Predicting errors in children’s production of verb morphology: Evidence from person/number marking in Finnish and Polish (1)

Introduction: Studies of naturalistic data have suggested that young children acquiring highly- inflected languages do so in a way that is largely error-free. However, low overall error rates in children’s production may hide higher error rates in certain parts of the paradigm (Rubino & Pine, 1998; Aguado Orea & Pine, 2015). To investigate this possibility, the current study examines children’s production of person/number marking in present tense verbs in two morphologically complex languages, Finnish and Polish, which differ in the complexity of the verb inflection pattern. Input-based accounts, which are based predominantly on English morphological learning (e.g., Matthews & Theakston, 2006), predict that both the input frequency of a particular word form (token frequency) as well as the density of a word’s phonological neighbourhood (PND), predict rates of learning. Therefore, we investigate not only whether children make errors in inflection, but also whether input-based accounts are able to predict where in the verb paradigm errors occur.

Method: Seventy-seven native Finnish-speaking children (46 females; mean age: 49.4 months; range: 35-63) and 81 native Polish-speaking children (43 females; mean age: 48.7 months; range: 35-59) participated in the study, which employed an elicited production paradigm. The stimuli consisted of 32 verbs in each language. The verbs were chosen across a range of surface form frequencies and from 8 (Polish) and 11 (Finnish) verb classes varying in PND, with counts taken from CDS corpora and standard grammar dictionaries respectively. Each verb was presented as an action in a video on a laptop computer. Children were shown animations of different characters (1st, 2nd, 3rd person; singular and plural) performing the various actions, and they produced both the pronoun and the inflected present-tense form of the verb (e.g., “Minä uin” I swim-1sg; “Sinä imuroit” You hoover-2sg).

Results: Analysis with mixed-effects models revealed that, for both languages, despite low overall error rates (7-8%), children made more errors with verb forms with lower token frequencies in the input (Finnish: β=0.38, SE=0.06, χ2(1)=39.14, p<0.0001; Polish: β=0.26, SE=0.05, χ2(1)=29.47, p<0.0001) and with verbs belonging to classes with lower PNDs (Finnish: β=0.17, SE=0.07, χ2(1)=5.78, p=0.016; Polish: β=0.21, SE=0.07, χ2(1)=8.36, p=0.004) (see Figures 1 and 2). In Finnish, but not in Polish, older children also produced fewer errors than did younger children (β=0.08, SE=0.03, χ2(1)=8.71, p=0.0032). The interaction between token frequency and PND was not significant in either language. Analysis of the children’s errors indicated that the types of errors made by children were influenced by the type of inflectional paradigm in the language.

Conclusion: These findings suggest that successful models of children’s acquisition of verb morphology need to be sensitive to the statistical properties of children’s input, i.e., both token frequency (reflecting children’s retrieval of individually stored verb forms) and PND (children’s use of phonological analogy), and the type of inflectional paradigm children are learning.