基于事件关联性的层次话题建模研究

互联网内容具有的多源异构、表达自由、观点丰富等特点给面向互联网的文本内容分析，包括事件抽取、话题发现与跟踪、情感分析、立场分析等核心任务，带来了巨大的挑战和机遇。本项目针对互联网话题的层次性与多面性，开展全面深入的层次话题建模研究，特别是针对目前主流话题模型存在的事件表示缺失及其导致的事件关联性计算缺失,提出基于事件关联性的层次话题建模方法体系。主要包括：1）基于神经网络模型学习事件的深度语义表示，提出一种基于谓词论元结构的事件表示学习方法；2）基于文本本身的词汇关联和基于谓词论元结构的语义关联，提出一种基于深度语义的事件关联性计算方法；3）基于事件关联性计算，将相关事件组织为子话题，并基于子话题之间的关联性，将相关的子话题组织为话题，实现一种基于事件关联性的层次话题建模方法；4）构建一个细粒度立场分析语料资源和应用平台，验证本项目提出的层次话题建模方法体系的有效性和实用性。

The multi-source heterogeneity, freedom of expression, and rich perspective of Internet content bring huge challenges and opportunities to the core tasks of Internet-oriented text content analysis, including event extraction, topic discovery and tracking, sentiment analysis, and stance analysis. This project aims at the hierarchical and multi-faceted nature of Internet topics, and conducts comprehensive and in-depth hierarchical topic modeling. Especially for the lack of event representation and the lack of event correlation computing caused by current mainstream topic models, this project proposes an event correlation-based hierarchical topic modeling methodology. It mainly includes: 1) learning deep semantic representation of events based on neural network models and proposing an event representation learning approach based on the predicate argument structure; 2) proposing a deep semantics-based event correlation computing approach according to text itself-based lexical association and predicate argument structure-based semantic association; 3) implementing a hierarchical topic modeling approach based on event correlation by first organizing related events into sub-topics based on event correlation computing and then organizing related sub-topics into topics based on the correlation between sub-topics; 4) constructing a fine-grained stance analysis corpus resource and application platform to verify the effectiveness and practicability of the hierarchical topic modeling methodology proposed in this project.