OpenNMT Forum

Does opennmt-py support "monotonic alignment" of local-attention? (from luong attention paper)

I know we can use luong attention when we set both --encoder_type and --decoder_type as “rnn”, but there is “monotonic alignment” algorithm in this original paper:

Is there a way to do this monotonic alignment in OpenNMT-py?

This is not implemented in OpenNMT-py.

It’s an important feature for NMT system.
Without alignment, we lost many words feature from source.

Really hope it’s implemented in soon furture.

I suggest reading the paper to learn what monotonic alignment is. Also, it’s typically used for speech recognition, not NMT.

Hi, guillaumekln; I mean, I hope the OpenNMT-py can implemented the alignment feature. It’s important.

And now I am using fast_align with OpenNMT-tf .