【中国新闻网】服务全球科研社区 中国团队推出新一代科学文献深度解析工具

Core Insights - The "AI+Science" research team from the Chinese Academy of Sciences has launched a new scientific literature parsing tool called "Panshi: Scientific Literature Parser," aimed at providing a truly intelligent parsing engine that understands scientific content for global researchers [1][2]. Group 1: Product Features - The Panshi Scientific Literature Parser 1.0 has been officially open-sourced and integrated as a core component of the ScienceOne model, serving the global research community [1]. - The tool achieves significant advancements in multi-modal parsing of scientific content, including formulas, text, and charts, through a dedicated algorithm training paradigm tailored for scientific literature [2]. - The parser's output includes high-precision text and formula recognition results, supporting various structured formats for seamless integration with downstream applications such as knowledge extraction and intelligent Q&A [3]. Group 2: Technical Innovations - The development of the parser is based on three technical pillars: comprehensive scientific data construction, multi-modal supervised fine-tuning strategies, and reinforcement learning optimization focused on scientific literature semantics [2]. - A systematic collection of training data covering three typical scientific writing forms—handwritten, digital typesetting, and scanned paper—provides a robust foundation for the model to understand the complexities of scientific expression [2]. - The model training employs a two-stage optimization strategy, initially using multi-modal supervised fine-tuning to grasp the joint representation of heterogeneous elements, followed by a gradient reinforcement learning strategy to enhance semantic understanding [2]. Group 3: Performance Evaluation - Systematic evaluations on multiple scientific literature datasets indicate that the Panshi Scientific Literature Parser demonstrates internationally leading performance in tasks such as chapter-level parsing and specialized formula recognition [3].