A Comparative Study of Generative Pre-trained Transformer-based Models for Chinese Slogan Generation of Crowdfunding

A Comparative Study of Generative Pre-trained Transformer-based Models for Chinese Slogan Generation of Crowdfunding

In recent years, language generation models have made significant progress and garnered extensive attention, aiming to generate diverse sentences across domains. However, effectively conveying deep semantics within constrained word limits and expressive formats remains a challenging endeavor. Therefore, we utilize GPT-2, GPT-3.5, and Bloom to generate slogan. Incorporating product descriptions, we have experimented, using metrics like ROUGE, BLEU, and semantic relevance for model evaluation. Overall, compared to product descriptions, GPT-3 demonstrates the best similarity in terms of vocabulary and meaning. In terms of human evaluation results, Bloom better captures the uniqueness of the slogan, while GPT-3 is more closely related to the description, and its sentences are the most fluent.