Building and Evaluating a Negation Detection System for Greek Biomedical Text

Introduction Negation detection is essential in biomedical NLP to correctly interpret medical records. Greek remains underrepresented compared to other languages, due to limited annotated corpora. This thesis develops and evaluates a negation detection system for Greek biomedical text using a deep learning approach, comparing it with rule based methods to improve automated biomedical text processing. Research Question This study examines: "How effectively can fine tuned transformer based models detect negation in Greek biomedical text, compared to rule based approaches?" BERT based models, fine tuned on Greek biomedical text, are evaluated against a NegEx adaptation, analyzing strengths and limitations in handling negation cues and scope. Method A quantitative experimental approach is applied, utilizing a Greek biomedical negation corpus to fine tune GreekBERT and XLM-RoBERTa for negation detection. Performance is assessed against a rule based NegEx system using precision, recall, and F1-score. Preprocessing includes tokenization, data augmentation, and other NLP techniques. Results The results demonstrate that the deep learning based approach offers a clear advantage over the rule based system in handling negation in Greek biomedical text. The transformer models demonstrate high accuracy and reliability, effectively capturing complex negation patterns that challenge traditional rule based methods. Discussion The findings confirm that transformer based models can robustly detect negation in Greek biomedical text. While performance is strong, handling complex or implicit negation is challenging, highlighting opportunities for future improvement.