search

ucfnlp / control-over-copying

(AAAI'20) The source code for the paper "Controlling the Amount of Verbatim Copying in Abstractive Summarization".
https://arxiv.org/pdf/1911.10390.pdf
Other
38 stars 9 forks source link

Is it possible to control amount of copying? #2

Closed aretius closed 4 years ago

aretius commented 4 years ago

Hey all,

I tried generating outputs of the following input through all available datasources: Hailey Baldwin has this to say after being accused of throwing shade at Selena Gomez's new song. but the output was Alec Baldwin accused of throwing shade at Selena Gomez.

Although the output is not correct, however what concerns me is the word Alec being present. Any thoughts on this?

KaiQiangSong commented 4 years ago

This is about the hallucination. Because we use Bert as our pre-trained model. It could introduce some knowledge from the pre-training data. Actually almost every abstractive summarization has this issue. We are finding a way to solve it.