search

aneesh-joshi / LSTM_POS_Tagger

A simple POS Tagger made using a Bidirectional LSTM using keras trained on the Brown Corpus
34 stars 19 forks source link

Training without embeddings #15

Open dutkaD opened 5 years ago

dutkaD commented 5 years ago

Is there an easy way to change the implementation to only train the model purely on the input sequences?

aneesh-joshi commented 5 years ago

Sure. You can remove the embedding layer.

You can use just a one hot vector. Or just the word index. Although I doubt that'd be much better.

If your problem is a lack of pretrained embeddings, you can just randomly initialize it and hope that it learns good embeddings.