2020/06/universal-sentence-encoder/ about blog-comments HOT 7 CLOSED

@PetroIvaniuk Sadly it's not clear on their end. As far as I understand, you can't retrain it with your own vocab.

You can read more in this thread

from blog-comments.

The multilingual USE model may also be of interest to you: https://arxiv.org/pdf/1907.04307.pdf

It has neat improvements over the earlier USE model, especially for QA tasks. Plus, semantic similarity across various languages.

from blog-comments.

@setu4993 Thank you for recommending that paper. Looks very interesting.

from blog-comments.

Appreciate the post.
But I wondering it is possible to add my own text set and re-train with to universal sentence encoder to receive vectors with account the data corpus features.

from blog-comments.

thank you Amit, really appreciate you writing this. It's very informative

from blog-comments.

Hi Amit, for one of my project 'Document Search', i have used this multilingual USE. but model is over trained for the most occurring word in the document. Suppose, if i ask question like "Why do we need ABC" ,it is throwing the response. but similarly, when i ask, "Need of ABC", I am expecting the same response. but it is not throwing the same response. i.e, it is throwing other responses of ABC with the highest cosine similarity score. Based on the semantic and context, i want it to be trained. Any help would be better..

from blog-comments.

Hi Amit, Thanks for such an informative article. In the DAN variant of universal sentence encoder you mentioned it feeds the average of words and bigrams token embeddings. Can you shed some light on how these token embeddings are generated for each word or unigram? It would be helpful if you could point to that respective paper.

from blog-comments.