I am trying to implement this paper but instead with T5 model. Very briefly, this is a controlled summarization paper where you modify the object of summarization from generating summary given the document to generating entity chain (list of entities) followed by summary. The paper further says that modifying the entity chain would lead to a modified summary. This is the part I am trying to implement with
T5Model but having trouble finding the API that would do this. Specifically I want to do
model.generate (thus beam search, sample or greedy) but given some decoder tokens (which is the modified entity chain).