AI Sanctions Screening Reduces False Positives
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for screening sanctioned entities in data transactions often result in overproduction of false alerts, leading to unnecessary resource expenditure and potential errors during secondary reviews due to resource constraints.
Innovation Solution
A method utilizing a neural network to process natural language questions generated from data envelopes, which includes transforming data into text format, extracting target information, and determining the presence of sanctioned entities based on context, thereby reducing false positives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional conservative scanning is used to screen sanctioned entities, then compliance with legal requirements is ensured, but false alerts are overproduced leading to unnecessary resource expenditure
Solution Approach 1:
The screening process is divided into two distinct stages: (1) an initial conservative scanning phase that ensures compliance by flagging all potential matches, and (2) a secondary AI-based review phase that processes only the flagged alerts. This segmentation allows the system to maintain high reliability in the first stage while reducing resource consumption in the second stage by limiting processing to a smaller subset of alerts.
Solution Approach 2:
The system intentionally performs excessive action in the first stage by flagging all potential matches including false positives, ensuring no compliance violations are missed. Then in the second stage, it applies partial action by using AI to review only the flagged alerts rather than all transactions, thereby reducing overall resource expenditure while maintaining compliance.
2Measurement precision
If all alerts are reviewed through secondary review process, then accuracy is improved, but resource constraints cause potential errors
Solution Approach 1:
The system creates a copy of the alert data in a standardized format suitable for AI processing, separating the initial scanning function from the review function. This allows the AI model to efficiently process copies of alert information without requiring access to the original complex transaction data, improving both accuracy and productivity.
Solution Approach 2:
The manual secondary review process is partially replaced with an AI-based automated review system that uses machine learning models to analyze flagged alerts. This substitution maintains high accuracy in identifying sanctioned entities while significantly increasing review throughput by automating the analysis of large volumes of alerts.
3Measurement precision
If computing resources are expanded to process all alerts, then false positives can be reduced, but resource costs increase
Solution Approach 1:
The system performs preliminary action by using conservative scanning to pre-filter and flag only the alerts that require further review. This preliminary step reduces the volume of data that needs to be processed in the secondary review stage, allowing the system to achieve high false positive reduction accuracy without requiring excessive computing resources for processing all alerts.
Solution Approach 2:
The system changes the parameter of alert processing by transforming alerts into a standardized text format with extracted target information and surrounding information. This parameter transformation enables the AI model to process alerts more efficiently by focusing on relevant features rather than raw transaction data, reducing computing resource requirements while maintaining high accuracy in false positive reduction.
Data Source
AI summary
A method and system for reducing false-positives using a neural network are disclosed. The method includes receiving a data envelope, scanning the data envelope and detecting a presence of data corresponding to a value in a list of sanctioned entities. The method further includes transforming format of the data envelope into a text format, identifying and extracting target information including the data corresponding to the value in the list of sanctioned entities and surrounding information, and reformatting the extracted target information and surrounding information into a natural language question. Once the natural language question is provided, the natural language question is processed by a neural network model for determining whether a sanctioned entity is included or not based on context provided by the natural language question.