Chinese Academic Text Title Generation Based on Open Source Large Language Models——Taking the Field of Humanities and Social Sciences as an Example
[Purpose/significance]As a compressed representation and the essence of the main idea of a dissertation,the title plays an important role in searching and citation.Taking the task of academic text title generation in the field of humanities and social sciences as an example,it provides a reference for the application of large language models in academic text mining.[Method/process]From an empirical perspective,we explore the effectiveness capability of the current open-source Chinese large language model Qwen-7B in the task of academic text title generation,and the feasibility of injecting the knowledge of academic text data into the open-source base large language model in the field of humanities and social sciences.Vocabulary-level recall and accuracy scores are performed using ROUGE and BLUE metrics,while utterance fluency and semantic relevance scores are performed using the ChatGPT intelligent dialog system.[Result/conclusion]It is found that injecting academic text knowledge in Chinese humanities and social sciences into the Qwen-7B base model does not effectively improve the model's ability in the title generation task,and the feature and semantic learn-ing ability of the open-source base large model Qwen-7B on Chinese needs to be further enhanced;the LLaMA2-7B model outper-forms the Qwen-7B model in the generation of Chinese academic text titles model.[Innovation/limitation]Based on the Qwen-7B model and academic full text data in the field of humanities and social sciences,the current mainstream open-source large language model in China is demonstrated to have the ability to be applied in the generation of academic text headings and the application paths,which provides theoretical and practical references for the academic full text mining and organization.The control models and training methods used in this paper are relatively homogeneous due to resource constraints,and need to be further extended to fully explore the boundaries of large language models in academic full text knowledge mining and organization.
natural language processingautomatic title generationacademic textlarge language modelsChatGPT
NETL
NSTL
万方数据
