facebookresearch/LASER

Laser support for non-official data

joshdehlong opened this issue · 2 comments

joshdehlong commented

Hello, we installed the laser instance via the official way. Our main task is to be able to compare meaning of 2 sentences in 2 different languages (basically what is laser meant to do).

We were able to successfully deploy vectors but when we tried to count the similarity between 2 sentences, the whole table was filled with zeros on our 2 different languages. When we used the official demo data (those articles from 2012), we actually got a regular score. Is this a purposeful implementation where LASER does not work with custom sentences or are we missing sth here?

heffernankevin commented

Hi @joshdehlong, thanks for getting in touch!

Is this a purposeful implementation where LASER does not work with custom sentences or are we missing sth here?

Yes, this will work with non-custom sentences. To help debug this, a couple of questions:

  1. Are you embedding using supported languages?
  2. When you say: "the whole table was filled with zeros on our 2 different languages", do you mean the vectors produced are all zeros? If so, can you maybe share the inputs you are trying?
heffernankevin commented

Closing due to inactivity (and hopefully issue is solved). Please re-open if needed!