“Grounding the Ungrounded: Estimating Locations of Unknown Place Names from Linguistic Associations and Grounded Representations”, 2014 (; backlinks; similar):
Spatial locations can be extracted from language statistics, based on the idea that nearby locations are mentioned in similar linguistic contexts, akin to Tobler’s first law of geography. However, the performance of language-based estimates is inferior to human estimates, raising questions about whether human spatial representations can actually be informed by such (inferior) statistics.
We show that alternative methods of computing co-occurrence statistics:
improve language-based estimates, illustrating that simple linguistic associations may in fact inform spatial representations. Most importantly, we show that by bootstrapping from grounded city locations, linguistic associations can be exploited to accurately estimate the locations of unknown cities, as well as human estimates of city locations.
These results support the hypothesis that (un-grounded) linguistic associations can be productively combined with pre-existing spatial representations to yield new grounded representations, shedding light on the issue of symbol grounding in cognition.
[Keywords: symbol grounding, geography, embodiment, symbolic cognition, embodied cognition, pointwise mutual association, latent semantic analysis]
See Also:
GeoLLM: Extracting Geospatial Knowledge from Large Language Models
SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery
Tile2Vec: Unsupervised representation learning for spatially distributed data
Deep Learning the City: Quantifying Urban Perception At A Global Scale
Can Language Models Encode Perceptual Structure Without Grounding? A Case Study in Color
Semantic projection recovers rich human knowledge of multiple object features from word embeddings