Automatic acquisition and use of lexical semantic information

最近的自然语言处理(NLP)应用--如网络搜索和信息提取--已经取得了令人印象深刻的结果，采用了非常肤浅地处理单词语义的方法。然而，我们经常感到沮丧，因为这些系统并没有真正理解我们的查询或正在处理的文档。构建能够在更深层次自动确定电子文本的含义的系统需要更复杂的关于单词的语言知识，但在广泛的范围内手动编码这种知识是不可行的。这里提出的研究有两个主要目标。第一个目标是开发自动获取词汇知识的方法，为了更好地捕捉词中编码的复杂信息以及它们之间的相互关系。我们采用了一种知识知晓的方法，这种方法的特点是将语言学和心理学的观察融入到新的统计语料库分析技术中，使我们能够缓解自然语言处理中的知识获取瓶颈。研究的第二个目标是将相同的知识知晓观点应用于词汇资源的处理，以确保计算词汇和词/概念本体可以更有效地使用，并且可以以更少的人工工作。从长远来看，从文本语料库中自动获取更深层次的词汇信息的研究，以及更有效地利用产生的计算词汇的技术，即使在短期内，这里提出的研究也将导致改进的自然语言处理应用，例如基于实际含义的网络搜索而不是简单的关键字匹配，以及可以应用于狭义范围内的被提取信息的限定类型的信息提取。

Recent natural language processing (NLP) applications---such as web searching and information extraction---have achieved impressive results by adopting `knowledge-lean' approaches that treat word semantics very superficially.  Nonetheless, we are frequently left with a feeling of frustration that these systems do not truly understand our queries or the documents being processed.  Building systems that can automatically determine the meaning of electronic text at a deeper level requires more sophisticated linguistic knowledge about words, yet hand coding such knowledge is infeasible on a broad scale.The research proposed here has two primary goals.  The first goal is to develop methods for the automatic acquisition of lexical knowledge, in order to better capture the complex information encoded in words, as well as their interrelationships.  We adopt a `knowledge-informed' approach that is distinguished by the incorporation of linguistic and psychological observations into novel techniques for statistical corpus analysis, enabling us to alleviate the knowledge-acquisition bottleneck in NLP.  The second goal of the research is to apply the same knowledge-informed perspective within the processing of lexical resources, to ensure that computational lexicons and word/concept ontologies can be used more effectively and with less manual effort.In the long run, research on automatic acquisition of deeper lexical information from text corpora, and techniques for more effectively using the resulting computational lexicons, will revolutionize the use of computers.  Even in the short term, the research proposed here will lead to improved NLP applications, such as web searching based on actual meaning rather than simple keyword matching, and information extraction that can be applied beyond circumscribed types of extracted information within a narrow domain.