Supervised Feature Extraction for Text Classification Accuracy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current text classification methods, such as latent semantic analysis and latent Dirichlet allocation, are inflexible and inefficient, struggling to accurately classify documents due to limited keyword coverage and reliance on large training datasets, making them suboptimal for specific topics like medical reports on lung diseases.

Innovation Solution

A method for determining supervised features in text classification using a classifier hyperplane in a multidimensional space, where each word's vector is trained with ground truth labels to generate accurate and flexible classifiers, allowing for iterative refinement and expansion of keywords for improved classification accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If manual keyword listing is used for text classification, then the method is simple and easy to implement, but the keyword coverage is limited and classification accuracy is sub-optimal

Engineering Contradiction:
Improveease of implementationVSAvoidclassification accuracy
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The system automatically generates keywords and features through unsupervised topic modeling and supervised feature determination, eliminating the need for manual keyword listing while improving classification accuracy through data-driven feature extraction

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent replaces the mechanical manual keyword listing process with automated computational methods including topic modeling algorithms and supervised learning, substituting human effort with machine-based feature extraction

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Adaptability or versatility

If unsupervised topic models are used to construct topics from text corpora, then the model can capture polysemy and produce semantically coherent word groups, but the topics are relatively inflexible and may not be of sufficient relevance to a desired topic

Engineering Contradiction:
Improvesemantic coherenceVSAvoidtopic relevance
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The system transitions from static unsupervised topic models to dynamic supervised feature determination, allowing the feature extraction process to adapt to specific classification tasks through training data while maintaining the semantic coherence benefits of unsupervised modeling

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent introduces supervised features as an intermediary layer between unsupervised topic models and final classification, using training data to bridge the gap between general semantic coherence and task-specific relevance

Inventive Principle:
Principle #24Intermediary (Mediator)

3Measurement precision

If the granularity of topics in the model is increased to improve classification accuracy, then the quality of individual topics improves, but the total number of topics increases which is inefficient

Engineering Contradiction:
Improveclassification accuracyVSAvoidnumber of topics
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts only the most relevant features for each class from the vocabulary, rather than creating topics for all possible granularities, thereby achieving high classification accuracy without proportionally increasing the number of topics or computational complexity

Inventive Principle:
Principle #2Taking out (Extraction)

4Measurement precision

If large training datasets are used to train topic models to generate accurate topics, then the classification accuracy improves, but the data requirement is necessarily large which is inefficient

Engineering Contradiction:
Improveclassification accuracyVSAvoidtraining data volume
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The system determines supervised features for each word in the vocabulary rather than requiring large amounts of training data, using a more targeted approach that achieves accurate classification with less data by focusing on feature determination at the word level

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3657354A1Supervised features for text classification
Publication Date: 2020.05.27 SIEMENS HEALTHINEERS AG
  • EP3657354A1 patent drawingFigure 1
  • EP3657354A1 patent drawingFigure 2
  • EP3657354A1 patent drawingFigure 3~4c

AI summary

A method of determining a supervised feature, for use in text classification, for each word of a vocabulary, is disclosed. The vocabulary comprises a plurality of words each associated with a word embedding comprising a vector representing the word in a multidimensional space. The method comprises obtaining training data comprising one or more input words, one or more word embeddings each associated with a respective one of the input words, and one or more ground truth labels, each associated with a respective one of the input words, and each indicating an association of the respective input word with a given class. The method comprises training, based on the training data, a classifier model thereby to generate a classifier hyperplane in the multidimensional space, and determining, based on the classifier hyperplane, for each word of the vocabulary, a supervised feature for use in text classification. An apparatus is also disclosed.