CRI:CRD Collaborative Research: General Techniques for Creating Treebanks with Multiple Representations: A Large-Scale Russian Application

来自宾夕法尼亚大学、科罗拉多大学、华盛顿大学和哥伦比亚大学的合作首席研究人员正在对创建一个包括依存结构和短语结构在内的多种表示的树库的可行性进行初步调查。感兴趣的问题包括：*依存结构和短语结构之间的转换在多大程度上可以自动化，以及在自动转换后需要在多大程度上进行手校正？*我们如何利用创建其他树库的经验？为了解决这些问题，合作小组正在召开一次国际学者工作会议，这些学者积极研究树岸，代表许多不同的形式和语言。在前人短语结构树库到依存结构、树邻接语法结构和组合范畴结构转换的基础上，进行了一系列实验，证明了从依存结构生成短语结构的可行性。这需要为所讨论的语言制定明确的短语结构和依存结构指南，目前仅限于英语。对于这两种形式中的一种，其他经过测试的语言包括汉语、韩语和阿拉伯语(短语结构)或捷克语、德语、荷兰语和俄语(依存结构)。该团队正在将他们的转换实验扩展到这些语言中的一种，在这个过程中创建了必要的额外指导方针。他们正在与研讨会上的同事分享对我们结果的定性评估，并讨论进一步试验和替代语言的适宜性，作为重新提交CRI提案的投入。

The collaborating Principal Investigators from the University of Pennsylvania, University of Colorado, University of Washington, and Columbia University are performing preliminary investigations into the feasibility of creating a treebank with multiple representations including both dependency structure and phrase structure.Issues of interest include: * To what extent can the conversion between dependency and phrase structure be automated, and to what extent is hand correction necessary after automatic conversion? * How can we leverage experience from the creation of other treebanks? To address these issues, the collaborating team is convening a working meeting of international scholars who have actively worked on treebanks representing many different formalisms and languages. Building on previous results in the conversion of phrase structure treebanks to dependency structures, tree adjoining grammar structures and combinatory categorical structures, a set of experiments is performed demonstrating the feasibiity of generating phrase structure from dependency structure. This requires well defined phrase structure and dependency structure guidelines for the language in question, currently limited to English. Other languages with tested guidelines for one of the two formalisms include Chinese, Korean, and Arabic (phrase structure) or Czech, German, Dutch, and Russian (dependency structure). The team is extending their conversion experiments to one of these languages as well, creating the necessary additional guidelines in the process. They are sharing a qualitative evaluation of our results with the colleagues at the workshop, and discussing the suitability of further experiments and alternative languages, as input into a resubmission of the CRI proposal.