Beyond Traditional Algorithms: Leveraging LLMs for Accurate Cross-Border Entity Identification

TL;DR

This paper investigates the use of Large Language Models to improve the accuracy of cross-border entity identification, addressing limitations of traditional string-matching algorithms in complex, multilingual financial contexts.

Contribution

It introduces LLM-based methods as a superior alternative for entity matching, demonstrating enhanced accuracy and reduced false positives over traditional techniques.

Findings

Traditional methods achieve over 92% accuracy but have high false positive rates.

LLM-based approaches surpass 93% accuracy and 96% F1 scores.

Interface-based LLMs significantly reduce false positives compared to traditional algorithms.

Abstract

The growing prevalence of cross-border financial activities in global markets has underscored the necessity of accurately identifying and classifying foreign entities. This practice is essential within the Spanish financial system for ensuring robust risk management, regulatory adherence, and the prevention of financial misconduct. This process involves a labor-intensive entity-matching task, where entities need to be validated against available reference sources. Challenges arise from linguistic variations, special characters, outdated names, and changes in legal forms, complicating traditional matching algorithms like Jaccard, cosine, and Levenshtein distances. These methods struggle with contextual nuances and semantic relationships, leading to mismatches. To address these limitations, we explore Large Language Models (LLMs) as a flexible alternative. LLMs leverage extensive training to interpret context, handle abbreviations, and adapt to legal transitions. We evaluate traditional methods, Hugging Face-based LLMs, and interface-based LLMs (e.g., Microsoft Copilot, Alibaba's Qwen 2.5) using a dataset of 65 Portuguese company cases. Results show traditional methods achieve accuracies over 92% but suffer high false positive rates (20-40%). Interface-based LLMs outperform, achieving accuracies above 93%, F1 scores exceeding 96%, and lower false positives (40-80%).