original Huang, MATLAB

[makrai]

http://www.socher.org/index.php/Main/ImprovingWordRepresentationsViaGlobalContextAndMultipleWordPrototypes

[gaebor]

@makrai Nincsen jogtiszta MatLab-om, sajnos a matek intézetnek lejártak a licenszei

[makrai]

megpróbálnád octave-val?

[makrai]

@DavidNemeskey in the paper, write about both the orig Huang et al and CMultiVec

[DavidNemeskey]

We had a look at the code, and found that

1. the training code is only for their pre-GloVe embedding, the k-means part is not included
2. the test code is basically CRelabelCorpus from CMultiVec
3. in README.txt from the latter:

The word representations use a dictionary of 100,232 words. 10 prototypes are used for
6,162 of the words, which roughly correspond to the most frequent words. To determine
which prototype to use given context, run run.m in matlab.

I am writing a mail to Huang to ask him how they arrived to those words.

[makrai]

Would you please also ask how they clustered the occurrences (to compute the prototypes)?

[makrai]

to write in the paper

• there is no code for the second part of the Huang paper
• CMultiVec is not adaptive

[DavidNemeskey]

No reply from Huang thus far, but I think it is not that important anyway. Once @kornai is done with the paper for today, I'll add those two sentences about this MSE.