On Referring to Sets of Objects Naturally

Several algorithms have been proposed for the generation of referring expressions to identify sets of objects, including incremental, best-first, and exhaustive searches. However, most of these algorithms produce uniform combinations of descriptors, some even including disjunctions and negations, which may not always be adequately expressible by natural language surface forms. In order to better support the production of descriptions considered natural, we apply and enhance a best-first searching procedure, which enables us to produce a larger variety of expressions of limited complexity. We incorporate restrictions which express constraints on preferred surface forms, and we enhance the repertoire of descriptions by compositions of partially identifying expressions and by descriptions of objects to be excluded. The results show that our algorithm produces reasonable specifications for surface expressions, with a significantly increased repertoire compared to its predecessors.