Authors
Willem M Otte, Archibald L H M van Wieringen, Bart J Koet
Published in
PloS one. Volume 21. Issue 11. Pages e0362816. Epub Nov 20, 2026.
Abstract
The grounding problem concerns how symbolic representations acquire meaning through connections to what they refer to. For word embeddings learned from text, a related empirical question is how much of the structure that generated the text they retain. This is difficult to test in natural corpora because that structure is generally unknown. We employed a microsimulation methodology to generate a controlled text corpus from a known spatial structure. A virtual language supermarket with 40 shelves organised into four departments (fruit, vegetables, beverages, extras) served as the generative environment. Product sequences from 80,000 simulated shoppers formed a corpus of 804,422 product selections, with each sequence unique. The GloVe algorithm was applied to this corpus to learn 50-dimensional word embeddings, which we compared with the network of shelves. Cosine distance between products correlated with their shortest-path distance in the network (Spearman ρ = 0.79 over all 7,140 product pairs, Mantel p = 0.001), and on average 61% of a product's network neighbours were among its nearest neighbours in the embedding, against 7% expected by chance. Over ten independent simulations these values were 0.80 ± 0.01 and 0.62 ± 0.02, and shuffling words across the corpus removed the association. Raw co-occurrence counts, PPMI, SVD and word2vec also captured global distance structure but retrieved fewer neighbours (29-50%). Recovery was best with a context window of one word (ρ = 0.93) and was weaker for rare products. K-medoid clustering of all 120 products separated the fruit and extras departments completely, while vegetables and beverages, which are interleaved in the network, were mixed (adjusted Rand index 0.82). UMAP projections followed the network closely (ρ = 0.92) but distorted some parts of it. These findings show that distributional word embeddings can recover much of the structure that a generating process puts into co-occurrence statistics. They do not show that the embeddings are grounded in, or understand, what the words refer to. This controlled simulation provides a reproducible way to measure structural recovery by word embeddings.
PMID:
43063378
Bibliographic data and abstract were imported from PubMed on 21 Nov 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 2
- Comments 0