Individual results

View docs

View in-depth performance of a single language model on a single test suite.

Region-by-region surprisal
Sample item for NP/Z Garden-path Ambiguity with Modifier (Overt Object)
Item
Condition
Start Transitive Verb Comma NP/Z Extension Verb Continuation
Item Condition Start Transitive Verb Comma NP/Z Extension Verb Continuation
1 no-obj_no-comma As the criminal shot the woman with her young daughters yelled at the top of her lungs
1 obj_no-comma As the criminal shot his gun the woman with her young daughters yelled at the top of her lungs
1 no-obj_comma As the criminal shot , the woman with her young daughters yelled at the top of her lungs
1 obj_comma As the criminal shot his gun , the woman with her young daughters yelled at the top of her lungs
Prediction performance for GPT-2 on NP/Z Garden-path Ambiguity with Modifier (Overt Object)
Accuracy
Formula
Description
AccuracyPredictionDescription
100.00% (562,no-obj_no-comma/6,Verb) > (561,no-obj_comma/6,Verb) No description provided.
100.00% (562,no-obj_no-comma/6,Verb) > (560,obj_no-comma/6,Verb) No description provided.
100.00% ((562,no-obj_no-comma/6,Verb) - (561,no-obj_comma/6,Verb) ) > ((560,obj_no-comma/6,Verb) - (559,obj_comma/6,Verb) ) We expect that the difference in surprisal between the Verb region in the no-obj_no-comma condition minus the no-obj_comma condition will be greater than the difference in the Verb region between the obj_comma and the no-obj_comma. This is because the comma unambiguously introduces a matrix clause.