Comparison of named entity recognition methodologies in biomedical documents.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Biology

Background:

• Biomedical Named Entity Recognition (Bio-NER) is crucial for extracting terms like RNA, proteins, and DNA from biomedical texts.
• Bio-NER is a foundational task for biomedical knowledge discovery.
• The study utilizes the BioNLP/NLPBA 2004 shared task dataset.

Purpose of the Study:

• To evaluate the effectiveness of different algorithms for Bio-NER.
• To explore the impact of word embeddings on Bio-NER performance.
• To assess methods for reducing the time and cost of creating labeled data for Bio-NER.

Main Methods:

• Implementation of Recurrent Neural Network (RNN) algorithms (Jordan-type and Elman-type).
• Application of Conditional Random Fields (CRF) as a machine learning algorithm.
• Utilizing pre-trained word embeddings: CCA, GloVe, and Word2Vec.

Main Results:

• Baseline Bio-NER achieved an F1 score of 70.09%.
• RNN algorithms yielded lower F1 scores (approx. 60.53% and 58.80%).
• CRF with word embeddings (CCA, GloVe, Word2Vec) achieved superior F1 scores ranging from 72.73% to 72.82%.

Conclusions:

• Unsupervised learning for word embeddings significantly improves Bio-NER performance.
• The use of pre-trained word embeddings reduces the need for extensive manual data labeling.
• This approach offers a more efficient and cost-effective method for building Bio-NER systems.