Statistical learning supports word learning and memory

• In Event: 1-153 - Retention of statistical learning

Integrative Statement

To build a vocabulary, infants must not only extract individual words from their ambient language, but they must link those newly segmented words with meaning. There is ample evidence that the ability to track statistical regularities (e.g., transitional probabilities, TP) in speech may help infants tackle these challenges by providing candidate object labels that are readily available to be linked to meaning (e.g., Graf Estes et al., 2007; Hay et al., 2011). However, extracting candidate words from speech does not guarantee that the infants will learn the meaning of those words because objects may not be in infants’ immediate environment when the label is spoken. Thus, while figuring out what a recently segmented word refers to in a particular situation is helpful, remembering it for future use may be equally important for successful word learning. Across two experiments we explore infants’ memory for newly segmented words.

In Experiment 1, 22- to 24-month-old English-speaking infants (n=32) were familiarized with an Italian corpus that had 4 embedded target words (casa, bici, fuga, melo): 2 high TP words (HTP words, TP=1.0) – syllables did not appear anywhere else in the corpus – and 2 low TP words (LTP words, TP=.33) – first and second syllables appeared many other times in the corpus. In a counterbalanced language, HTP and LTP words were switched. After a 10-minute delay, infants were trained and tested on 2 HTP and 2 LTP label-object associations, using a Looking-While-Listening paradigm (Fernald et al., 2008). Infants showed above chance accuracy to look at the labeled object on HTP trials, t(31)=5.08, p<.001, but not on LTP trials (p >.7)(see Figures 1 & 2). Further, infants were significantly more accurate on HTP than LTP trials, t(31)=4.03 p<.001. These findings contrast with a recent study from our lab that demonstrate successful mapping of both HTP and LTP words when the referent training phase and testing occurred immediately after familiarization. Thus, when trained and tested immediately after familiarization both HTP and LTP words make good candidate object labels, however, when trained and tested following a delay, only HTP words are mapped to meaning. Together our findings suggest that words with strong internal statistics are more robustly stored in long-term memory than words with weaker internal statistics, and remain available to support word learning after a delay.

In a follow-up experiment we asked whether infants show differential memory for label-object associations that are formed immediately following initial segmentation, based on differences in the labels’ sequential statistics. Here, we replicate Experiment 1, but instead insert a delay between referent training and test. Now, infants (n=32) showed similar above chance accuracy to look at labeled objects on both HTP, t(31)=4.48, p<.001 and LTP test trials, t(31)=3.57, p<.01, suggesting that when infants were given the opportunity to map both HTP and LTP words to meaning immediately after familiarization, the meaning representations were successfully retained following a 10-minute delay. Together, these findings suggest that statistical information may be play a greater role in supporting memory for sound patterns than memory for meaning.

Authors

• Ferhat Karaman, University of Tennessee
  Non-Presenting Author

• Jessica Hay, University of Tennessee, Knoxville
  Presenting Author