Enhancing Distant Analogical Retrieval via Generating Abstract Redescriptions of the Target

VERONICA D'ANGELO; MAXIMO TRENCH

Simposio; AAAI FALL SYMPOSIUM CONCEPTUAL ABSTRACTION AND ANALOGY IN NATURAL AND ARTIFICAIL INTELLIGENCE; 2020

ASSOCIATION FOR THE ADVANCEMENT OF ARTIFICIAL INTELLIGENCE

Several empirical studies demonstrate the difficulties experienced by humans and machines for retrieving semantically distant analogs from memory. This bottleneck in analogical processing has been explained in terms of the computational cost implied in performing a full structural match between the target and every candidate representation in LTM. Accordingly, computer simulations (e.g., MAC/FAC; Forbus et al., 1995) divide retrieval in two distinct phases: a computationally cheap filter that selects a handful of candidates based on the presence of overlapping concepts, and a more costly structural matcher that compares such candidates with the target. Based on the vector normalization process assumed to operate on the base and target content vectors whose dot products constitute the initial phase of the retrieval algorithm, Gentner et al. (2009) obtained behavioral and computational evidence for the idea that removing object attributes from the target can enhance the retrieval of base analogs lacking surface similarity (the "late abstraction principle"). Given that in the behavioral and computational demonstrations of this idea the abstract targets were obtained by providing a second analogous situation and comparing (or intersecting) it with the original target, it is crucial to assess whether humans and machines can capitalize on the late abstraction principle without being externally provided with additional information about the target. In a first experiment, we presented a group of participants with an instructional material focused on arriving at abstract representations of situations by means of replacing concrete nouns such as car or policeman by indefinite pronouns such as something or somebody, respectively. In line with the hypothesis that removing object properties from the target representation could enhance distant retrieval by suppressing surface content in favor of relational structures, we predicted that the deployment of indefinite pronouns would render distant retrieval more likely. However, results showed that even though participants somehow managed to apply this technique to redescribe the target problem, the retrieval of the source situation did not prove statistically superior to that of a control condition who was not asked to elaborate on the problem's representation.We speculate that the replacement of nouns by indefinite pronouns might have suppressed not only irrelevant features of the problem's entities, but also properties of such entities that are relevant for the role they play in the target situation (Falkenhainer, 1990). On the other hand, the replacement activity might have been so mechanical that students could have performed it in a localist manner, that is, without maintaining the whole structure of the target active in working memory. A handful of computer models of analogical reasoning had stressed that the activation of large propositional structures in working memory is neither psychologically plausible (Hummel & Holyoak, 1997) nor advantageous for interdomain retrieval (e.g., Finlayson & Winston, 2006). In view of the above considerations, a more efficient search for superficially-dissimilar analogs as complex as those typically used in the problem-solving literature would consist in fostering a more concise representation of the target problem. In a second experiment, we assessed the effectiveness of asking participants to briefly characterize the role played by key objects within the particular situation depicted by the target problem. To this end, two groups of participants read the base problem plus two distracter stories. After a contextual separation, participants in the abstraction condition received a brief instructional material on how to describe central objects of a problem according to the situation-specific role they played in relation to the other elements of the situation. Before attempting to solve the target problem, they were asked to generate relational descriptions of key objects of the target problem, as well as to think of other known situations involving objects that are amenable to similar descriptions.Participants in a control condition received the target problem after completing an unrelated task whose time demands were similar those of the instructional material received in the abstraction condition. In contrast with the results of Experiment 1, the retrieval rate of the experimental condition (43.6%) was higher than that of the control condition (20.7%).The computational and instructional implications of this finding are discussed.