Pipeline paraextração automática de informações em documentos fiscais utilizando técnicas de inteligência artificial

Abstract

The automatic extraction of information from fiscal documents, such as invoices and billing sta- tements, presents challenges due to layout variability and the lack of national standardization. In this context, this work proposes the development of a pipeline for automated document reading that combines Optical Character Recognition (OCR), Natural Language Processing (NLP), and spatial text features to locate and extract data without relying on individual mappings through rigid templates. The methodology included image preprocessing with the Hough Transform and Unsharp Masking for alignment and noise correction, followed by the identification of label des- criptors using approximate search algorithms (Fuzzy Matching) with sliding n-grams. To extract the values, a semantic approach based on the BERT model juridics/bertimbau-base-portuguese- sts-scale was used to infer the types of entities expected for each label descriptor. In addition, a Named Entity Recognition (NER) model with fine-tuning was applied to identify, within spatially coherent text blocks (candidates), the information that could correspond to these labels. The combination of these components resulted in a decision system that integrates positional, se- mantic, and logical aspects. The results showed that preprocessing reduced the Character Error Rate (CER) by up to 45.2% in degraded documents. Label identification achieved an F1-Score of 96.00%, while the BERT model obtained 90% accuracy in semantic classification, and the fine-tuned NER model increased the F1-Score from 0.04 (baseline) to 0.71. The proposed solu- tion proves to be a resilient architecture operable locally (on-premises), capable of adapting to the heterogeneity of Brazilian fiscal documents and optimizing the efficiency of data acquisition.