Abstract
This paper argues that if phonological and phonetic phenomena found in language data and in experimental data all have to be accounted for within a single framework, then that framework will have to be based on neural networks. We introduce an artificial neural network model that can handle stochastic processing in production and comprehension. With the “inoutstar” learning algorithm, the model is able to handle two seemingly disparate phenomena at the same time: gradual category creation and auditory dispersion. As a result, two aspects of the transmission of language from one generation to the next are integrated in a single model. The model therefore addresses the hitherto unsolved problem of how symbolic-looking discrete language behaviour can emerge in the child from gradient input data from her language environment. We conclude that neural network models, besides being more biologically plausible than other frameworks, hold a promise for fruitful theorizing in an area of linguistics that traditionally assumes both continuous and discrete levels of representation.
Highlights
What will be the ultimate model of phonology and phonetics and their interactions? It will have to be a model that accounts for at least four types of valid behavioural data, namely 1) the generalizations that phonologists have found within and across languages, 2) the phenomena that psycholinguists and speech researchers have found by observing speakers, listeners, and language-acquiring children, 3) the mergers, splits, chain shifts and other sound change phenomena found by historical phonologists and dialectologists, and 4) the phenomena that have been observed when languages come in contact, such as loanword adaptations
We provide a first proposal of a neural network model that can handle two important aspects of the transmission of a sound system from one generation to the namely category creation and auditory dispersion, and we simulate the model on a range of synthetic data
If the model contains levels of representation, it may look like Figure 1, which can be thought of as containing the minimum number of levels needed for a sensible description: phonetics seems to require at least an Auditory Form (AudF, specifying a continuous stream of sound) and an Articulatory Form (ArtF, specifying muscle activities), and phonology seems to require at least an Underlying Form (UF, containing at least lexically contrastive material) and a Surface Form (SF, containing a whole utterance divided up in prosodic structure such as syllables); the Morpheme level connects the phonology to the syntax and the semantics in the lexicon
Summary
The weights turn out to have become the conditional probabilities of SF given UF (as in Figure 3), so outstar learning exhibits the probability-matching behaviour that we wanted; the sum of the weights going out from each UF node is 1 This could have been predicted theoretically, by realizing that in the equilibrium situation 0 = 〈ai aj − ai wi j〉 = 〈ai aj〉 − 〈ai〉wi j, so if learning converges, it must move the weights asymptotically toward (14). Neural network models for phonology and phonetics has some specificity, and it is even a bit frequency-dependent in both directions (because instar and outstar are both frequency-dependent in one direction) It has the additional advantage over both instar and outstar learning that it is symmetric in input and output: the formula stays the same if i and j are swapped, i.e. the inoutstar learning rule does not care about the direction of processing.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
