Phonotactic illegality and probability in speech perception : evidence from second language listeners

Abstract

Phonotactic knowledge is knowledge of the sound combinations of a language. Previous research showed that frequent sound combinations are perceived more easily than illformed ones, while illegal combinations are often filtered out or adapted. Phonotactics can be described in categorical terms (i.e., a sound combination is legal or illegal), or in gradient terms (i.e., a combination has a certain level of wellformedness). In addition, phonotactics can both be described in negative or positive terms, by defining when combinations are legal vs.\ when they are illegal. This thesis present experimental research on the mental representation of phonotactics that is used for speech perception. A cross-modal priming experiment on second language listening showed that consonant clusters that are frequent in the second language are easier to process, in spite of perceptual epenthesis, an illusion caused by illegality of the same clusters in the first language. This disjunction between a categorical and a gradient effect suggest separate representations of both categorical and gradient phonotactics. However, the dichotomy between categoricalness and gradience is confounded with the dichotomy between negative and positive representations. The categorical filtering effects are based on negative knowledge, while the gradient facilitation effect is based on positive knowledge. Therefore, phonotactic effects on speech segmentation were investigated next. In segmentation, the effect of negative categorical knowledge is facilitatory, instead of inhibitory. Illegal combinations cue word boundaries, because illegal combinations cannot occur word-internally. An eye-tracking method was developed to test segmentation effects independently of word activation. An experiment with Dutch native listeners showed that phonotactic knowledge has an effect on segmentation in the absence of full lexical recognition. A follow-up experiment showed that illegal segmentations are inhibited; there might additionally be a bias toward more wellformed segmentations. In the same experiment, second language listeners of Dutch were more likely to split consonant clusters that are illegal in Dutch, but legal in their native language. Nevertheless, this effect was more categorical for native listeners. When other, more wellformed segmentations fail to yield words, second language listeners make segmentations that leave illegal clusters intact, which native listeners hardly do. These results show the existence of categorical effects of illegality and gradient effects of wellformedness. All second language learners studied before showed transfer of categorical phonotactic illegality from their first language. To test if categorical knowledge is always fossilised, the segmentation of English and Dutch by native Dutch listeners, proficient in English, was studied. The listeners applied English categorical phonotactics to English, showing that categorical phonotactics can be acquired. The empirical data presented in this thesis suggest that phonotactic representations can be intrinsically positive and negative, while their effects are categorical and gradient, depending on the task and the strength of the representations. A model of phonotactic grammar is proposed. This grammar is based on constraints representing positive or negative knowledge. It is applied to speech before lexical access takes place and can yield both gradient and categorical effects on speech perception.