TrustHLT/ECtHR_dataset

Dataset description

This published dataset repository contains 2 different datasets, namely ecthr dataset and uniformly distributed (according to the violation existence labels) subsample of it.

ECTHR dataset

We scrape HUDOC database to create the ECtHR dataset. Our dataset does not include ambiguous cases in terms of violation existence. What are those ambiguous scenarios: - Inadmissible cases: The cases which did not make their way to the court to be investigated; - Lack of jurisdiction: When the court does not have enough sources to decide on the case; - Preliminary objection allowed (or partially allowed) - Inadmissibility or lack of jurisdiction to be applied either to the full document or some part of it.

Labels

• We already labelled dataset based on violation existence task for our experiments. However, we provide following information that can be used for further experiments:
  • violated articles: Articles or protocols of European Convention of Human Rights were found violated for the case;
  • non-violated articles: Articles claimed to be violated by the applicant, but no violation was found by the court based on the case's merits;
  • non-analyzed articles: Articles were not taken into consideration by the court (i.e., do not exist in the conclusion of the case).
• As a result of filtering the HUDOC data, we collect 16.7k ECtHR cases, where 15.2k cases have at least one article violated, and 1.5k cases have no violation.

---

Violation existence dataset

• For our experiments, we subsample the dataset defined above based on the minority subset (i.e., negative samples);
• As a result we obtain 3.1k cases for the violation existence classification task;
• As the negative samples have the non-violation of article 6 and 8 more than the other articles, we subsample the positive cases accordingly, to maintain the article distribution across categories.
• We also maintain temporal distribution based on decades, as year-based distribution is too sparse.