search

ericxsun / fastText

Library for fast text representation and classification.
Other
15 stars 9 forks source link

Word embeddings (vectors) lost their 'sense' after incr training #5

Open agoeroeg opened 6 years ago

agoeroeg commented 6 years ago

Hi, I downloaded the german wiki word vectors from fasttext: https://fasttext.cc/docs/en/pretrained-vectors.html (bin+text) I copy-pasted some political text in a training file (around 180MB), cleaned with: cat train.txt | sed -e 'y/[]/()/' -e "s/([.!?,'/()])/ \1 /g" | tr "[:upper:]" "[:lower:]" | sed -e 'y/ÖÜÄ/öüä/' > train_clean.txt Trained the wiki.de.bin ./fasttext skipgram -input train_clean.txt -inputModel wiki.de.bin -output wikiplus -incr Extracted the vectors from the *.vec files Made a cluster analysis and was amazed, how stupid my new vectors are. E.g. the clustering with the original vectors would cluster bus and train together, the new vectors make not much sense. Do you have any idea, why?

ericxsun commented 6 years ago

That's weird, I'll take a look.

agoeroeg commented 6 years ago

Thanks. I extracted these vectors: egrep '^(afd|merkel|macron|meinung|rosen|aprikosen|pegida|komisch|erdbeeren|xdf6hkj|mirnixdirnix|fußballer|beschlussauffassung|geschäftsordnung|bus|bahn|fahren|kunde) ' wikiplus.de.vec > vectors_trained.txt

(translation: afd|merkel|macron|opinion|roses|apricots|pegida|strange|strawberries|xdf6hkj|mirnixdirnix|footballer|decision|opinion|bus|train|ride|customer) My training text will differ from yours, but still, I would expect, that if I train a political text with no roses, apricots, football, bus and train ride, then these will keep their 'sense'/neighbourhood.