Adam Albright, du MIT, donnera le vendredi 6 février à 15h en salle DS-1950 (UQAM, Pavillon DeSèves, métro Berri-UQAM) une conférence intitulée : "Sources of phonotactic well-formedness : statistical, phonetic, and phonological biases"
Abstract : Phonotactic probability (i.e., the difference between common vs. uncommon combinations of sounds, such as English `brat’ (high probability) vs. `twerp’ (low probability)) has been shown to influence many types of phonological processing, including word segmentation, recognition, learning, and wordlikeness judgments. In the psycholinguistics literature, phonotactic effects are often modeled as learned statistical knowledge about local co-occurrence probabilities (Bailey and Hahn 2001 ; Vitevitch and Luce 2004). At the same time, characterizing differences between possible and impossible sequences of sounds is also a mainstay of phonological theory. Under one dominant approach (Optimality Theory ; Prince and Smolensky 2004), differences in grammatical well-formedness are attributed to innate constraints on phonological representations. Given the success of statistical models in capturing phonotactic effects, however, it is natural to wonder whether some or all grammaticality judgments may attributed to learned probabilities, without the need for innate constraints (Hayes and Wilson 2008), and even without innate phonological representations or grammar.
In this talk, I discuss several ways in which grammatical structure and constraints may in fact constrain and complement statistical phonotactic knowledge. Based on acceptability ratings of nonce words, I show that a model of phonotactic well-formedness benefits from incorporating structured phonological representations, in the form of phonological features. The results show distinct effects of probabilities stated over "surface" structures (perceptual categories) and "grammatical" structures (phonological features). Furthermore, statistical learning is not enough : speakers prefer some unattested sequences (such as /bn/) over others (such as /bz/) in a way that cannot be fully explained by probabilities over either surface categories or feature combinations. I show that a model which incorporates phonetically motivated biases for certain sequences over others. Finally, it appears that speakers prefer some combinations of rare sequences over others. In particular, words containing two very rare consonant clusters (e.g., ’snalt’) are deemed much worse than we would expect based on the acceptability of the sub-parts. I show that this effect, too, may be modeled as the contribution of a distinct grammatical level of evaluation.