The independent effect of transitional probability on verbal statistical learning

Abstract: In a typical SL task, participants are first exposed to a nonsensical artificial language for 5~10 mins and then asked to finish a 2 alternative forced choice task (2AFC). Transitional probabilities (TPs), the core concept in SL, represent the predictably between syllables. In a given artificial language, syllables within a target word occur together more frequently, resulting in higher TPs than those of the syllables that span across word boundaries. The latter is referred to as partwords and consist of lower TPs. After the exposure phase, participants enter the test phase and are presented with a partword and target word in each trial of the 2AFC task. If the accuracy across participants is higher than chance level, it is assumed that learning has occurred. However, studies have also shown that factors other than TPs, such as word token frequency, word length variations (or the lack thereof) also impact SL performance in such tasks. To date, these factors as well as their interactive effects remain under studied.
In the experiment one, we aimed to investigate whether TPs affect SL learning performance when controlling for target words’ and partwords’ token frequencies. In doing so, we created the artificial language by randomizing the order of two trisyllabic words and two disyllabic words. During the 2AFC task, three types of items (target word, partword, and nonword) were paired together, with two items in equal length in each trial. There were 24 trials in the test. 40 native Mandarin monolinguals participated in the experiment; they first listened to the artificial language for 5mins and then finished the 2AFC task. In the experiment two, an artificial language was generated with 10 syllables and presented in exposure phase, to examine whether experiment one’s learning effect came from the TP or participants’ prior language bias.
Results in experiment one showed that the accuracy of all trials was significantly higher than chance (0.5) at the group level, suggesting that participants were able to segment the artificial language of mixed word lengths. Participants were also marginally better at choosing target words over partwords, and partwords over nonwords. To investigate the independent effect of TP in SL, we subset the data by word length and found that participants’ accuracy choosing trisyllabic target words over partwords was marginally lower than their choosing disyllabic target words over partword, which suggests that disyllabic words confer advantage in SL for this group of participants. In addition, participants’ accuracy in choosing trisyllabic partwords over nonwords was significantly higher than that of disyllabic target partwords over nonwords. In the experiment two, there were no significant learning effect in any levels when the statistical information was absent.
A series of results across two behavior experiments highlight the unique contribution from TPs alone, since accuracy was assessed by controlling for word token frequency and word length. Thus, the present study suggests that TP exerts effect on verbal SL performance independent of word token frequency. Further studies should take into account more types of statistical rule such as mutual information and backward TP.