关键词: 语音统计学习, 转换概率, 词长期待, 韵律

Abstract:

Statistical Learning (SL) has long been established as a powerful mechanism in language learning and development. Within this framework, transitional probability (TP) of various levels have been shown to confer differing task performance for adults. Recent studies have also highlighted the role of linguistic experience in the process of SL. However, it remains unclear whether different word lengths as well as varying levels of TPs may impact the segmentation of continuous speech.

In the low TP condition, the superior outcome of disyllabic contrasts might stem from the Mandarin speakers' prior linguistic experience—their expectation that words should be of two syllables. For the trisyllabic contrasts, lower TPs may provide relatively weakened statistical regularities for tracking word boundaries, which may in turn lead to difficulty extracting words. Importantly, our findings show that when both factors present difficulties (e.g., trisyllabic contrasts in the low TP condition), such that the word length violates the listeners' expectation and the TPs do not provide high levels of consistency, word segmentation can no longer be supported. The current study showed for the first time that when combining TP and word length in the speech material, these two factors impact word segmentation in a complex manner. This study offers new insight for future SL designs as well as potentially informative directions in exploring how individual differences based on linguistic backgrounds may manifest itself in word segmentation tasks.

Sixty native Mandarin monolinguals participated in a word segmentation task. An artificial language was designed with the same flat tone paired with 13 syllables, resulting in two disyllabic and three trisyllabic monotonic words. While only the segmental tier offered reliable information to segmentation, information from the suprasegmental level ensured that each word was phonologically legal in Mandarin. The words were then combined into two conditions of a monotonic artificial language: for the hTP language, all TPs within words were 1.0; whereas they were 0.6 in the lTP condition. Two types of nonwords (trisyllabic and disyllabic) were created for the test phase, then paired with target words of equal length in each trial. Adults were first exposed to the monotonic artificial language and then tested in a 2 alternative forced-choice task (2AFC) to decide whether a word or a nonword sounded more familiar.

The mixed two-way ANOVA with word length (disyllables vs. trisyllables) as a within-subject factor and TP (high TP vs. low TP) as a between-subjects factor yielded nonsignificant effects for either word length or TP levels. There was also no significant interaction. A series of one-sample t-tests were conducted between the participants' average accuracy and chance level (.5). Participants preferred words over nonwords in most conditions except for trisyllabic contrasts in the low TP condition. Additional planned contrasts among the conditions revealed that when the TPs were low, participants performed significantly better in the disyllabic condition than in the trisyllabic condition; under the trisyllabic condition, high TP yielded better performance than low TP.

Key words: verbal statistical learning, expectation for word length, transitional probabilities, prosody