<p>Notarized documents serve as certified records of acts and contracts overseen by notaries, conferring unchallengeable legal authority. These documents significantly reduce transaction costs in goods and services markets, with millions produced and utilized annually by stakeholders, including banks, insurance companies, and government agencies. Extracting information from these unstructured documents is essential for automating business processes reliant on their content. However, this task is challenging due to the distinct characteristics of these legal documents and the limited availability of datasets and language models in Spanish. This paper addresses the issue of Information Extraction from Spanish notarized documents. We propose a framework that relies solely on end-to-end data usage, enabling the extraction of legal elements without labeled datasets. Our framework is effective in legal domains, particularly with notarized documents, where text extraction from scanned documents is common and often affected by optical character recognition noise. We created a new dataset from a corpus of notarized documents and fine-tuned several pretrained multilingual language models, including mT5, mBART, and the large language model BLOOMZ. Additionally, we developed a model from scratch based on the T5 architecture, utilizing a corpus of Spanish notarized documents. The promising results, with an F1 score of 96.13, validate the effectiveness of our proposed framework compared to existing methods in legal document information extraction.</p>