Comparing Rating Scales and Preference Judgements in Language Evaluation

Rating-scale evaluations are common in NLP, but are problematic for a range of reasons, e.g. they can be unintuitive for evaluators, inter-evaluator agreement and self-consistency tend to be low, and the parametric statistics commonly applied to the results are not generally considered appropriate for ordinal data. In this paper, we compare rating scales with an alternative evaluation paradigm, preferencestrength judgement experiments (PJEs), where evaluators have the simpler task of deciding which of two texts is better in terms of a given quality criterion. We present three pairs of evaluation experiments assessing text fluency and clarity for different data sets, where one of each pair of experiments is a rating-scale experiment, and the other is a PJE. We find the PJE versions of the experiments have better evaluator self-consistency and interevaluator agreement, and a larger proportion of variation accounted for by system differences, resulting in a larger number of significant differences being found.