使用事件向量解析新聞

Abstract

新聞在生活中扮演重要的角色，而新聞中的事件卻是一個抽象的概念，我們定義事件是一群實體的集合。利用 Wikipedia 的資訊，我們整合三個 entity linking 系統 Wiki, PBoH 和 RandomWalk，並將新聞中的實體名字對應到 Wikipedia 的實體。並使用 TransE 模型將知識圖譜向量化，保存實體和實體之間的關係。最後，我們從新聞中建立一實體圖，用辨識圖型中重疊社群 (overlapping community) 的方法，使用 Nonnegative Matrix Factorization 的模型，辨識出新聞中的事件。

Nowadays, news is an important part of our daily life. But the event in news is an abstract concept. We define an event as a group of entities. By leveraging the information from Wikipedia, we integrate three entity linking system Wiki, PBoH and RandomWalk to link the entity mention in documents to correspondent entities in Wikipedia. We also use TransE model to embed each entity as a vector, which contains the relational information between entities. Finally, we build an entity graph from news document and use Nonnegative Matrix Factorization approach to find the overlapping community in our entity graph. The community is a group of vertices(entities) in the entity graph, which corresponds to our entity definition.