Closed matthias-samwald closed 7 years ago
yes, it does take word bi-grams into account. though the bigrams are more sparse and often get rather small weights. did you observe the same also with one of our pre-trained bi-gram models?
Hi Matthias, Thanks for pointing out. :) It seems I hadn't added the code for adding n-grams in nnSent and analogiesSent. Can you try it now? I will try to resolve the other issue ASAP. It is most probably due to the commit https://github.com/epfml/sent2vec/pull/8 .
Your recent commits have solved the issue, different word orders now lead to different results, as expected. Thanks!
I trained the model with wordNgrams set to 2. I tried the same input sentence with permuted word order. I would expect the results to be at least slightly different, since the word n-grams are different, but they are precisely the same. Are the word n-gams not taken into account here?
(I will reply regarding the sorting issue soon)