Hierarchical Text Segmentation for Medieval Manuscripts

Proceedings of the 28th International Conference on Computational Linguistics, pages 6240–6251 Barcelona, Spain (Online), December 8-13, 2020 6240 Hierarchical Text Segmentation for Medieval Manuscripts Amir Hazem* B´eatrice Daille* Louis Chevalier§ Dominique Stutzmann§ Christopher Kermorvant†‡ * LS2N - Universit´e de Nantes, Nantes § IRHT - Institut de recherche et d’histoire des textes, CNRS, Paris † TEKLIA, Paris ‡ LITIS - Universit´e de Rouen-Normandie, Rouen {amir.hazem, beatrice.daille}@ls2n.fr {dominique.stutzmann, louis.chevalier}@irht.cnrs.fr {kermorvant}@teklia.com Abstract In this paper, we address the segmentation of books of hours, Latin devotional manuscripts of the late Middle Ages, that exhibit challenging issues: a complex hierarchical entangled structure, variable content, noisy transcriptions with no sentence markers, and strong correlations between sections for which topical information is no longer sufficient to draw segmentation boundaries. We show that the main state-of-the-art segmentation methods are either inefficient or inapplicable for books of hours and propose a bottom-up greedy approach that considerably enhances the segmentation results. We stress the importance of such hierarchical segmentation of books of hours for historians to explore their overarching differences underlying conception about Church. 1 Introduction Text segmentation is essential in many downstream applications including document understanding and navigation, summarization, information retrieval and discourse parsing (Purver, 2011; Riedl and Bie- mann, 2012; Li et al., 2018). Traditional unsupervised approaches assume a high correlation between segments and subtopics. Therefore, based on a prior text decomposition, two adjacent segments are merged if they are highly correlated. On the contrary, if their similarity is below a certain threshold, a shift is determined (Hearst, 1997; Riedl and Biemann, 2012). When sufficient topically annotated train- ing data are available, deep neural approaches based on CNN (Wang et al., 2017) or LSTM (Koshorek et al., 2018) can be efficiently applied (Li et al., 2018; Arnold et al., 2019). Until now, text segmentation methods have exclusively addressed data sets lying within the scope of narrative and expository texts or user dialogues texts and sometimes artificially generated data (Choi, 2000; Jeong and Titov, 2010; Glavaˇs et al., 2016; Koshorek et al., 2018). In this paper, we address transcriptions of ancient devotional manuscripts (thereafter also “MS”), from the Middle Ages, known as “books of hours”. Books of hours were used by lay people as a guidance in their daily prayers. They represent an important source of information on the late Middle Ages’ religious and social practices, and provide opportunities for historical analysis in order to better understand the cultures and faiths of the European society. More than 10,000 manuscripts of ca. 300 pages in aver- age are preserved: they are a specific, standardized, and proto-industrial production (Stutzmann, 2019). However, their textual content is still scarcely studied, because of the lack of transcriptions, the com- plexity of their liturgical content and their standard appearance. They look very uniform at first glance, but each of them is unique. The content differ from one copy to the other as a large choice of devotional readings are available. Conversely, some prayers are used in several hours of the day and several times within …

Content truncated to 500 words