SBIR Phase I: Geographic Information Retrieval for Arabic

MetaCarta的SBIR第一阶段研究项目提出引入一种新的标注技术，即并行引导，以利用现有数据集创建高质量的地名提取和解析训练材料。能够处理阿拉伯语的信息检索（IR）系统已经存在，但是没有执行地理信息检索（GIR）。MetaCarta用户的经验表明，几乎不可能改造标准的基于关键字的红外系统来执行高水平的GIR，因此实现阿拉伯语GIR功能的唯一方法是从GIR系统开始。高质量英语GIR系统的可用性使得通过创新的并行引导技术来解决机器学习项目的最大瓶颈，即缺乏人工真实的训练数据。许多消歧，以及通常从文本中提取语义内容，仍然是由基于规则的系统执行的，这些系统总结了一个领域的专家知识。相比之下，MetaCarta采用了结合隐马尔可夫和最大熵方法的机器学习技术。对于阿拉伯语，我们建议将基于规则的组件限制为形态学分析，后期阶段，特别是地名的提取和消歧将由经过真实阿拉伯语文本训练的系统执行。虽然现在有大量的普通（不真实的）阿拉伯文本，特别是由语言数据联盟（LDC）制作的阿拉伯语Gigaword语料库，但标记材料的数量要少得多，而且提取和消除歧义所需的详细真值需要手工注释来创建，这是非常费力的。MetaCarta将使用LDC 2004T17和T18平行语料库作为输入，通过现有的MetaCarta系统运行英文部分以产生深入的地名注释，并将该注释投影到阿拉伯语部分。这项技术对有兴趣将GIR扩展到阿拉伯语文档的客户具有广泛的吸引力。有代表性的客户对限于狭窄地理范围的活动非常感兴趣，而且许多提供关于具有关键战略重要性的中东地区的信息的文件只有阿拉伯文。部署阿拉伯语GIR还将使分析人员能够更迅速地集中注意有关文件。

This SBIR Phase I research project by MetaCarta proposes to introduce a novel annotation technique, parallel bootstrapping, to take advantage of the existing data sets in creating high quality training material for toponym extraction and resolution. Information Retrieval (IR) systems that can deal with Arabic already exist, but perform no Geographic Information Retrieval (GIR). As the experience of MetaCarta's users shows, it is practically impossible to retrofit standard keyword-based IR systems to perform GIR at a high level, so the only way to achieve Arabic GIR capability is to start with a GIR system. The availability of a high quality English GIR system makes it possible to address the greatest bottleneck of machine learning projects, the lack of manually truthed training data, by an innovative parallel bootstrap technique. Much of disambiguation, and in general, the extraction of semantic content from text, is still performed by rule-based systems that summarize expert knowledge of a domain. In contrast, MetaCarta employs machine-learning techniques that combine Hidden Markov and Maximum Entropy methods. For Arabic, we propose to restrict the rule-based component to morphological analysis, with later stages, in particular the extraction and disambiguation of toponyms to be performed by systems trained on truthed Arabic text. While plain (untruthed) Arabic text is now available in large quantities, see in particular the Arabic Gigaword corpus produced by the Linguistic Data Consortium (LDC), the amount of tagged material is considerably less, and the detail truth values required for toponym extraction and disambiguation are extremely labor-intensive to create by manual annotation. MetaCarta will use as input the LDC 2004T17 and T18 parallel corpora, running the English side through the existing MetaCarta system to produce the in-depth toponym annotation, and projecting back this annotation on the Arabic side.This technology has broad appeal to customers that have an interest in extending GIR to Arabic documents. Representative customers are highly interested in activities restricted to narrow geographic confines, and many of the documents providing information about Middle Eastern areas of key strategic importance are available only in Arabic. Deploying Arabic GIR would also enable the analysts to more rapidly focus on the relevant documents.