Automated Extraction of Sentencing Decisions from Court Cases in the Hebrew Language

Abstract

We present the task of Automated Punishment Extraction (APE) in sentencing decisions from criminal court cases in Hebrew. Addressing APE will enable the identification of sentenc ing patterns and constitute an important step ping stone for many follow up legal NLP ap plications in Hebrew, including the prediction of sentencing decisions. We curate a dataset of sexual assault sentencing decisions and a manuallyannotated evaluation dataset, and implement rulebased and supervised models. We find that while supervised models can iden tify the sentence containing the punishment with good accuracy, rulebased approaches outperform them on the full APE task. We con clude by presenting a first analysis of sentenc ing patterns in our dataset and analyze com mon models' errors, indicating avenues for fu ture work, such as distinguishing between pro bation and actual imprisonment punishment. We will make all our resources available upon request, including data, annotation, and first benchmark models.