Assessing the Importance of Frequency versus Compositionality for Subword-based Tokenization in NMT
This work clarifies a fundamental design choice in tokenization for NLP practitioners, though it is incremental in nature.
The paper investigated the relative importance of frequency versus compositionality in subword tokenization for neural machine translation, finding that frequency alone accounts for 90%-95% of BPE's performance scores in experiments across multiple language pairs.
Subword tokenization is the de facto standard for tokenization in neural language models and machine translation systems. Three advantages are frequently cited in favor of subwords: shorter encoding of frequent tokens, compositionality of subwords, and ability to deal with unknown words. As their relative importance is not entirely clear yet, we propose a tokenization approach that enables us to separate frequency (the first advantage) from compositionality. The approach uses Huffman coding to tokenize words, by order of frequency, using a fixed amount of symbols. Experiments with CS-DE, EN-FR and EN-DE NMT show that frequency alone accounts for 90%-95% of the scores reached by BPE, hence compositionality has less importance than previously thought.