google-research / pegasus

Apache License 2.0
1.61k stars 316 forks source link

New Named Entities generated in summary #210

Open Vindhya-Singh opened 2 years ago

Vindhya-Singh commented 2 years ago

Hey! I am trying to use the pegasus summarizer in my code where for a given news article, the model should generate a summary and this summary is then reflected on a Kibana dashboard. For a few articles, I have found that pegasus generates new names that have never been mentioned in the text. For example,

Comedy has moved from films and poetry sessions to cafes, clubs, bars, corporate shows, festivals, TV, YouTube, and streaming services, writes Sudha G Tilak. Comedy has moved from films and poetry sessions to cafes, clubs, bars, corporate shows, festivals, TV, YouTube, and streaming services.

is the output generated for this news article from BBC: article here. In the above summary the name 'Sudha G Tilak' appears while it never appears in the original text. Any explanation to why this is happening? Any workaround to this?

P.S: I am not fine-tuning the model but using all the original parameters.