Computational modelling of phonological acquisition: Simulating error patterns in nonword repetition tasks

Abstract

Nonword repetition tasks (NWRTs) are employed widely in various studies on language development and are often relied upon as diagnostic tools. However, the mechanisms that underlie children's performance in NWRTs are very little understood. In this paper we present NWRT data from typically developing 5- to 6-year-olds (5:4–6:8) and examine the pattern of their phonological errors within the syllabic domain. We show that the children display a strong tendency for errors at the syllable onset, with fewer errors in coda position. We then show how the same pattern can be simulated by a computer model, thus shedding some light on the cognitive mechanisms that underlie specific error patterns as well as general phonological development.

Keywords:

• Nonword repetition
• Computational modelling
• Syllabic position
• Onset
• Coda

Acknowledgements

The authors would like to thank the Leverhulme Trust who funded this research in the form of a Research Grant (ref F/01 374/G) to the second author. We would also like to thank Sarah Watson for collecting and transcribing the data, Gabrielle Le Geyt for recording the nonword stimuli, Hannah Witherstone for helping with inter-rater reliability for the children's nonwords, and Chloe Marshall and an anonymous reviewer for their valuable comments and suggestions on earlier versions of this paper. Our thanks are also due to the schools and children who participated in the study.

Notes

¹Throughout the paper, the term “syllabic position” refers to the position of phonemes within syllables.

²Metsala and Walley (1998) and Walley, Metsala & Garlock (2003) have incorporated this as part of their Lexical Restructuring Model.

³Kirk and Demuth's study included comparison of two groups of clusters. Word initially: /s/ + stop and /s/ + nasal; word-finally: stop + /s/ and nasal + /z/. Only these consonants were considered since, as far as English is concerned, they are the only consonants that can combine to form both word-initial and word-final sequences.

⁴For the purposes of this example, only five individual phonemes are illustrated below the top node. However, as mentioned above, the model is programmed to know the whole phonemic inventory of English prior to the beginning of the learning procedure.

⁵The frequency of a bi-phone sequence is multiplied by 5, that of a tri-phone sequence is multiplied by 25, while that of a quadri-phone sequence is multiplied by 125 (5×5×5), and so on for longer sequences.

⁶And, in turn, having had the opportunity to practise articulation of a phoneme sequence is more advantageous than having practised its component phonemes in separate phonological contexts.

⁷For the purposes of this operation, the consonant that is being substituted is temporarily removed from the inventory.

⁸However, vowel articulation errors can still occur in cases where the single phoneme contained within a chunk is a vowel.

⁹These were the first author and a second researcher not involved in this project but experienced in coding nonword repetitions.

¹⁰These were added to represent the increase in input during the first year of schooling.

¹¹In fact, we also analysed primacy/recency effects for all three syllable nonwords used in the nonword tests presented here (three syllable nonwords were used because both nonword sets contain this length of stimuli and this length also allows for the examination of primacy and recency). No primacy and recency effects were observed (F(2,50) = 2.52, p = .091, = .09). As discussed above, this is most likely because the nonwords were not designed in order to examine primacy and recency effects, which are only likely to emerge when other contributing factors are controlled for.

¹²As far as the model is concerned, vowel errors will not be discussed as they were deliberately inhibited (see section on articulating an input sequence). We leave the question of simulating consonantal vs. vocalic errors for further research.

¹³A reviewer points out that, under onset maximisation, it might be the sequence stop + liquid that is responsible for the high rate in onset errors rather than the onset position itself. Although the type of sequence may well play a part in the distribution of errors, this possibility is hard to evaluate since the manner in which the CNRep was constructed does not allow to control for differences in melodic sequences. There is, however, at least one set of nonwords in which the type of onset sequence does not account for the observed onset effect. For the 3syllable nonwords there are three stop + liquid sequences word-initially and none word-medially. Nevertheless, our data show a tendency for w-medial onset errors (34 vs 22 w-initial).

¹⁴Treiman & Danis (1988) reported a tendency for codas to attract more errors than onsets in an experiment that tested subjects’ ability to repeat lists of nonwords. However, this was not a standard NWRT, as it was concerned with lists rather than individual nonwords, and it presumably tapped on a slightly different set of abilities particularly in relation to the interaction between short-term memory and phonological performance.

¹⁵Vowel errors are not discussed as they were deliberately inhibited due to the fact that at this stage we are primarily concerned with the modelling of consonantal errors. We leave the question of consonantal vs. vocalic errors for further research.

¹⁶In fact, this simplicity could be viewed as a further strength of the EPAM-VOC model, as it has been argued that simpler models are preferable to more complex ones, as the latter are less readily falsifiable, and thus have less explanatory power (Fum et al. 2007, Myung, 2000).

¹⁷This might be different if the model received more input, for example if it was to simulate older children.

¹⁸The […] indicate that the second vowel may either end up on its own or be chunked up with another phoneme.

¹⁹Its frequency is 269,780, which is twice as much as the second most frequent vowel.

²⁰Only the most recurrent chunking patterns are given here.

²¹Chunks that contained both elements were excluded from the count.

²²Please note that, unlike most dialects, the English spoken in Nottingham and the Midlands still realises velar nasal-plosive sequences word-finally, hence the transcription in examples such as , where most English speakers would produce .