Probing Classifiers: Promises, Shortcomings, and Advances

Abstract Probing classifiers have emerged as one of the prominent methodologies for interpreting and analyzing deep neural network models natural language processing. The basic idea is simple—a classifier trained to predict some linguistic property from a model’s representations—and has been used examine wide variety properties. However, recent studies demonstrated various methodological limitations this approach. This squib critically reviews probing framework, highlighting their promises, shortcomings, advances.