Weighted Abduction for Discourse Processing Based on Integer Linear Programming

This chapter explores the logical framework called weighted abduction as applied to solving discourse-processing tasks. Weighted abduction incorporates a cost propagation mechanism allowing us to estimate the likelihood of the obtained abductive proofs. We use a tractable implementation of weighted abduction based on Integer Linear Programming and a large knowledge base generated automatically. We first perform an experiment on plan recognition using the dataset originally developed for Ng and Mooney's system [39]. Then we apply our discourse processing pipeline for predicting whether one text fragment logically entails another one (Recognizing Textual Entailment task). The study we describe is the first attempt to apply tractable inference-based natural language processing on a large scale.