Predicting Ad Hoc Categories with Word Embeddings

Authors

Alina Dracheva, Dartmouth CollegeFollow
Yiran Jiang, Dartmouth CollegeFollow
Alexandra Anderson, Dartmouth CollegeFollow

Student Co-presenter Names

Yiran Jiang, Alexandra Anderson

Files

Student Class

2026

Student Affiliation

Junior Research Scholar

First Advisor

Jonathan Phillips

First Advisor Department

Linguistics and Cognitive Science Program

Second Advisor

Alina Dracheva

Second Advisor Department

Linguistics and Cognitive Science Program

Description

The goal of the project is to develop a system which predicts what comes to mind for novel ad hoc categories using large language models. Ad hoc categories, formed in response to situational demands, reflect the flexibility of human thinking. Building on prior research into how word embeddings are situated in human feature spaces and representational techniques of language models, this project aims to investigate if language models can emulate the human process of creating ad hoc categories. Ad hoc category formation operates in a multidimensional feature space, where items with similar scores are clustered along contextually relevant dimensions. As such, word embeddings may model human conceptual representations, which our experiment aims to test. To test this, participants will interact with a web platform to generate members from novel prompts that combine base words and modifiers, for example, “A zoo animal you can bring on the plane”. Responses will be situated in a human-derived feature space and compared to word embeddings created using fastText. Dissimilarity matrices derived from the embedding space and the human derived feature space will be used to compare the structure of relationships between items in each space. This approach will reveal the dimensions most predictive of an ad hoc category and could be generalized to other bases. Findings would have broad implications on the representational potentials of large language models and their alignment with human conceptual organization.

Publication Date

2025

Keywords

computational linguistics, high-dimensional feature spaces, semantic projection

Disciplines

Cognitive Psychology | Computational Linguistics

Dartmouth Digital Commons Citation

Dracheva, Alina; Jiang, Yiran; and Anderson, Alexandra, "Predicting Ad Hoc Categories with Word Embeddings" (2025). Wetterhahn Science Symposium Posters 2025. 2.
https://digitalcommons.dartmouth.edu/wetterhahn_2025/2