A Digital ‘Metal Detector’ for Classical Philology

A Digital ‘Metal Detector’ for Classical Philology
10 July 2020, 5.30pm - 7.00pm

Harry Tanner, Galway

This event will be held online at https://www.youtube.com/watch?v=EF1LDSePIIg&feature=youtu.be

Given the vast time frame we study in Classics, it is inevitable that words will change their meaning during the course of the material we read. But how can a digital tool be used to track those changes and pinpoint areas for Classical philology to investigate? This paper presents a new method for Classics: by training a neural network on partitioned time intervals (e.g. 800BCE to 600BCE; 600BCE to 400BCE) of the Diorisis corpus of Greek text, a word embedding for each word in each time period is built. In studies on English Word embeddings (x dimensional vectors which describe a word’s distribution in text), they have been shown to capture an approximation of meaning: searching for the closest vector to the embedding for ‘King’ will output ‘Palace’, ‘Throne’, ‘Queen’, and ‘Crown’. By calculating the cosine distance between a word in each successive time period, a rough measure of a word’s rate of change is calculated. Were ‘King’ to change its meaning substantially, you would expect to see a high cosine distance between its use in the first period and the second. All words in the ancient Greek corpus can then be ranked by their rates of change (either high or low). The objective is to build a philological ‘metal detector’ which can alert researchers to previously under-researched words which may significantly change their meaning from one period to another. This paper interrogates the results of the model for two words which the computer has identified as significantly different in meaning, and then subjects them to manual philological analysis.


Valerie James
020 7862 8716