BERT in Plutarch’s Shadows

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6071 - 6080 December 7-11, 2022 ©2022 Association for Computational Linguistics BERT in Plutarch’s Shadows Ivan P. Yamshchikov Max Planck Institute for Mathematics in the Sciences Leipzig, Germany CEMAPRE, University of Lisbon, Portugal ivan@yamshchikov.info Alexey Tikhonov Inworld.AI Berlin, Germany altsoph@gmail.com Yorgos Pantis Max Planck Institute for Mathematics in the Sciences Leipzig, Germany Charlotte Schubert University of Leipzig Leipzig, Germany Jürgen Jost Max Planck Institute for Mathematics in the Sciences Leipzig, Germany Abstract The extensive surviving corpus of the ancient scholar Plutarch of Chaeronea (ca. 45-120 CE) also contains several texts which, according to current scholarly opinion, did not originate with him and are therefore attributed to an anony- mous author Pseudo-Plutarch. These include, in particular, the work Placita Philosopho- rum (Quotations and Opinions of the Ancient Philosophers), which is extremely important for the history of ancient philosophy. Little is known about the identity of that anonymous author and its relation to other authors from the same period. This paper presents a BERT language model for Ancient Greek. The model discovers previously unknown statistical prop- erties relevant to these literary, philosophical, and historical problems and can shed new light on this authorship question. In particular, the Placita Philosophorum, together with one of the other Pseudo-Plutarch texts, shows similar- ities with the texts written by authors from an Alexandrian context (2nd/3rd century CE). "I do not need a friend who changes when I change and who nods when I nod; my shadow does that much better." (Plutarch, Quomodo adulator ab am- ico internoscatur 53b 10) 1 Introduction Authorship attribution through some form of statis- tical inference dates at least half of a century back, when Mosteller and Wallace used statistics of short, frequent words to estimate authorship of Federalist Papers, disputed by Mosteller and Wallace (1963). The physicist Fucks (1968) was the first who sys- tematically developed the methods that relied on statistical patterns like word or sentence length that have been found to be able to distinguish between different authors. While undeniably successful in many cases, they do not require any deeper under- standing of the texts in question and therefore have their natural limitations. For a detailed review of various authorship attribution methods developed further, we refer the reader to (Stamatatos, 2009). These methods rely on statistical patterns like word or sentence length that have been found to be able to distinguish between different authors. While undeniably successful in many cases, they do not require any deeper understanding of the texts in question and therefore have their natural limita- tions. Transformer artificial neural networks have shown spectacular success in machine translation and answering search queries by internally extract- ing, after extensive pretraining, abstract patterns, and long-range dependencies in human language samples through encoder hierarchies (Vaswani et al., 2017). Therefore, it seems natural to ap- ply such schemes to other tasks that traditionally depended on human language understanding. Thus, this paper uses versions of BERT to investigate questions …

Content truncated to 500 words