Data augmentation in natural language processing: a novel text generation approach for long and short text...

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning (ML)
• Data Science

Background:

• Training data development is crucial for ML classifier performance.
• Data augmentation methods artificially create training data to improve classifiers.
• NLP faces challenges in creating universal text transformation rules for new linguistic patterns.

Purpose of the Study:

• To present and evaluate a novel text generation method for enhancing ML classifiers.
• To assess the method's effectiveness on both short and long text classification tasks.
• To analyze performance gains, particularly in low-data regimes and real-world applications.

Main Methods:

• Development and application of a new text generation technique for data augmentation.
• Evaluation across 11 diverse datasets, including constructed low-data regimes and real-world tasks.
• Comparative analysis against baseline (no augmentation) and an existing augmentation technique.

Main Results:

• Promising performance improvements observed for both short and long text tasks.
• Significant accuracy gains (up to 15.53% and 3.56%) in constructed low-data regimes.
• Substantial improvements in real-world low-data tasks, with up to +4.84 F1-score gains.

Conclusions:

• The proposed text generation method effectively enhances ML classifier performance, especially in low-data scenarios.
• The method demonstrates applicability across various datasets, though its suitability varies.
• Understanding dataset characteristics is key for successful implementation and maximizing performance gains.