NLP for AML Review Analysis Qorus-Infosys Finacle Banking Innovation Awards 2025- Nominated

Concept and objectives: Our project leverages Natural Language Processing (NLP) to automatically assess the similarity between the descriptions of anti-money laundering (AML) alerts and the explanatory notes provided by branch managers. It also extracts information from these notes using topic modeling. The primary objective is to efficiently identify discrepancies and non-compliant justifications. Reasons behind: The current manual verification of all AML alerts and their related notes is time-consuming and resource-intensive, and sometimes not fully comprehensive. This often leads to auditors spending significant effort on adequately justified reports. Our solution aims to filter out well-reasoned cases, allowing auditors to focus exclusively on potential issues. State of competition: While generic text similarity tools exist, there is a lack of a specific, off-the-shelf solution tailored to the unique characteristics of AML alert descriptions and branch explanations within the Italian banking sector. Our project fills this gap by combining a pre-trained data corpus (Wikipedia, Facebook) with one created from proprietary data, focused on the banking sector, anti-money laundering, and the Italian context. Our approach stands out further due to its innovative nature. Specifically, our topic modeling activities on the proprietary corpus have enabled us to extract information from the text of explanatory notes with an unprecedented level of detail. This work has been recognized for its originality and scientific rigor, leading to the publication of a paper presented at the Italian Conference on Computational Linguistics. Sources of inspiration: The project draws inspiration from advancements in NLP for semantic analysis and text comparison, coupled with the operational challenges faced by the bank's audit function in processing AML alerts efficiently. Initially, we analyzed topic modeling to aggregate information from the explanatory notes, subsequently evolving towards more advanced NLP techniques, recognizing significant opportunities for streamlining colleagues' work. Departments involved: The key departments involved are Internal Audit (the team of analysts and network auditors, which identified the need and will be the primary user); Anti-Money Laundering Compliance, whose contribution was crucial in defining the requirements for justifications; and potentially the Commercial Network, specifically branch managers, should a review of non-compliant notes be necessary. Main results so far: We have developed twoNLP models: one that calculates the similarity score between alert descriptions and branch notes and another that extracts the main topics from branch justifications. The first few months of backtesting have shown an algorithm accuracy of 70%. We are refining the processing of the model's outcomes, anticipating a quarterly involvement of auditors for verification.