Towards Improved Scientific Knowledge Proliferation: Leveraging Large Language Models on the Traditional Scientific Writing Workflow

Authors:

Abstract
Technological advances in Natural Language Processing have brought forth language models capable of advanced response delivery. For humans, inputting natural language to a software system, and getting natural language as a response, is intuitive to learning and development. Scientific papers are traditionally written manually by human researchers, but with the advent of mainstream Large Language Models, e.g., OpenAI’s ChatGPT, it is of increasing concern to scientists and academics that content in scientific papers may be generated by Artificial Intelligence (AI). Wishing to stop this is a losing attitude, as large-scale generative AI only becomes more powerful and accessible as time progresses. Taking the more tenable position of cautious adaptation, this paper argues that there exists a taxonomy, as yet implicit, in the structure of scientific papers, and that language models can be used by scientific researchers to bootstrap scientific writing. Furthermore, AI can augment their own writing workflow to traverse the academic publishing pipeline with greater efficiency. Despite the shortcomings of language models, e.g., hallucination, when prompted appropriately with sufficient constraints and input data, language models are extremely accurate and efficient content providers. In this work, the canonical scientific paper is broken down into its taxonomy of parts, where it is then considered how each part can benefit from language models, e.g., in generating abstracts and keywords, reformatting sections, theorizing titles, etc. A theoretical system for the implementation of the proposed idea using GPT-4 is provided. Finally, a call for consensus among the academic and scientific communities regarding the use of language models in the scientific writing workflow is established.

This publication has 17 references indexed in Scilit: