2018 Friday Session A 1000

Friday, November 2, 2018 | Session A, East Balcony | 10am

The Learnability of Semantic Distinctions: Insights from an Artificial Language Learning Experiment on Evidentiality
D. Saratsli, S. Bartell, A. Papafragou

It is often assumed that cross-linguistically more prevalent distinctions are easier to learn (Typological Prevalence Hypothesis – TPH).1 Prior work supports this hypothesis in phonology, morphology and syntax2,3,4 but has not addressed semantics. Furthermore, tests of the TPH with children are complicated (e.g., because of the potential role of cognitive development). Here we ask whether the TPH predicts the relative learnability of semantic distinctions in a domain that is not grammaticalized in English and can be taught to adults without native language interference: evidentiality (the encoding of information source).

Cross-linguistically, there are three common types of evidential morpheme: Direct (firsthand/perceptual evidence), Inferential (inference based on evidence), and Reportative (hearsay).5 In general, evidential systems mark Reportative or Inferential access (systems that only mark Direct access are rare)6 ; the most widespread evidential system involves only Reportative morphemes.5 According to TPH, Reportative-marking systems should be the most learnable, while Direct-marking systems the least learnable. We test this prediction using Artificial Language Learning (ALL). A previous learnability study7 on evidentiality offered preliminary support for TPH. However, that study used static pictures where Reportative access alone was marked with a salient visual cue that could have boosted system learnability. The current study uses dynamic videos whose visual characteristics are consistent across systems.

English speakers (n=101) were exposed to an “alien” language that was similar to English but had a novel verb-final morpheme, ga, and had to figure out what ga meant. They were shown 21 videos in which a girl gained access to an event through observation of someone’s action (Direct), inference from visual clues (Inferential), or report (Reportative; 7 videos per access type). For each video, the girl’s access to the event was controlled by a third character (Fig.1). At the end of each video, the girl produced a sentence with or without ga. There were three between-subject conditions depending on system (whether ga marked Direct, Inferential or Reportative access). Participants later completed a Production task: they watched 12 new videos (4 per access type) and had to complete the girl’s sentences with ga if appropriate. They also completed a Comprehension task: they watched 36 videos (12 per access type), on half of which the girl made errors in the use of ga (50% misses, 50% incorrect inclusions), and had to say whether ga was used correctly or not. A one-way ANOVA conducted on composite Production and Comprehension scores revealed an effect of System (F(2,98)=6.535, p<0.01; Fig.2). Pairwise comparisons (Bonferroni) revealed a significant advantage of the Reportative system over the Direct (p=.004) and the Inferential system (p=0.01).

Our data support the TPH, since the typologically prevalent Reportative evidential system was learned best and the rare Direct worst. Furthermore, our data support the conjecture that, cross-linguistically, indirect sources seem to be marked preferentially (and acquired more easily) compared to direct sources. We discuss this pattern in terms of the pragmatic need to mark indirect (and potentially more unreliable) over direct sources of information.