any how to ingest a relationship/analogy data into prep.py?

[gahu...@gmail.com]

Hi,

I have a question regarding Swivel, and it maybe a stupid question.

I am trying to train the model so that it will only print out the related word.

I have ingested quite a small data into prep.py and ran swivel.py, text2bin.py and nearest.py to find the related words according to the target words that I have specified.

However, not all of the printed out words were related to the specified word.

For example, I have ingested information regarding drone, and on nearest.py my specified word was 'system'.

I was expecting the word 'drone' to pop out as one of the top on the list since the word drone is always next to the word system.

But it did not give out the expected result.

Just like the 'Word2Vec' model, shouldn't there be a list of analogy that I should ingest to help the model to know the relationship?

Any thoughts??

[Chris Waterson]

Hi Jason. Just like word2vec or Glove or SVD, Swivel takes a corpus of text and computes vectors based on the co-occurrences of words in the text.  Each of these algorithms is essentially computing a low-rank factorization of a matrix whose cells contain the pairwise mutual information of the words. More text usually leads to better vectors, since it results in better estimates for the PMI between two words.

The vectors that are produced by these algorithms have the property that they can be used to solve analogies: the analogy data is not part of the training data.

Hope that helps.

--
You received this message because you are subscribed to the Google Groups "Swivel Embeddings" group.
To unsubscribe from this group and stop receiving emails from it, send an email to swivel-embeddings+unsubscribe@googlegroups.com.
To post to this group, send email to swivel-embeddings@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/swivel-embeddings/9d22f45e-38c5-4bf6-bb01-52acc927eb1f%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.