No Abstract so we have included the conclusions from the chapter.
Press here to buy this book!
or see complete book contents
The past decades have seen thousands of studies investigating cerebral asymmetries, with much of this work resulting in well formed theoretical models. It is somewhat surprising that only one computational model was found in our investigation of the literature, and it does not deal with the variety of semantic effects found in the literature. Part of this paucity of computational models of semantics has to be due to what has, until recently, been the rather intractable problem of representing semantics in ways to capture both the richness of the words and concepts used in language, as well as the stimuli used in the diversity of experiments that have been conducted.
Our goal has been twofold for this chapter. First, we wanted to present a model of concept acquisition in which any word could have an accompanying representation that could then be used empirically. Elsewhere (Burgess & Cottrell, 1995; Burgess & Lund, in press; Lund & Burgess, in press-a; Lund et al., 1995, 1996; also see Landauer & Dumais, 1996) we have suggested that our Hyperspace Analogue to Language, HAL, accomplishes this. However, HAL is not a complete model of semantics. We propose that it richly captures the initial, bottom-up, component of semantic activation that can be characterized as automatic. There is nothing in the representations themselves that encodes higher-level information processing such as problem solving or attention. Secondly, we wanted to implement HAL's vector representations in a processing mode, such that the initial bottom-up component of semantic activation as a function of semantic similarity could be integrated, at least to some degree, with collocation frequency. Collocation frequency appears to be a reasonable candidate to account for some of the variance for which one might typically invoke some attentional component. We think that semantic similarity and collocation frequency go a long way toward developing the representational component of the model. A long series of experiments have implicated timecourse of activation as an important aspect of memory retrieval. The processing component was implemented by incorporating timecourse with activation onset and activation decay rates. By evolving a set of parameters for each hemisphere and stimuli set such as can be done with this implementation (see Equation 3), one can simulate any DVF single-word priming experiment. There are currently limitations on implementing procedural aspects of an experiment such as masking primes or targets or modeling word-length effects.
We presented empirical results from three experiments with very different stimuli and their accompanying simulations of the representational and processing components. The simulations involved representing the stimuli actually used in the original experiments and simulating the relevant processing component. Some experiments were selected because they provided a good opportunity to replicate with HAL several of the important findings in the literature, namely ambiguity retrieval and direct and mediated priming. The age of acquisition experiment was selected because it highlighted aspects of HAL's semantic representations that required, for a successful simulation, a different pattern of results than the earlier experiments. This experiment demonstrated that when cerebral asymmetries did not occur with human participants, they did not occur with the simulation either.
An advantage of HAL's vector representations of words is that stimuli from virtually any word-based experiment can be retrieved from the matrix. An assumption of the current implementation, and potentially a limitation, is that the two hemispheres do not differ in the nature of their representations. This is the first time we have incorporated collocation frequency as part of a HAL simulation. We recognize that the hemispheres might well differ in the nature of their representations, and that there are many ways in which that may happen. We suggest that collocation frequency might be an important candidate, as well as the semantic representations themselves.
There is initial evidence with deep dyslexic patients that the semantic representations used in HAL are better predictors of semantic paralexias than word association norms (Buchanan, Burgess, & Lund, 1996). An untested aspect of these representations is that they can be lesioned to model brain damage. Varying proportions of vector elements could be eliminated, and experiments that have evaluated various processing deficits could have their hypotheses tested with this approach.
The semantic representations that are acquired by HAL would appear to mimic stimuli from a broad range of experiments including the three presented in this chapter. We suspect that the continued evaluation of this processing account of cerebral asymmetries using HAL will bring both successes and failures. However, we are even more confident that developing computational approaches such as this one will result in an increase in testable representational and processing theories of cerebral asymmetries.