Spreading the Word: The Theory of Distributed Representations in Speech and Natural Language Processing

传播信息：语音和自然语言处理中的分布式表示理论

• 批准号: RGPIN-2015-04069
• 负责人: Penn, Gerald
• 金额: $ 3.13万
• 依托单位: University of Toronto
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2016
• 资助国家: 加拿大
• 起止时间: 2016-01-01 至 2017-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=612758
• 关键词: Spreading; Word; Theory; Distributed; Representations

Motivated by recent progress in the use of neural networks in areas such as speech recognition and machine translation, this grant will fund research into the application of neural networks to a wide variety of other tasks in language technology. Although much of the popular excitement over neural networks arises from their alleged correspondence to human cognitive mechanisms, what excites engineers and computer scientists about them is their utility as very compact sequence classifiers - tools that can accurately label long sequences of data, such as speech sounds or words of text, given only examples to train on. Undertaken by one of the research labs that pioneered the use of neural networks in the area of speech recognition, the projects funded by this grant will consider not merely the transcription of speech into text, but the understanding of what was said or written. This emphasis on meaning will enable a new generation of language technology applications that can, for example, summarize large newspaper articles into their most relevant 10 or 20 percent. A key challenge to doing this is determining how individual representations of words combine to form representations of phrases and whole sentences. There is a branch of applied mathematics that studies this sort of combination, called mathematical linguistics, which will play a crucial role in this research. We will study the use of neural networks in mimicking those combinations.

由于在语音识别和机器翻译等领域使用神经网络的最新进展，该赠款将资助对神经网络应用于语言技术的其他多种任务的应用。 尽管对神经网络的许多流行兴奋源于它们与人类认知机制的相应，但激发工程师和计算机科学家的兴奋是它们的实用性，它们是非常紧凑的序列分类器 - 可以准确地标记出长长的数据序列，例如语音声音或文本词，只有文本词，仅给出了训练的示例。 由一项研究实验室进行的，该实验室率先在语音识别领域中使用神经网络，该赠款资助的项目不仅会考虑将语音转录到文本中，而且还会考虑对所说或写作的理解。对意义的这种强调将使新一代的语言技术应用程序可以将大量的报纸文章汇总到其最相关的10％或20％中。做到这一点的关键挑战是确定单词的单个表示如何结合形成短语和整个句子的表示。有一个应用数学分支研究了这种组合，称为数学语言学，这将在这项研究中起着至关重要的作用。我们将研究神经网络在模仿这些组合中的使用。