Using variability to guide dimensional weighting: associative mechanisms in early word learning

Abstract

At 14 months, children appear to struggle to apply their fairly well-developed speech perception abilities to learning similar sounding words (e.g., bih/dih; Stager & Werker, 1997). However, variability in nonphonetic aspects of the training stimuli seems to aid word learning at this age. Extant theories of early word learning cannot account for this benefit of variability. We offer a simple explanation for this range of effects based on associative learning. Simulations suggest that if infants encode both noncontrastive information (e.g., cues to speaker voice) and meaningful linguistic cues (e.g., place of articulation or voicing), then associative learning mechanisms predict these variability effects in early word learning. Crucially, this means that despite the importance of task variables in predicting performance, this body of work shows that phonological categories are still developing at this age, and that the structure of noninformative cues has critical influences on word learning abilities.

Figure 1

A) Results of learning with non-varying pitch values (e.g. Stager & Werker, 1997). VOT values become strongly predictive of visual target identity, while trained F0 values become strongly associated with both visual targets. Note: this figure only presents F0 data for simplicity; indexical values would be expected to mirror F0 data. B) Results of learning with variable F0 values (e.g. Rost & McMurray, 2009). VOT values are strongly predictive of visual target identity, while F0 values are very weakly and diffusely related to visual targets.

Figure 2

Schematic of the SST and RM models.

Figure 3

Expectation values generated from testing the RM and SST models. The recognition threshold must fall between the activation of the match and mismatch items in the RM model; both match and mismatch activation in the SST will exceed any threshold between these values.

Figure 4

Expectation values generated from the lif and neem model. When trained on buk and puk with the SST methodology, both units are activated above the recognition threshold from the RM buk and puk simulations. However, for lif and neem, only the matching unit exceeds the recognition threshold, while the mismatching unit falls below the threshold.

Figure 5

Expectation values generated by the RM model, compared against various iterations of the SST model with different variability in F0. The lower panel displays histograms of the distribution of F0 values in the different conditions. In the SST model with no F0 variability and with small F0 variability, both the match and mismatch expectation values exceed the match value for the RM model. When a large amount of F0 variability was added, the SST mismatch activation began to enter thethreshold range from the RM simulation.

Figure 6

Results from parameter manipulations. The vast majority (86%) of models showed the empirical pattern of results, in which match was greater than mismatch activation for both RM and SST models; and both match and mismatch activation for the SST model were greater than match activation for the RM model. Of models that failed, 6% failed because they were incapable of learning the items at all (often due to very small learning rates); 8% failed because the threshold for learning in the RM model also predicted learning in the SST model.

Figure 7

Expectation values generated from simulations of Rost and McMurray (2010) compared to expectation values from the RM model. In the Rost and McMurray (2010) simulations, F0 and indexical values do not vary while VOT around prototypical values; both match and mismatch activations in this simulation will exceed the recognition threshold for the RM model.

Figure 8

Schematic of the model used to simulate Ballem and Plunkett (2005).

Figure 9

Results of simulations of Ballem and Plunkett (2005). When presenting a correctly pronounced target, activation of the target exceeds that of the competitor: the model correctly activates the correct target. When a mispronunciation is presented, this difference decreased, representing children’s decreased looking preference for mispronounced targets.