From text mining to knowledge extraction

语义网向全世界承诺提供大量现成可用的信息。像谷歌这样的公司最近致力于向用户提供不仅仅是文本信息，还有一些语义内容。由于将这些信息手动编码为某种正式的知识表示格式是一项巨大的工作，因此从文本中自动提取信息变得与此相关。提取信息对上下文太敏感，不能应用于Web本身。因此，简洁的域，如企业内部网，可能会证明提供更现实的应用基础。此外，由于对这样的公司库的查询通常是面向目标（或任务）的，与一般的浏览查询相反，查询-应答系统受益于组成这些库的文本的分类可能是有用的。当手动构建时，这些分类也太昂贵而无法实现。因此，在这种情况下，文件是（部分）分类方面的确切主题，这项研究计划提出：1）研究将各种特征选择算法集成到知识提取过程中，基于我们以前在特征选择、文本分类和知识提取方面的一些工作，2）研究了本体自动构建工具对知识抽取过程本身的影响; 3）提出了一些概率评估算法，用于评估抽取的知识结构与查全率和查准率的相关性; 4）提出了一种知识索引模式，以加快从相关文档中过滤掉不相关文档所需的查找时间。

The Semantic Web promises a wealth of ready-to-use information to the world. Companies like Google recently engaged in such an endeavour to provide more than just textual information to its users, but some semantic content as well. Since it is a colossal work to manually encode that information into some formal knowledge representation format, automatic information extraction from text becomes therefore relevant for that purpose. Extracting information is too context sensitive to be applicable on the Web per se. Therefore, concise domains, like corporate Intranets, may prove to offer more realistic application grounds. Also, since queries to such corporate libraries are usually goal (or task) oriented, as opposed to general browsing queries, it may be useful for the query-answering system to benefit from a classification of the texts that compose these libraries. When manually built, these classifications are also too costly to be achieved. Consequently, in that context where documents are (partly) classified with regard to precise subject matters, this research program proposes: 1) to study the integration of various feature selection algorithms into the knowledge extraction process, based on some of our previous work in feature selection, text classification and knowledge extraction, 2) to study the impact of automatic ontology construction tools on the knowledge extraction process itself, 3) to propose some probabilistic assessment algorithm for the relevance of the extracted knowledge structures pertaining to recall and precision, and 4) to propose some knowledge indexing schema that would speed up the look up time needed to filter out the irrelevant documents from the relevant ones.