Panels 3, 4 and 5 are computed per model. Switch to see them disagree.
WordNet has no entry for nuanced. That is exactly the gap the model side fills: curated dictionaries lag behind how language is actually used, while a model learns whatever its training data contained.
Definitions come from WordNet, a hand-curated dictionary.
This is the tokenizer splitting text, not the model understanding it.
WordNet lists none for this word.
Shade shows how close the model puts each word to nuanced.
WordNet lists none for this word.
Shade shows how close the model puts each word to nuanced.
This model splits nuanced into 2 pieces, so it has no vector of its own here: this is the average of its fragments, and the results below are correspondingly rough.
WordNet lists none for this word.
Shade shows how close the model puts each word to nuanced.
Model neighbours are distributional, not dictionary synonyms. Two words can be close because they appear in similar sentences, which is why an antonym can outscore a synonym here.
This model splits nuanced into 2 pieces, so it has no vector of its own here: this is the average of its fragments, and the results below are correspondingly rough.
Scores are projections onto axes we defined from anchor words, not labels the model assigns.
These sit just outside the closest neighbours in panel 3, and no dictionary lists any of them as related to nuanced. That gap is the model’s own learned association.
These sit just outside the closest neighbours in panel 3, and no dictionary lists any of them as related to nuanced. That gap is the model’s own learned association.
This model splits nuanced into 2 pieces, so it has no vector of its own here: this is the average of its fragments, and the results below are correspondingly rough.
These sit just outside the closest neighbours in panel 3, and no dictionary lists any of them as related to nuanced. That gap is the model’s own learned association.
These are statistical associations in the training data, not the model thinking.