Sihan Wu, Susanna Gordon, Martin Boeker, Luise Modersohn
The approach achieves strong NER performance after optimization. Better data quality or applying HPT improve performance comparably, but combining both yields only marginal gains additionally.
INTRODUCTION: Clinical text in electronic health records contains valuable information for medical research. Named entity recognition (NER) and relation extraction (RE) are key tasks for structuring this information, but existing deep learning approaches are often computationally expensive.
METHODS: We develop a lightweight, pipeline-based NER-RE system using neural networks with contextualized embeddings based on spaCy and Thinc. The impact of data preprocessing and hyperparameter tuning (HPT) on NER is evaluated. RE is performed using a separate classifier on candidate entity pairs. Experiments are conducted on i2b2 and n2c2 datasets.
RESULTS: The approach achieves strong NER performance after optimization. Better data quality or applying HPT improve performance comparably, but combining both yields only marginal gains additionally.
DISCUSSION: Lightweight models are effective for NER but limited for RE due to error propagation. Our NER results also suggest that HPT can compensate for data quality issues.