2018 Friday Session A 0930

Friday, November 2, 2018 | Session A, East Balcony | 9:30am

How the input shapes the acquisition of verb and noun morphology: computational modeling across three highly inflected languages

F. Engelmann, J. Kolak, S. Granlund, V. Vihman, B. Ambridge, J. Pine, A. Theakston, E. Lieven

Both usage-based and recent rule-based accounts of morphological language acquisition assume that learners rely to some extent on rote storage/retrieval of phonological forms and phonological analogy. The majority of previous computational explorations of appropriate mechanisms has focused on simple systems like English or investigated only a small part of a paradigm (e.g., [1-2], but see [3]).

We present neural network simulations of how children acquire inflectional morphology for the full paradigm of present-tense person/number marking and case marking in three, morphologically complex languages. Three-layer networks with 200 hidden units were trained on 2000-3000 verb (Finnish, Polish) or noun forms (Finnish, Polish, Estonian) from natural, child-directed speech. The results are compared in detail with five large-scale elicited-production studies with children between the ages of 2;9 and 5;3.

The input to the model consisted of phoneme sequences representing verb stems (verb models) or nominative noun forms (noun models) and a code for the target person/number or case context, respectively. The models were trained using backpropagation to output the correct phoneme sequence of the target form. Inputs were presented probabilistically according to their token frequencies in child-directed speech corpora.

The models acquired adult-like mastery of the system after about 3 million training trials (presenting one form per trial) and could generalize (i.e., produce the correct target for untrained items) to about 85% of the test items used in the elicited-production experiments. In addition, the models yielded the key phenomena observed in the experiments: effects of token frequency and phonological neighborhood density (inflectional class size) of the target form, and a general error pattern that involved the replacement of low frequency targets by higher-frequency forms of the same lemma, or forms with the correct person/number or case, but with a suffix from an inappropriate inflection class. The results also revealed cross-linguistic differences, which will be discussed with regard to stem changes, syncretism and suffix variation.

Only when using a continuous phonological measure of error instead of a binary correct/incorrect distinction, the modeling additionally indicated that in all languages the effect of phonological neighborhood was smaller for items of higher token frequency (as predicted by storage+analogy, but confirmed experimentally for Polish nouns only), suggesting that the binary measure used in the experiments may be insufficiently sensitive for some effects to be detected.

Hierarchical clustering of the models’ internal representations revealed that lemmas were grouped on the basis of phonological similarities that included items from a different class (example in Figure 1). Errors could therefore be better predicted when defining phonological neighborhood in terms of friends (neighbors with the same inflection) and enemies (neighbors with a different inflection).

Our findings demonstrate that acquisition of even highly complex systems of inflectional morphology can be accounted for by an input-based theoretical model that assumes rote storage and phonological analogy. In addition, the simulations suggest that future behavioral and computational studies should explore the possibility of more sensitive, non-binary dependent variables, and use more sophisticated measures of phonological neighborhood density as a predictor of errors in children’s speech.