面向语言处理的藏语概念框架语义知识库构建研究

With the improvement and development of various artificial intelligence algorithms, as well as neural networks, deep learning has become a hot spot, those powerful neural architectures performing in big data processing has made Breakthroughs progress. However, there is still a search in progress for whether if could neglect a large number of specific projects for specific tasks or whether to incorporate more language structures into deep learning. As for a small language, under the circumstance of that data size is too small and the big data is empty, the support and assistance of the Knowledge base will be become a key resource, and the importance will never be replaced. This topic has referenced theories like Frame Semantics, Valence Grammar, Case Grammar, and also referenced Berkeley Frame Net, Word Net by Princeton University's ,and the Mind Net by Microsoft, Beside those ,we also referenced the 905 Semantic Knowledge Engineering, the How Net, the CCD by Peking University (Chinese Concept Dictionary) , Shanxi University English Frame Net and more research methods. The research is based on the Tibetan grammatical theories, to explore the laws and rules of Tibetan the framework of semantic analysis, to adjusting and enriches the methods of Tibetan framework semantic analysis, and to establish the Tibetan frame semantic repository. We set the Tibetan actual text as the study object, for the reason to own a better affection of description and analysis on Tibetan Semantic Framework System, we established Tibetan Words Cell Base, Tibetan Semantic Framework Base, and more Tibetan syntactic knowledge bases for each domain.

随着人工智能各种算法的兴起，神经网络、深度学习等成为热点，在大数据处理中用简单而强大的神经架构执行复杂任务，取得了突破性进展。但在是否可以忽略大量针对特定任务的特征工程，或是否将更多的语言结构融入深度学习方面仍在探索。就小语种而言，在没有大数据的情况下知识库的支撑和协助成为一个关键资源，其重要性不可替代。本课题借鉴框架语义学、依存语法、格语法等理论，参考FrameNet、WordNet、MindNet和我国的905语义知识工程、HowNet、北大CCD及山西大学汉语框架语义知识库等的研究方法，立足藏语语法理论，探索藏语框架语义分析的规律，调整充实藏语框架语义系统分析方法，建立藏语概念框架语义资源库。以藏语真实文本为研究对象，建立各个域的藏语词元库、语义框架库、句法知识库，从而对藏语概念框架语义系统进行描写与分析。该资源库的建成将为藏语句法分析、语义理解、文本理解、智能检索等发挥关键作用。

本课题借鉴框架语义学、依存语法、格语法等理论以及迁移学习前沿方法，立足藏语语法虚词理论，探索藏语依存语法、框架语义分析的规律，调整充实藏语框架语义系统分析方法和藏语依存知识表示方法，融合藏语语法知识，利用深度学习方法在数据预处理、无监督、半监督、双语迁移学习等方法构建依存树库、框架知识库等，取得了一定的成效。.在理论层面，项目在句法语义标记体系建构方面，不仅考虑了不同语种的依存和框架语义模式，同时创新性利用了藏语固有的语法特点，将虚词融入句法、语义两个层面并对其赋予架构作用和一致性鲁棒作用，成效显著。.在方法层面，项目在初期依存数据集建构方面，不仅利用人工进行探索性标注，同时利用虚词知识库和虚词功能作用自动建构前期粗标注数据集，并进行人工审校后利用神经网络方法进行语言建模，从数据收集、预处理、半监督数据到有监督数据建构，研究呈体系化，具有一定的创新性。.在资源方面，项目最终建成多个知识库和5万句依存树库、框架语义知识库，该树库是目前国内最大的藏语依存树库。