BRONCO

BRONCO provides a manually annotated corpus for information extraction from German oncological discharge summaries, supporting development and evaluation of named entity recognition, entity normalization, and negation and speculation detection.

Key Features:

• Data composition: The corpus comprises 200 manually deidentified oncological discharge summaries totaling 11,434 sentences and 89,942 tokens.
• Annotations: The dataset contains 11,124 annotations for medical entities (including diagnoses, treatments, and medications) and 3,118 annotations for attributes such as negation and speculation.
• Annotation quality: A structured, quality-controlled annotation process was carried out involving two groups of medical experts to ensure consistency and comprehensiveness.
• Data structure: Seventy-five percent of the corpus is published as a set of shuffled sentences while 25% is reserved as a held-out dataset for evaluation.
• Baseline results: Baseline evaluations using state-of-the-art techniques on the held-out dataset report F1-scores ranging from 0.10 to 0.90 across different tasks.

Scientific Applications:

• Information extraction from German medical texts: The corpus supports development and benchmarking of IE systems on German oncological discharge summaries.
• Named entity recognition (NER): Annotations enable training and evaluation of NER models for diagnoses, treatments, and medications.
• Entity normalization: Annotated entity mentions support mapping to normalized concepts or terminologies.
• Negation detection: Negation annotations enable development and testing of negation detection methods.
• Speculation detection: Speculation annotations enable development and testing of speculation/uncertainty detection methods.

Methodology:

The corpus was created by manual deidentification of 200 discharge summaries followed by structured annotation by two groups of medical experts; the dataset is split into 75% shuffled sentences and a 25% held-out set, and baseline evaluations with state-of-the-art techniques yielded F1-scores of 0.10–0.90.