Query about the function "encode_label" in autoprompt/utils.py about autoprompt

[XM-WANG]

Many thanks for your help!

[taylorshin]

Only single token objects are considered for the label/prediction target. That error must mean that somehow one of the relation object labels is being split into multiple tokens. Do you mind sharing more details of your error so that we can pinpoint which label is causing it?

[XM-WANG]

Many thanks for your reply!

Here I list some labels that will cause this error.

• (P17): ValueError: Label "Romania" gets mapped to multiple tokens.
• (P19): ValueError: Label "Constantinople" gets mapped to multiple tokens.
• (P20): ValueError: Label "Edinburgh" gets mapped to multiple tokens.
• (P27): ValueError: Label "Iceland" gets mapped to multiple tokens.

I wrote a script to automatically run the experiment with different relations. Nealy all relations contain some labels that trigger this error. I doubt that it is because of some wrong arguments I used.

This problem happened when I added the argument "--perturbed" for the relation extraction experiments with RoBERTa.

Could you please provide your command of the perturbed relation extraction experiments with RoBERTa?

[taylorshin]

For experiments with RoBERTa, we only used examples where the label is a single token for both BERT and RoBERTa. I added code that handles the aforementioned ValueError so if you pull the latest code from the master branch you should be able to run your command for creating triggers for perturbed relation extraction (with RoBERTa) without any errors.