Closed ijyliu closed 11 months ago
It would be very hard for me to obtain good email response chain data.
I would need permission to use emails from some sort of institutional setting where people have been engaging in back and forth, assigning tasks and doing and replying to those emails.
Emails often have a query, a response, and, importantly, a thank you or other grading of the quality of the response.
Note one interesting technique may be somehow feeding the model the query and the evaluation/thanks messages and having the model fill in the middle, the response.
Look into bidirectional encoders (BERT) as a way to learn from both sides of input/fill in the middle?