Can Transformers Learn $n$-gram Language Models?

TL;DR

This paper investigates whether transformer models can learn n-gram language models, comparing their performance to traditional methods on different types of n-gram tasks, revealing conditions where transformers excel or underperform.

Contribution

The study empirically evaluates transformers' ability to learn random n-gram language models, contrasting their performance with classical methods and identifying scenarios where transformers outperform specialized n-gram learners.

Findings

Transformers outperform classical methods on n-gram models with shared parameters.

Classical estimation techniques outperform transformers on arbitrary probability n-gram models.

Transformers excel in learning n-gram models with shared parameters, surpassing traditional methods.

Abstract

Much theoretical work has described the ability of transformers to represent formal languages. However, linking theoretical results to empirical performance is not straightforward due to the complex interplay between the architecture, the learning algorithm, and training data. To test whether theoretical lower bounds imply \emph{learnability} of formal languages, we turn to recent work relating transformers to $n$-gram language models (LMs). We study transformers' ability to learn random $n$-gram LMs of two kinds: ones with arbitrary next-symbol probabilities and ones where those are defined with shared parameters. We find that classic estimation techniques for $n$-gram LMs such as add-$\lambda$ smoothing outperform transformers on the former, while transformers perform better on the latter, outperforming methods specifically designed to learn $n$-gram LMs.