The role of textual coherence in incremental analogical mapping

Tate T. Kubose, Keith J. Holyoak, John E. Hummel
2002 Journal of Memory and Language  
The LISA model of analogical reasoning (Hummel & Holyoak, 1997 ) assumes that mapping is performed incrementally within limited-capacity working memory, and that processing is guided by principles of text coherence. Predictions of the model derived by computer simulations were tested in four experiments, using both semantically-impoverished structural analogs and semantically-rich story analogs. Results for college students revealed a grouping effect. Processing multiple propositions together
more » ... nerated more accurate mappings than did processing individual propositions, but only when propositions that jointly provided strong structural constraints were grouped together. Other experiments revealed asymmetries in mapping and inference accuracy: mappings and inferences generated from a more coherent analog to a less coherent analog were more accurate than those made in the reverse direction. Implications for computational models of analogical reasoning and for education are discussed. Kubose et al. 4 But although incremental analogical mapping is psychologically plausible, many of the incremental models do not provide a strong basis for the underlying assumption that WM capacity is inherently limited. For example, although I-SME can extend initial mappings if additional information is provided, it operates on whatever information about the analogs is given at once. Thus if entire analogs are initially provided (as in most experiments in the literature), I-SME becomes identical to SME, a model that has no particular capacity limit. IAM is more constrained, as it breaks down the information in the source analog into subsets, and begins by mapping only the subset with the most higher-order structure. However, IAM does not specify any limit on the size of that subset (i.e., the subset with the most structure can be arbitrarily large). The STAR model (Halford et al., 1994) does posit inherent capacity limits on analogical reasoning; however, that model does not handle mappings between hierarchicallystructured analogies of the sort that commonly arise in everyday problem solving and comprehension (Holyoak & Thagard, 1995) . The focus of the present paper is on the LISA model Hummel & Holyoak, 1997) . This model can map hierarchical structures, and is based on a type of architecture that provides a basis for estimating the maximum capacity available for mapping. LISA distinguishes WM from a broader active memory (cf. long-term working memory; Ericsson & Kintsch, 1995) , which is the active subset of long-term memory. During analogical mapping, the two analogs and the emerging mappings between them are assumed to reside in active memory. Following Cowan (1995), LISA assumes that active memory has no firm capacity limit, but rather is time-limited, with activation decaying in roughly 20 sec unless it is reactivated by allocation of attention. LISA assumes that within active memory, at any time a very small number of role bindings are in WM, and that these constitute the immediate focus of attention. In LISA's WM, distributed representations of predicates and their arguments are dynamically bound into Kubose et al. 5 propositional structures by synchrony of firing: Units representing the semantic features of predicate roles fire in synchrony with units representing the features of the fillers of those roles, and separate role-filler bindings fire out of synchrony with one another. For example, to represent the proposition Abe loves Cathy, units representing Abe (e.g., human, male, adult, etc.) fire in synchrony with units representing "lover" (e.g., emotion, positive, strong), while units for Cathy fire in synchrony with units for "beloved". Crucially, the Abe-as-lover set must fire out of synchrony with the Cathy-as-beloved set. The collection of active but mutually desynchronized role bindings is referred to as the phase set, and corresponds to the model's working memoryi.e., the collection of role-bindings it is thinking about "right now". The size of the phase set-i.e., the capacity of WM-is determined by the number of role-filler bindings (phases) it is possible to have simultaneously active but mutually out of synchrony. This number is necessarily limited (assuming the underlying neural substrate is subject to random noise), and its value is proportional to the length of time between successive peaks in a single phase (the period of the oscillation) divided by the duration of each phase (at the level of small populations of neurons) and/or temporal precision (at the level of individual spikes; see also Lisman & Idiart, 1995). There is evidence that binding is accomplished by synchrony in the 40 hz (gamma) range, meaning a neuron or population of neurons generates one spike (or burst) approximately every 25 ms (see Singer & Gray, 1995, for a review). If the temporal precision of spike timing is in the range of 4 -6 ms (Singer & Gray, 1995) , then with a 25 ms period, the capacity of WM ought to be approximately 25/5 = 5 role bindings (see also
doi:10.1016/s0749-596x(02)00011-6 fatcat:smi5wnjf2fclzcb7zexgo3byaa