How does `-dropout_input` work when word features are used?

The doc says that the dropout_input option applies dropout between the word embeddings and the first layer of the encoder. It does not specify if the feature-embeddings are also dropped out.

I suspect that dropout is applied to the concatenation of word-embedding and feature-embedding, but I’d like to make sure.

This is correct. It is applied to the input of the first recurrent layer.

1 Like