Learning relations between concepts: classification and conceptual combination Barry Devereux (Barry.Devereux@ucd.ie), Fintan Costello (Fintan.Costello@ucd.ie) Department of Computer Science, University College Dublin, Belfield, Dublin 4, Ireland. Abstract People interpret noun-noun compounds like “wind power” by inferring a relational link between the compound’s two constituent concepts. Various studies have examined how people select the best relation for a compound from a set of candidate relations. However, few studies have investigated how people learn such relations in the first place. This paper describes an experiment examining how people learn which relations are possible between concepts. Participants in this experiment learned artificial, laboratory controlled relations between pairs of items and then judged how likely those relations were for new pairs of items. The results showed that people’s judgement of relation likelihood was reliably influenced by the presence of facilitating features for relations and by the diagnosticity of features for relations. A simple exemplar-based model of classification, using both diagnostic and facilitating features, was applied to people’s judgements of relation likelihood. This model accurately predicted people’s judgements of relation likelihood in the experiment, using no free parameters to fit the data. Introduction When, in everyday discourse, people encounter noun-noun compounds such as “mountain stream” or “lake boat”, they interpret those compounds by inferring a relation that can be used to combine the two constituent concepts (inferring that a “mountain stream” is a stream that flows down a mountain, that a “lake boat” is a boat that sails on a lake). In theoretical accounts of conceptual combination, this process involves selecting the best relation for a compound from a set of candidate relations. Some theories give a standard set of candidate relations to be used in all compounds (Gagné Shoben, 1997), while others derive candidate relations from the internal structure of the concepts being combined (Costello Keane, 2000; Wisniewski, 1997). Many studies have investigated how people select the best relation for a given compound (e.g. Costello Keane, 2001, Wisniewski, 1996). However, there have been very few studies investigating how people learn and form these relations in the first place. In this paper we aim to fill this gap. The paper describes an experiment investigating how people learn relations between two sets of novel concepts. In the experiment we designed four different relations that could hold between artificial, laboratory-generated ‘beetle’ and ‘plant’ concepts. Participants learned these relations from sets of examples, with each example showing one sort of relation holding between one type of plant and one type of beetle. After learning, participants were shown new pairs of plants and beetles, and asked to say which of the four learned relations could hold between those two items. This experiment was designed to examine two different possible factors in people’s learning of relations between concepts: the presence of diagnostic features for those relations, and the presence of facilitating features. By diagnostic features for a relation we mean features of a constituent concept that are strongly associated with a particular relation. Diagnostic features are most familiar in the case of single concepts: for example “has four legs” and “is made of wood” are diagnostic features for the single concept chair: most things that are chairs have those features, and most things that are not chairs do not. Similarly, the feature “has a flat surface raised off the ground” might be diagnostic for the relation is-sat-on-by: most instances of the is-sat-on-by relation have that feature; most instances of other relations do not. In the experiment we asked whether people would use the diagnostic features for relations when selecting likely relations for beetle-plant pairs. By facilitating features we mean the features of a pair of concepts that are necessary for a given relation to be possible, and without which that relation cannot hold. For example, while the compound “steel chair” can easily be interpreted using the made-of relation, the compound “kitchen chair” cannot possibly be interpreted as “a chair made of kitchens” simply because kitchens are not a type of substance. Being a substance is a necessary facilitating feature for an item to take part in the made-of relation. Again, in the experiment we asked how people would use such facilitating features when selecting likely relations for beetle-plant pairs. This paper is organised as follows. In the next section we discuss the representation of relations in terms of sets of examples, as used in our experiment. We then describe the experiment in detail. To foreshadow the results, we found that both diagnostic and facilitating features had a reliable influence on people’s selection of likely relations for pairs of items. We then describe how an exemplar-based model of concept conjunction (Costello, 2000, 2001) can be applied to the results of this experiment, giving a close fit to people’s judgements of relation likelihood. Finally, we conclude by discussing the implications of our findings for theories of conceptual combination. Learning Relations from Exemplars Our primary assumption is that the relations selected during conceptual combination are essentially categories, just as the concepts that they link are essentially categories. We use an exemplar representation to describe these relational categories. Exemplar theories of classification, which propose that a category is represented as the set of remembered instances of that category and that new items are classified on the basis of their similarity to those instances (e.g. Medin Schaffer, 1978; Nosofsky, 1984), have successfully accounted for a number of patterns seen in people’s learning of single categories. We extend the exemplar approach to allow both relations and the concepts that they link to be represented by sets of instances.
|Title of host publication||Proceedings of the Cognitive Science Society|
|Publication status||Published - 01 Jan 2004|