Abstract
Annotation of unlabeled textual corpora with linguistic features is a fundamental technology in the computational study of cultural heritage materials and digital humanities (DH). Recent advances in NLP have leveraged pretrained language models to offer simultaneous tokenization, tagging, and dependency parsing of unlabeled text, using annotation schemes like Universal Dependencies (UD). The capability of LMs for continued pretraining on unlabeled data different from the training data has also made them suitable for adaptation to disparate domains. However, the adoption of these methods into DH workflows remains on the periphery. This can be attributed to differences in research objectives: NLP is centered on prediction performance and assigns less attention to proposals that enable helpful evaluation by scholars. This thesis proposes the application of novel substitution rules that are designed to improve UD output generated by pretrained LMs in the case of historical text and investigates the use of pretrained LMs as a mechanism for refining the proposed rule-based system. The research is conducted using Modern Historical Japanese (MHJ) corpora as a case study. The thesis begins by formulating the use of TEI XML as output for markup of Japanese text, using features generated by NLP annotation tools. An alignment strategy is proposed for markup of branching and segmented text and a conformant TEI XML customization is introduced for markup using linguistic attributes. The thesis then proposes the use of novel substitution rules: these are designed so that dependencies are parsed using substituted terms that are present in pretrained vocabularies, and then restoring historical forms in the parsed sentence. Using the rule-based system as a testbed, the thesis presents the application of domain adaptation methods to develop a model that can flag instances of inaccurate UD output from a pretrained LM. These instances facilitate development of new rules and a domain-adapted LM trained with the masked language modeling objective is applied for generation of potential candidates.