Using statistical smoothing to estimate the psycholinguistic acceptability of novel phrases

Native speakers of languages perceive differences in the acceptability of phrases even when those phrases are both grammatical and novel (previously unseen). We suggest that smoothing, a statistical technique used by natural language processing engineers, provides several candidate mechanisms for investigating this phenomenon. We describe the creation of a large data set of predictions from several smoothing algorithms about the acceptability of unseen grammatical phrases and a novel experimental method for the pairwise comparison of these models. We use this method to compare three smoothing methods and consider the results in light of the differences among the models. We argue that the data support the idea that similarity in this domain is best thought of as a form of asymmetric representational distortion and that the informational basis over which such estimates are made is broad, rather than narrow, as has been previously suggested.