In this experiment, we set out to investigate whether lexical knowledge provides the basis for learning about structure. We hypothesized that infants would be able to learn from a brief exposure period to visually-presented grammars XYX or XYY when they were composed of items of which they knew the name (Known condition) but not when they did not or could not know the name of the item (Unknown condition and Nonsense conditions).
Our results showed that a very specific subset of the infants were able to generalize, namely boys in the Known condition, familiarized with XYY, in Block 1. These are admittedly weak results. However, the fact that these boys looked longer to the inconsistent XYX grammar than to the consistent XYY grammar indicates that learning from the familiarization phase was completed by the time the test phase started (looking to novel items typically indicates a completion of learning during familiarization). On the other hand, for the girls, learning during the familiarization phase did not seem to be completed, as seen by their by their longer looking times overall during the test coupled with their lack of discrimination between the grammars. While these results indicate that some learning seems to be possible, they are not robust in terms of power, and generalization of interpretation should be done with caution.
In addition, our findings in the Unknown and Nonsense conditions point to variation as a result of individual differences. In the Unknown condition, infants varied on whether they were "long lookers" or "short lookers" overall. Those infants who maintained their looking during the familiarization continued to maintain their looking during the test, although this maintenance in attention did not result in any preferences with respect to the test items. In the Nonsense condition, boys and girls differed, and this was influenced by whether they had a high or a low vocabulary. However, neither in the Nonsense condition, nor in the Unknown condition did these differences translate into preferences with respect to test items. Taken together, the results of the three conditions give us an indication that while in the Unknown and Nonsense condition, data varied with respect to participants’ individual differences, in the Known condition, data varied more with respect to the experimental manipulation.
These differences between the Known condition and the other two condi- tions indicate that lexical knowledge could in fact be playing a role in learning about the pattern. This may be because the lexical knowledge of the items helps the infants to analyze the stimuli, as they can potentially name the items covertly as they appear on screen. Lexical knowledge could thus be helping to direct infants’ attention to the pattern, to categorize the stimuli, and to memorize the regularity during familiarization.
3.5.1
Conclusion and future work
Our results, although not robust, indicate that lexical knowledge of items pre- sented in a rule learning task could aid in pattern generalization. The experi- ments must be replicated, and extended upon, to verify that the effect found with the specific set of participants here was not a chance finding. Nevertheless, the results of this study help us direct our attention for future work.
What we can say with more certainty is that in the conditions in which familiarization triads consisted of familiar items with unknown names, or of nonsense items, there was no evidence that infants learned an underlying XYX or XYY rule. Thus, rather than the mere familiarity of items, the crucial fa- cilitating factor in learning rules with visual stimuli may be the availability of a lexical specification of the pictured item. If learning in Saffran et al. (2007) simply occurred because of the familiarity of the pictured items, infants in our Known and Unknown conditions, which both contained items that were famil- iar to them, should have performed comparably. Instead, we found that infants had much longer looking times in the Known condition, both during familiar- ization and during test, than in the other conditions. We found an indication of learning only in this Known condition, although both exposure grammar (XYY) and sex (male) of the participant played a role as well.
In order to confirm the findings presented here, more infants should be tested particularly in the Known condition. In addition, several further consid- erations should be mentioned with respect to follow-up work.
The role of repetition
The only significant results we found were in boys trained with the XYY gram- mar. In much of the literature, and also in our own results presented in chapter 2, we find that repetition is a highly facilitating factor. Gómez, Gerken, and Schvaneveldt (2000) argue that repetition is the basis of transfer in AGL tasks. In their work with adults, they found that participants could only generalize from a familiarized finite state grammar to novel input on the basis of the pres- ence of repetition. Experiments looking specifically at Marcus-type grammar learning with both infants (S. P. Johnson et al., 2009) and adults (Endress et al., 2007; Chen et al., 2015) have shown that repetition is a highly salient cue for rule learning. Repetitions, particularly in edge positions, have been identified as "perceptual or memory primitives" that form one of the principal mecha- nisms by which we are able to attend to and learn artificial grammars (Mehler et al., 2008; Endress et al., 2009). It may thus be that learning rules containing adjacent repetitions is the default, and that extra support is necessary to move beyond this perceptual bias.
In our Known condition, the fact that infants lost interest over time when being exposed to the XYY grammar, but not to the XYX grammar, can be interpreted as evidence that some level of familiarization with the repetition pattern had taken place. In addition, just as in S. P. Johnson et al. (2009)’s
findings with eight- and 11-month-old infants, boys in our Known condition were unable to generalize after exposure to the XYX pattern but showed evi- dence of generalization from the XYY grammar. Our findings thus add to the claim that the presence of the repetition during familiarization is a facilitating factor.
Differences by sex
While we specifically counterbalanced the number of boys and girls per exper- imental manipulation to make sure there would not be a confound of sex, with the numerous post-test removals of data points and participants, we did not always have an equal number of data points per factor for each sex. However, there is a potential difference between sexes worth exploring in future (visual) rule learning studies. Literature on differences between sexes in infant studies is scarce, making it difficult to integrate our results with previous findings. Nev- ertheless there is some evidence from linguistic tasks that young infants may already show a difference in preference and processing of patterns depending on their sex. For example, Shi (2007) showed that eight-month-old girls preferred frequent function words, while boys preferred infrequent ones. Ter Haar and Levelt (2018) found that 12-month old boys, but not girls, tended to prefer sets of frequent over infrequent CV syllables.
Differences between the sexes have also been found in previous rule learn- ing studies. In an ERP study, Mueller, Friederici, and Mannel (2012) found that three-month-old girls showed a mismatch negativity to stimuli violating a rule, while boys showed a positivity. While both detected the mismatch be- tween the standard and deviant, the authors concluded that the girls, whose response was similar to adults’, were more mature at this age with respect to their processing. Lany and Gómez (2008) also found that at 12-months-old, girls were better than boys at learning adjacent dependencies between word categories and generalizing the dependencies to non-adjacent positions. These studies indicate that there are differences between sexes in their processing of structured information that require further investigation.
Role of variety
Visual inspection of our stimuli (see Figures 3.1, A1) and those found in Saffran et al. (2007) reveal that, compared to the minimally different pictures of dogs and cats in the triads used in Saffran et al., the X and Y images we used within a triad were very different from each other. The minimal differences in Saffran’s stimuli may have played an important role in infants’ ability to learn rules. The X and Y pictures in our triads varied on multiple dimensions: shape, color, orientation, and identity value (i.e., items were not different versions of the same object). While this may have been interesting for infants overall, it may have impeded them from discovering an underlying rule. We know from previous research that there is a "sweet spot" of complexity with respect to the
stimuli presented to infants and their ability to learn at various levels (e.g., "The Goldilocks Effect" of Kidd, Piantadosi, & Aslin, 2010); learning cannot occur if stimuli are too complex, nor if they are too simple and are not informative enough. The uniformity of the triads in Saffran et al. (2007), consisting only of very similar dogs (or cats), might have helped the infants to understand that the triad formed some sort of unity (according to the Gestalt principles of perception and similarity), within which there was a pattern of same and different (XXY or XYX; Newport & Aslin, 2004). The triads in our study, on the other hand, did not form a clear unity, or Gestalt, in any of the conditions. If both lexical knowledge and Gestalt principles of similarity are facilitating factors, this would account for the stronger learning effects in Saffran et al. (2007). Further research looking at the effect of minimal differences between items forming a rule, and the interaction with lexical knowledge is the necessary next step in understanding the success of infants in Saffran et al. (2007).