2016년도 제 10차 언어학 콜로퀴엄
발표자: Adam Albright (MIT)
일 시: 2016년 11월 25일(금요일) 오후 4시30분
장 소: 관정도서관 3층 양두석홀
제 목: Learning biases in the lab and in the mind
초 록: A central goal of modern linguistic theory is to explain typology: why do some patterns recur frequently, while others are rare or unattested? Within generative linguistics, a common strategy has been to posit that unattested patterns correspond to `impossible’ grammars. However, attested but rare patterns pose a challenge: clearly, grammars that derive them must be possible, so what accounts for their low frequency? One common response is to hypothesize that some grammatical preferences are biases, rather than absolute restrictions (Wilson, Moreton, Hayes, White, Do, Green, and others). However, it is also likely that many patterns are rare for learnability or diachronic reasons (Blevins 2004, Stanton 2016), or other non-grammatical reasons such as colonialization and language contact. In order to test the hypothesis that there are universal biases, we need additional, converging evidence that these restrictions are `synchronically active’, for example, by studying how language is learned. If we can show using controlled comparisons that some phonological patterns are learned more slowly, less accurately, or using different mechanisms than others, then we would have evidence that humans are indeed biased towards certain patterns over others.
In this talk, I discuss a series of Artificial Grammar experiments, carried out in collaboration with Youngah Do (Hong Kong University), designed to test several different phonological preferences. The first concerns a bias against phonological alternations: when presented with phonological alternations that apply 100% of the time, participants frequently nonetheless prefer forms that obey paradigm uniformity. By varying the amount of data that participants receive, we can show that this is an untrained preference; learners bring to the task a prior assumption that paradigms will be uniform. A second type of preference is a generality bias: by training participants on alternations involving some segments and withholding data about others, or by giving participants conflicting data about different segments, we can show that learners nonetheless assume that processes target broad classes of phonologically similar segments. Computational modeling confirms that a model that incorporates paradigm uniformity and generality preferences provides the best match to participants’ preferences.
The final type of preference is a substantive preference for certain phonological processes over others. By presenting participants with two phonological alternations simultaneously, it is possible to compare how quickly or accurately they are learned. The results show that participants prefer certain alternations, such as final devoicing of voiced stops and intervocalic voicing of voiceless stops, over others, such as final nasalization of voiced obstruents and intervocalic spirantization of voiceless stops. These preferences mirror observed typological asymmetries: final devoicing of obstruents is well attested, but there are few languages with final nasalization. At the same time, there are interesting discrepancies between the preferences we observe in the lab and the typological data: in fact, intervocalic spirantization is well attested typologically. I discuss possible sources of these discrepancies, and ways that further experiments may help to shed light on the nature of substantive biases.