rohit-upadhya/lexclipr

Dataset Division

• Unique_query : contains unique queries that are not present in train, val, test split. Used for testing unseen understandability of models.
• Train_all : contains the unsplit train, test, val datapoints.
• train : train split
• val : val split
• test : test split

Original Paper : LexCLiPR: Cross-Lingual Paragraph Retrieval from Legal Judgments

Bibtext:

@inproceedings{upadhya-t-y-s-s-2025-lexclipr,
    title = "{L}ex{CL}i{PR}: Cross-Lingual Paragraph Retrieval from Legal Judgments",
    author = "Upadhya, Rohit  and
      T.y.s.s, Santosh",
    editor = "Che, Wanxiang  and
      Nabende, Joyce  and
      Shutova, Ekaterina  and
      Pilehvar, Mohammad Taher",
    booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.acl-long.683/",
    doi = "10.18653/v1/2025.acl-long.683",
    pages = "13971--13993",
    ISBN = "979-8-89176-251-0",
    abstract = "Efficient retrieval of pinpointed information from case law is crucial for legal professionals but challenging due to the length and complexity of legal judgments. Existing works mostly often focus on retrieving entire cases rather than precise, paragraph-level information. Moreover, multilingual legal practice necessitates cross-lingual retrieval, most works have been limited to monolingual settings. To address these gaps, we introduce LexCLiPR, a cross-lingual dataset for paragraph-level retrieval from European Court of Human Rights (ECtHR) judgments, leveraging multilingual case law guides and distant supervision to curate our dataset. We evaluate retrieval models in a zero-shot setting, revealing the limitations of pre-trained multilingual models for cross-lingual tasks in low-resource languages and the importance of retrieval based post-training strategies. In fine-tuning settings, we observe that two-tower models excel in cross-lingual retrieval, while siamese architectures are better suited for monolingual tasks. Fine-tuning multilingual models on native language queries improves performance but struggles to generalize to unseen legal concepts, highlighting the need for robust strategies to address topical distribution shifts in the legal queries."
}