README.md

metadata

dataset_info:
  features:
    - name: _id
      dtype: int64
    - name: lai_yuan
      dtype: string
    - name: lai_yuan_id
      dtype: int64
    - name: biao_ti
      dtype: string
    - name: fa_yuan
      dtype: string
    - name: fa_ting_guan_dian
      dtype: string
    - name: shen_li_jie_duan
      dtype: string
    - name: ri_qi_s31
      dtype: string
    - name: ri_q1_s41
      dtype: string
    - name: fu_yan
      dtype: string
    - name: an_you
      sequence: string
    - name: dang_shi_ren
      sequence: string
    - name: an_jian_lei_xing
      dtype: string
    - name: wen_shu_bian_hao
      dtype: string
    - name: shen_li_qing_kuang
      dtype: string
    - name: guan_jian_ci
      sequence: string
    - name: ting_shen_guo_cheng
      dtype: string
    - name: ren_yuan
      dtype: string
    - name: pan_jue_jie_guo
      dtype: string
    - name: label
      struct:
        - name: classification
          dtype: string
        - name: reason
          dtype: string
    - name: charge
      dtype: string
    - name: concept
      dtype: string
  splits:
    - name: train
      num_bytes: 51203688
      num_examples: 2914
    - name: test
      num_bytes: 51902129
      num_examples: 2928
  download_size: 42728249
  dataset_size: 103105817
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
license: apache-2.0
task_categories:
  - text-classification
  - text-generation
language:
  - zh
tags:
  - legal
pretty_name: Legal Concept Entailment
size_categories:
  - 1K<n<10K

Legal Concept Entailment Dataset

Dataset Description

This dataset is released as part of our ACL 2025 main conference paper: Automating Legal Interpretation with LLMs: Retrieval, Generation, and Evaluation. It is designed for the Legal Concept Entailment (LCE) task, which evaluates the quality of legal interpretations by assessing a model's ability to understand and apply vague legal concepts to specific, unseen cases.

The core idea is that a high-quality interpretation of a legal concept should improve a model's performance on determining whether that concept applies to the fact of an unseen case. This dataset serves as a benchmark for this evaluation.

Legal Concept Entailment

The LCE task is a dual-part task designed to test a model's understanding of legal concepts.

• Binary Classification: Given the fact description of a case and a relevant vague legal concept, the model must predict a binary label (Yes/No) indicating whether the concept applies to the facts of the case.
• Reason Generation: The model must also generate a textual reason explaining its classification decision. The quality of this reason is evaluated for consistency against a "gold" reason derived from the court's actual judgment.

An example of the task is shown below:

Languages

The data is in Chinese (zh), as it is sourced from China Judgments Online.

Dataset Structure

Data Splits

The dataset is divided into two splits:

• train: Contains 2914 instances for generating legal concept interpretation.
• test: Contains 2928 instances for testing.

Data Instances

Each instance in the dataset corresponds to a legal case and a specific legal concept.

• label: Contains the classification label (Yes or No) and the gold reason for the classification extracted from the court view.
• concept: The vague legal concept being evaluated.
• ting_shen_guo_cheng: The fact description of the case.
• fa_ting_guan_dian: The court view on the case, which includes the legal interpretation of the concept.