Erik Chan

Thanks! I was initially also using bleualign, but had too many documents to align and using google translate is too expensive for my project.

Finally got this working. Currently training a Chinese -> English model with 3gb parallel corpora 80mb dictionary. It's already been a few hours, how long should I expect it to...

+1

Also, do you think it is better to create a model that only labels one term e.g B-TERM ? Ultimately, I want to be able to identify multiple terms B-TERM1,...

I'm thinking out aloud here... since ultimately this is a transformer model I presume the sequence needs to be aligned. Would I structure the data this way? ``` # TWO...

Thanks, I will try this. How do you suggest I convert the second input into numeric feature? I would want to keep B-TERM and I-TERM to capture the beginning and...

What about [SEP], [CLS], [BOS] tokens in a numeric sequence? Do I keep them the same? Or also convert to a digit?

So it looks like my labeled data includes examples where B-TERM can go as high as B-TERM26 in one sentence (26 different terms tagged in the source and target text)....

Yes, I agree only one term is much simpler for the model to learn. However, in the case where I have a sentence that has 26 different terms (most sentences...

Posting my model and evaluation results below. Seems like something went wrong. Val_accuracy dropped significantly on epoch 7 and 10 (from 0.4062 -> 0.0117 -> 0.0078). My model basically didn't...