Research of text paraphrase generation based on self-contrastive learning.

Area of Science:

• Natural Language Generation
• Computational Linguistics
• Artificial Intelligence

Background:

• Paraphrase generation is crucial for Natural Language Generation (NLG) tasks like machine translation and dialogue systems.
• Current methods struggle to produce paraphrases that are both high-quality and diverse.
• This limitation hinders the effectiveness of NLG applications.

Purpose of the Study:

• To enhance the quality and diversity of text paraphrase generation.
• To introduce novel self-contrastive learning models for improved paraphrase generation.
• To address the limitations of existing paraphrase generation techniques.

Main Methods:

• Developed Contrastive Generative Adversarial Network (ContraGAN) for supervised learning.
• Introduced Contrastive Model with Metrics (ContraMetrics) for unsupervised learning.
• ContraGAN uses a discriminator for quality refinement; ContraMetrics employs multi-metric filtering and keyword prompts for diversity.

Main Results:

• Both models demonstrated significant improvements over state-of-the-art methods on benchmark datasets.
• ContraGAN improved semantic fidelity (BERTScore +0.46) and fluency (perplexity -1.57).
• ContraMetrics enhanced diversity and lexical richness (iBLEU +0.37, P-BLEU +3.34).

Conclusions:

• The proposed self-contrastive learning models effectively address key challenges in paraphrase generation.
• ContraGAN and ContraMetrics offer practical and improved solutions for NLG applications.
• The models contribute to advancing the state-of-the-art in diverse and high-quality text generation.