Evin: Building A Knowledge Base Of Events

We present EVIN: a system that extracts named events from news articles, reconciles them into canonicalized events, and organizes them into semantic classes to populate a knowledge base. EVIN exploits different kinds of similarity measures among news, referring to textual contents, entity occurrences, and temporal ordering. These similarities are captured in a multi-view attributed graph. To distill canonicalized events, EVIN coarsens the graph by iterative merging based on a judiciously designed loss function. To infer semantic classes of events, EVIN uses statistical language models. EVIN provides a GUI that allows users to query the constructed knowledge base of events, and to explore it in a visual manner.