Same-relation word offsets are parallel above a shuffled baseline

- On pre-trained Google News word2vec and the 40 BATS relations, the pairing-consistency score (AUC of true vs shuffled offset cosine similarities, chance 0.5) is 0.851 inflectional, 0.679 derivational, 0.559 encyclopedic, 0.539 lexicographic; all randomized analogy sets sit at about 0.500. [fournier-etal-2020-analogies-minus-analogy-test] - Offset concentration (mean cosine between normalized offsets) is only 0.295 / 0.156 / 0.198 / 0.031 on the same four relation types, and randomized pairs also give non-zero values, so parallel offsets alone do not show a coded relation. [fournier-etal-2020-analogies-minus-analogy-test] - The classic arithmetic analogy test is dominated by within-pair similarity (e.g. France~Paris), so it under-reports this offset regularity. [fournier-etal-2020-analogies-minus-analogy-test] - The same pattern (inflectional highest, semantic relations lowest) is reported for GloVe, dict2vec, ConceptNet Numberbatch and the static BERT and GPT-2 token embeddings; GPT-2 token embeddings reach 0.732 on derivational morphology versus 0.679 for word2vec. [fournier-etal-2020-analogies-minus-analogy-test]