基于跨语言信息的大规模稀缺资源知识库自动构建技术研究

知识库构建可视为人工智能一项核心技术，被广泛运用于搜索引擎、问答系统、智能对话系统等任务中。英文、中文等语言下具有充足的资源用以支撑大规模知识库的构建。 然而，资源稀缺的语言缺少丰富的标注数据资源和高质量的基础语言处理工具，从而很难精确的从文本中挖掘知识以构建大规模的知识库。基于此，本项目面向稀缺资源场景，提出一套弱监督框架下的跨语言资源利用体系，帮助构建稀缺资源语言下的知识库。具体包括：1）提出一种基于双语的语义映射方法，用于解决稀缺资源下实体表示不充分的问题，从而从稀缺资源下识别命名实体。2）提出一种基于弱监督框架下的抗噪关系挖掘模型，通过引入多层注意力机制过滤噪声以获取准确的实体间关系。3）鉴于稀缺资源缺少足够的标注数据和标注工具，需要将上述资源丰富语言下挖掘出的实体和关系翻译到稀缺资源语言下，基于此，提出一种引入张量模型的消歧翻译模型，用于关系和实体在多语言间切换。

The automatic construction of knowledge can be seen as a key technique of artificial intelligence, which can be widely applied to various tasks, including search engine, question answering system and intelligent dialogue system. Languages like English and Chinese have abundant resources for construction of knowledge graph at large scale. However, low-resource languages suffer from lacking adequate annotated data and missing fundamental language processing tools with high quality, which make it hard to mining knowledge from text accurately for constructing knowledge graph at scale. On account of this, for the scenario of low-resource, this program proposes a weakly supervised system for utilizing cross-lingual resources, to help construct knowledge graph from low-resource languages. This specifically includes:.1) to propose a bilingual semantic mapping-based method to solve the problem that entities being inadequately represented in low-resource language scenarios, so as to recognize named entities in low-resource languages..2) to investigate a weakly supervised anti-noise relation mining model, which incorporates multi-layer attention mechanism to filter noise in order to capture accurate relationships between entities..3) Considering the fact that low-resource languages lack enough annotated data and annotation tools, it's necessary to translate entities and relations extracted from rich-resource language to low-resource language. Based on this, we introduce a tensor-based disambiguity translation model to help entities and relations transfer among languages.