Parse Me if You Can: Artificial Treebanks for Parsing Experiments on Elliptical Constructions

Parse Me if You Can: Artificial Treebanks for Parsing Experiments on Elliptical Constructions Kira Droganova,1 Daniel Zeman,1 Jenna Kanerva,2 Filip Ginter2 1Charles University, Faculty of Mathematics and Physics, Praha, Czechia 2TurkuNLP Group, University of Turku, Finland 1{droganova, zeman}@ufal.mff.cuni.cz, 2{jmnybl, figint}@utu.fi Abstract In this work we focus on a particular linguistic phenomenon, ellipsis, and explore the latest parsers in order to learn about parsing accuracy and typical errors from the perspective of elliptical constructions. For this purpose we collected and processed outputs of several state-of-the art parsers that took part in the CoNLL 2017 Shared Task. We extended the official shared task evaluation software to obtain focused evaluation of elliptical constructions. Since the studied structures are comparatively rare, and consequently there is not enough data for experimentation, we further describe the creation of a new resource, a semi-artificially constructed treebank of ellipsis. Keywords: Ellipsis, Syntactic parsing, Evaluation, Universal Dependencies 1. Introduction Ellipsis, i.e. omission of linguistic content that is silently understood by both the speaker and the addressee, is a phe- nomenon present—in various forms—in many natural lan- guages. Ellipsis obviously makes natural language under- standing harder; but sometimes it also complicates syntac- tic parsing of the content that is not omitted. In depen- dency syntax (which is the framework within which we op- erate), a parent node may be missing while its dependents are present. One might either create an “empty” node for the missing word, or choose a substitute parent among the words that are not missing. Both options make parsing dif- ficult: in the former case, the parser must learn where to generate empty nodes; in the latter, relations are drawn be- tween nodes that would not be connected otherwise, hence they are not easily learned from data. In any case, modern dependency parsers are data-driven and they can hardly account for those types of ellipsis that are not represented in training data. If the data contains empty nodes, the parser can try to learn generating them. If the data does not contain any specific annotation of ellipsis, we have to hope that the parser learns to occasionally attach dependents to strange parents, even without knowing that it is ellipsis what caused the lack of better options. In this study we focus on elliptical constructions in the so-called basic representation of Universal Dependencies (UD) (Nivre et al., 2016). The annotation style of UD does not mark ellipsis explicitly when it does not have to: most types are solved by simply promoting one orphaned depen- dent to the position of its missing parent. Admittedly, there are treebanks that overtly annotate a wider range of ellip- tical structures. Our main reason for working with UD is practical: substantial data is available in this annotation style for several dozens of languages, and state-of-the art parsers have been trained and tested on UD. The one exception where UD explicitly marks ellipsis are certain types of gapping and stripping (Droganova and Ze- man, 2017), where multiple orphaned dependents of a miss- ing predicate …

Content truncated to 500 words