Investigating Information Extraction and Language Models in Medical Domain Text Processing

This dissertation demonstrates that carefully adapted language-model pipelines can transform unstructured clinical-trial and pharmacological prose into reliable, low-latency structured data. Four interconnected studies support this claim.Tri-AL platform. An open-source dashboard ingests all 440 k+ ClinicalTrials.gov records—including every historical revision—into a normalized schema and parses the 20 GB XML archive over 10x faster than a Beautiful Soup baseline, while exposing hooks for demographic analytics and supporting integration of user-defined modules. Clinical trial summarization. An encoder–decoder model is trained on 57k description–summary pairs to condense clinical trials into a few sentences. ROUGE evaluation shows a 20% improvement over the baseline, while graph-based evaluation indicates the model preserves 71% of critical biomedical entities, yielding concise yet informative summaries suitable for evidence scans. MoA classification. A collection of models—including traditional classifiers (decision trees, random forests, XGBoost) and contrastively fine-tuned masked-language-model variants—achieves a macro F1 of 97%, effectively handling class imbalance and drug-class sparsity while also providing interpretable insights. Scalable medical NER pipeline. A dynamic and scalable pipeline is introduced for training lightweight Named Entity Recognition (NER) models adaptable to different entity types. Knowledge distillation compresses the large teacher model into a 110M-parameter student that retains 70% of gold-label accuracy (F1=0.61) while running 1000x faster and consuming just 6% of the memory. Collectively, these contributions provide scalable tools and empirical evidence that domain-specific NLP methods can be integrated to accelerate trial discovery, enhance drug-development analytics, and support data-driven clinical decision-making.