Toy mechinterp: take a word embedding, reflect it across all axes (i.e. negate the vector), then "back-project" by finding the nearest word in the embedding vocabulary.

This is intentionally dumb — it's a simple baseline for exploring what "opposite" means in distributional vector spaces.

uv sync
uv run oppositeday meek
uv run oppositeday meek --origin mean

With the default glove-wiki-gigaword-50 embeddings, the naive axis-reflection trick produces something very non-semantic (this is part of the point):

word:  meek
model: glove-wiki-gigaword-50
origin: zero
opposite (top-1): aiport  (sim=0.6693)

Mean-centered reflection (reflect about the global mean embedding vector) doesn't magically fix it, but is a useful variant to compare against:

word:  meek
model: glove-wiki-gigaword-50
origin: mean
opposite (top-1): sulabh  (sim=0.7326)

If the embedding is a vector v ∈ R^d, then we define:

v_opposite = -v

Then we search the embedding vocabulary for the nearest neighbor to v_opposite (cosine similarity by default).

• By default this repo uses the glove-wiki-gigaword-50 embeddings via gensim.downloader.
• The first run will download the model into your ~/gensim-data cache.
• The experiment write-up (negative result) is in EXPERIMENT_LOG.md.