How to improve translator in C1W4?

Hey everyone,

I implemented en-to-fr translator in w4, and ended with questions about this model.

The accuracy of model is 0.556. Which basically means that only half of words will be translated corrected. How can I improve this score?

  • Will adding more train examples take some effect?
  • Is there any more efficient cost function for gradient?
  • Or maybe such approach is not efficient for translation and should not be used in real applications

Also…

  • How to deal with translation duplicates when same en word has few translation in fr?
  • Where can I get full en/fr word embeddings?
  • And how did this word embedding obtain? What actual meaning of each of 300 features? Some resource can be useful for overall understanding since vector space lectures explain what is word embedding but not how we can calculate it for real application.

Can somebody help me with these questions?

Have a good day.

Hi @webskyter

These are good questions and most of them will be answered later in the specialization, so I would suggest to revisit them after finishing the specialization.

For example, writing down the questions you have after finishing week or course and coming back to them later after finishing the specialization (or even later), trying to answer them, I think is a good way of learning. Of course there are questions that you should get answers right away (like code errors, lecture explanations etc.) before continuing but your particular questions, I think, falls into the former category.

I know I didn’t answer your questions and that could be slightly disappointing :slight_smile: but I guess I wanted to advise not to get ahead of yourself and the answers you would get now from the community would be somewhat redundant to the lectures later in the course.

Happy learning!