NER for Court Cases

This solution tags entities in U.S. court case records as either government or person to support criminal justice research. Working with national case data from PACER spanning 2002 to 2020, the team built models to identify which parties in court proceedings are individuals versus government entities. The system helps researchers studying resource allocation for courts, prison conditions, and other criminal justice topics by making it easier to understand who is involved in cases and their outcomes. The team sampled 100,000 observations from a dataset of 3.4 million records provided by the challenge sponsor.

The team developed three approaches to entity classification. First, they created a rule-based model using twelve ordered rules based on patterns like presence of formal titles, phrases such as 'department of' or 'city of', and token counts. A Decision Tree model applied to these rule-based labels achieved 99% accuracy with 0.6% false negatives. Second, they tested GLiNER_medium-v2.1, a zero-shot model from HuggingFace, on a 2,000-record sample, which matched the rule-based target 59% of the time. Third, they began work on a custom spaCy small English language pipeline trained with labeled examples of formal government titles and person suffixes, though computational limitations prevented full development. The team tokenized text, analyzed token length patterns, created binary features, and used the extra_info field to classify edge cases like John Doe entries.

The solution addresses weak national data collection across U.S. state criminal justice systems, where the country spends approximately $264 billion annually. By properly tagging entities involved in court cases, the work supports research efforts focused on court resource allocation and evaluating prison conditions. The models help identify who was involved in cases and case outcomes, information that is currently difficult to obtain across fragmented state systems.