Adverse News Detection and Classification System for Anti-Money Laundering Investigations

Abstract

Adverse media screening is an important component of Anti-Money Laundering (AML) compliance, requiring the analysis of large volumes of unstructured news data. This project presents the development of an automated adverse media detection system using Natural Language Processing techniques to support AML-focused risk analysis. The system collects multilingual news articles from publicly available Really Simple Syndication feeds, applies Named Entity Recognition using spaCy models for English and Spanish, and incorporates rule-based filtering and normalization to improve entity extraction consistency. A keyword-based risk-scoring methodology assigns article-level risk scores, which are later aggregated to evaluate entity-level exposure to adverse media indicators. A Streamlit based web application was developed to facilitate interactive exploration of ranked entities and associated articles. The system achieved 90.66% precision in entity extraction, with a 95% confidence interval ranging from 85.55% to 94.09%, demonstrating the effectiveness of the proposed approach for multilingual adverse media analysis within AML contexts.