Discovery of Event Classifications and Causal Constraints from Earthquake Articles

Many researches of Text Mining (TM)[6] and Knowledge Discovery in Texts (KDT)[4], which are applications of discovery of association rules and sequence detection [1, 8] to text databases, are successful and promising. However, their main aims are concentrated on “loose” associations of documents and information in them. For example, association rule hammer -> nails [s,c] means that “documents on hammer” can be associated with “documents on nails” with support s and confidence c. On the other hand, as noticed in Information Extraction (IE, see MUC[3] conferences), documents contains explicit and implicit information about events and their causal relations. In such tasks, “deeper” natural language processing techniques and information models are required. This paper will propose a method of discovery of ontologies of domain events, in particular, earthquakes, i.e., their classifications and causal constraints. Firstly the concept of event classifications that is based on Barwise-Seligman [2]’s Channel Theory and their extraction method which exploits information extraction from documents, will be introduced. Extractable information will be based on event model that will be newly introduced here, and the method will be customized to earthquake articles. As the result, we can get an event table from an earthquake article. Secondly the notion and discovery method of causal constraints will be introduced. Causal constraints are constraints supported by event classifications in the sense of Barwise-Seligman [2], but their premises are temporally restricted to the ones preceded to each of their consequents. Thirdly the notion of occurrences and Occurrence Calculus will be introduced. An occurrence is a token restricted to its appearance in a document, and its types have the frequency of its tokens’ occurrences. By both notions, we can mine causal constraints with parameters: its support and its confidence, as association rules in data mining.