gensim-data
gensim-data copied to clipboard
piskvorky
Historical word embeddings
…by Stanford, https://nlp.stanford.edu/projects/histwords/
We released pre-trained historical word embeddings (spanning all decades from 1800 to 2000) for multiple languages (English, French, German, and Chinese). Embeddings constructed from many different corpora and using different embedding approaches are included.
Paper: Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change Code: Github License: Public Domain Dedication and License
@piskvorky can you be more concrete, which embeddings need to be added (there are many)?
@piskvorky problem: each "zip" contains many models named like 1800-w.npy + 1800-vocab.pkl, 1810-w.npy + 1810-vocab.pkl, all this makes sense if we give all the embeddings at once for the user (which is now impossible).
It is probably worth closing this issue (because it does not apply to us)
@piskvorky for example, All English (1800s-1990s) (from Google N-Grams eng-all) http://snap.stanford.edu/historical_embeddings/eng-all_sgns.zip
This archive contains many files (pairs of matrix + vocab)
Archive: eng-all_sgns.zip
sgns/1860-vocab.pkl
sgns/1850-w.npy
sgns/1900-vocab.pkl
sgns/1930-w.npy
sgns/1880-w.npy
sgns/1870-w.npy
sgns/1910-w.npy
sgns/1970-vocab.pkl
sgns/1810-vocab.pkl
sgns/1970-w.npy
sgns/1810-w.npy
sgns/1920-vocab.pkl
sgns/1840-vocab.pkl
sgns/1990-vocab.pkl
sgns/1950-w.npy
sgns/1880-vocab.pkl
sgns/1980-w.npy
sgns/1830-w.npy
sgns/1830-vocab.pkl
sgns/1950-vocab.pkl
sgns/1890-vocab.pkl
sgns/1820-vocab.pkl
sgns/1800-w.npy
sgns/1940-vocab.pkl
sgns/1960-w.npy
sgns/1930-vocab.pkl
sgns/1850-vocab.pkl
sgns/1990-w.npy
sgns/1820-w.npy
sgns/1940-w.npy
sgns/1980-vocab.pkl
sgns/1920-w.npy
sgns/1890-w.npy
sgns/1960-vocab.pkl
sgns/1800-vocab.pkl
sgns/1840-w.npy
sgns/1870-vocab.pkl
sgns/1910-vocab.pkl
sgns/1900-w.npy
sgns/1860-w.npy
i.e. this file contains 20 distinct models (same situation for other links). To use these models for their intended purpose, they are needed all at once (they do not make sense separately). In our case, adding 20 models (which are useless apart from each other) is a very bad idea (moreover, it is extremely inconvenient for the user how to use all at once).
I see what you mean, but don't see it as a problem. Why couldn't the dataset loader just return a dictionary of models?
You suggest to join all of this to one large pickle (dict of KeyedVectors) and return it to the user, am I right?
No, I mean a dictionary where the key is a particular model name string (year?) and value the relevant Python object (Word2Vec or whatever).
If, as you say, the models are worthless in isolation, then we should return them all in bulk.
We can store only one gz file for model right now, for this reason, I talked about large pickle before.
Aha, I see. Yes, that is a possibility -- if the models are sufficiently small, we could pickle everything as a single dict (no separate .npy files etc).
Sorry for exhuming an old issue, but I was wondering if adding these pre-trained historical word embeddings is still under consideration. These would be very valuable to research I am conducting. Thank you.
