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

What it is about

The paper defines Automated Punishment Extraction (APE): from an Israeli court sentencing decision in a sexual-offence case, find the sentence that states the actual imprisonment and give its length in months. The corpus has 1043 Magistrate and District Court decisions (1990–2021, from the Nevo database), about 3 million words; 100 cases (13k sentences) are annotated with help from legal experts. Hebrew uses the same word for total imprisonment, actual imprisonment and suspended (probation) imprisonment, which makes the task hard. The system first picks the sentence (keyword filter, then a rule scorer, an SVM or a random forest), then reads the number. The rule-based scorer wins the full task (APE F1 0.65 against 0.58 for random forest and 0.50 for SVM). The best model extracts the correct term in 68% of cases, with an average error of about 5 months. The most common error is taking the probation term for the actual one. The corpus is available only on request.

Ideas

• I1. A deterministic check for compound sentences: when a sentence has exactly three numbers and the total equals the actual term plus the suspended term, take the actual term. This is arithmetic over numerals the parser already tags, so it needs no model.
• I2. Weighted cue lists with an explanation: strong and moderate cues for the judicial verb (present tense, infinitive used with past meaning), negative cues for docket brackets and for requests. Score each sentence, pick a threshold on a dev set, and print which cues fired. This fits Mova's "say why" rule.
• I3. The Hebrew phrase meaning "and the rest" as a test: it marks the actual term when it follows the number. Write 20–30 sentences with and without the phrase as a test set in the register of legal Hebrew.

For Mova

I1 belongs in Nouns, adjectives, pronouns and numerals, in a new numerals section (Mova has no numerals chapter yet). I2 is a rule table in prag. I3 goes into 8. Lexicon and register as a legal-register test.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/2021.nllp-1.4

@inproceedings{wenger2021automated,
  title = {{Automated Extraction of Sentencing Decisions from Court Cases in the Hebrew Language}},
  author = {Wenger, Mohr and Kalir, Tom and Berger, Noga and Chalamish, Carmit Klar and Keydar, Renana and Stanovsky, Gabriel},
  booktitle = {Proceedings of the Natural Legal Language Processing Workshop 2021},
  year = {2021},
  doi = {10.18653/v1/2021.nllp-1.4},
  url = {https://doi.org/10.18653/v1/2021.nllp-1.4}
}

References

2 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Nikolaos Aletras, Dimitrios Tsarapatsanis, Daniel PreoţiucPietro, and Vasileios Lampos. 2016. Pre dicting judicial decisions of the european court of human rights: A natural language processing per spective. PeerJ Computer Science, 2:e93. *Figure 3: The distribution of predicted punishments in our corpus. These were extracted using our rule-based model which performed best on this task. Median predicted punishment is 3 years, while more than a third of the punishments are below 15 months.* L Karl Branting, Craig Pfeifer, Bradford Brown, Lisa Ferro, John Aberdeen, Brandy Weiss, Mark Pfaff, and Bill Liao. 2021. Scalable and explainable legal prediction. Artificial Intelligence and Law, 29(2):213238. find
2. Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis, Nikolaos Aletras, and Ion An droutsopoulos. 2020. Legalbert: The mup pets straight out of law school. arXiv preprint arXiv:2010.02559. ↗

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/aclbulk-2021-wenger-automated-extraction-sentencing-decisions-court/