Combining Frequent and Discriminating Attributes in the Generation of Definite Descriptions

The semantic content determination or attribute selection of definite descriptions is one of the most traditional tasks in natural language generation. Algorithms of this kind are required to produce descriptions that are brief (or even minimal) and, at the same time, as close as possible to the choices made by human speakers. In this work we attempt to achieve a balance between brevity and humanlikeness by implementing a number of algorithms for the task. The algorithms are tested against descriptions produced by humans in two different domains, suggesting a strategy that is both computationally simple and comparable to the state of the art in the field.