German Legal Named Entity Recognition With 19 Semantic Classes

Aggregating approximately 67,000 sentences with over 2 million tokens for Named Entity Recognition in German federal court decisions. It includes 54,000 manually annotated entities mapped to 19 fine-grained semantic classes and over 35,000 automatically annotated TimeML-based time expressions.

Use Cases

• Train a Named Entity Recognition model to identify 19 semantic classes such as person, judge, lawyer, and law within German legal text.
• Develop a sequence labeling system for extracting entities like court, company, and organization from over 2 million tokens of legal sentences.
• Build a time expression parser using the over 35,000 TimeML-based annotations to analyze temporal references in court decisions.
• Fine-tune a legal NER service for German documents, focusing on fine-grained classes including ordinance, European legal norm, and contract.

Strengths

• Large scale of 67,000 sentences and over 2 million tokens provides substantial text for model training.
• Includes 54,000 manually annotated entities ensuring high-quality labels for 19 distinct semantic classes.
• Augmented with over 35,000 automatically annotated TimeML-based time expressions, adding a temporal analysis dimension.

Limitations

• The dataset is specific to German federal court decisions, limiting generalizability to other legal systems or languages.
• Temporal staleness as the last update was in February 2021, potentially missing recent legal terminology or rulings.
• No information is provided on potential class imbalance among the 19 semantic entity types.

Provenance

Source

German federal court decisions, developed for the EU project Lynx.

Collection Method

Sentences were manually annotated for entities and automatically annotated for time expressions.

Freshness

2021-02-05

Geography

Germany