Team UDEL KBGen 2013 Challenge
Keith Butler, Priscilla Moraes, Ian Tabolt, Kathy D. McCoy · 2013
This document describes the University of Delaware’s entry into KBGen 2013 Challenge which provided teams with input data representation from the AURA knowledge base (KB), developed in the context of the HALO Project at SRI International, along with a lexicon mapping for concepts present on those input files. Training sentences were also provided. The task was to accurately generate an English sentence depicting the information from a set of triples from the knowledge base. Our initial generator implementation utilized SimpleNLG in a java wrapper. Our tack was to focus on the realization of major sentence types, generally identified by the presence of a particular function in the KB triples, e.g. has-function, subevent, plays. These functions provided the main verb and sentence structures, and other KB relations were fit into this structure (in subject/object position or as adjuncts) in a rulebased way. For instance, Figure 1 shows a triples file from the testing data that was identified under the cluster has-function, along with the sentence generated by our system and the rule used to realize the cluster for this relation type. 1