Extraction and categorization of transition information from large volume of texts using patterns and machine learning

The information on transition of a thing is important for acquiring knowledge on the thing. In this study, we extracted transition information from a large number of sentences using pattern-based methods. We obtained an F-measure of 0.86 for extraction of transition information by this method. In order to improve the results, we then combined the pattern-based methods with supervised machine-learning methods. We obtained F-measures of 0.91 and 0.89 for extraction of transition information by support vector machine and maximum entropy methods, respectively. Thus, we confirmed that the combined approach outperformed the pattern-based method. We also categorized transition information. In the experiments, we obtained an F-measure of 0.6 for transition information categorization for categories containing many events in the training data set. Furthermore, we examined sentences in terms of conceptual changes in their transition information. We classified transition information in various ways. The categories by the classification provide a theoretical basis for future studies on transition information.