LoginSign Up
← Back to all sparks
W

word2vec

AI-ASSISTANTS
Velocity0.0

Distributed Representations of Words

github.com/bnosac/word2vec ↗

word2vec for R spent its 0.4 release proving two training paths give identical embeddings

nlpembeddingsword2vecr-packagebnosac
◆Current state
word2vec is a standalone C++ word2vec implementation wrapped for R, part of the bnosac NLP family. Version 0.4.0 made word2vec() a generic with character and list methods, so models can be trained from a list of tokenised sentences instead of only from a file on disk, and reordered the vocabulary so both paths produce identical embeddings given identical tokenisation. The 2025 release is documentation and a DESCRIPTION DOI note.
◆Where it's heading
Development has been about widening the input surface and the comparison surface rather than the algorithm: encoding arguments, cosine as an alternative to dot similarity, doc2vec applied to already-trained models, and finally in-memory tokenised input. The vocabulary sorting change in 0.4.0 is the notable one — it altered embeddings slightly for everyone upgrading, in exchange for reproducibility between the two training paths. Since then the package has moved only when the wider bnosac set does.
◆Prediction
With both training paths unified and the recent release confined to packaging, there is no visible thread pointing at further feature work; the next release most likely arrives with the next CRAN sweep across the sibling packages.

◆Recent moves

  1. 8mo ago

    Documentation braces and arXiv DOI note

    Documentation wording in word2vec_similarity and a DESCRIPTION DOI note. Part of the same CRAN sweep that touched doc2vec and BTM the same day.

    View source ↗
  2. 2y ago

    Train from tokenised sentence lists; word2vec becomes generic

    Makes word2vec() a generic with character and list methods so models can be trained from in-memory tokenised sentences, and sorts the vocabulary by frequency then token so both paths yield the same embeddings. Users upgrading get slightly different vectors as a result — a deliberate trade for reproducibility across input types.

    View source ↗
  3. 5y ago

    Cosine similarity option in word2vec_similarity

    Adds a type argument allowing cosine similarity alongside the default dot product, answering a user request. Small but changes what the comparison function can express.

    View source ↗
  4. 5y ago

    doc2vec usable on trained models; txt_clean_word2vec added

    Allows doc2vec to run on an already-trained word2vec model and adds txt_clean_word2vec for corpus preparation. Extends the package past training into downstream use of the resulting vectors.

    View source ↗
  5. 5y ago

    Conditional udpipe example; encoding argument

    Guards an example on udpipe availability, bundled with the earlier addition of an encoding argument. Minor packaging and input handling.

    View source ↗
  6. 5y ago

    doc2vec support added

    Adds doc2vec to the package, extending it from word-level to document-level vectors. The first step of the broadening that 0.3.3 and 0.4.0 continue.

    View source ↗