Classification techniques with minimal labelling effort and application to medical reports

Area of Science:

• Natural Language Processing
• Machine Learning
• Information Retrieval

Background:

• Text document classification is crucial for organizing large datasets.
• Traditional supervised methods require extensive labeled data, limiting their real-world applicability.
• Partially Supervised Classification (PSC) offers a promising alternative to reduce labeling costs.

Purpose of the Study:

• To evaluate the effectiveness and efficiency of Partially Supervised Classification (PSC) for real-world text document classification.
• To compare different methods within the PSC framework, specifically for medical documents.
• To investigate the impact of feature selection on PSC performance.

Main Methods:

• Employed a two-step Partially Supervised Classification (PSC) strategy to minimize manual labeling.
• Evaluated various methods for each step of the PSC process.
• Conducted experiments using real-world medical documents, comparing Expectation-Maximization (EM) with Support Vector Machines (SVM).
• Assessed the influence of careful feature subset selection on classification accuracy.

Main Results:

• The Expectation-Maximization (EM) algorithm for building classifiers demonstrated superior performance compared to Support Vector Machines (SVM).
• Careful selection of document feature subsets significantly enhanced classification performance.
• The PSC approach proved effective and efficient for classifying real-world medical documents.

Conclusions:

• Partially Supervised Classification (PSC) is a viable and efficient strategy for text document classification, particularly in domains with limited labeled data.
• EM-based classifiers are more effective than SVMs within the evaluated PSC framework for medical text.
• Feature selection is a critical component for optimizing PSC performance in practical applications.