Vagueness as Cost Reduction: An Empirical Test

NLG systems that generate text from numerical data must decide between alternative linguistic forms of the given numerical content, such as whether to use a precise or a vague expression. Currently there is little empirical data for these systems to draw on when making these decisions. We performed experiments with human readers in which participants responded to instructions in the form of referring expressions, where we manipulated whether the instruction used a vague or a crisp referring expression, in order to test the hypothesis that vagueness reduces processing costs for the comprehender. Results indicate that people respond more quickly and accurately to vague linguistic expressions than to crisp numerical expressions, but that this benefit also accrues to precise terms that avoid numbers.