How BPE Affects Memorization in Transformers

TL;DR

This paper investigates how the size of the subword vocabulary in Byte-Pair Encoding influences the memorization behavior of Transformer models, affecting their ability to memorize and reproduce training data.

Contribution

It reveals that larger BPE vocabularies increase memorization and vulnerability to attacks, highlighting the importance of vocabulary size in model training and deployment.

Findings

Larger BPE vocabularies lead to more memorization in Transformers.

Models with bigger vocabularies are more susceptible to membership inference attacks.

Increasing BPE vocabulary reduces sequence length, impacting memorization behavior.

Abstract

Training data memorization in NLP can both be beneficial (e.g., closed-book QA) and undesirable (personal data extraction). In any case, successful model training requires a non-trivial amount of memorization to store word spellings, various linguistic idiosyncrasies and common knowledge. However, little is known about what affects the memorization behavior of NLP models, as the field tends to focus on the equally important question of generalization. In this work, we demonstrate that the size of the subword vocabulary learned by Byte-Pair Encoding (BPE) greatly affects both ability and tendency of standard Transformer models to memorize training data, even when we control for the number of learned parameters. We find that with a large subword vocabulary size, Transformer models fit random mappings more easily and are more vulnerable to membership inference attacks. Similarly, given a prompt, Transformer-based language models with large subword vocabularies reproduce the training data more often. We conjecture this effect is caused by reduction in the sequences' length that happens as the BPE vocabulary grows. Our findings can allow a more informed choice of hyper-parameters, that is better tailored for a particular use-case.