Recovering Registers from Leveled Wordlists

Abstract

For vocabulary learning in language acquisition, it is desirable for learners to acquire words that they are likely to need in the language environments they will encounter. Such language environments are referred to as “registers” in general corpora, which are typically designed to include diverse registers. However, the proportion of registers included, that is, which registers are included and to what extent, is determined by the circumstances under which each general corpus was compiled and is not necessarily optimized for language learning. To bridge this gap, various leveled wordlists have been created in language education using linguistic resources other than word frequency, such as expert judgment and learner responses. However, it has not been quantitatively clear what gap in register proportions in general corpora these leveled wordlists were designed to fill. This study proposes a method that, given a leveled wordlist and a general corpus, estimates the register ratio that best aligns the frequency ordering of words across registers with the leveled wordlist. This makes it easier for learners and educators to interpret which wordlists are appropriate for particular learning goals. Our method is formulated as a linear programming problem and yields a globally optimal solution. Unlike neural networks, it is less susceptible to variation due to initial values or approximation and is therefore easier to interpret. We evaluated the proposed method on two languages, English and Japanese, through a range of experiments. We further show that it can also be used to evaluate vocabulary lists created for specific contexts, such as those generated by Large Language Models like ChatGPT.

Resources

Details