Analyzing transfer learning impact in biomedical cross-lingual named entity recognition and normalization.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• The exponential growth of biomedical literature necessitates efficient information extraction.
• Named Entity Recognition (NER) is vital for acquiring knowledge from unstructured biomedical texts.
• Limited Spanish biomedical word embeddings hinder current NER approaches.

Purpose of the Study:

• To develop novel Spanish biomedical word representations.
• To introduce deep learning models for recognizing biomedical entities in Spanish texts.
• To evaluate the performance of Bi-LSTM-CRF and BERT-based architectures for NER.

Main Methods:

• Development of several Spanish biomedical word representations.
• Implementation of a Bi-LSTM-CRF deep learning model.
• Implementation of a BERT-based deep learning architecture.
• Evaluation on PharmaCoNER and CORD-19 datasets for entity identification, classification, and normalization.

Main Results:

• The BERT model achieved an F-score of 88.80% for entity identification and classification on PharmaCoNER.
• The BERT model achieved an F-score of 79.97% for entity normalization on PharmaCoNER.
• Both models demonstrated strong performance on the CORD-19 dataset, with BERT achieving 78.86% F-measure.

Conclusions:

• Deep learning models trained on in-domain data significantly enhance NER performance.
• Contextualized representations are crucial for handling the complexity and ambiguity of biomedical texts.
• Developing language-specific embeddings is essential for advancing NER in non-English languages.