Probing the naturalness bias in learning stress A cross-linguistic study Matías Guzmán Naranjo & Joanna Zaleska (University of Leipzig)

In the present paper we report on a cross-linguistic artificial grammar learning study, whose aim was to probe Carpenter’s (2010) observation that it is easier to learn an artificial language in which stress is attracted to low vowels than one where stress occurs on high vowels. The results of our reserch do not confirm Carpenter’s findings and suggest that artificial language experiments can be affected by the native language of the participants even when it does not contain the relevant structures. Background Carpenter’s (2010) study compared the acquisition of two artificial languages which differed only in terms of their stress pattern. In language A, stress fell on the leftmost syllable with a low vowel, and otherwise on the leftmost syllable. The stress rule in language B differed from language A only along the substantive axis: stress was attracted by high, rather than low vowels. Though formally similar, the two rules differed in terms of ‘naturalness’. Only rules such as those in language A are cross-linguistically attested and conform to a universally-fixed constraint ranking (Stress Prominence Hierarchy (Smith 2002), based on Prince & Smolensky’s (1993) Peak Prominence Hierarchy). In Carpenter’s experiment, conducted with native speakers of English and Quebec French, participants in both groups were found to learn the ‘natural’ pattern significantly better than the ‘unnatural’ one. She concluded that her findings support the claim that language-specific language mechanisms aid the learners in distinguishing and learning phonological patterns. Two research questions lay at the core of our study. They were: 1. Can the results of Carpenter (2010) be replicated using new materials and different participants? 2. Can the native language of the participants be a confounding factor? To explore these questions, we conducted a study similar to Carpenter’s and ran it with speakers of three different languages: English, Polish, and Spanish. Experiments involving two additional groups of speakers (German and Czech) are currently being carried out. Stimuli We constructed two artificial languages for the purpose of the study, differing only in terms of their stress pattern. In the ‘natural’ language (L ANGUAGE A), stress fell on the rightmost nonhigh vowel, and otherwise on the rightmost vowel. In the ‘unnatural’ language (L ANGUAGE B), the stress rule was the mirror image of the natural version, with stress attracted by high vowels. Materials for all stages of the experiment were open trisyllabic words in a CVCVCV pattern. The consonant inventory was drawn from the set [p t k f s m n l], while the vowels were drawn from the set [a O i u]. Using these inventories, two sets of words, each made up of 32 items, were created, serving both the natural and unnatural language. Participants Three groups of speakers (between 20 and 36 years old), randomly assigned to one of the two conditions (natural vs. unnatural), were recruited for the experiments: • The total number of Polish speakers was 14 for language A and 16 for language B. • The total number of Spanish speakers was 15 for language A and 9 for language B. • The total number of English speakers was 10 for language A and 13 for language B. Procedure The experiment was coded using Experigen (Becker & Levine 2010) and accessed online. The subjects were randomly assigned to learn either the ‘natural’ or ‘unnatural’ language. In each condition there were three experimental stages: familiarization, and two test phases. In the familiarization stage, participants were exposed to the language for about 6 minutes, without explicit instruction. They heard the 32 familiarisation items, each repeated four times, accompanied by a photographic image corresponding to each item. Each sound/image combination was shown for 2.5 seconds. All tokens appeared in random order for all participants. The participants were then tested on their knowledge of the artificial language by having to select forms which agree with the rules of the language they have learnt. The first test phase examined participants’ knowledge of familiar items, while in the second test phase, unfamiliar items were used. Each test phase consisted of 32 trials. Results The results of the experiment for Polish speakers did not show any advantage for the natural pattern over the unnatural pattern for the first test (mean proportion of correct answers language A = 0.66, language B = 0.68, W = 123, p-value = 0.87), but there was a small but statistically significant advantage at the 0.05 confidence level for the unnatural pattern for the second test of the experiment

test_1 0.9

test_2

proportion of correct answers

0.8

0.7 NativeLanguage English 0.6

Polish Spanish

0.5

0.4

A

B

A

B

language

Figure 1: Box plots for each test for each language type for participants of all three languages. The dots represent the mean proportion of correct answers for each participant. They have been jittered on the x axis to avoid overlapping.

(mean proportion of correct answers language A = 0.53, language B = 0.62, W = 69, p-value = 0.04478). Overall, participants performed better during the first test than in the second test. For Spanish speakers there was no significant difference between the natural and unnatural patterns for either the first test (mean proportion of correct answers language A = 0.65, language B = 0.66, W = 84, p-value = 0.8997) nor the second test (mean proportion of correct answers language A = 0.57, language B = 0.61, W = 36, p-value = 0.1444). For the Spanish speakers the results also show that they performed considerably better during the first test than during the second test. For English speakers there was again no significant differences between the natural and the unnatural pattern for the first test (mean proportion of correct answers language A = 0.63, language B = 0.63, W = 65, p-value = 0.6948) nor for the second test (mean proportion of correct answers language A = 0.50, language B = 0.53, W = 49, p-value = 0.3434). Once more speakers performed considerably better during the first test than during the second test. A mixed-effect logistic regression fitted to the whole data set with participant as a random effect and NATIVE L ANGUAGE and PATTERN (natural-unnatural) as predictors revealed that the effect of NATIVE L ANGUAGE was in fact significant, with both Polish and Spanish speakers performing better than English speakers, and a small effect in terms of PATTERN showing that the unnatural pattern was overall easier to learn than the natural pattern. These results do not support the naturalness bias hypothesis, and they do not replicate the results presented by Carpenter. Furthermore, these results strongly support the hypothesis that native language can have a significant effect on artificial language learning experiments and should be controlled for. References Becker, Michael & Jonathan Levine. 2010. Experigen: an online experiment platform. Available at https://github. com/tlozoot/experigen Carpenter, Angela C. 2010. A naturalness bias in learning stress. Phonology 27, 345-392. Prince, Alan & Paul Smolensky. 1993. Optimality Theory: Constraint interaction in generative grammar. Technical Report, Rutgers University Center for Cognitive Science and Computer Science Department, University of Colorado at Boulder. Smith, Jennifer L. 2002. Phonological augmentation in prominent positions. Doctoral dissertation, University of Massachusetts, Amherst.