Supervised training without using label + quantization

[userFT]

Hi! I wanted to reduce the size of the pre-model (.bin or .vec file, several GB for these files), I started to look at the quantization method proposed by FastText. I tried to quantize a .bin file, however it was not possible because it only works for supervised models. Therefore I use the supervised training method with cc.en.300.vec file in input and quantize the model obtained from the training. The size was greatly reduced (from 4,4 GB to ~300MB) and it seems to work like the fullsized "raw" model. The quantized model is working, I'm able to retrieve logical results from a user input.

But how is it possible to use the .vec file to train through the supervised method without using label ?
I thought : supervised method = text classification and unsupervised method = word representation

Can someone enlighten me ?

[antisrdy]

Dear @userFT,
I'm trying to quantize pre-trained vectors as well. Would you mind providing further details on how you were eventually able to quantize the .bin file (leveraging the supervised training method)?

[userFT]

Hi! I used the .vec file from this and trained the model using the supervised training method. Then with the model obtained, I quantized it.
However by doing this (and from my observation), the model will lose the ability to find semantically close words (for example: wheel/tyres, strawberry/raspberry, etc.), it will only works based on the wordNgrams defined.

Therefore I don't recommend using this technique.