Natural language processing for speech information retrieval and computer-assisted language learning

现有的搜索引擎允许用户从互联网上的文本中检索有用的信息，但不能从诸如广播新闻或其他记录的音频文档中检索有用的信息。用户必须浏览较长的音频记录才能找到相关部分。尽管在商业应用和处理自然语言的研究原型方面都取得了很大进展，但它们在信息检索系统中的性能仍然需要改进。我们提出了符号和统计方法相结合的新方法来处理单词和句子的意义。我们计划将它们应用于几个尚未尝试的任务中。特别是，我们建议将语义相似度方法应用于信息检索、自动语音识别和计算机辅助语言学习中的任务。我们专注于多种语言和跨语言技巧。

Existing search engines allow a user to retrieve useful information from texts on the Internet, but not from audio documents such as broadcast news or other recordings. The user has to browse through long audio recordings to find relevant parts. Despite much progress both in commercial applications and in research prototypes that deal with natural language, their performance of information retrieval systems still needs to be improved. We propose new ways of combining symbolic and statistical methods for dealing with the meaning of words and sentences. We plan to apply them in several tasks where they have not yet been tried. In particular, we propose to apply semantic similarity methods to tasks in Information Retrieval, Automatic Speech Recognition, and Computer Assisted Language Learning. We focus on multiple languages and on cross-language techniques.