skip to main content

This content will become publicly available on January 1, 2023

Title: Predicting Human Judgments of Relational Similarity: A Comparison of Computational Models Based on Vector Representations of Meaning
Computational models of verbal analogy and relational similarity judgments can employ different types of vector representations of word meanings (embeddings) generated by machine-learning algorithms. An important question is whether human-like relational processing depends on explicit representations of relations (i.e., representations separable from those of the concepts being related), or whether implicit relation representations suffice. Earlier machine-learning models produced static embeddings for individual words, identical across all contexts. However, more recent Large Language Models (LLMs), which use transformer architectures applied to much larger training corpora, are able to produce contextualized embeddings that have the potential to capture implicit knowledge of semantic relations. Here we compare multiple models based on different types of embeddings to human data concerning judgments of relational similarity and solutions of verbal analogy problems. For two datasets, a model that learns explicit representations of relations, Bayesian Analogy with Relational Transformations (BART), captured human performance more successfully than either a model using static embeddings (Word2vec) or models using contextualized embeddings created by LLMs (BERT, RoBERTa, and GPT-2). These findings support the proposal that human thinking depends on representations that separate relations from the concepts they relate.
Authors:
; ; ;
Award ID(s):
1827374
Publication Date:
NSF-PAR ID:
10330145
Journal Name:
Proceedings of the 44th Annual Meeting of the Cognitive Science Society
Sponsoring Org:
National Science Foundation
More Like this
  1. By middle childhood, humans are able to learn abstract semantic relations (e.g., antonym, synonym, category membership) and use them to reason by analogy. A deep theoretical challenge is to show how such abstract relations can arise from nonrelational inputs, thereby providing key elements of a protosymbolic representation system. We have developed a computational model that exploits the potential synergy between deep learning from “big data” (to create semantic features for individual words) and supervised learning from “small data” (to create representations of semantic relations between words). Given as inputs labeled pairs of lexical representations extracted by deep learning, the modelmore »creates augmented representations by remapping features according to the rank of differences between values for the two words in each pair. These augmented representations aid in coping with the feature alignment problem (e.g., matching those features that make “love-hate” an antonym with the different features that make “rich-poor” an antonym). The model extracts weight distributions that are used to estimate the probabilities that new word pairs instantiate each relation, capturing the pattern of human typicality judgments for a broad range of abstract semantic relations. A measure of relational similarity can be derived and used to solve simple verbal analogies with human-level accuracy. Because each acquired relation has a modular representation, basic symbolic operations are enabled (notably, the converse of any learned relation can be formed without additional training). Abstract semantic relations can be induced by bootstrapping from nonrelational inputs, thereby enabling relational generalization and analogical reasoning.

    « less
  2. The ability to generate and process semantic relations is central to many aspects of human cognition. Theorists have long debated whether such relations are coarsely coded as links in a semantic network or finely coded as distributed patterns over some core set of abstract relations. The form and content of the conceptual and neural representations of semantic relations are yet to be empirically established. Using sequential presentation of verbal analogies, we compared neural activities in making analogy judgments with predictions derived from alternative computational models of relational dissimilarity to adjudicate among rival accounts of how semantic relations are coded andmore »compared in the brain. We found that a frontoparietal network encodes the three relation types included in the design. A computational model based on semantic relations coded as distributed representations over a pool of abstract relations predicted neural activities for individual relations within the left superior parietal cortex and for second-order comparisons of relations within a broader left-lateralized network.« less
  3. Analogy problems involving multiple ordered relations of the same type create mapping ambiguity, requiring some mechanism for relational integration to achieve mapping accuracy. We address the question of whether the integration of ordered relations depends on their logical form alone, or on semantic representations that differ across relation types. We developed a triplet mapping task that provides a basic paradigm to investigate analogical reasoning with simple relational structures. Experimental results showed that mapping performance differed across orderings based on category, linear order, and causal relations, providing evidence that each transitive relation has its own semantic representation. Hence, human analogical mappingmore »of ordered relations does not depend solely on their formal property of transitivity. Instead, human ability to solve mapping problems by integrating relations relies on the semantics of relation representations. We also compared human performance to the performance of several vector-based computational models of analogy. These models performed above chance but fell short of human performance for some relations, highlighting the need for further model development.« less
  4. Fitch, T. ; Lamm, C. ; Leder, H. ; Teßmar-Raible, K. (Ed.)
    Is analogical reasoning a task that must be learned to solve from scratch by applying deep learning models to massive numbers of reasoning problems? Or are analogies solved by computing similarities between structured representations of analogs? We address this question by comparing human performance on visual analogies created using images of familiar three-dimensional objects (cars and their subregions) with the performance of alternative computational models. Human reasoners achieved above-chance accuracy for all problem types, but made more errors in several conditions (e.g., when relevant subregions were occluded). We compared human performance to that of two recent deep learning models (Siamesemore »Network and Relation Network) directly trained to solve these analogy problems, as well as to that of a compositional model that assesses relational similarity between part-based representations. The compositional model based on part representations, but not the deep learning models, generated qualitative performance similar to that of human reasoners.« less
  5. The ability to recognize and make inductive inferences based on relational similarity is fundamental to much of human higher cognition. However, relational similarity is not easily defined or measured, which makes it difficult to determine whether individual differences in cognitive capacity or semantic knowledge impact relational processing. In two experiments, we used a multi-arrangement task (previously applied to individual words or objects) to efficiently assess similarities between word pairs instantiating various abstract relations. Experiment 1 established that the method identifies word pairs expressing the same relation as more similar to each other than to those expressing different relations. Experiment 2more »extended these results by showing that relational similarity measured by the multi-arrangement task is sensitive to more subtle distinctions. Word pairs instantiating the same specific subrelation were judged as more similar to each other than to those instantiating different subrelations within the same general relation type. In addition, Experiment 2 found that individual differences in both fluid intelligence and crystalized verbal intelligence correlated with differentiation of relation similarity judgments.« less