Accurate Deep Syntactic Parsing of Graphs: The Case of French

Accurate Deep Syntactic Parsing of Graphs: The Case of French Corentin Ribeyre◦ Eric Villemonte de la Clergerie⋆ Djamé Seddah⋆ ⋄ ◦Université de Genève ⋆Alpage, INRIA ⋄ Université Paris-Sorbonne corentin.ribeyre@unige.ch eric.de_la_clergerie@inria.fr djame.seddah@paris-sorbonne.fr Abstract Parsing predicate-argument structures in a deep syntax framework requires graphs to be predicted. Argument structures represent a higher level of abstraction than the syntactic ones and are thus more difficult to predict even for highly accurate parsing models on surfacic syntax. In this paper we investigate deep syntax parsing, using a French data set (Ribeyre et al., 2014a). We demonstrate that the use of topologically different types of syntactic features, such as dependencies, tree fragments, spines or syntactic paths, brings a much needed context to the parser. Our higher-order parsing model, gaining thus up to 4 points, establishes the state of the art for parsing French deep syntactic structures. Keywords: deep syntax, graph parsing, syntactic features 1. Introduction The availability of many manually annotated syntactic cor- pora allows data-driven syntactic parsing to perform effi- ciently on English (routinely reaching 92% of labeled at- tachment accuracy) as well as on languages with a richer level of morphology (85% to 88%) (e.g. French, German, Arabic). However, despite this impressive level of perfor- mance, it has now become clear that such surface syntac- tic parses are often insufficient for semantically-oriented tasks such as question-answering systems (Berant et al., 2013). This is because many predicate-argument depen- dencies, such as those arising for instance in control-verb constructions, it-cleft constructions, participle clauses and so on, are lacking from the annotation schemes underlying most data set used by surface syntactic parsers. Representing such constructs in a dependency scheme of- ten leads to graph representations, which are a real chal- lenge to predict, as shown, for example, by the perfor- mance obtained in Abstract Meaning Representation pars- ing (Banarescu et al., 2013; Flanigan et al., 2014; Artzi et al., 2015) or in graph-based semantic dependency pars- ing (Oepen et al., 2014). Even though some recent works have made good progresses in parsing full-fledged semantic structures (Beschke et al., 2014; Berant and Liang, 2014), they mostly focus on English.1 We propose to dig into that direction by relying on the re- cent release of a deep syntactic graphbank for French, the DEEPFTB, (Ribeyre et al., 2014a), whose deep syntactic graphs were automatically annotated on top of the depen- dency version of the French Treebank (Abeillé and Barrier, 2004; Candito et al., 2010). In order to draw meaning- ful comparisons with recent results for English, we use the same types of features as Ribeyre et al. (2015), which were extracted from the DM corpus (Oepen et al., 2014), an En- glish predicate-argument structure corpus sharing similari- ties with the DEEPFTB. Despite the differences of language and annotation scheme, we observe that the same combination of different topolog- 1An exception is the work of Ballesteros et al. (2014) on deep syntacting parsing for Spanish, but their work is restricted to tree structure parsing. ical syntactic information leads to the …

Content truncated to 500 words