Abstract
This paper presents the first automated system for extracting sentencing decisions from Hebrew-language court cases, specifically focusing on sexual assault litigation. Hebrew presents unique challenges as a low-resource language with complex morphology and unstructured legal documentation. We propose a two-stage pipeline for Automated Punishment Extraction (APE) that combines expert-crafted rules with machine learning. Our results demonstrate that a rule-based approach outperforms standard supervised models, achieving a 68% accuracy in identifying actual imprisonment terms with a mean error of only five months.
Introduction
Legal Natural Language Processing (NLP) has advanced significantly in English, yet many languages remain "low-resource" due to a lack of specialized datasets and pre-trained models. Hebrew is particularly difficult because of its morphological complexity and the unstructured nature of Israeli court decisions. A significant hurdle is the ambiguity of sentencing language; for instance, the term Ma’asar (imprisonment) refers to both actual jail time and conditional sentences (probation). Furthermore, judges often cite past cases or prosecution requests within the same document, making it difficult for automated systems to distinguish the final verdict. This lack of structured data has historically limited empirical research into sentencing patterns for sensitive crimes like sexual assault.
Methods
We developed a two-stage pipeline for Automated Punishment Extraction (APE). The first stage, Imprisonment Sentence Detection, filters the hundreds of sentences in a case down to a handful of candidates using keyword-based filtering. We compared a rule-based scoring system against supervised models like Support Vector Machines (SVM) and Random Forests (RF). The second stage, Duration Extraction and Normalization, uses custom Hebrew number-parsing logic to convert linguistic expressions into a standardized number of months.
Datasets
The study utilizes a new corpus of 1,043 sexual assault sentencing decisions from Israeli Magistrate and District Courts spanning 1990–2021, totaling 3 million words. A subset of 100 cases (13,000 sentences) was manually annotated by legal experts from the Association of Rape Crisis Centers in Israel (ARCCI) to serve as a gold standard for evaluation.
Key Results
The evaluation revealed that the rule-based model outperformed "black box" machine learning approaches in this specific domain. As shown in the performance comparison below, the rule-based model achieved the highest F1-scores and the lowest average error.
The rule-based model correctly extracted the exact punishment in 68% of cases, with a mean error of 5 months—significantly lower than the 11.6 months for Random Forest. Analysis of the full dataset indicated a median predicted punishment of 3 years, though over a third of sentences were below 15 months, suggesting a gap between actual sentences and statutory maximums.
Discussion and Limitations
The primary bottleneck remains linguistic ambiguity, specifically the confusion between actual imprisonment and probation terms, which often share identical units and terminology. Supervised models showed high recall but poor calibration, frequently selecting references to past cases or prosecution requests instead of the final verdict. Future work will explore the integration of Hebrew-specific transformer models like Alephbert and anchor-based learning to better distinguish between different types of sentencing mentions within a single document.
Summary updated