Differentia compositionem facit. A Slower-Paced and Reliable Parser for Latin

Differentia compositionem facit A Slower-paced and Reliable Parser for Latin Edoardo Maria Ponti‡, Marco Passarotti⋆ ‡University of Pavia, ⋆CIRCSE Research Centre – Università Cattolica del Sacro Cuore Corso Strada Nuova 65 – 27100 Pavia – Italy, Largo A. Gemelli 1 – 20123 Milan – Italy edoardomaria.ponti01@universitadipavia.it, marco.passarotti@unicatt.it Abstract The Index Thomisticus Treebank is the largest available treebank for Latin; it contains Medieval Latin texts by Thomas Aquinas. After experimenting on its data with a number of dependency parsers based on different supervised machine learning techniques, we found that DeSR with a multilayer perceptron algorithm, a right-to-left transition, and a tailor-made feature model is the parser providing the highest accuracy rates. We improved the results further by using a technique that combines the output parses of DeSR with those provided by other parsers, outperforming the previous state of the art in parsing the Index Thomisticus Treebank. The key idea behind such improvement is to ensure a sufficient diversity and accuracy of the outputs to be combined; for this reason, we performed an in-depth evaluation of the results provided by the different parsers that we combined. Finally, we assessed that, although the general architecture of the parser is portable to Classical Latin, yet the model trained on Medieval Latin is inadequate for such purpose. Keywords: Combination, Dependency Parsing, Latin 1 Introduction Latin language displays a rich fusive morphology. Mor- phemes may carry several grammatical meanings, syntac- tic constituents are often discontinuous, and word order is quite free, as it can be determined more on pragmatic grounds than by syntax. Such a free word order status for Latin results in a high number of non-projective pairs of nodes in dependency trees, which in turn affects negatively parsing accuracy (Nivre and Nilsson, 2005). This partly ex- plains why parser performances for Latin lag behind other languages. Furthermore, Latin suffers from lacking sufficiently large language resources, and in particular treebanks. Also, texts written in Latin are scattered across a wide span of periods (covering more than two millennia), territories and styles. Since no selection of texts can be representative enough of Latin as a whole, any experiment run on textual data in this language is only able to deal with its particular variety rep- resented by the specific data in question. In this respect, Passarotti and Ruffolo (2010) showed that, while testing on Classical Latin a model for syntactic parsing trained on Me- dieval Latin data, the quality of results dramatically drops. So far, the highest accuracy rates for dependency parsing of Latin are those reported by Passarotti and Dell’Orletta (2010), who trained a parsing model on the Index Thomisti- cus Treebank1 (IT-TB; Passarotti (2011)). Indeed, the IT- TB is larger than other comparable resources for Latin, like for instance the Latin Dependency Treebank (LDT) for Classical Latin, developed by Bamman and Crane (2007). Moreover, the texts of the IT-TB are written in quite a formal variety of Medieval Latin. In this variety, non- projectivity rate is milder than in Classical Latin: according to Passarotti …

Content truncated to 500 words