Portal de Periódicos da CAPES

Spontaneous Analogy by Piggybacking on a Perceptual System

2013; Wiley; Volume: 35; Issue: 35 Linguagem: Inglês

1551-6709

Marc Pickett, David W. Aha,

Topic Modeling

Spontaneous Analogy by Piggybacking on a Perceptual System Marc Pickett David W. Aha NRC/NRL Postdoctoral Fellow Washington, DC 20375 marc.pickett.ctr@nrl.navy.mil Navy Center for Applied Research in Artificial Intelligence Naval Research Laboratory (Code 5510); Washington, DC 20375 david.aha@nrl.navy.mil Abstract Source Most computational models of analogy assume they are given a delineated source domain and often a specified target domain. These systems do not address how analogs can be isolated from large domains and spontaneously retrieved from long-term memory, a process we call spontaneous analogy. We present a system that represents relational structures as feature bags. Us- ing this representation, our system leverages perceptual algo- rithms to automatically create an ontology of relational struc- tures and to efficiently retrieve analogs for new relational struc- tures from long-term memory. We provide a demonstration of our approach that takes a set of unsegmented stories, constructs an ontology of analogical schemas (corresponding to plot de- vices), and uses this ontology to efficiently find analogs within new stories, yielding significant time-savings over linear ana- log retrieval at a small accuracy cost. Target (a) Mapping 1 Spontaneous Analogy In our day-to-day experience, we often generate analogies spontaneously (Wharton, Holyoak, & Lange, 1996; Clement, 1987). That is, with no explicit prodding, we conjure up analogs to aspects of our current situation. For example, while reading a story, we may recognize a plot device that is anal- ogous to one used in another story that we read long ago. The shared plot device may be a small part of each story, it is usually not explicitly delineated for us or presented in iso- lation from the rest of the story, and we may recognize the analogy of the plot device even if the general plots of the two stories are not analogous. Somehow, we segment out the plot device and retrieve the analog 1 from another story in long- dormant memory. Spontaneous analogy is the process of ef- ficiently retrieving an analog from long-term memory given an unsegmented source domain such that part of the source shares structural similarity with the analog, though they might not share surface similarity. This process differs from stan- dard models of analogy, which are given a delineated source concept, and often a target concept. Given a pair of analogs, analogical mapping is relatively straightforward. The more difficult problem is finding the analogs to begin with. As Chalmers, French, and Hofstadter (1992) argue “when the program’s discovery of the correspondences between the two situations is a direct result of its being explicitly given the appropriate structures to work with, its victory in finding the analogy becomes somewhat hollow”. 1 In our terminology, an analog is substructure of a domain that is structurally similar to a substructure of another domain, and an analogical schema is a generalization of an analog. For example, an input domain might be the entire story of Romeo & Juliet, an analog would be the part of the story where Romeo kills Tybalt, who killed Romeo’s friend, Mercutio (like in Hamlet where Ham- let kills Claudius, who killed Hamlet’s father), and an analogical schema would be the generalized plot device of a “revenge killing”. Pterodactyls! Canyon (b) Spontaneous Retrieval Figure 1: An analog of Analogical Mapping vs. Sponta- neous Analogy. In Analogical Mapping (a), we are given an explicit source and target, free from interfering context. In spontaneous analogy (b), the analogs are spontaneously re- trieved from long-term memory. The process of spontaneous analogy shares some proper- ties with low-level perception, as exemplified in Figure 1. Within seconds of being presented with a visual image of a pterodactyl flying over a canyon, one can typically describe the image using the word “pterodactyl”, even if one has had no special explicit recent priming for this concept, indeed even if one has not consciously thought about pterodactyls for several years. For us to produce the word “pterodactyl”, we must segment the pterodactyl from the canyon and retrieve the “pterodactyl” concept from the thousands of concepts stored in memory. We must have learned the “pterodactyl” concept to begin with from unsegmented images. This perceptual pro- cess is robust to noise: The pterodactyl in the image could be partially occluded, ill-lit, oddly colored, or even drawn as a cartoon, and we are still able to correctly identify this shape (to a certain point). Likewise, many details of the plot devices from the above story example could be altered or obfuscated, but this analogy would degrade gracefully. Our primary technical contribution in this paper is an algo- rithm called Spontol 2 that solves the problem of spontaneous analogy: efficient parsing, storage, and retrieval of analogs from long-term memory. That is, given a corpus of many large unsegmented relational structures, Spontol discovers analog- ical schemas that are useful for characterizing the corpus and efficiently retrieves analogs given a new structure. E.g., given a set of narratives in predicate form, Spontol discovers plot 2 Spontol is short for “spontaneous analogy using the Ontol on- tology learning and inference algorithm”.