Linguistic annotation of cuneiform texts using treebanks and deep learning

Abstract

We describe an efficient pipeline for morpho-syntactically annotating an ancient language corpus which takes advantage of bootstrapping techniques. This pipeline is designed for ancient language scholars looking to jump-start their own treebank projects, which can in turn serve further pedagogical research projects in the target language. We situate our work in the field of similar ancient language treebank projects, arguing that our approach shows that individual humanities scholars can leverage current machine-learning tools to produce their own richly annotated corpora. We illustrate this pipeline by producing a new Akkadian-language treebank based on two volumes from the online editions of the State Archives of Assyria project hosted on Oracc, as well as a spaCy language model named AkkParser trained on that treebank. Both of these are made publicly available for annotating other Akkadian corpora. In addition, we discuss linguistic issues particular to the Neo-Assyrian letter corpus and data-encoding complications of cuneiform texts in Oracc. The strategies, language models, and processing scripts we developed to handle both linguistic and data-encoding issues in this project will be of special interest to scholars seeking to develop their own cuneiform treebanks.