arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2406.04306cs.LGcs.AI
收藏

通过语义多样化语言生成改进不确定性估计

Improving Uncertainty Estimation through Semantically Diverse Language Generation

ELLIS Unit Linz and LIT AI Lab, Institute for Machine Learning, Johannes Kepler University Linz(林茨ELLIS单元和LIT AI实验室,机器学习研究所,林茨约瑟夫·胡纳克大学);NXAI GmbH(NXAI公司)

Lukas Aichberger, Kajetan Schweighofer, Mykyta Ielanskyi, Sepp Hochreiter

AI总结:

针对LLM因预测不确定性导致幻觉的问题,提出语义多样化语言生成(SDLG)方法,通过引导模型生成多样化替代文本来精确量化不确定性,在问答任务中性能优于现有方法且计算效率最高。

摘要:

大型语言模型(LLMs)在生成文本时可能会产生幻觉。这些幻觉使得 LLMs 变得不可信,从而阻碍了其在社会和工业中的应用。当前的 LLMs 以自回归的方式,通过预测并附加文本 token 来生成文本。当 LLM 对将要生成的下一个 token 的语义含义不确定时,它很可能开始产生幻觉。因此,已有观点认为,预测不确定性是产生幻觉的主要原因之一。我们引入了语义多样化语言生成(SDLG)来量化 LLMs 中的预测不确定性。SDLG 引导 LLM 为初始生成的文本生成语义上多样化且可能成立的替代文本。该方法提供了对偶然性语义不确定性的精确度量,能够检测初始文本是否可能为幻觉。在问答任务上的实验表明,SDLG 始终优于现有方法,同时计算效率最高,为 LLMs 中的不确定性估计设定了新标准。

英文摘要:

Large language models (LLMs) can suffer from hallucinations when generating text. These hallucinations impede various applications in society and industry by making LLMs untrustworthy. Current LLMs generate text in an autoregressive fashion by predicting and appending text tokens. When an LLM is uncertain about the semantic meaning of the next tokens to generate, it is likely to start hallucinating. Thus, it has been suggested that predictive uncertainty is one of the main causes of hallucinations. We introduce Semantically Diverse Language Generation (SDLG) to quantify predictive uncertainty in LLMs. SDLG steers the LLM to generate semantically diverse yet likely alternatives for an initially generated text. This approach provides a precise measure of aleatoric semantic uncertainty, detecting whether the initial text is likely to be hallucinated. Experiments on question-answering tasks demonstrate that SDLG consistently outperforms existing methods while being the most computationally efficient, setting a new standard for uncertainty estimation in LLMs.

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。

↑