T5-base model create spelling mistake is summary


I am using T5-base model for abstractive summarization, results are good but I am getting newly generated spelling mistakes in the summary which were not actually present in input text.
Can anyone tell me why these spelling mistakes occuring and how can I solve this?

I think it’s due to your min_output size, for example if you have forced the model to generate results at minimum more than 50 sequence, and somehow the prediction length predicted only 40 sequences, I think it will start to generate random tokens just to reach the 50 seq.

1 Like

Hi Zack, hope you are doing well !!
Thank you for your reply.

Actually it is not generating random tokens, but it is misspelling them.

For e.g a word “productive” in input text is spelled as “priductive”

1 Like