Generation of Dutch referring expressions using the D-TUNA corpus

This paper describes our research into generating Dutch noun phrases as descriptions of furniture objects or people. This is usually done in two steps: attribute selection and realisation. This research focuses only on the realisation step: generating a noun phrase from given attributes. The research is done on the Dutch version of the TUNA-corpus, which contains annotated human-produced descriptions. Three algorithms were developed for this task, each an improvement over the last. We extracted the lexical choice from the D-TUNA corpus, and used templates generated from the corpus which specified the order of attributes. The algorithms were then evaluated for string similarity with the original human descriptions from the corpus, and a human evaluation was carried out which tested clarity and fluency. A steady improvement of scores in both the automatic and human evaluation was observed for each new version of the algorithm.