利用大型语言模型（LLM）在文本分类任务中的随机性：在法律事务中定位特权文件
Exploiting the Randomness of Large Language Models (LLM) in Text Classification Tasks: Locating Privileged Documents in Legal Matters
AI总结：
本文研究了在法律事务中利用LLM随机性提高特权文件识别准确性的方法，并探讨了其在合规流程中的应用价值。
摘要：
在法律事务中，文本分类模型最常用于在大规模数据集中筛选出符合特定预选标准（如与特定主题相关的文件）的文档，例如法律特权通信和律师指派文件。在此背景下，大型语言模型已表现出强大的性能。本文通过实证研究探讨了在律师-客户特权文件检测中，随机性在基于LLM的分类中的作用，重点研究四个关键维度：（1）LLM在识别法律特权文件的有效性，（2）随机性控制参数对分类输出的影响，（3）其对整体分类性能的影响，（4）一种利用随机性来提高准确性的方法。实验结果表明，LLM能够有效识别特权文件，随机性控制参数对分类性能影响较小，而且我们开发的利用随机性的方法对提高准确性有显著影响。值得注意的是，这种利用随机性的方法在整合到企业制裁合规流程中时，也能够增强企业对LLM输出的信心。随着越来越多的组织依赖LLM来增强合规流程，减少输出的变异性有助于建立内部和监管层面对LLM衍生制裁筛查决策的信心。
英文摘要：
In legal matters, text classification models are most often used to filter through large datasets in search of documents that meet certain pre-selected criteria like relevance to a certain subject matter, such as legally privileged communications and attorney-directed documents. In this context, large language models have demonstrated strong performance. This paper presents an empirical study investigating the role of randomness in LLM-based classification for attorney-client privileged document detection, focusing on four key dimensions: (1) the effectiveness of LLMs in identifying legally privileged documents, (2) the influence of randomness control parameters on classification outputs, (3) their impact on overall classification performance, and (4) a methodology for leveraging randomness to enhance accuracy. Experimental results showed that LLMs can identify privileged documents effectively, randomness control parameters have minimal impact on classification performance, and importantly, our developed methodology for leveraging randomness can have a significant impact on improving accuracy. Notably, this methodology that leverages randomness could also enhance a corporation's confidence in an LLM's output when incorporated into its sanctions-compliance processes. As organizations increasingly rely on LLMs to augment compliance workflows, reducing output variability helps build internal and regulatory confidence in LLM-derived sanctions-screening decisions.
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。
