Abstract:
This paper describes the design and empirical evaluation of statistical models that use domain and lexical knowledge to organize new semantic options in interfaces for editing knowledge bases. We employ the models in a system that allows a domain expert to perform languageneutral knowledge editing by interacting with natural language text generated by a natural language generation system. This editing produces a knowledge base used for natural language generation in multiple languages. To create statistical models, we use natural language techniques to automatically process a large corpus and then collect statistics based on the frequency of objects and verb-object pairs. The models are used to produce methods for organizing new semantic options based on 1) overall object frequency and 2) verb-object pair frequency. We compare these two innovative methods to the more traditional method of organizing options alphabetically. We conduct a set of experiments to evaluate these three methods for organizing new semantic options and analyze subjects’ interactions with the different methods in terms of speed and accuracy.