Structured Extraction and Hierarchical Parsing of Addresses Using Large Language Models

In the era of digitalization and cross-border trade, address inconsistencies present a critical challenge across sectors such as taxation, fraud prevention, law enforcement, and public service delivery. Unstructured and inconsistent address formats-characterized by spelling errors, missing postal codes, incorrect or absent city and street names, and improper punctuation-hinder the effectiveness of automated systems. These issues are particularly detrimental to efforts aimed at combating tax fraud in an European region or international context, where precise address data is essential for accurate entity verification, transaction tracking, and fraud detection. Despite initiatives by tax authorities, postal agencies, and government bodies to enforce standardized address formats, real-world data continues to exhibit substantial variability and noise. Traditional rule-based solutions fall short in addressing the complexity of such data, while existing AI-based methods are often not tailored to the nuances of European address formats. In this research, we propose an approach to transform unstructured address data into structured formats using pre-trained large language models (LLMs), with a focus on the European region. Our experimental results demonstrate that the LLM-based approach significantly outperforms conventional methods in generating structured addresses, and these structured outputs have shown strong performance when evaluated using our rule-based script. This hybrid solution offers a promising tool in the broader effort to ensure regulatory compliance and enforce tax authorities in the fulfillment of their legal mandate against illegal activities.