search

yumeng5 / WeSTClass

[CIKM 2018] Weakly-Supervised Neural Text Classification
Apache License 2.0
77 stars 30 forks source link

Use spherical text embeddings? #3

Closed theashworld closed 4 years ago

theashworld commented 4 years ago

Has anyone tried using https://github.com/yumeng5/Spherical-Text-Embedding instead of w2v in the code?

yumeng5 commented 4 years ago

Hi,

Thanks for the question. Personally I haven't tried to replace word2vec with my spherical embedding in this WeSTClass framework. It is completely reasonable to do so, but I wouldn't expect the results to be significantly different since both algorithms are still unsupervised embeddings and they are not designed to distinguish the categories during embedding learning.

To learn discriminative embeddings from weak supervision (label names or keywords), I would suggest taking a look at another embedding framework CatE. I did try to replace word2vec with CatE in WeSTClass and observed better classification results (described in the CatE paper).

Let me know if you have any other questions!

Best, Yu