How Can We Investigate Ancient Greek Categories Without the Influence of Our Own? Exploring Kinship Terminology Using Word2Vec

Abstract

Recent advances in Natural Language Processing (NLP) continue to open up new lines of enquiry for lexicography, especially for research areas concerned with semantic categorisation. The aims of this paper fall into two parts. In the first half, the NLP tool Word2Vec is introduced and discussed with respect to unsupervised category formation for Koine Greek, using kinship terms as a demonstration of the model’s utility. The latter half of the paper employs a prototype theory framework to analyse the model’s distributional information in the construction of new dictionary entries for each kinship term. Such NLP-based research highlights the importance of encyclopaedic information for the construction of lexical entries; it also significantly improves upon earlier attempts at categorisation for Koine Greek (Louw and Nida 1988), which fall victim to using the L1 categories of the lexicographer in the construction of semantic domains for the ancient language.

Applying Distributional Semantic Models to a Historical Corpus of a Highly Inflected Language: the Case of Ancient Greek

“…Perrone et al (2021) compare the results of two context-predict models to a dynamic Bayesian mixture model for the task of detecting semantic change, but conclude that the latter approach delivers superior results over the context-predict models. List (2022) investigates how Word2Vec models can be used for lexicographic purposes. Finally, recently various transformer models have been trained for Ancient Greek, including Singh et al (2021) (BERT), Yamshchikov et al (2022) (BERT) and Spanopoulos (2022) (RoBERTa).…”