An unsupervised approach to identifying causal relations from relevant scenarios
Mehwish Riaz · Illinois Digital Environment for Access to Learning and Scholarship (University of Illinois at Urbana-Champaign) · 2010
Semantic relations between various text units play an important role in natural language understanding, as key elements of text coherence. The automatic identification of these semantic relationships is very important for many language processing applications. One of the most pervasive yet very challenging semantic relations is cause-effect. In this thesis, an unsupervised approach to learning both direct and indirect cause-effect relationships between inter- and intra-sentential events in web news articles is proposed. Causal relationships are leaned and tested on two large text datasets collected by crawling the web: one on the Hurricane Katrina, and one on Iraq War. The text collections thus obtained are further automatically split into clusters of connected events using advanced topic models. Our hypothesis is that events contributing to one particular scenario tend to be strongly correlated, and thus make good candidates for the causal information identification task. Such relationships are identified by generating appropriate candidate event pairs. Moreover, this system identifies both the Cause and Effect roles in a relationship using a novel metric, the Effect-Control-ratio. In order to evaluate the system, we relied on the manipulation theory of causality