arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2402.10978cs.LGcs.AIcs.CL
收藏

具有保形事实性保证的语言模型

Language Models with Conformal Factuality Guarantees

Stanford University(斯坦福大学)

Christopher Mohri, Tatsunori Hashimoto

AI总结:

本文提出保形事实性框架,通过连接语言建模与保形预测将输出正确性转化为不确定性量化问题,利用回退算法为任意黑盒语言模型提供高概率正确性保证,并在闭卷问答与推理任务中实现了80-90%的正确性保证且保留了大部分原始输出。

摘要:

保证语言模型(LM)输出的正确性和事实性是一个重大的开放问题。在本研究中,我们提出了保形事实性(conformal factuality),这是一个通过连接语言建模和保形预测来为LM提供高概率正确性保证的框架。我们观察到,LM输出的正确性等价于一个不确定性量化问题,其中不确定性集合被定义为LM输出的蕴含集合。利用这一联系,我们证明了语言模型中的保形预测对应于一种回退算法,该算法通过逐步降低LM输出的具体性(并扩展相关的不确定性集合)来提供高概率的正确性保证。这种方法适用于任何黑盒LM,并且只需很少的人工标注样本。我们在闭卷问答(FActScore, NaturalQuestions)和推理任务(MATH)上对我们的方法进行了评估,结果表明,我们的方法能够提供80-90%的正确性保证,同时保留LM原始输出的大部分内容。

英文摘要:

Guaranteeing the correctness and factuality of language model (LM) outputs is a major open problem. In this work, we propose conformal factuality, a framework that can ensure high probability correctness guarantees for LMs by connecting language modeling and conformal prediction. We observe that the correctness of an LM output is equivalent to an uncertainty quantification problem, where the uncertainty sets are defined as the entailment set of an LM's output. Using this connection, we show that conformal prediction in language models corresponds to a back-off algorithm that provides high probability correctness guarantees by progressively making LM outputs less specific (and expanding the associated uncertainty sets). This approach applies to any black-box LM and requires very few human-annotated samples. Evaluations of our approach on closed book QA (FActScore, NaturalQuestions) and reasoning tasks (MATH) show that our approach can provide 80-90% correctness guarantees while retaining the majority of the LM's original output.

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。

↑