Neural Language Model Vector Retrieval for Noisy Transaction Data

Solution Overview

Problem

Conventional information retrieval systems face challenges in handling noisy data, particularly in online transaction records, where machine processing alters data, making it opaque and difficult to interpret, leading to inaccuracies and errors that hinder the identification of merchants or relevant information.

Innovation Solution

The use of a neural language model to generate mathematical representations of search terms and data records, weighted by inverse document frequency, allows for accurate retrieval by comparing similarity metrics, even with noisy or non-standard data, effectively identifying merchant information in transaction records.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If machine processing is used to handle transaction data, then data processing speed and automation are improved, but data accuracy and interpretability deteriorate due to noise and opacity

Engineering Contradiction:

Improvedata processing speedVSAvoiddata accuracy

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The patent introduces an information retrieval system as an intermediary between noisy machine-processed transaction data and the need for accurate merchant identification. The system uses search queries derived from transaction data to retrieve merchant information from a corpus, acting as a mediator that translates noisy data into meaningful results without requiring direct interpretation of the original noisy fields

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent replaces traditional mechanical data matching approaches with a language model-based information retrieval system. Instead of using rigid field-by-field comparison mechanisms, the system uses natural language processing and semantic search to match transaction data with merchant information, enabling accurate retrieval despite noise in the data

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Device complexity

If conventional search methods are used on noisy transaction data, then system simplicity is maintained, but retrieval accuracy deteriorates due to data noise

Engineering Contradiction:

Improvesystem simplicityVSAvoidretrieval accuracy

Core Design Contradiction:

Device complexityVSMeasurement precision

Solution Approach 1:

The patent introduces an information retrieval system as an intermediary between noisy machine-processed transaction data and the need for accurate merchant identification. The system uses search queries derived from transaction data to retrieve merchant information from a corpus, acting as a mediator that translates noisy data into meaningful results without requiring direct interpretation of the original noisy fields

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent changes the parameters of the search process by using language models to generate semantic representations of both query and corpus data. Instead of searching based on exact field matches, the system transforms data into vector representations and searches based on semantic similarity, fundamentally changing how the retrieval process operates to handle noise

Inventive Principle:

Principle #35Parameter changes

3Speed

If exact match searching is used, then search speed is improved, but adaptability to noisy data deteriorates

Engineering Contradiction:

Improvesearch speedVSAvoidadaptability to noisy data

Core Design Contradiction:

SpeedVSAdaptability or versatility

Solution Approach 1:

The patent changes the parameters of the search process by using language models to generate semantic representations of both query and corpus data. Instead of searching based on exact field matches, the system transforms data into vector representations and searches based on semantic similarity, fundamentally changing how the retrieval process operates to handle noise

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent performs preliminary processing of both the transaction data and the merchant corpus using language models before the actual search. By pre-computing vector representations and storing them in an indexed format, the system prepares the data in advance to enable fast semantic search without requiring complex computations during the actual retrieval operation

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS11640417B2System and method for information retrieval for noisy data

Publication Date: 2023.05.02 Q2 SOFTWARE

AI summary

Embodiments of systems and methods for information retrieval are disclosed. Embodiments of such systems and methods may perform information retrieval based on a language model that is used to generate a single vector for the search terms of a query. Similarly, a single vector representation of each of the data records to be searched is obtained and the single vector representing the search terms of the query compared to the single vector of each data record to determine a similarity metric. The resulting similarity metrics associated with each of the data records can be used to rank, present or return one or more data records.