Automatic generation of highlights for academic paper via prompt-based learning

What the paper says

Purpose Highlights in academic papers serve as condensed summaries of the author’s key work, allowing readers to quickly grasp the paper’s focus. However, many journals do not currently offer highlights for their articles. To address this gap, some scholars have explored using supervised learning methods to extract highlights from academic papers. A significant challenge in this approach is the need for substantial amounts of training data. Design/methodology/approach This study examines the effectiveness of prompt-based learning for generating highlights. We develop task-specific prompt templates, populate them with paper abstracts and use them as input for language models. We employ both locally inferable pre-trained models, such as GPT-2 and T5, and the ChatGPT model accessed via API. Findings By evaluating the model’s performance across three datasets, we find that the ChatGPT model performed comparably to traditional supervised learning methods, even in the absence of training samples. Introducing a small number of training samples further enhanced the model’s performance. We also investigate the impact of prompt template content on model performance, revealing that ChatGPT’s effectiveness on specific tasks is highly contingent on the information embedded in the prompts. Originality/value This study advances the field of automatic highlights generation by pioneering the application of prompt learning. We employ several mainstream pre-trained language models, including the widely used ChatGPT, to facilitate text generation. A key advantage of our method is its ability to generate highlights without the need for training on domain-specific corpora, thereby broadening its applicability.