arXiv 2606.160742026-06-16cs.CLcs.AI

PVminerLLM2：通过偏好优化改进患者声音的结构化提取

PVminerLLM2: Improving Structured Extraction of Patient Voice via Preference Optimization

Yale School of Medicine（耶鲁大学医学院）；Yale School of Public Health（耶鲁大学公共卫生学院）；Texas State University（德克萨斯州立大学）

Samah Fodeh, Linhai Ma, Ganesh Puthiaraju, Srivani Talakokkul, Afshan Khan, Elyas Irankhah, Sreeraj Ramachandran, Ashley Hagaman, Sarah Lowe, Aimee Roundtree

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

提出PVminerLLM2，通过偏好优化和令牌级门控稳定项、混淆感知偏好对构建等技术，解决监督微调难以处理的细粒度错误，在患者声音结构化提取任务上优于基线模型。

摘要：

动机：患者生成的文本包含关于患者生活经历、社会背景和护理参与的关键信息，但大多是非结构化的，限制了其在以患者为中心的结果研究中的应用。先前的工作引入了PV-Miner基准和PVMinerLLM模型用于结构化提取。然而，仅靠监督微调（SFT）难以处理罕见、细粒度且分布不均的错误，尤其是在令牌关键的结构化输出中。结果：我们提出了PVminerLLM2，一组改进的用于结构化患者声音提取的LLM，它应用偏好优化来解决监督微调无法处理的令牌级错误。我们的方法引入了（i）带有令牌级门控稳定项的偏好目标，防止在偏好优化下绝对令牌似然的退化，以及（ii）混淆感知的偏好对构建，以更好地捕捉低分离度的区分。我们进一步引入了令牌重要性加权和逆频率重加权，以解决令牌不平衡和类别偏斜问题。在多种模型规模下，PVMinerLLM2始终优于强基线，在代码、子代码和跨度上分别获得了高达4.43%、3.50%和1.55%的提升，并且优于使用现有偏好优化方法训练的基线LLM。可用性和实现：PVminerLLM2的补充材料、代码、评估脚本和训练模型公开于：https://github.com/Data-Mining-Lab-Yale/PVminerLLM2

英文摘要：

Motivation: Patient-generated text contains critical information on patients' lived experiences, social context, and care engagement, but remains largely unstructured, limiting its use in patient-centered outcomes research. Prior work introduced the PV-Miner benchmark and PVMinerLLM models for structured extraction. However, supervised fine-tuning (SFT) alone struggles with rare, fine-grained, and unevenly distributed errors, particularly in token-critical structured outputs. Results: We present PVminerLLM2, an improved set of LLMs for structured patient voice extraction that applies preference optimization to address token-critical errors beyond the reach of supervised fine-tuning. Our method introduces (i) a preference objective with token-level gated stabilization term that prevents degradation of absolute token likelihood under preference optimization, and (ii) confusion-aware preference pair construction to better capture low-separation distinctions. We further incorporate token-importance weighting and inverse-frequency reweighing to address token imbalance and class skew. Across multiple model sizes, PVMinerLLM2 consistently outperforms strong baselines, achieving gains of up to 4.43% (Code), 3.50% (Sub-code), and 1.55% (Span), and outperforms baseline LLM trained with existing preference optimization methods. Availability and Implementation: The supplementary material, code, evaluation scripts, and trained models for PVminerLLM2 are publicly available at: https://github.com/Data-Mining-Lab-Yale/PVminerLLM2

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。