Probabilistic Approaches to Learning the Semantics and Syntax of Words

我们的指尖上有数十亿字的在线文本，然而有效处理这些文本的工具——不仅仅是寻找文档，而是真正理解它们——却非常缺乏。一个主要的障碍是单词固有的灵活性。人们不断扩展单词的含义，并以前所未有的方式将它们有效地组合起来。然而，当前的自然语言处理（NLP）系统在很大程度上依赖于静态词汇资源——电子词典和本体——这些资源无法支持理解文本所需的灵活性和适应性。该建议侧重于学习丰富的语义和句法知识的计算方法，使用鲁棒概率表示来捕捉单词的灵活性并适应新用法。

We have billions of words of online text available at our fingertips, yet tools for effectively processing it - going beyond merely finding documents, to truly understanding them - are sorely lacking. A primary obstacle is the inherent flexibility of words. People are continually extending the meaning of words and combining them productively in previously unseen ways. Current natural language processing (NLP) systems, however, rely largely on static lexical resources - electronic dictionaries and ontologies - that are unable to support the needed flexibility and adaptability for understanding text. This proposal focuses on computational methods for learning rich semantic and syntactic knowledge about words, using robust probabilistic representations that capture the flexibility of words and are adaptable to new usages. We approach this problem from two complementary perspectives. First, we ask how it is that very young children learn the complex information about words so effortlessly, and yet such understanding has proven elusive to NLP systems. We develop computational models of child word learning that help us to better understand this uniquely-human ability. Our models play an important role in the scientific study of cognition, by contributing precise, falsifiable theories of human language-learning mechanisms, and also contribute potential algorithms for use in NLP systems. Second, we develop novel techniques for automatically building and extending large-scale lexical resources for NLP. We adapt advanced statistical machine learning techniques to our tasks by incorporating linguistic and cognitive knowledge, enabling us to automatically infer richer information about words that can support more flexible and adaptable NLP tools than is currently possible. We also integrate linguistic and visual information in the processing of multimodal documents to contribute improved algorithms for automatic annotation of images with keywords to support more efficient image search.