Temporal Expression Recognition in Legal Transcripts

Abstract

In litigation, trial transcripts provide verbatim records of witness testimony, primarily given in response to attorney questioning. To effectively analyze these transcripts, lawyers must often reconstruct events in chronological order—a task that begins with identifying dates associated with testified facts. This paper introduces two datasets for temporal expression extraction from legal transcripts: a primary dataset derived from a lengthy 1995 U.S. criminal trial, and a smaller robustness-testing dataset drawn from seven other legal proceedings. We evaluate semi-supervised approaches for date entity recognition, fine-tuning neural models on weakly labeled training data, and benchmarking them against both small and large language models. Our best-performing models achieve 83% F1-score on the primary dataset (FLAIR rule-modified) and 72% F1-score on the cross-domain, small test set (BERT-cased). These results, alongside our annotated datasets and corresponding experiments, provide a foundation for developing robust date extraction and temporal ordering tools for speech-derived legal text. Moreover, we identify unique challenges for state-of-the-art NER models on legal transcripts, including legal terminology and multiple anchor date resolution.

Resources

Details