Larger scale monolingual resource for target language

(Vincent Nguyen) #1

Hello again :slight_smile:

I am not sure if this is feasible or not so feel free to close this thread if not relevant.

I am wondering if instead of doing a n-best list based rescoring, it would be possible to interpolate the output scores / weights with another rnnlm stack trained on a much larger monolingual dataset, hence improving the likelihoods at output.

similar stuff has been done in smt systems of course.


(srush) #2

Yes. This can be done and will be easier to do with the new beam search code. Let’s revisit once that is landed.