Long short-term memory RNN for biomedical named entity recognition.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Natural Language Processing

Background:

• Biomedical Named Entity Recognition (BNER) is vital for biomedical information extraction.
• Traditional BNER systems often rely on complex, hand-crafted features.
• Machine learning models like Conditional Random Fields (CRFs) have been applied to BNER.

Purpose of the Study:

• To develop a novel Recurrent Neural Network (RNN) framework for BNER.
• To evaluate the effectiveness of word and character embeddings as features.
• To achieve state-of-the-art performance in BNER without manual feature engineering.

Main Methods:

• A Recurrent Neural Network (RNN) architecture was developed.
• Bidirectional Long Short-Term Memory (LSTM) units were employed to capture contextual information and long-range dependencies.
• A CRF layer was utilized for joint decoding of sentence labels.
• Word and character embeddings served as the primary features.

Main Results:

• The bidirectional LSTM-RNN (BLSTM-RNN) model achieved state-of-the-art performance.
• An F1 score of 86.55% was obtained on the BioCreative II gene mention (GM) corpus.
• An F1 score of 73.79% was achieved on the JNLPBA 2004 corpus.
• Performance improved with domain-specific pre-trained word embeddings and character-level representations.

Conclusions:

• The proposed neural network architecture effectively performs BNER without manual feature engineering.
• The model demonstrates superior performance, outperforming previous top systems on the JNLPBA corpus.
• The study highlights the benefits of using pre-trained word embeddings and character-level features.
• Source code is available for public use.