Search Abstracts | Symposia | Slide Sessions | Poster Sessions
A hierarchical categorical structure of abstract concepts
Poster Session A - Sandbox Series, Thursday, October 24, 10:00 - 11:30 am, Great Hall 3 and 4
This poster is part of the Sandbox Series.
Rutvik Desai1, Nandish Chokshi2, Sidney Crouse1, Fereshteh Kavandi1, Parth Desai3, Sophie Arheix-Parras1; 1University of South Carolina, 2BITS Pilani, Goa, India, 3University of California, Berkeley
Categories of concepts are classically divided into superordinate, basic level, and subordinate categories. Most work on conceptual categories is based on concrete concepts. It is not clear if similar hierarchical structure exists for abstract concepts, and if so, what the categories are. A few studies have examined this question using a relatively small number of abstract concepts that are typically restricted to nouns. Here, we used distributional vector embeddings to derive categories and hierarchical structure for a large, comprehensive set of abstract concepts across grammatical categories. A common characteristic of popular distributional embeddings such as GloVe or Word2Vec is that they assign high similarity to concepts that are thematically associated, such as cow-milk or dog-bone. Since we are interested in taxonomic similarity, this represents a confound. Here, we used embeddings that were developed to highlight taxonomic similarity. In these representations, taxonomically similar concepts (e.g., dog-wolf) receive greater similarity than associated concepts (dog-bone). We used approximately 10,000 abstract words for this study from the Brysbaert et al. concreteness rating database. One-third of the words with the lowest rating were selected to represent abstract words. Similarly, 1/3 of the words with the highest concreteness rating were selected as concrete words. We submitted taxonomic vectors corresponding to each word to hierarchical clustering, for both abstract and concrete sets independently. Words in categories at each level were then submitted to a Large Language Model (ChatGPT 4.0) to facilitate generation of labels for each category, which were then manually examined and edited. The results reveled hierarchical structure for abstract concepts with 27 levels. After broad top-level classes, several major categories were revealed. These included emotional and social concepts (morality, humanism, fraud, treason), qualities and states (persistence, insight, skill, democracy), emotional actions and states (tempt, flaunt, frustrate, obsess), neutral actions and states (devise, govern, develop, exist), thinking and communication (allege, explain, doubt, predict), intensifiers (fiercely, richly, gloriously, shockingly), properties (original, opulent, secret, decent), qualifiers (adaptive, invasive, fiscal, proven), manner or degree adverbs (willfully, negatively, poetically, profitably), and negative traits or beliefs (cheapness, obviousness, radicalism, imprecision). These large categories were further subdivided into multiple other categories of increasing specificity. Surprisingly, closed class words (are, with) did not form a separate category, but were included in various other categories. As a comparison and ‘sanity check,’ the concrete words also revealed an extensive hierarchy that matched many of the intuitive and known categories such as animals, tools, food items, and musical instruments. These results provide hypotheses regarding possible organization of abstract concepts in the mind and brain. We hope to test these hypotheses using behavioral methods such as masked priming, and with neuroimaging methods. The categories generated by these taxonomic embeddings can be compared against those generated with other distributional models, or with manually created multi-dimensional representations such as experiential ratings, and validated through behavioral and neuroimaging methods. Such large scale studies can provide a more comprehensive picture of representation and organization of abstract and even concrete concepts that goes beyond superordinate, basic level, and subordinate categories.
Topic Areas: Meaning: Lexical Semantics,