Probing the Probing Paradigm: Does Probing Accuracy Entail Task Relevance?
This work critically examines the probing paradigm in NLP analysis, highlighting potential misinterpretations in evaluating model representations, which is important for researchers in interpretability and model analysis.
The paper investigates whether high accuracy on probing tasks indicates that linguistic properties are relevant to a model's main task, finding that models can encode such properties even when unnecessary for the task, with pretrained embeddings playing a key role and synthetic tasks showing encoding above chance from random noise.
Although neural models have achieved impressive results on several NLP benchmarks, little is understood about the mechanisms they use to perform language tasks. Thus, much recent attention has been devoted to analyzing the sentence representations learned by neural encoders, through the lens of `probing' tasks. However, to what extent was the information encoded in sentence representations, as discovered through a probe, actually used by the model to perform its task? In this work, we examine this probing paradigm through a case study in Natural Language Inference, showing that models can learn to encode linguistic properties even if they are not needed for the task on which the model was trained. We further identify that pretrained word embeddings play a considerable role in encoding these properties rather than the training task itself, highlighting the importance of careful controls when designing probing experiments. Finally, through a set of controlled synthetic tasks, we demonstrate models can encode these properties considerably above chance-level even when distributed in the data as random noise, calling into question the interpretation of absolute claims on probing tasks.