Title: Leveraging Language Models for Automated Research Paper Title Generation

Abstract:

A research paper’s title serves as a concise summary of its core concepts and, at times, its key findings. Because crafting an effective title can be difficult for authors, automated generation tools offer valuable support. This study introduces a method for deriving paper titles from abstracts by utilizing open-weight pre-trained large language models. To facilitate this research, we employed the CSPubSum and LREC-COLING-2024 datasets, alongside a newly curated dataset named SpringerSSAT, which was compiled from four social science journals published by Springer. Furthermore, we assessed the performance of GPT-3.5-turbo operating in a zero-shot configuration. The models were evaluated using a comprehensive suite of metrics, including ROUGE, METEOR, MoverScore, BERTScore, and SciBERTScore. Our experimental results indicate that the fine-tuned PEGASUS-large model surpasses other contenders, such as the zero-shot GPT-3.5-turbo and the fine-tuned LLaMA-3-8B, across the majority of these evaluation measures. Additionally, we found that ChatGPT is capable of producing inventive titles for academic papers. In conclusion, the study suggests that titles generated by artificial intelligence are generally suitable and dependable.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC