Event Recognition from News Webpages through Latent Ingredients Extraction

We investigate the novel problem of event recognition from news webpages. “Events” are basic text units containing news elements. We observe that a news article is always constituted by more than one event, namely Latent Ingredients (LIs) which form the whole document. Event recognition aims to mine these Latent Ingredients out. Researchers have tackled related problems before, such as discourse analysis and text segmentation, with different goals and methods. The challenge is to detect event boundaries from plain contexts accurately and the boundary decision is affected by multiple features. Event recognition can be beneficial for topic detection with finer granularity and better accuracy. In this paper, we present two novel event recognition models based on LIs extraction and exploit a set of useful features consisting of context similarity, distance restriction, entity influence from thesaurus and temporal proximity. We conduct thorough experiments with two real datasets and the promising results indicate the effectiveness of these approaches.