Learning precise partial semantic mappings via linear algebra

Abstract

In natural language interfaces, having high precision, i.e., abstaining when the system is unsure, is critical for good user experience. However, most NLP systems are trained to maximize accuracy with precision as an afterthought. In this thesis, we put precision first and ask: Can we learn to map parts of the sentence to logical predicates with absolute certainty? To tackle this question, we model semantic mappings from words to predicates as matrices, which allows us to reason efficiently over the entire space of semantic mappings consistent with the training data. We prove that our method obtains 100% precision. Empirically, we demonstrate the effectiveness of our approach on the GeoQuery dataset.