Pseudo-Document Generation for Text Classification Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current text classification technologies face challenges in accurately classifying documents with scarce training data, as existing methods are inefficient and often rely on assumptions that do not hold in enterprise applications, leading to suboptimal performance, especially for Bayesian classifiers like Naïve Bayes.
Innovation Solution
The implementation of Jackknife Document Replication (JDR), a data-centric technique that generates pseudo-documents by subsampling original documents, combined with model-centric approaches such as adjusting the smoothing parameter, to enhance the accuracy of Bayesian classifiers in low-data scenarios, and the use of adjointness to unify these approaches.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional text classification methods are used with scarce training data, then classification accuracy deteriorates, but generating more training data through manual collection is time-consuming and cost-ineffective
Solution Approach 1:
The patent creates pseudo-documents by copying and subsampling from existing original documents. The Jackknife Document Replication technique generates multiple pseudo-documents from a single original document by systematically subsampling portions of the text, thereby creating artificial training data without manual collection efforts
Solution Approach 2:
The patent performs preliminary data preparation by creating an expanded training set through pseudo-document generation before the actual classification task. By pre-generating synthetic training data using Jackknife subsampling, the system prepares sufficient training material in advance, eliminating the need for time-consuming manual data collection later
2Reliability
If Bayesian classifiers like Naïve Bayes are used with scarce training data, then classification performance becomes suboptimal, but collecting more original training documents is expensive and impractical
Solution Approach 1:
The patent generates multiple pseudo-documents from limited original documents through systematic subsampling. Each original document produces several pseudo-documents that maintain the topical characteristics while providing additional training samples, effectively multiplying the available training data quantity
Solution Approach 2:
The patent adjusts the smoothing parameter in Bayesian classifiers to optimize performance with the generated pseudo-documents. By tuning model parameters alongside data generation, the system adapts the classifier to work effectively with the synthetic training data distribution
3Measurement precision
If more original training documents are collected to improve classification accuracy, then the quantity of training data increases, but the cost and time required for data collection increase exponentially
Solution Approach 1:
The patent creates synthetic training data by copying and subsampling from existing documents. One original document can generate multiple pseudo-documents, providing an efficient multiplication of training data that avoids the exponential cost of manual data collection
Solution Approach 2:
The system performs self-service data augmentation by automatically generating pseudo-documents from its own existing training corpus. The Jackknife Document Replication process is automated, allowing the system to expand its training data without external human intervention or additional data collection efforts
Data Source
AI summary
A method of generating training documents for training a classifying device comprises, with a processor, determining a number of sub-samples in a number of original documents, and creating a number of pseudo-documents from the sub-samples, the pseudo-documents comprising a portion of the number of sub-samples. A device for training a classifying device comprises a processor, and a memory communicatively coupled to the processor. The memory comprises a sampling module to, when executed by the processor, determine a number of sub-samples in a number of original documents, a pseudo-document creation module to, when executed by the processor, create a number of pseudo-documents from the sub-samples, the pseudo-documents comprising a portion of the number of sub-samples, and a training module to, when executed by the processor, train a classifying device to classify textual documents based on the pseudo-documents.