Harnessing the Power of Pretrained Models for Scientific Paper Title Generation

In the rapidly evolving landscape of scientific literature, crafting practical and informative titles for research papers plays a pivotal role in enhancing discoverability and reader engagement. This paper delves into the realm of title generation from abstracts. In this work, we explore the practical application of BART (Bidirectional Aligner for Representation Transformers) and T5 (Text-to-Text Transfer Transformer), two state-of-the-art pretrained language models, in the context of title generation for research papers. We present a comparative analysis of the capabilities of these models to generate informative and compelling titles. To evaluate our models, we use metrics: F1 score and Rouge scores, alongside human assessments by expert reviewers. Our results show that machine-generated metrics favored T5 for its close resemblance to the original title pattern and text-to-text approach. However, human evaluation favored BART for its superior understanding of the paper’s core content, facilitated by bidirectional pre-training and denoising capabilities.