Extracting Semantic Relations with Supervision

Abstract

In this chapter, we focus on the supervised learning of semantic relations. Supervised learning is a general machine learning methodology in which a predictive model is “trained” by generalizing from examples of data that are provided along with their true label. In the context of relation classification, this requires the provision of text that has been annotated with the relations it expresses. Supervised learning can perform very well, but there are specific difficulties for this paradigm to overcome. First, annotated data must be available, which can be a serious bottleneck. Second, learning good models requires the design and acquisition of a representative set of features. We will describe in this chapter some of the available datasets, and we will pay particular attention to how the relation arguments and the surrounding context are described. Depending on how shallow or deep the features describing a relation are, such approaches may or may not be useful in open information extraction, where, due to the large amount of data to be processed, fast—and therefore shallow—methods are used. The techniques presented in this chapter are more appropriate for the semantic analysis of individual texts, where we can afford to compute interesting features and possibly apply models for different relation types to a given test instance in order to determine the relation that holds. Chapter 4 will review methods that are more suitable to open information extraction or to processing large text collections.