Scholarlink

Address standardization is a key task in spatial data management, particularly in last-mile logistics, where address inconsistencies from multiple sources directly affect delivery operations. Traditional address parsing methods, including rule-based and machine learning approaches, struggle with linguistic variability, typographical errors and ambiguous formatting. Recent advances in Large Language Models (LLMs) offer a promising alternative, as they can leverage contextual understanding without requiring task-specific training. This study evaluates the effectiveness of four pre-trained LLMs (Qwen2.5-72B-Instruct, LLaMA 3, Mistral Large and DeepSeek-R1) in extracting structured information from Spanish delivery addresses. Their capability to handle real-world variability is assessed by analyzing their performance in identifying key address components. The results indicate that while LLMs achieve high accuracy in structured elements like street number, they exhibit inconsistencies in more variable components, such as street type. Furthermore, challenges related to task comprehension and reasoning are identified, highlighting areas for future improvement. The findings underscore the transformative potential of LLMs in automating address parsing, offering a scalable, readily available alternative to traditional methods.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Address Parsing in the Era of LLMs: A Comparative Analysis

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ana Pegado-Bardayo,
  • Jesús Muñuzuri,
  • Alicia Robles-Velasco,
  • Juan Carlos Ragel-Bonilla

摘要

Address standardization is a key task in spatial data management, particularly in last-mile logistics, where address inconsistencies from multiple sources directly affect delivery operations. Traditional address parsing methods, including rule-based and machine learning approaches, struggle with linguistic variability, typographical errors and ambiguous formatting. Recent advances in Large Language Models (LLMs) offer a promising alternative, as they can leverage contextual understanding without requiring task-specific training. This study evaluates the effectiveness of four pre-trained LLMs (Qwen2.5-72B-Instruct, LLaMA 3, Mistral Large and DeepSeek-R1) in extracting structured information from Spanish delivery addresses. Their capability to handle real-world variability is assessed by analyzing their performance in identifying key address components. The results indicate that while LLMs achieve high accuracy in structured elements like street number, they exhibit inconsistencies in more variable components, such as street type. Furthermore, challenges related to task comprehension and reasoning are identified, highlighting areas for future improvement. The findings underscore the transformative potential of LLMs in automating address parsing, offering a scalable, readily available alternative to traditional methods.