THE PROBLEM
Client, a FinTech software provider is developing a treasury automation product targeted towards SME customers which profiles the banking transactions of customers to generate future cash flow predictions. Profiling of banking transactions required the extraction of critical information from banking transaction descriptions (like payment mechanism, currency pairs, counterparties, nature of transaction etc.) and classification of the transaction into predefined categories. Client had already developed a module for the same but the coverage quality, and types of information being extracted were not satisfactory and wanted to explore additional mechanisms to enhance the module.
INXITE OUT APPROACH
Automated Ingestion & Secure Storage of Transactions
Banking transactions for a subscribed customer is ingested automatically in a scheduled manner from Open Banking APIs and stored securely for further processing.
Preliminary Information Extraction
Context aware regex-based rules were developed to extract information like payment mechanism, counterparties, currency pairs, conversion rates etc., and to categorize the transaction based on the nature of transaction and currency-pairs involved.
Knowledge Graph based Information Extraction
Google Knowledge Graph (enhanced by a contracting n-gram wrapper) along with Local Knowledge Graph (based on historical model outputs on transactions in the past) were used for extraction of information such as Transaction type (hierarchical categories over 3 levels of granularity), Parties involved in the transaction etc.
RESULT
Module developed achieved the following improvements over the existing approach:
• Increased coverage (the number of transactions from which information could be extracted) by ~32%.
• Enhanced the accuracy of transaction categorization and information extraction by ~42%.
• Doubled the breadth of information extracted for each transaction.
