Non-linear Text Classification via Dimension Reduction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current automated machine learning techniques for text classification are inefficient and costly due to inaccuracies and the need for human review, especially when dealing with large volumes of text samples, such as user-generated content.

Innovation Solution

The use of non-linear classifiers like random forest and Nyström/Fisher classifiers, combined with dimension reduction techniques, to classify text samples by generating features through n-grams and applying hashing operations, reducing the feature space and improving classification accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If automated machine classification is used to classify text samples, then productivity is improved, but reliability deteriorates due to inaccuracies and false positives

Engineering Contradiction:
Improveclassification throughputVSAvoidclassification accuracy
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent segments the classification task into multiple stages: an automated machine learning classifier handles the majority of text samples to maintain high productivity, while a secondary review process specifically targets samples with low confidence scores or predicted errors. This segmentation allows the system to process large volumes of text efficiently while maintaining high accuracy by applying human review only where needed.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system implements feedback mechanisms where the automated classifier's predictions are evaluated against ground truth labels, and error patterns are used to refine the classification model. Additionally, human reviewer corrections feed back into the training data, continuously improving the classifier's accuracy over time while maintaining high throughput.

Inventive Principle:
Principle #23Feedback

2Reliability

If human review is used to classify text samples, then reliability is improved, but productivity deteriorates due to high costs and slow processing

Engineering Contradiction:
Improveclassification accuracyVSAvoidclassification throughput
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

Human review is segmented to handle only specific subsets of text samples—those with low confidence predictions, ambiguous content, or samples identified as potential errors by the automated system. This selective segmentation maintains high reliability for critical cases while preserving overall productivity by avoiding manual review of all samples.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Instead of applying full human review to all text samples, the system applies partial action by using the automated classifier for the majority of samples and reserving human review for only the necessary portion. This partial application of human expertise maintains acceptable accuracy levels while dramatically improving throughput compared to complete manual review.

Inventive Principle:
Principle #16Partial or excessive action

3Productivity

If feature space is reduced through dimension reduction operations, then productivity is improved by faster processing, but measurement precision may deteriorate

Engineering Contradiction:
Improveprocessing speedVSAvoidfeature representation accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The system extracts and retains only the most informative features from the original feature space through dimension reduction operations. By identifying and removing redundant or less informative features, the system reduces computational complexity and improves processing speed while preserving the essential characteristics needed for accurate text classification.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentUS9342794B2Non-linear classification of text samples
Publication Date: 2016.05.17 BAZAARVOICE INC

AI summary

Non-linear classifiers and dimension reduction techniques may be applied to text classification. Non-linear classifiers such as random forest, Nyström/Fisher, and others, may be used to determine criteria usable to classify text into one of a plurality of categories. Dimension reduction techniques may also be used to reduce feature space size. Machine learning techniques may be used to develop criteria (e.g., trained models) that can be used to automatically classify text. Automatic classification rates may be improved and result in fewer numbers of text samples being unclassifiable or being incorrectly classified. User-generated content may be classified, in some embodiments.