Degree with T5

[gonconist]

Hello,

I am running some tests with t5-small, and t5-base as pretrained model for Degree ED, but its behavior during generation is strange. More specifically, the generated text seems to be in a different language than English, and sometimes it is concatenated with the word "is" that is present in the template 'Event trigger is <Trigger>', for example: pred_texts = ['Event trigger isurmatoarele.', 'Event trigger is imobil.', 'Event trigger is betreut.', ... ].

Looking at the source files for Degree, I have noticed that with you've already experimented with t5 as encoder model. Do you have any suggestions as to how configure the generation/decoding steps for t5 to properly function?

Thanks!