Watch and learn : event‑domain term extraction from social networks

Abstract

Event tracking algorithms detect and track, but they do not understand what happens in events. Term extraction research has studied the concepts of general domains-computer science, medicine, law-but not What happens in event domains, and not from noisy social networks, where they are popularly narrated. The event structure, the message and its form distinguish event domains from general domains, and formal text from user-generated content. In this article, we present the Event-Aware Term Extractor (EVATE), the first term extractor built for event domains, and the first built for user-generated content. EVATE learns semantically: it tracks events to extract terms that describe What happens, and then ranks them with a termhood statistic designed for event domains. We compared our novel approach with four traditional term extractors in three disparate event domains on data from Twitter (now X). Because EVATE learns semantically, its lexicons described What happens in events better than standard approaches. Even when the term extractors could not adapt to unorthodox event domains, our novel method propped up the others as a semantic re-ranker. The results show that we need algorithms designed for event domains and for user-generated content. Crucially, they also show that we only need one semantic extractor like EVATE to adapt traditional algorithms.