Evaluating Subword Tokenization: Alien Subword Composition and OOV Generalization Challenge

TL;DR

This paper introduces a new evaluation framework for subword tokenization, combining intrinsic morphological classification with extrinsic OOV generalization tests, revealing that alien tokenization hampers model performance.

Contribution

It proposes a novel combined intrinsic-extrinsic evaluation framework for subword tokenization, including a new tool and benchmark for assessing morphological versus alien tokenization effects.

Findings

UniMorph Labeller achieves 98% accuracy

Alien tokenization results in poorer OOV generalization

Morphological tokenization improves semantic compositionality

Abstract

The popular subword tokenizers of current language models, such as Byte-Pair Encoding (BPE), are known not to respect morpheme boundaries, which affects the downstream performance of the models. While many improved tokenization algorithms have been proposed, their evaluation and cross-comparison is still an open problem. As a solution, we propose a combined intrinsic-extrinsic evaluation framework for subword tokenization. Intrinsic evaluation is based on our new UniMorph Labeller tool that classifies subword tokenization as either morphological or alien. Extrinsic evaluation, in turn, is performed via the Out-of-Vocabulary Generalization Challenge 1.0 benchmark, which consists of three newly specified downstream text classification tasks. Our empirical findings show that the accuracy of UniMorph Labeller is 98%, and that, in all language models studied (including ALBERT, BERT, RoBERTa, and DeBERTa), alien tokenization leads to poorer generalizations compared to morphological tokenization for semantic compositionality of word meanings.