Understanding priors in language generation models

了解语言生成模型中的先验

• 批准号: 2878914
• 金额: --
• 依托单位: University of Edinburgh
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2023
• 资助国家: 英国
• 起止时间: 2023 至 无数据
• 项目状态: 未结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-2878914
• 关键词: Understanding; priors; language; generation; models

The objective of my research is what makes it possible for algorithms to generate text, both by understanding the underlying grammar of language and the semantic relationship between words. The focus is on the exploration of different neural network architectures.One case study is that of hallucinations in machine trainslation.In machine translation, hallucinations are generations that are "completely semantically incorrect and also grammatically viable". It is important to understand in depth this phenomenon, not only because hallucinations can harm the confidence of users of commercial systems. Hallucinations in NMT are surprisingly well-defined cases of highly pathological outputs and any insight on how or why they happen will contribute towards answering the question why do transformers generate language (that is coherent to humans)?

我的研究目标是，如何使算法能够通过理解语言的基本语法和单词之间的语义关系来生成文本。重点是探索不同的神经网络体系结构。一个案例是机器训练中的幻觉。在机器翻译中，幻觉是“在语义上完全不正确，在语法上也是可行的”的几代人。深入了解这一现象很重要，不仅因为幻觉会损害商业系统用户的信心。NMT中的幻觉是非常明确的高度病理输出的案例，任何关于它们是如何或为什么发生的洞察将有助于回答为什么变形金刚产生语言(这与人类是一致的)这个问题。