Multi Modal Distance - An Approach to Stemma Generation With Weighting

Multi Modal Distance - An Approach to Stemma Generation with Weighting Armin Hoenen CEDIFOR Goethe University Frankfurt hoenen@em.uni-frankfurt.de Abstract Stemma generation can be understood as a task where an original manuscript M gets copied and copies – due to the manual mode of copying – vary from each other and from M. Copies M1, .., Mk which survive historical loss serve as input to a mapping process estimating a directed acyclic graph (tree) which is the most likely representation of their copy history. One can first tokenize and align the texts of M1, .., Mk and then produce a pairwise distance matrix between them. From this, one can finally derive a tree with various methods, for instance Neighbor-Joining (NJ) (Saitou and Nei, 1987). For computing those matrices, previous research has applied unweighted approaches to token similarity (implicitly interpreting each token pair as a binary observation: identical or different), see Mooney et al. (2003). The effects of weighting have then been investigated and Spencer et al. (2004b) found them to be small in their (not necessarily all) scenario(s). The present approach goes beyond the token level and instead of a binary comparison uses a distance model on the basis of psycholinguistically gained distance matrices of letters in three modalities: vision, audition and motorics. Results indicate that this type of weighting have positive effects on stemma generation. Keywords: stemmatology, weighted distance, multi modal Figure 1: First modern stemma by Schlyter, 1827, from O’Hara (1996) with texts= nodes, copy processes= edges. 1. Introduction Stemma generation is the process of determining the most likely tree1 with manuscripts being represented by nodes in the usually directed acyclic graph2 and edges representing copy processes or chains of such, see Figure 1. Since the late 1950ies computational methods have been applied to stemmatological tasks (Ellison, 1957) and in the last decade evaluation against benchmark datasets (or artificial tradi- tions) has been conducted (Baret et al., 2004; Spencer et al., 2004a; Roos and Heikkil¨a, 2009; Hoenen, 2015a). These datasets have been generated by first giving one text (root) 1Insights and historical considerations on the tree as a pre- ferred technical model for stemmatology and concurrent graph theoretical considerations are discussed for instance in Hoenen et al. (2017), Flight (1994) and sources therein. 2Note, that we target stemmata for closed traditions (Pasquali and Pieraccioni, 1952) that have no multiple originals, as is prob- able for orally transmitted epics (Lord, 1960). Hoenen (2017) has attempted to reconcile tree and network perspectives on stemmata. to volunteers to be handcopied (or dictated). Its copies have then been handcopied again and so forth. The true vor- lage3-copy relations (edges in the true stemma) have been recorded by the authors, so that we know the entire true tree with all edges and the position of root. Those manuscript texts have been digitized and manually aligned for a stem- matology challenge (Roos and Heikkil¨a, 2009) where from a subset of manuscripts several teams attempted to recon- struct – manually or automatically – the true tree.4 …

Content truncated to 500 words