Extraction of specific entities from documents

This thesis explores information extraction, which involves identifying and extracting specific information from unstructured documents such as Images, Word, Excel, Outlook messages, and PDFs. Despite the availability of established algorithms to perform these tasks, the field has gained more attention with the growth of deep learning, transfer learning, and self-supervised learning approaches. The primary focus of this thesis is the study and fine-tuning of LayoutLMv3, a pre-trained multimodal transformer for Document AI that is capable of handling text and image data. The research investigates the potential of this state-of-the-art model for extracting specific entities from logistic documents in the supply chain management domain. The performance of the fine-tuned LayoutLMv3 model is compared to its predecessor, LayoutLMv2 while investigating the preparation of the dataset from PDFs, preprocessing images, post-processing, and inference algorithm. The objective is to enhance efficiency, particularly in the sensitive sector of supply chain management. The project was developed as part of an internship at Wenda SRL, a startup based in Bologna that is expanding into supply chain management. The work followed the CRISP-DM methodology typical of data science projects and aims to improve efficiency in the SCM sector. Due to company policy, the dataset used in this thesis is confidential. The deployment phase uses Hugging Face to prepare models for implementation as accessory products for Wenda SRL’s customers