Memorisation versus Generalisation in Pre-trained Language Models

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7564 - 7578 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Memorisation versus Generalisation in Pre-trained Language Models Michael Tänzer Imperial College London m.tanzer@imperial.ac.uk Sebastian Ruder∗ Google Research ruder@google.com Marek Rei Imperial College London marek.rei@imperial.ac.uk Abstract State-of-the-art pre-trained language models have been shown to memorise facts and per- form well with limited amounts of training data. To gain a better understanding of how these models learn, we study their generali- sation and memorisation capabilities in noisy and low-resource scenarios. We find that the training of these models is almost unaffected by label noise and that it is possible to reach near-optimal results even on extremely noisy datasets. However, our experiments also show that they mainly learn from high-frequency patterns and largely fail when tested on low- resource tasks such as few-shot learning and rare entity recognition. To mitigate such lim- itations, we propose an extension based on prototypical networks that improves perfor- mance in low-resource named entity recogni- tion tasks. 1 Introduction With recent advances in pre-trained language mod- els (Peters et al., 2018; Devlin et al., 2019; Liu et al., 2019; He et al., 2020), the field of natural lan- guage processing has seen improvements in a wide range of tasks and applications. Having acquired general-purpose knowledge from large amounts of unlabelled data, such methods have been shown to learn effectively with limited labelled data for downstream tasks (Howard and Ruder, 2018) and to generalise well to out-of-distribution examples (Hendrycks et al., 2020). Previous work has extensively studied what such models learn, e.g. the types of relational or linguis- tic knowledge (Tenney et al., 2019; Jawahar et al., 2019; Rogers et al., 2020). However, the process of how these models learn from downstream data and the qualitative nature of their learning dynam- ics remain unclear. Better understanding of the learning processes in these widely-used models is ∗ Work done prior to joining Google. needed in order to know in which scenarios they will fail and how to improve them towards more robust language representations. The fine-tuning process in pre-trained language models such as BERT (Devlin et al., 2019) aims to strike a balance between generalisation and memo- risation. For many applications it is important for the model to generalise—to learn the common pat- terns in the task while discarding irrelevant noise and outliers. However, rejecting everything that oc- curs infrequently is not a reliable learning strategy and in many low-resource scenarios memorisation can be crucial to performing well on a task (Tu et al., 2020). By constructing experiments that allow for full control over these parameters, we are able to study the learning dynamics of mod- els in conditions of high label noise or low label frequency. To our knowledge, this is the first quali- tative study of the learning behaviour of pre-trained transformer-based language models in conditions of extreme label scarcity and label noise. We find that models such as BERT …

Content truncated to 500 words