2018 Friday Poster 6662

Friday, November 2, 2018 | Poster Session I, Metcalf Small | 3pm

Predicting Gender Assignment in the Acquisition of Icelandic
S. Bjoernsdottir, S. Sigurjonsdottir

Summary Using Icelandic as a case study, we propose that children learn grammatical gender by discovering predictive patterns of assignment, a process which can be quantitatively modeled by the Tolerance Principle [1].

Background Grammatical gender systems differ cross-linguistically. The assignment may rely on either semantic or formal criteria but may also be lexically arbitrary [2-3]. The acquisition of gender is also language specific: some show very early mastery [4-5] where others follow protracted developments [6-7]. Yet currently there is no mechanistic account of how children derive the language-specific rules of gender assignment.

Methods Icelandic has a 3-gender system (Masculine, Feminine, and Neuter) which interacts with case marking, thereby following formal criteria [2-3]. We propose that the morphological endings of the noun in the Nominative singular, by far the most frequent inflectional form, is predictive of Icelandic gender assignment. This process is accounted for by the Tolerance Principle [1, 8], an independently motivated model for detecting productive linguistic patterns: a rule applicable to N words is productive iff the number of exceptions is below the threshold of N/lnN. Because a productive rule can relatively few exceptions, the Tolerance Principle predicts that in some cases, no default gender exists if no gender is statistically dominant.

Results Our data draws from a half-million-word corpus of child and child-directed Icelandic [9; please give age range]. The distribution of nouns by gender and morphological ending can be found in Table I. The gender for nouns with inflectional endings -i, -a, and -r is nearly perfectly predictable, trivially meeting the Tolerance threshold. Indeed, these noun are target consistent from before 2;0 with no errors found.

Of particular interest are nouns with a morphologically uninformative (null) ending which are syncretic across the three classes. As Table 2 shows, there are 198 such nouns, with frequencies of 29 (M), 35 (F), and 134 (N) respectively. Despite N being a super-majority (134/198), the number of M and F (29+35=64) nouns exceed the tolerance threshold of 37 exceptions (198/ln198=37). Thus we predict the gender of null-marked nouns to be lexically arbitrary. This is borne out in the corpus data; see (1). Syncretic nouns constitute about 18% (38/212) that are attested in the child production data. Of the 38 syncretic nouns, 7 errors of overgeneralization are attested, with no systematic direction of errors–in contrast to cases with a gender default, which leads to over-regularization [6].

Conclusion Our methods can be applied to languages with semantic or phonological criteria for gender assignment and can make clear predictions about child language development on the child-directed input statistics. The no-default prediction under the Tolerance Principle also has important implications for other controversial topics in morphosyntactic acquisition as well (e.g., the no-default case of Polish masculine singular genitive; 1, 10).

References

1. Yang, C. (2016) The Price of Linguistic Productivity. MIT.

2. Corbett, G. (1991). Gender. Cambridge UP.

3. Comrie, B. (1999). J. Psycholing. Res.. 28(5), 457-466.

4. Gvozdev, N. (1961). Moscow: APN RSFSR.

5. Pérez-Pereira, M. (1991). J. Child Lg. 18(3), 571-590.

6. Rodina, Y. & Westergaard, M. (2015). J. Germanic Ling. 27(2), 145-187.

7. Szagun, G. (2004). Child Lg. 31(1), 1-30.

8. Schuler, K., Yang, C., & Newport, E. (2016). The 38th Cognitive Society Annual Meeting.

9. Sigurjonsdottir, S.(2007). The Fia Corpus.

10. Dąbrowska, E. (2001). J. Child Lg. 28(3), 545-574.