Pareto Probing: Trading Off Accuracy for Complexity

Abstract

The question of how to probe contextual word representations for linguistic structure in a way that is both principled and useful has seen significant attention recently in the NLP literature. In our contribution to this discussion, we argue for a probe metric that reflects the fundamental trade-off between probe complexity and performance: the Pareto hypervolume. To measure complexity, we present a number of parametric and non-parametric metrics. Our experiments using Pareto hypervolume as an evaluation metric show that probes often do not conform to our expectations-e.g., why should the non-contextual fastText representations encode more morpho-syntactic information than the contextual BERT representations? These results suggest that common, simplistic probing tasks, such as part-of-speech labeling and dependency arc labeling, are inadequate to evaluate the linguistic structure encoded in contextual word representations. This leads us to propose full dependency parsing as a probing task. In support of our suggestion that harder probing tasks are necessary, our experiments with dependency parsing reveal a wide gap in syntactic knowledge between contextual and non-contextual representations.

Low-Complexity Probing via Finding Subnetworks

“…Other works take an informationtheoretic view: Voita and Titov (2020) measure the complexity of the probe in terms of the bits needed to transmit its parameters, while Pimentel et al (2020b) argue that probing should measure mutual information between the representation and the property. Pimentel et al (2020a) propose a Pareto approach where they plot accuracy versus probe complexity, unifying several of these goals. We use these proposed metrics to compare our probing method to standard probing approaches.…”

“…We first vary the complexity of each probe, where for subnetwork probing we associate multiple encoder weights with a single mask, 4 and for the MLP probe we restrict the rank of the hidden layer. We then plot the resulting accuracy-complexity curve (Pimentel et al, 2020a).…”

“…First, we find that the neuron subnetwork probe has both higher accuracy on pre-trained models and lower accuracy on random models, so it is both better at finding properties of interest and less able to learn the tasks on its own. Next, we measure complexity as the bits needed to transmit the probe parameters (Pimentel et al, 2020a;Voita and Titov, 2020). Varying the complexity of each probe, we find that subnetwork probing Pareto-dominates MLP probing in that it achieves higher accuracy given any desired complexity.…”

Does My Representation Capture X? Probe-Ably

“…Early probing studies in NLP include (Zhang and Bowman, 2018) and (Tenney et al, 2019c), the former being an early example of the importance of comparing with randomized representations or labels. Further discussion has introduced control tasks and the selectivity metric (Hewitt and Liang, 2019), formalized notions of ease of extraction (Voita and Titov, 2020) and described other strategies for taking model complexity into account (Pimentel et al, 2020a).…”

Intrinsic Probing through Dimension Selection

“…Contours belong to our probe. Conneau et al, 2018) is one prominent method, which consists of using a lightly parameterized model to predict linguistic phenomena from intermediate representations, albeit recent work has raised concerns on how model parameterization and evaluation metrics may affect the effectiveness of this approach (Hewitt and Liang, 2019;Pimentel et al, 2020b;Maudslay et al, 2020;Pimentel et al, 2020a). Most work in intrinsic probing has focused in the identification of individual neurons that are important for a task (Li et al, 2016;Kádár et al, 2017;Li et al, 2017;Lakretz et al, 2019).…”