent upon the Bi-LSTM gates are introduced to handle the long and complex range dependencies in biomedical entities. BioCreative II GM Corpus, Pubmed, Gold-standard dataset, and JNLPBA dataset are used in this research work.