Embeddings

Paper Refs. :

• Mikolov et al 2013 - Distributed Representations of Words and Phrases (SGNS)
• Rong, Xin 2014 - word2vec Parameter Learning Explained
• Levy & Goldberg 2014 - Neural Word Embedding as Implicit Matrix Factorization

Results

• Nearest neighbours: strong (e.g. france → spain, italy, germany)
• Analogies: ~14% top-1 (semantic > morphological; limited by the small 17M-token corpus)

Limitations

Small corpus → weak on analogies (esp. capital-country, morphology). For better analogy accuracy, train on a larger corpus (enwik9+). Lowercased English only; drops OOV.

Usage snippet

from huggingface_hub import hf_hub_download
import torch
import torch.nn.functional as F

path = hf_hub_download(repo_id="ocdbytes/embeddings", filename="embeddings_200.pt")
# weights_only=False because the checkpoint bundles Python dicts (word2idx/idx2word),
# which the default restricted loader (torch>=2.6) may reject.
ck = torch.load(path, map_location="cpu", weights_only=False)

syn0 = ck["syn0"]
word2idx, idx2word = ck["word2idx"], ck["idx2word"]
emb = F.normalize(syn0, dim=1)

def neighbours(word, n=10):
    i = word2idx[word]
    sims = emb @ emb[i]
    top = sims.topk(n + 1).indices.tolist()
    return [idx2word[j] for j in top if j != i][:n]

def analogy(a, b, c, n=5):
    t = F.normalize(emb[word2idx[b]] - emb[word2idx[a]] + emb[word2idx[c]], dim=0)
    sims = emb @ t
    ban = {word2idx[a], word2idx[b], word2idx[c]}
    top = sims.topk(n + len(ban)).indices.tolist()
    return [idx2word[j] for j in top if j not in ban][:n]

print(neighbours("king"))                      # -> ['viii', 'elizabeth', 'queen', ...]
print(analogy("france", "paris", "germany"))   # -> ['berlin', ...]

Code

• ocdbytes-ai/embeddings