The Internal Structure of Semantic Relations: Effects of Relational Similarity and Typicality

Abstract

We examined whether the processing of semantic relations shows typicality effects similar to those found for the processing of entity concepts. Participants performed four relational processing tasks with the same set of word-pair stimuli: relational exemplar generation; similarity ranking; analogical verification; and a paired-associate learning task. In the similarity ranking task, we gathered separate rankings for relational, role and semantic similarity between word pairs. We found significant correlations at the item level among relational generation frequencies, analogical verification RTs/accuracy and relational luring in associative memory. Relational similarity predicted exemplar generation frequencies, analogical verification RTs/accuracy, and relational luring in associative memory. Role similarity predicted exemplar generation frequency, and analogical verification RTs, but not relational luring. Semantic similarity did not predict any of the tasks, after controlling for the other two factors. Contrary to current theories which posit that semantic similarity is more important for retrieving relevant analogues, and that analogical mapping is based on role-filler bindings, relational similarity was the strongest predictor across all tasks. These results suggest that just like entity concepts, semantic relations have an internal structure that gives rise to typicality effects across a variety of tasks, which could provide constraints for testing competing theories of relational representation.

Predicting patterns of similarity among abstract semantic relations.

“…The present findings address a longstanding question concerning the mental and neural representation of semantic relations: whether relations are simply implicit (e.g., defined by the vector distance between the representations of the entities being related, as assumed by Word2vec-diff), or have explicit representations based on the meanings of specific relations (Chaffin, 1992;Chaffin & Hermann, 1988b;Popov et al, 2017Popov et al, , 2020. Several lines of evidence now converge on the conclusion that semantic relations are coded explicitly by a distributed code, as postulated by the BART model.…”

“…Several lines of evidence now converge on the conclusion that semantic relations are coded explicitly by a distributed code, as postulated by the BART model. A distributed representation created by learning from examples is able to account for human behavioral data both for within-relation structure—the typicality gradient observed for semantic relations (Chaffin & Herrmann, 1988a; Jurgens et al, 2012; Lu et al, 2019; Popov et al, 2020)—and also for between-relation structure—the pattern of similarity judgments across a diverse set of relations, as shown here. In addition, the same basic model can predict patterns of neural similarity at the item level within a frontoparietal network for relations computed during a verbal analogy task (Chiang et al, 2020).…”

“…Human intuitions regarding semantic relations exhibit several complexities. Much like object categories (Rosch, 1975); examples of semantic relations form typicality gradients rather than simply being “all or none” (Chaffin, 1992; Chaffin & Herrmann, 1988a; Popov et al 2020). For example, hot–cold is considered a better example of the contrast relation than is warm–cool (Jurgens et al, 2012).…”

“…A key obstacle to evaluating models of relational similarity is a lack of systematic data on human judgments. Here we report human judgments of similarity between semantic relations (i.e., judgments involving multiple pairs of words), which have been shown to strongly predict human performance on a number of relational reasoning tasks including relation exemplar generation frequency, analogical verification accuracy and response time, and confusability of word pairs in episodic memory judgments (i.e., relational luring in associative memory; Popov et al, 2020). Here we use relational similarity data to evaluate alternative theoretical measures of relational similarity ultimately based on distributional semantics.…”

La Teoría de la Asignación Categorial Sobre el Pensamiento Analógico: un Desafío a la Teoría de la Proyección de Estructura

“…Entendemos que la propuesta de que los predicados relacionales pueden incluir todo tipo de elementos (e.g., entidades y propiedades de entidades, VEDs de un hecho) vacía de significado el concepto de relación, quedando reducido el concepto a "aquello que el sistema cognitivo percibió como una relación y que dispuso entonces como un predicado en una proposición de dos o más argumentos". Una caracterización de esta naturaleza evade una tarea que es fundamental en cualquier teoría del pensamiento analógico, a saber, brindar una caracterización de qué es una relación en términos psicológicos (Popov et al, 2020). Entendemos que, al menos para el tipo de analogías que estamos considerando, sería más productivo cambiar el concepto de relación (entendido como predicado de dos o más argumentos) por un concepto más amplio como, por ejemplo, estructura relacional, dentro del cual pudiera reconocerse un papel relacional a las entidades, sus PERs, las dependencias semánticas entre entidades y relaciones, etc.…”