A common task in legal Information Retrieval (IR) is to find relevant legal sources from case-law collections. While legal practice often requires pinpoint citations (pincites) to specific case paragraphs, most existing public legal IR datasets lack paragraph-level citation annotations. Yet, publicly available datasets with such information contain data leakage in the query text and exclude paragraphs that are neither citing nor cited from the corpora, creating an unrealistic and oversimplified retrieval setting, potentially leading to inflated performance. To address these limitations, we contribute a large-scale legal IR dataset constructed from Court of Justice of the European Union (CJEU) judgments. The dataset contains: (i) masked case/paragraph queries, with removed citation information; (ii) a corpus that includes all paragraphs; and (iii) case- and paragraph-level ground truth citations, with partial human expert validation. Our dataset supports both the development and rigorous evaluation of legal IR methods, at multiple query-document levels (case-to-case, paragraph-to-case, and paragraph-to-paragraph retrieval). Link to dataset and code: https://huggingface.co/datasets/theresiavr/legalpincite

The collection of historical notarial documentation from Bologna is a valuable source, providing deep insights into the city’s institutional, legal, and socio-economic history. However, many of these manuscripts have sustained physical damage during centuries of conservation, rendering the text incomplete. To address this, we explored the restoration of these Latin notary documents using encoder-based pre-trained language models (PLMs) under the assumption that the length of missing text is known by estimation from the physical damage. We address the structural misalignment between the physical lacuna of the manuscript and the subword tokenization schemes of PLMs by designing an iterative decoding strategy to align model predictions with the known physical dimensions of lacuna. We also compared the efficacy of monolingual versus multilingual pre-training. Our strategy significantly outperforms baselines consist of standard decoding methods. Furthermore, stratified analysis across different text sections reveals that while monolingual models achieve better performance in general, multilingual models show a suggestive advantage in lexically dense segments, though this finding is not statistically significant. Overall, the best performance achieved by our method is a Hit@1 rate of 35.47% in the short-span setting and 18.75% in the long-span setting. While fully autonomous restoration remains an open challenge, our system provides a useful assistive tool for paleographers.