Question about Seq2SeqIntentModel about nlp-architect

[peteriz]

You are correct. Seq2SeqIntentModel model only tags for intent slots (entities).

[peteriz]

The SNIPS NLU benchmark can be used as an intent dataset since it contains annotations of intent type and entities for each intent type.

[hkvision]

So my question is that if I use another dataset with entities only, can I use Seq2SeqIntentModel?

[peteriz]

It is possible, but you will still have to provide intent labels as the model expects to see 2 annotations (outputs). You can mark the label of the intent the same label for all samples and the network should learn it pretty quickly.
Another options would be to replicate the model (own your own) without the intent prediction part of the network.

[hkvision]

But in the model definition here: https://github.com/NervanaSystems/nlp-architect/blob/master/nlp_architect/models/intent_extraction.py#L337 , there is only one output, namely tags, and also in your SNIPS example here: https://github.com/NervanaSystems/nlp-architect/blob/master/examples/intent_extraction/train_seq2seq_model.py#L83 , train_i is never used. Thus I suppose intent is not needed at all?
I understand that for the MultiTaskIntentModel if I only have entities, I can follow your way to use the model. I'm asking about the Seq2SeqIntentModel.

[github-actions]

Issue marked as stale, closing.