ITR/SY(CISE) Learning Syntactic/Semantic Information for Parsing

ITR/SY(CISE) 学习用于解析的句法/语义信息

• 批准号: 0112435
• 负责人: Eugene Charniak
• 金额: $ 44.94万
• 依托单位: Brown University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2001
• 资助国家: 美国
• 起止时间: 2001-08-15 至 2006-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0112435&HistoricalAwards=false
• 关键词: ITR; SY; CISE; Learning; Syntactic

This research concerns the unsupervised learning of structural information about English that is not present in current tree-banks (specifically the various Penn tree-banks). That is, one wants a machine to learn this information without having to create a corpus in which the information is annotated. The structural information to be learned often falls at the boundary between syntax and semantics; for example, does the fact that the "New York Stock Exchange" has as part of the name the location "New York" fall under syntax or semantics? What about the similarity between the expressions "[to] market useless items" and "the market for useless items"? The intention is to learn this kind of information in a form that current statistical parsers can use so that they can output more finely structured parses. But this is not meant to suggest that parsing is the sole use for this sort of information. More and more systems for automatically extracting information from free text use coreference detection and "named-entity recognition" (e.g., recognizing that "New York" is a location, but "New York Stock Exchange" is an organization). There is evidence to suggest that both coreference and named-entity recognition can be improved with the finer level of analysis to be made possible by this research. Or again, "language models" (programs that assign a probability to strings in a language) are standard parts of all current speech-recognition systems; there is evidence that suggests that finer grained syntactic analysis can improve current language models. Thus, this research will enable a wide variety of systems to make better use of language input and so make these systems more accessible to a diverse user pool.

这项研究涉及当前树库（特别是各种宾夕法尼亚树库）中不存在的英语结构信息的无监督学习。 也就是说，人们希望机器能够学习这些信息，而不必创建一个对信息进行注释的语料库。 所要学习的结构信息往往处于语法和语义的边界； 例如，“纽约证券交易所”名称中包含“纽约”这一事实是否属于语法或语义？ “营销无用物品”和“无用物品市场”这两个表述有何相似之处？ 目的是以当前统计解析器可以使用的形式学习此类信息，以便它们可以输出更精细结构的解析。 但这并不意味着解析是此类信息的唯一用途。 越来越多的自动从自由文本中提取信息的系统使用共指检测和“命名实体识别”（例如，识别“纽约”是一个位置，但“纽约证券交易所”是一个组织）。 有证据表明，通过本研究实现更精细的分析水平，共指和命名实体识别都可以得到改善。 或者，“语言模型”（为语言中的字符串分配概率的程序）是所有当前语音识别系统的标准部分；有证据表明，更细粒度的句法分析可以改进当前的语言模型。 因此，这项研究将使各种系统能够更好地利用语言输入，从而使这些系统更容易被不同的用户群访问。