OpenNMT Forum

Are word features supported in opennmt-tf?

Hi there,

It was unclear to me whether the word features (token|feature1|feature2) kind of model supported by the opennmt-tf framework. I was unable to find any documentation on how I do this using opennmt-tf. Kindly help me out.


Yes they are supported, but it requires separate and parallel input files instead of the pipe-based syntax.

See this model definition with multiple word features as an example. The expected data configuration is also described in this link.

Thank you… Is there anyway I can extend this approach to the target side features as well?

Target side features are not implemented in OpenNMT-tf.

I see… but I can still use them in opennmt-py?

There is an open pull request for that, but it is not merged yet:

Great… Thank you. I will keep a watch.