Analogical Learning for Natural Language Processing

自然语言处理的类比学习

• 批准号: 249630-2012
• 负责人: Langlais, Philippe
• 金额: $ 1.6万
• 依托单位: Université de Montréal
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2015
• 资助国家: 加拿大
• 起止时间: 2015-01-01 至 2016-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=570668
• 关键词: Analogical; Learning; Natural; Language; Processing

A proportional analogy is a relation between 4 entities noted [x:y::z:t] which reads ``x is to y as z is to t''. Verbal analogies (e.g. [student: class : juror: jury]) and geometrical ones are used in many IQ tests, such as SAT college entrance tests. The ability to establish such relations is often recognized to lie at the core of human cognition. Analogical proportions have also been studied extensively in Artificial Intelligence. Formal proportional analogies are those involving a relation at a formal level, such as [reader : unreadable :: doer : undoable] or [he loves his books : she does not love her book :: he likes his teachers : she does not like her teacher]. Formal proportional analogies have a long history in linguistics, in which the work of de Saussure in the early 20th century was highly influential. Analogical learning realizes the automatic mapping of an input space representation (e.g. sequence of words) into an output space representation (e.g. parse trees) without requiring to learn an alignment between the two levels of representation. This is done thanks to the multi-dimensional paradigmatic views analogies capture in the data (e.g. plural/singular, active/passive, etc.). Furthermore, this mapping is learnt without decomposing the training data (e.g. sentences) into smaller parts (e.g. words), making formal analogy good at capturing long dependencies in the data. Those properties are departing from most approaches in NLP (aka classification). In this research proposal, we will put formal analogy at test in a number of key NLP tasks, among which Machine Translation (MT) and syntactical parsing.

比例类比是4个实体之间的关系，记为[x：y：：z：t]，读作“x对y，z对t”。语言类比（如[学生：班级：陪审员：陪审团]）和几何类比被用于许多智商测试，如SAT大学入学考试。建立这种关系的能力通常被认为是人类认知的核心。 类比比例在人工智能中也得到了广泛的研究。形式比例类比是那些涉及形式层面上的关系的类比，例如[读者：不可读：：行为者：不可做]或[他喜欢他的书：她不喜欢她的书：：他喜欢他的老师：她不喜欢她的老师]。形式比例类比在语言学中有着悠久的历史，其中世纪初的索绪尔在这方面的研究具有重要影响。 类比学习实现了输入空间表示（例如单词序列）到输出空间表示（例如解析树）的自动映射，而不需要学习两级表示之间的对齐。这要归功于数据中捕获的多维范式视图类比（例如复数/单数，主动/被动等）。此外，这种映射是在不将训练数据（例如句子）分解为更小的部分（例如单词）的情况下学习的，这使得形式类比善于捕捉数据中的长依赖关系。这些属性背离了NLP（又名分类）中的大多数方法。 在这项研究计划中，我们将在一些关键的NLP任务中测试形式类比，其中包括机器翻译（MT）和句法分析。