Transformer-based approach for symptom recognition and multilingual linking

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Biomedical Natural Language Processing

Background:

• Named Entity Recognition (NER) and entity linking are crucial for extracting clinical information.
• Existing methods may struggle with the nuances of clinical text and multilingual contexts.

Purpose of the Study:

• To develop and evaluate a transformer-based approach for Spanish symptom NER.
• To implement and assess a multilingual entity linking system for clinical texts.

Main Methods:

• Fine-tuning a RoBERTa-based model with BiLSTM-CRF for Spanish NER.
• Employing a hybrid dictionary-based approach with SapBERT and GPT-3.5 for multilingual entity linking.

Main Results:

• Achieved an F1 score of 0.73 for Spanish symptom NER.
• Obtained 0.73 accuracy for multilingual entity linking and 0.6123 for Spanish entity linking, surpassing previous benchmarks.

Conclusions:

• The proposed transformer-based methods demonstrate strong performance in clinical NER and multilingual entity linking.
• This approach enhances the extraction and standardization of clinical information across languages.