Scientific questions serve as the starting point of scientific inquiry， determining the depth， breadth， and impact of research endeavors. However， amidst the exponential growth of global scientific publications， identifying high-value research gaps from the vast volume of literature has become an overwhelming cognitive burden for researchers. Consequently， developing automated methodologies to generate research questions from large-scale literature is of critical importance.

To address this need， this paper proposed the Automatic Generation Method of Scientific Questions（AGMSQ）， a novel framework leveraging Large Language Models（LLMs）. By tailoring the generation process to specific question types， AGMSQ guided LLMs to produce high-quality research questions that were structurally rigorous and deeply grounded in the literature context. The method comprised three core modules： the Scientific Question Classification Module， the Generation Template Design Module， and the LLM Generation Module. First， the Classification Module categorized questions into five types： descriptive， explanatory， methodological， evaluative， and normative. This fine-grained taxonomy enabled the model to capture the distinct logical patterns and semantic requirements inherent to different modes of scientific inquiry， thereby enhancing the precision of generation. Second， the Template Design Module constructed element-generation templates based on the structural principles of each question type. It integrated key element triplets extracted from “Future Work Sentences”（FWS） with domain extension search topics， which were matched to the triplets via semantic distance. Finally， the LLM Generation Module utilized parameter-fine-tuned models—including ChatGPT-4， ChatGPT-3.5， Claude 3 Sonnet， and Gemini Pro—to synthesize research questions based on the combined input elements. Additionally， the study introduced two quantitative indicators—the Utilization Rate of Prompts（URP） and the Occupancy Rate of New Words（ORN）—to evaluate and optimize the generation performance of the LLMs.

The experiments utilize an FWS dataset sourced from the natural language processing domain， specifically targeting the generation of methodological questions. Expert evaluations indicate that the research questions generated by AGMSQ demonstrate favorable performance in terms of clarity， originality， feasibility， and academic value. Notably， among the evaluated models， Claude 3 Sonnet exhibits the superior generation performance. Furthermore， quantitative analysis based on URP and ORN metrics corroborates the expert findings， confirming that the optimized prompts effectively reduce semantic redundancy and increase the efficient utilization of input information. These findings validate the capability of LLMs to generate methodological questions within the natural language processing domain， offering empirical evidence and valuable insights for future exploration across diverse disciplines and question types. Overall， this study offers new insights and tools for automating research topic selection， representing a concrete practice of the “AI for Science” paradigm.