Extract word matrix and context matrix from output of trainWordEmbedding / word2vec

When I use trainWordEmbedding on a set of documents to train a word embedding that I can then use word2vec with, I get an object "emb" as output that I can input into word2vec. Using word2vec I then get, for each word, the vectors that I can then further process.

However, I would like to also receive as output the underlying word matrix and context matrix (as well as the value of the loss of the training). Does anyone know how I can access these data?

[Christopher Creutzig]

What exactly do you mean by “word matrix” and “context matrix”?

I guess the “context matrix” is what (some) other people call the cooccurrence matrix in the skip-gram model? We do not currently have a way to compute that.

Answer 1

[Jayanti]

Hi Daniel,

By word matrix I assume you want the unique words in the document. When you use “trainWordEmbedding” to train a word embedding model on a set of documents, it returns an object called “emb”. This object includes a property named “Vocabulary”, which contains the unique words from the model, stored as a string vector. You can access these unique words using the following code:

emb = trainWordEmbedding(filename);
words = emb.Vocabulary;

By context matrix I assume you mean cooccurrence matrix. However, I couldn't find specific documentation on accessing a co-occurrence matrix directly through the “trainWordEmbedding” or “word2vec”.

Hope this will be helpful!