“One” in 768 Dimensions: Using Language Models to Study the Usage of <i>yi</i> 一 on a Large Scale

Abstract With the availability of large corpora that can be queried using computational methods, the importance of considering large amounts of usage data to accurately describe the behavior of a word has increasingly been recognized. The most common method to conduct such research is the reading of concordances, i.e. compilations of a word in its surrounding contexts sampled from a corpus. In this study, an unsupervised way of clustering concordances using high-dimensional contextual word embeddings that capture semantic as well as syntactic information about a word is proposed, and tested on yi 一 (one) in a large corpus of pre-modern Chinese. The results show that the approach yields clusters that can be fruitfully explored in a Construction Grammar framework but might be inconsistent in the sense that instances of what a researcher would consider one phenomenon can be split over more than one cluster.