Selective Annotation of Modal Readings: Delving into the Difficult Data

This work presents an alternative approach to the question of how to  define taxonomic categories in the word sense disambiguation problem  of determining the reading of a modal auxiliary verb. We focus, in this  paper, on uses of could , which we found to be the most difficult of  all of the modal auxiliary data. Rather than classifying uses of could  according to taxonomic labels such as ability or deontic , we classify  uses with respect to where the reading situates the eventuality being  described relative to the speech time. For example, the sentence ‘You  could swim.’ is about a swimming eventuality in the past leading up  to the time of speech, if it is read as being an ability. The sentence is  about a swimming eventuality in the future of the time of speech, if  it is read as being a polite suggestion. We classify the uses according  to specific temporal properties of the eventuality that the modal auxiliary  describes. We achieve between 0.614 kappa  and 0.769 kappa inter-annotator  agreement. The temporally-inspired categories we propose have immediate  relevance for models of modal auxiliary meaning and provide the  groundwork for improved classification into finer-grained, inference-rich  semantic categories.