关键词: 知识抽取, 多模态知识, AI4S, 科技文献, 知识单元, 大语言模型

Abstract:

[Purpose/Significance] The deep integration of the latest technological revolution and industrial transformation has created an urgent demand for high-quality multimodal corpora for artificial intelligence-driven scientific discovery (AI4S) and large language models. Traditional coarse-grained knowledge organization methods based on documents have become insufficient for deep knowledge services. This study aims to construct a toolchain for extracting multimodal and multigranular knowledge units from scientific and technological literature, enabling the systematic mining of structured knowledge units from massive literature and enhancing the depth and efficiency of knowledge services. [Method/Process] This study conducted a systematic review of mainstream knowledge extraction tools, both domestical and international, and performed a comparative analysis and screening on dimensions such as technical principles, functional characteristics, application advantages, existing limitations, and processing efficiency. An application demand system was constructed from four levels: identification of research subjects, context tracing, content analysis, and evidence localization. Taking the field of rice breeding as an empirical scenario, a knowledge representation model for multimodal information was constructed based on the physical organizational logic of literature. Documents were divided into four major categories of 22 knowledge units: basic information subjects, structural support, material systems, and academic descriptions. The boundaries between knowledge units are clear, and there are abundant associative relationships. Integrating the extraction needs of various types of scientific and technological literature knowledge units with tool research results, a pipeline-style extraction process framework for multimodal and multigranular knowledge units has been designed. This framework implemented a pipeline-style processing framework for the entire process of document acquisition, physical structure analysis, logical structure reconstruction, multimodal content extraction, and knowledge unit fusion and storage, constructing a cascading processing toolchain from PDF original documents to semi-structured data, and then to structured knowledge. To address three major issues: insufficient accuracy of basic information, chaotic structure of academic statements, and missing information in supporting materials, GROBID domain-adaptive retraining, XML and Markdown fusion parsing, and DeepSeek large model hierarchical extraction instructions were optimized and integrated into a full-chain toolchain. [Results/Conclusions] Preliminary experiments on the toolchain have achieved good extraction of multimodal and multigranular data. In optimization experiments, overall micro-average F1 score of the header model increased by nearly 3 percentage points, significantly enhancing the model's balance and generalization ability when processing documents in diverse formats. The problems of chaotic distribution and weakened structure of academic statement information were successfully solved, achieving robust structured extraction of more than ten types of statement information such as acknowledgements, conflicts of interest, and data availability. The introduction of the large language model DeepSeek enabled deep mining and association of chart titles, formal citation sentences, and related discussion sentences in literature. The model achieved an F1 score greater than 0.99 for extracting chart titles and greater than 0.93 for recognizing formal citation sentences. Verification through the SciWatch platform demonstrates the extraction, presentation, knowledge association, and contextual coherence of charts, supporting deep literature understanding and cross-validation. The multimodal knowledge extraction toolchain for scientific literature constructed in this paper has been able to efficiently and accurately complete the automated extraction and structured application of various knowledge units in scientific literature, covering a complete toolchain, including preprocessing, multimodal information recognition, relation extraction, knowledge fusion, and storage. The research results provide a scalable solution and practical reference for the evolution of domain knowledge mining and knowledge service technology.

Key words: knowledge extraction, multimodal knowledge, AI4S, scientific literature, knowledge unit, large language model