Text Classification Using Positive-Only Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automatic text classification systems are inefficient and costly due to their reliance on substantial amounts of training data, especially in scenarios with highly imbalanced text where only a small percentage needs to be classified, and they cannot function without training data, particularly when limited to only positive labeled data.
Innovation Solution
A method that calculates a metric for each text document using a sentiment score, word function score, and distance score, allowing for preliminary classification and ranking without negative labeled data, and iteratively improves by adding new keywords to enhance classification accuracy, enabling efficient classification with limited positive training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional automatic text classification algorithms are used, then classification accuracy can be maintained, but substantial amounts of training data (both positive and negative examples) are required, making the system prohibitively expensive and inefficient
Solution Approach 1:
The patent segments the classification process into two distinct stages: a first classification stage that processes all text documents using a computational model trained on limited positive examples, and a second classification stage that applies to documents meeting specific criteria from the first stage. This segmentation allows the system to achieve accurate classification while requiring minimal training data, as the first stage filters documents to reduce the burden on the second stage.
Solution Approach 2:
The patent applies preliminary action by performing the first classification stage on all text documents before applying the more resource-intensive second classification stage. The first stage uses a computational model trained on limited positive examples to identify documents that meet specific criteria, thereby preparing and filtering the data set before the second stage processes it. This preliminary filtering reduces the quantity of documents requiring extensive processing in the second stage.
2Adaptability or versatility
If text classification is performed in highly imbalanced scenarios where only a small percentage of text should be classified in a particular way, then the system must handle sparse positive examples, but current systems cannot perform without substantial training data with both positive and negative examples
Solution Approach 1:
The patent inverts the traditional classification approach by training the computational model primarily on positive examples rather than requiring balanced positive and negative training data. The first classification stage is designed to identify documents meeting specific criteria using this model trained on limited positive examples, thereby inverting the conventional requirement for substantial negative training data. This inversion enables the system to effectively handle highly imbalanced text scenarios.
3Reliability
If traditional classification systems are used, then comprehensive classification can be achieved, but the process is time-consuming and computationally expensive
Solution Approach 1:
The patent segments the classification process into two stages where the first stage quickly filters documents using a computational model trained on limited data, and the second stage applies more intensive processing only to documents meeting specific criteria. This segmentation maintains classification completeness by ensuring all documents are evaluated through the first stage, while reducing processing time by limiting the second stage to a subset of documents.
Solution Approach 2:
The patent applies partial action by performing the computationally intensive second classification stage only on documents that meet specific criteria from the first stage, rather than applying it to all documents. This partial application of the second stage reduces overall processing time and computational cost while maintaining reliability, as the first stage ensures comprehensive initial evaluation of all documents.
Data Source
AI summary
Systems, method, and computer-readable mediums for automated text classification, and particularly a mechanism for performing binary classification using only a set of positive labeled data as training data and having a large set of unlabeled data, where the algorithm can function without any information regarding the negative class. The disclosed classification systems and methods may use a text classification process which automatically classifies text based on the current positive training data available, but identifies additional words which can be added to the positive training data such that future iterations of the text classification can better identify the positive class of text.