LMGAN: Linguistically Informed Semi-Supervised GAN with Multiple Generators.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Semi-supervised learning is crucial for text classification when labeled data is scarce.
• Existing generative adversarial network (GAN) approaches face challenges in generating realistic data distributions for dynamic text data.
• The generator in conventional GANs struggles to adapt to evolving real data distributions, limiting its effectiveness.

Purpose of the Study:

• To propose a novel generative adversarial network (GAN) model, LMGAN, for enhanced semi-supervised text classification.
• To address the limitations of existing GANs in producing accurate fake data distributions for text.
• To improve classification performance using limited labeled data by enriching fake data generation.

Main Methods:

• Developed Linguistically Informed Semi-Supervised GAN with Multiple Generators (LMGAN).
• Utilized Bidirectional Encoder Representations from Transformers (BERT) and the GAN-BERT discriminator.
• Employed multiple generators and leveraged linguistically meaningful hidden layers of BERT to enrich fake data distributions.

Main Results:

• LMGAN demonstrated the ability to generate well-distributed fake data, outperforming previous GAN-based methods.
• The model achieved improved performance in semi-supervised text classification, particularly with extremely limited labeled data (down to 20.0%).
• LMGAN effectively reduced the discrepancy between fake and real data distributions by utilizing BERT's hidden layers.

Conclusions:

• LMGAN offers a robust solution for semi-supervised text classification by improving fake data generation quality.
• The proposed model shows significant advantages when dealing with minimal labeled datasets.
• Leveraging linguistic information from BERT's hidden layers is key to enhancing GAN-based text classification performance.