Guide T5 summarization with additional features

Models
1

Hi,

I am trying to train a T5-base model to summarize/generate texts in a guided fashion.
What is the standard way to feed additional features to this kind of model?

My intuition led me to the following input template:
“”"
summarize: Input text
feature A: 1
feature B: 0
“”"

I did a simple experiment to test this out. I only used one extra feature that should have a high impact on the generated output. The model performed a bit better with the additional input data. But it did not work as I expected: I played with the trained model and it seems that toggling the extra feature (0/1) has little to no impact on the predicted text.

Thanks!

Tom

1 Like
2

Hi Tom, did you ever find any good resources for this? I have a similar use case.

4

I‘m not completly sure if I understand the problem but it sounds like contraints could help on Generation.