You should evaluate your language model on marginal likelihood over tokenisations

TL;DR

This paper proposes evaluating language models using marginal likelihood over tokenisations instead of a single tokenisation, showing it improves out-of-domain performance and better captures tokeniser uncertainty.

Contribution

It introduces methods to estimate marginal likelihood over tokenisations and demonstrates its advantages over traditional single tokenisation evaluation.

Findings

Marginal perplexity can be significantly lower than one-best perplexity.

Tokeniser entropy correlates with differences in perplexity.

Marginal likelihood estimation is feasible with a manageable number of samples.

Abstract

Neural language models typically tokenise input text into sub-word units to achieve an open vocabulary. The standard approach is to use a single canonical tokenisation at both train and test time. We suggest that this approach is unsatisfactory and may bottleneck our evaluation of language model performance. Using only the one-best tokenisation ignores tokeniser uncertainty over alternative tokenisations, which may hurt model out-of-domain performance. In this paper, we argue that instead, language models should be evaluated on their marginal likelihood over tokenisations. We compare different estimators for the marginal likelihood based on sampling, and show that it is feasible to estimate the marginal likelihood with a manageable number of samples. We then evaluate pretrained English and German language models on both the one-best-tokenisation and marginal perplexities, and show that the marginal perplexity can be significantly better than the one best, especially on out-of-domain data. We link this difference in perplexity to the tokeniser uncertainty as measured by tokeniser entropy. We discuss some implications of our results for language model training and evaluation, particularly with regard to tokenisation robustness.