Old French Dependency Parsing: Results of Two Parsers Analysed from a Linguistic Point of View

Old French Dependency Parsing: Results of Two Parsers Analysed from a Linguistic Point of View Achim Stein Institut f¨ur Linguistik/Romanistik, Universit¨at Stuttgart achim.stein@ling.uni-stuttgart.de Abstract The treatment of medieval texts is a particular challenge for parsers. I compare how two dependency parsers, one graph-based, the other transition-based, perform on Old French, facing some typical problems of medieval texts: graphical variation, relatively free word order, and syntactic variation of several parameters over a diachronic period of about 300 years. Both parsers were trained and evaluated on the Syntactic Reference Corpus of Medieval French (SRCMF), a manually annotated dependency treebank. I discuss the relation between types of parsers and types of language, as well as the differences of the analyses from a linguistic point of view. Keywords: dependency parsing, medieval texts, Old French, error analysis 1. Introduction 1.1. Previous and related work The treatment of Medieval texts in general, and Old French (OF) in particular is a challenge for NLP. The first issue is variation. Previously I have shown that the problem of syntactic variation in a treebank of OF texts spanning over 300 years is not unsurmountable, since a global model trained on the totality of the texts is almost as accurate as specific models trained for subcorpora (Stein, 2014). The results of the graph-based mate tools parser (Bohnet, 2010; Bj¨orkelund et al., 2010) were quite satisfac- tory (LAS: 82.62%). The second issue is the relatively free word order in OF, in combination with the null-subject property, i.e. sentences need not have an overt subject (like e.g. Modern Italian or Spanish). The third issue is that with regard to inflection, OF is closer to e.g. Modern German than to Modern English. Verbs are marked for person, number, and tense/mood; nouns, pro- nouns, and adjectives are marked for number and case (OF has a two-case system), see also Table 3. 1. Issue 1 is not the main focus of this paper, but com- pared to the previous experiment (Stein, 2014), a dif- ferent part-of-speech tagger was selected in order to improve the morphological analysis despite of the or- thographical variation (see section 3.2.). 2. With regard to issue 2, it has been argued that transition-based parsers are assumed to be more suit- able for languages with relatively free word order. This is due to the fact that transition-based parsers are more “dynamic”, in that they learn sequences of tran- sitions that are applied successively to the words of a sentence by taking words from a queue, and adding processed words to a stack. Graph-based parsers are more “static”, in that they learn a model over complete dependency graphs by summing up all the attachment scores in a sentence. 3. With regard to issue 3, “joint morphological and syn- tactic disambiguation is especially important for richly inflected languages, where there is considerable inter- action between morphology and syntax such that nei- ther can be fully disambiguated without considering the other” (Bohnet et al., 2013). Further properties of graph-based vs transition-based approaches are dis- cussed …

Content truncated to 500 words