I’m a bit confused about what I should be focusing on I believe that improving the existing NN model is the best thing to do. A seq2seq model wouldn’t guarantee better performance and is much heavier both efficiency and space wise when compared to a CBOW model. I want to start working on my proposal, should I just focus on how I can improve the existing NN approach and working on converting ngram pairs to the NN approach?
Any update on this?
How much USD in total would you need to run your project on AWS?
Hi,
It would depend on the direction of the project a seq2seq model takes much longer to run and needs much more memory as compared to a simple Continuous bag of words model. I’m also trying to obtain permission from my college to give me the authentication to make use of some of the clusters available. So I can’t be too sure at this point. If I’m expanding on the current NN model I might not need AWS at all.
1 Like
Just converting confusion sets to the new approach might not be enough, as it seems to be a task that requires a lot of work (collecting data etc) that isn’t very, well, interesting. Or am I missing something?
Can you tell me more about this project : “Train a statistical tagger for English” from my understanding the tags used by LT and by the Penn Tree Bank are different and we need to find a way of disambiguating these differences. LT also wants a better POS tagger. Am I missing anything else?
I have recently worked on a POS tagger for Norwegian and English as a part of my NLP class using HMM’s ,Conditional Random Fields and a bidirectional LSTM. Would this experience be useful for this project and is it a feasible project to work on in the summer?
This is originally from @MarcinMilkowski, maybe he can comment on it?
I’d think so, but it also feels like quite an experimental project. It might not be clear if it actually works until a lot of work has been put into it.
His last answer on the forum was in May 2017, do you think he would reply? Or is there a way I can contact him directly.
He should be notified now that his username was mentioned. You could also try to contact him at http://marcinmilkowski.pl/en/contact-me, but yes, he’s probably very busy.