Exploiting semantic constraints for estimating supersenses with CRFs

The annotation of words and phrases by ontology concepts is extremely helpful for semantic interpretation. However many ontologies, e.g. WordNet, are too fine-grained and even human annotators often have disagreements about the precise word sense. Therefore we use coarse-grained supersenses of WordNet. We employ conditional random fields (CRFs) to predict these supersenses taking into account the interaction of neigboring words. As the annotation of training data is costly we modify the CRF algorithm to process lumped labels, i.e. a set of possible labels for each training example, one of which is the correct label. This information can be derived automatically from WordNet. By this and the use of new features we are able to increase the f-value for about 8% compared to previous results. If we use training data without annotations it turns out that the resulting F-value is only slightly lower than for a fully labeled training data. Therefore unlabeled data may be employed in an unsupervised way to increase the quality of predicted supersenses.