A Tree Extension for CoNLL-RDF

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 7161–7169 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 7161 A Tree Extension for CoNLL-RDF Christian Chiarcos, Luis Glaser Applied Computational Linguistics Lab (ACoLi) Goethe University Frankfurt, Germany {chiarcos, lglaser}@informatik.uni-frankfurt.de Abstract The technological bridges between knowledge graphs and natural language processing are of utmost importance for the future development of language technology. CoNLL-RDF is a technology that provides such a bridge for popular one-word-per-line formats with tab-separated values (TSV) that are widely used in NLP (e.g., the CoNLL Shared Tasks), annotation (Universal Dependencies, Unimorph), corpus linguistics (Corpus WorkBench, CWB) and digital lexicography (SketchEngine): Every empty-line separated table (usually a sentence) is parsed into an graph, can be freely manipulated and enriched using standard graph technology, and then be serialized back into in a TSV format, RDF or other formats. An important limitation is that CoNLL-RDF provides native support for word-level annotations only. This does include dependency syntax and semantic role annotations, but neither phrase structures nor text structure. We describe the extension of the CoNLL-RDF technology stack for two vocabulary extensions of CoNLL-TSV, the bracket notation used in earlier CoNLL Shared Tasks and the extension with XML markup elements featured by CWB and SketchEngine. In order to represent the necessary extensions of the CoNLL vocabulary in an adequate fashion, we employ the POWLA vocabulary for representing and navigating in tree structures. Keywords: linguistic annotation, data formats, interoperability, tab-separated values, RDF, graphs, trees, CoNLL-RDF, POWLA 1. Introduction & Motivation Since their original conception in the early 1990s, one- word-per-line formats, mostly based on tab-separated val- ues (TSV), have become enormously influential in cor- pus linguistics (Evert and Hardie, 2011, Corpus Work- bench/CWB),1 digital lexicography (Kilgarriff et al., 2014, SketchEngine),2 natural language processing (Schmid, 1994, TreeTagger), and as exchange format of various an- notation projects (Universal Dependencies (Nivre et al., 2016), PropBank, UniMorph).3 The main reasons for their continuing popularity are the apparent simplicity of tables and tab-separated values, and their expressive power that allows for the flexible and seamless extension of annota- tions with additional columns and novel annotations, mak- ing these formats capable of expressing and combining any kind of word-level annotation in a compact, human- readable, and machine-processable format. Yet another factor contributing to the popularity of TSV for- mats in natural language processing is their application as exchange and evaluation format(s) of the Shared Tasks of the SIGNLL Conference on Computational Natural Lan- guage Learning (CoNLL)4, conducted annually since 1999. With CoNLL data sets as an important basis for future eval- uation of most NLP tasks, their formats are supported by most existing NLP tools, and various CoNLL-TSV formats (‘dialects’) have seen a steady increase in popularity for linguistic data in NLP applications since. Even when cre- ated independently from CoNLL Shared Tasks, novel TSV formats following CoNLL conventions are now usually re- ∗ The authors contributed equally to the submission. 1http://cwb.sourceforge.net/ 2http://www.sketchengine.eu 3http://universaldependencies.org/, http://propbank.github.io/, http://unimorph. github.io/ 4http://www.signll.org/conll ferred to …

Content truncated to 500 words