Embedding models poorly capture real world measurements
Embedding Models Measure in Peculiar Ways
Computation and LanguageMachine Learning
Summary
Embedding models are designed to represent meaning, such as similarity between words or concepts, in a numeric space. This paper shows that when it comes to representing real-world physical measurements like mass, distance, time, and volume, these embeddings do not reflect these quantities accurately or meaningfully. Instead, the models seem to rely more on superficial features like how the measurement values look as text strings. Attempts to fix this misalignment only made small improvements.
What this means in practice
- •For natural language processing engineers: Avoid relying on embedding-based similarity for tasks involving physical measurements due to weak alignment with real-world quantities.
- •For data quality analysts: Spot potential misinterpretations of physical quantity data embedded in semantic models by recognizing their sensitivity to textual form rather than measurement accuracy.
Authors
Juri Opitz, Andrianos Michail
Abstract
Embedding spaces define notions of semantic similarity and distance. We study whether those embeddings reflect physical measurements of mass, distance, time and volume, which admit a unique, objective notion of semantic equivalence and distance. We find that physical measurement is only weakly modeled in the embedding space, and that instead quite peculiar measurement patterns can be observed. Further analysis indicates that embedding representations of physical measurements are strongly influenced by superficial string similarity, and recalibration of similarity does not substantially improve the alignment.