Can vectors show us where to go? A tentative exploration of how artificial intelligence can be used to unfold textual traditions
Can vectors show us where to go? A tentative exploration of how artificial intelligence can be used to unfold textual traditions

Summary/Abstract: Identifying how extant manuscript copies of a text are related is a central, yet highly time- and labour-intensive task in philology. In our paper, we investigate how digital methods can support textual scholarship in accelerating this process. Using a parallel corpus of Greek and Church Slavonic sentences, we fine-tuned a multilingual large language model to compute sentence embeddings for these two varieties. We used this model for intra- and interlingual comparisons of manuscripts preserving the Greek and Slavonic versions of the Dioptra. This work’s textual tradition has been extensively studied, and it thus provides a benchmark for evaluating our model. Our results are encouraging: in two of three test cases, we reproduce the findings of traditional textual research, thus providing an initial validation of our approach. Experiments with noise augmented data suggest that our approach also works with imperfect automatic manuscript transcriptions. The correct identification of the relationship among the Slavonic witnesses proved to be the most challenging task. In the discussion, we outline current limitations and future improvements, including orthographic simplification and scaling to larger and more balanced training corpora to address the problem of linguistic and orthographic heterogeneity of Medieval Church Slavonic texts.

• Details
• Contents