Text Classification via LLM Ensemble Queries

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current text classification methods, such as keyword search and machine learning models, are limited in detecting sensitive information due to their sensitivity to phrasing variations and require large labeled datasets, which hampers their effectiveness in applications like spam filtering and sentiment analysis.

Innovation Solution

A system and method utilizing an ensemble of closed-ended questions combined with a conversational language model to classify textual content, generating queries from the text and feeding inference values into a decision model for accurate classification, which can handle phrasing variations and improve accuracy with weaker language model assumptions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If keyword search and template-based text filters are used, then the system is simple and fast, but it fails to interpret phrasing variations and is style sensitive

Engineering Contradiction:

Improvesimplicity and speedVSAvoidability to interpret phrasing variations

Core Design Contradiction:

Ease of operationVSAdaptability or versatility

Solution Approach 1:

The patent introduces an intermediary component (image generation module) that transforms text into visual representations. This intermediary allows the system to bypass the limitation of direct text matching while maintaining simplicity, as the image-based representation can capture semantic meaning beyond exact keyword matches.

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent replaces the mechanical keyword-matching system with an image-based semantic representation system. Instead of relying on exact text string matching, the system converts text to images and uses image analysis to determine semantic equivalence, thereby achieving robustness to phrasing variations.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

If machine learning models are used for language classification, then classification accuracy improves, but large sets of labeled data are required

Engineering Contradiction:

Improveclassification accuracyVSAvoidamount of labeled data

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent inverts the traditional approach by not directly classifying text using machine learning models. Instead, it converts text to images and uses image analysis to infer text properties. This inversion allows the system to leverage image processing capabilities while avoiding the data hunger of traditional NLP models.

Inventive Principle:

Principle #13The other way round (Inversion)

Solution Approach 2:

The patent substitutes the traditional machine learning-based text classification mechanism with an image-based analysis mechanism. By converting the problem domain from text to images, the system can achieve accurate classification without requiring large labeled text datasets, as image analysis models can be trained on or transfer from image data.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentUS20240281668A1Document classification using large language models

Publication Date: 2024.08.22 NEC CORPOATION OF AMERICA

AI summary

A system and a method for classifying text from a document using an ensemble of close ended questions and a neural network based large language model, which might have been trained for different purposes. The method comprising feeding the language model with queries based on the text and the ensemble and post processing output of the language model using a knowledge representation rule based model or an additional machine language model. The method may be used as spam or phishing filter, detect sensitive contents due to various criteria, and/or the like.