Learning Effect Axioms via Probabilistic Logic Programming

Abstract

In this paper we showed how we can automatically learn the structure and parameters of probabilistic effect axioms for the Simple Event Calculus (SEC) from positive and negative example interpretations stated as short dialogue sequences in natural language. We used the cplint framework for this task that provides libraries for structure and parameter learning and for answering queries with exact and inexact inference. The example dialogues that are used for learning the structure of the probabilistic logic program are parsed into dependency structures and then further translated into the Event Calculus notation with the help of a simple ontology. The novelty of our approach is that we can not only process uncertainty in event recognition but also learn the structure of effect axioms and combine these two sources of uncertainty to successfully answer queries under this probabilistic setting. Interestingly, our extension of the logic-based version of the SEC is completely elaboration-tolerant in the sense that the probabilistic version fully includes the logic-based version. This makes it possible to use the probabilistic version of the SEC in the traditional way as well as when we have to deal with uncertainty in the observed world. In the future, we would like to extend the probabilistic version of the SEC to deal -- among other things -- with concurrent actions and continuous change.