Related to : #198
Right now, edit action transducers are a niche part of the repo. However, we could technically generalize the logic to any sequence to sequence model with teacher forcing and no longer need the transducer to be a standalone task. (Basically you would just replace the vocabulary with a set of edit actions and have a model learn an edit action sequence.) Would be interesting to have this be a default choice across architectures to compare simple auto-regressive string prediction against a transduction sequence. As a side point, I think it would alleviate some of the conflicts we have when updating since it would make it a generalizable feature across models.
Related to : #198
Right now, edit action transducers are a niche part of the repo. However, we could technically generalize the logic to any sequence to sequence model with teacher forcing and no longer need the transducer to be a standalone task. (Basically you would just replace the vocabulary with a set of edit actions and have a model learn an edit action sequence.) Would be interesting to have this be a default choice across architectures to compare simple auto-regressive string prediction against a transduction sequence. As a side point, I think it would alleviate some of the conflicts we have when updating since it would make it a generalizable feature across models.