Granular causality for learning by reading

2011-10-26It has long been the vision of AI researchers to build systems that are able to learn and understand causal patterns in discourse as they read input text, so that new inferences can be made on the input discourse, and numerous causal patterns can be extracted from texts that may be from relatively different domains. Discovering causal relations has proved to be a challenging research problem. One reason for this is that causal markers are dependent on the domains and genres of English discourse that they exist in. For instance, causal markers in football articles, are different from causal markers in bio-medical articles. In this thesis I prove the domain and genre dependence of causal markers. Most previous work for discovering causality has focused on either limited domains, or the top most frequent patterns representing causality in language. Both of these approaches have the shortcomings of being able to extract only a small set of causal relations. To have a wider coverage, there is an urgent need to develop systems that are able to discover domain dependent causal markers as well as low frequency causal markers in text. This thesis aims at solving this problem. ❧ I introduce a Theory of Granular Causality as it exists in natural language discourse. Although the phenomenon of granular causality is very common in discourse, there has been very limited work for understanding it, and no work done to extract such relations from discourse. In this thesis, I propose a Theory of Granular Causality, and prove the features of this theory using a human annotation study. Next, elaborate on how this theory can be used to discover causal markers from small domains of discourse, addressing the problems raised by the domain dependent nature of causal markers. Additionally, I extend this theory to discover causal relations from discourse which are not marked by a causal marker, i.e., causality between two sentences. Finally, I apply this theory for answering ”how” style causal questions, which are the second most popular questions on the web after factoid (when, when) questions. ❧ This dissertation is the first look at granular causality as a phenomenon in natural language, and applications of this theory for solving challenging problems