Class Vector Adjustment for NLP False Positive Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automated document classification methods using natural language processing face challenges in accurately distinguishing similar documents, such as non-disclosure and non-compete agreements, due to high similarity in terminology, leading to low classification accuracy and the need for dedicated neural networks for each class, which is time-consuming and costly to train and maintain.
Innovation Solution
A method involving pre-specified class vectors and a test set of documents with embedding vectors, where misclassified documents adjust the class vectors by calculating an adjustment vector based on the mean of misclassified documents' vectors, improving classification accuracy without requiring new neural networks for each class.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional neural network approaches are used to classify similar document types, then classification coverage is improved, but classification accuracy deteriorates due to high similarity in terminology
Solution Approach 1:
The patent segments the classification task by creating separate specialized classifiers for each document type (non-disclosure agreement classifier, non-compete agreement classifier, etc.). Each classifier is trained independently on its specific document type, allowing them to specialize in recognizing subtle distinctions within their domain while maintaining overall system versatility through the ensemble of multiple classifiers.
2Measurement precision
If dedicated neural networks are trained for each document class, then classification accuracy for similar documents is improved, but training time and cost increase significantly
Solution Approach 1:
The patent performs preliminary action by pre-training each specialized classifier on its specific document type before deployment. This upfront training investment creates ready-to-use specialized models that can quickly and accurately classify new documents without requiring retraining when new documents arrive, thus reducing ongoing time costs.
Solution Approach 2:
The patent uses copying by creating multiple specialized classifiers that follow the same architectural pattern and training methodology. Once a classifier is successfully trained for one document type, the same approach can be copied to create classifiers for other document types, reducing the overall time and effort required compared to developing each classifier from scratch.
3Adaptability or versatility
If the classification system needs to add or remove document classes, then system adaptability is improved, but retraining effort and complexity increase
Solution Approach 1:
The patent implements universality by creating a modular classification system where each document type has its own independent classifier. This allows the system to universally handle multiple document types through the same ensemble framework, making it easy to add or remove document classes by simply adding or removing individual classifiers without affecting the overall system architecture or requiring retraining of other classifiers.
4Ease of manufacture
If conventional approaches are used, then implementation simplicity is maintained, but false positive rates increase for similar document types
Solution Approach 1:
The patent implements feedback by having multiple specialized classifiers vote on document classification. Each classifier provides its assessment based on its specialized training, and the ensemble combines these assessments to reach a final classification decision. This feedback mechanism from multiple specialized sources reduces false positives by cross-validating classifications, while the modular nature maintains implementation simplicity.
Data Source
AI summary
Embodiments improve a document classification system by adjusting the class vector of each class based on embedded vectors of documents known to have been misclassified by the system. Embodiments methods of adjusting the class vector may be described as machine learning.