AI Account Summary Generation via Feature Extraction and Rule-Based Validation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence-based approaches for generating account-related summaries in anti-money laundering (AML) reports are inefficient, prone to factual hallucination, and fail to accurately present transaction details.
Innovation Solution
A computer-implemented method and system that access historical transaction data and a rule generator file to generate a set of transaction features, extract relevant features using a meta-learning model, and create a structured report template with natural language sentences, which are then populated with corresponding feature values to produce accurate account-related summaries.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual analysis is used to generate AML reports, then accuracy and thoroughness of analysis is improved, but time consumption and labor requirements increase
Solution Approach 1:
The system segments the AML report generation process into multiple independent modules: data collection module, feature extraction module, risk assessment module, and report generation module. Each module handles a specific aspect of the analysis, allowing parallel processing and reducing overall time while maintaining comprehensive coverage through coordinated operation of all segments.
Solution Approach 2:
The patent introduces an intermediary risk assessment model that acts as a mediator between raw transaction data and the final AML report. This intermediary layer processes and interprets complex transaction patterns, extracting meaningful risk indicators and translating them into actionable insights, thereby reducing the time required for manual analysis while maintaining high accuracy.
2Extent of automation
If existing AI models are used to generate account summaries, then automation level is improved, but factual accuracy and reliability deteriorate due to hallucination
Solution Approach 1:
The system implements a feedback mechanism where the generated account summary is automatically validated against the source transaction data and risk assessment results. Any discrepancies or factual errors detected in the summary are flagged and corrected by the system, ensuring high factual accuracy while maintaining full automation. This feedback loop prevents hallucination by continuously verifying output against ground truth data.
Solution Approach 2:
The patent replaces conventional AI language models with a rule-based expert system that uses predetermined risk assessment criteria and transaction analysis algorithms. This substitution of mechanical/AI-based systems with deterministic rule-based logic eliminates the hallucination problem inherent in neural networks, as the system only generates summaries based on explicit rules and verified data, ensuring factual accuracy while maintaining automation.
3Loss of information
If detailed transaction data is processed, then completeness of information is improved, but processing complexity and computational resources increase
Solution Approach 1:
The system extracts only the relevant and necessary features from the detailed transaction data using a feature extraction module. It identifies and isolates key risk indicators such as transaction amounts, frequencies, counterparty information, and temporal patterns, while filtering out redundant or irrelevant data. This extraction process reduces processing complexity by focusing computational resources only on the most informative features, maintaining information completeness for risk assessment purposes.
Solution Approach 2:
The patent applies local quality by differentiating the processing depth for different types of transaction data. High-risk transaction categories (e.g., cross-border transfers, cash transactions) receive detailed analysis and comprehensive data processing, while low-risk transactions are processed with simpler algorithms. This localized approach to data processing intensity ensures that computational resources are allocated efficiently based on the inherent complexity and risk level of each data type, maintaining information completeness where needed while reducing overall processing complexity.
Data Source
AI summary
Embodiments provide artificial intelligence based methods and systems for generating account-related summaries. Method performed by server system include accessing rule generator file and historical transaction data corresponding from database. The method includes generating a set of transaction features based on the rule generator file. The method includes extracting via a first machine learning model, a subset of relevant transaction features from the set of transaction features based on the historical transaction data. The method includes generating via second machine learning, a structured report template based on the subset of relevant transaction features. The structured template report includes a plurality of natural language sentences embedded with the subset of relevant transaction features. The method includes generating an account-related summary for the account holder by substituting each of the subset of relevant transaction features embedded in the plurality of natural language sentences with a corresponding feature value from the historical transaction data.