Modelling subphonemic information flow: an investigation and extension of Dell's (1986) model of word production

Abstract

Dell (1986) presented a spreading activation model which accounted for a number of early speech error results, including the relative proportions of anticipations, perseverations and exchanges found in speech error corpora, the lexical bias effect, the phonological similarity effect, and the effect of speech rate on error rate. This model has had an immense influence on the past 20 years of research into word production, with the original paper being cited over 1,000 times. Many studies have questioned how activation should flow between words and phonemes in this model. This thesis aimed to clarify what current speech error evidence tells us about how activation flows between phonemes and subphonemic representations, like features. Does activation cascade from phonemes to features, and does it feed back? The work presented here extends previous modelling investigations in two ways. Firstly, whereas previous modelling research has tended to evaluate model behaviour using arbitrarily chosen parameter settings, we illuminate the influence of the parameters on model behaviour and propose methods to draw general conclusions about model behaviour from large numbers of simulations at orthogonally varied parameter settings. Secondly, we extend the scope of the simulations to consider output at a subphonemic level, modelling recent data acquired via acoustic and articulatory measurements, such as voicing onset time (VOT), electropalatography (EPG) and ultrasound, alongside older transcribed speech error data. Throughout the thesis, we consider whether parameter settings which lead the model to capture individual results also permit other results to be accounted for and do not cause otherwise implausible behaviour. Through manipulating parameter settings in Dell's (1986) original model, we find that increasing the number of steps before selection generally does not decrease the error rate, but rather increases it, contrary to results reported by Dell (1986). This calls into question the claim that an increase in steps before selection provides a good model of a slower speech rate. We also demonstrate that the model captures the negative correlation reported by Dell, Burger, and Svec (1997) between error rate and the ratio of anticipations to perseverations, and further predicts that there should be a negative correlation between this ratio and the proportion of errors which are non-contextual. However, our results show that no parameter setting allows the model to generate enough exchanges to match even minimum estimates from a reanalysis of multiple speech error corpus reports, without falling foul of other constraints; in particular, limits on the overall number of errors generated. We suggest that the exchange completion triggering mechanism proposed by Dell (1986) is not strong enough, and that current corpus evidence provides little support for his account of word sequencing. Focusing on single word production therefore, the second part of the thesis investigates behaviour of models with output at a subphonemic level. We find that, provided sufficient contextual errors occur at the featural level, a model in which only the identity of the selected phoneme is conveyed to the featural level can account for: (i) the phonological similarity effect found in transcribed records of speech errors (whereas in models with output at the phoneme level, feedback from features to phonemes is required); (ii) detectable influences of intended phonemes in VOT measurements of unintended phonemes, as well as the effect of error outcome lexicality on these results ( findings presented in support of cascading from phonemes by Goldrick & Blumstein, 2006); and (iii) increased similarity of EPG measurements of articulations to reference measurements of competing articulations when production of the competing onset would result in a word (McMillan, Corley, & Lickley, 2009). Initial results appear to con firm however that, in contrast, phonological similarity effects on the relationship of articulatory and acoustic measurements of productions to reference measurements (McMillan, 2008) can only be accounted for in an architecture with feedback from features to phonemes. To strengthen conclusions about articulatory evidence of lexical bias and phonological similarity effects, future work needs to consider the extremely strong effects of frequency observed in these simulations. The results presented in this thesis contribute to a greater comprehension of the behaviour of Dell's (1986) influential model, and further demonstrate that the model can be extended to account for new instrumental evidence, whilst clarifying the constraints on activation flow between phonemes and features which this new evidence imposes.