This highly idiosyncratic and specific binary classifier is designed for the sole purpose of helping linguists find instances of the English Preposing in PP (PiPPs) construction in corpora. PiPPs are unbounded dependency constructions like "Happy though we were with the idea, we decided not to pursue it". This model does a good job of classifying sentences for whether or not they contain an instance of the construction.
The model is a fine-tuned
bert-base-cased model. The fine-tuning data are available as
annotated/pipp-labels.csv in
this project repository. All the annotations were done by Christopher Potts for the project "Characterizing English Preposing in PP constructions".
The sole purpose of the model is to try to identify sentences containing PiPPs. I assume that one is first filtering sentences using very general regexs, and then this model helps you find the gems as you go through examples by hand.
The model is useless for really anything except this linguistically motivated for task. And, even from the perspective of theoretical linguistics, this is a highly niche application!
