Probability and Natural Language Processing

计算机理解自然语言的主要问题之一是一直存在的歧义。单词有多种含义，句子可以以几种不同的方式进行语法分析，名词短语可以引用文本中的几个可能的对象，等等。通常不可能做出正确的选择，因为其他选择在逻辑上是不可能的。相反，正确的选择只是比其他选择更有可能或更有可能。为了更好地解决这些和相关的问题(如计划识别，根据角色的行动来决定故事角色的计划)，本研究人员将把概率论应用到语言理解中。特别是，所提出的系统将使用来自句子的各种线索以及上下文，来建立概率分布的贝叶斯网络表示。正如这项研究目前的工作一样，概率将被用来指导处理，并最终完全剔除不太可能的替代方案。

One of the major problems in computer understanding of natural language is the ever present ambiguity. Words have multiple meanings, sentences can be grammatically parsed several different ways, noun phrases can refer to several possible objects in the text, etc. Typically it is not possible to make the correct choice because the others are logically impossible. Rather the correct choice is simply more plausible, or probable, than the others. This investigator will integrate the use of probability theory into language comprehension in order to better solve these and related problems (like plan recognition deciding on a story character's plans on the basis of the character's actions). In particular, the proposed system will use the various clues from the sentence as well as the context, to build up a Bayes network representation of the probability distribution. As in current work by this research the probabilities will be used to guide processing and eventually to completely clip less likely alternatives.