Explaining sonority projection effects

Finished · Phonological Learning

Summary

The paper conducts a nonce-word experiment showing that sonority projection effects (distinctions speakers make between unattested phonological sequences in terms of well-formedness based on sonority sequencing) are present for speakers. Using these results, the paper compares the success of a handful of phonotactic learners in correctly evaluating well-formedness of unattested onset clusters. Results show that phonotactic learners MUST be equipped with (1) syllabification details of the language and (2) some sort of sonority mechanism that generalizes beyond segments (e.g., feature classes) in order to accurately predict relative well-formedness of unattested clusters.

One additional interesting finding is that the models that had both of the above properties did not do well in predicting nonce words with clusters that ARE attested in the language of participants. Authors note that other models equipped with predicting well-formedness based on lexical type frequency did much better on attested forms. Their future recommendation is that phonotactic models should incorporate both syllabification and sonority generalization features with analysis of forms based on lexical type frequency in order to model the entire range of data.

One final note, the Generalized Neighborhoud Model (GNM) is based on exemplar theoretic ideas and did a poor job at modeling unattested cluster well-formedness. This is unsurprising, given that linguistic structure is said to emerge from use and experience. Without the ability to model unused, nonce words and clusters, these models are limited in their efficacy. In essence, they need to incorporate some type of mechanism that can generalize sonority sequencing across segments. The GNM has this but still fails for other reasons. It is a bit confusing to me why the GNM incorporates featural information, however, since exemplar theories typically argue that exemplars smaller than a morpheme/word/phrase do not exist. Maybe, the models encode exemplars as words made up of features? In any case, it seems necessary to have some representational element smaller than the word and morpheme to accurately model nonce well-formedness of clusters.