WikiBank: Using Wikidata to Improve Multilingual Frame-Semantic Parsing

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4183–4189 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4183 WikiBank: Using Wikidata to Improve Multilingual Frame-Semantic Parsing Cezar Sas1, Meriem Beloucif2, Anders Søgaard1 1Department of Computer Science, University of Copenhagen 2Department of Informatics, Universit¨at Hamburg cezar.sas@gmail.com, beloucif@informatik.uni-hamburg.de, soegaard@di.ku.dk Abstract Frame-semantic annotations exist for a tiny fraction of the world’s languages, Wikidata, however, links knowledge base triples to texts in many languages, providing a common, distant supervision signal for semantic parsers. We present WIKIBANK, a multilingual resource of partial semantic structures that can be used to extend pre-existing resources rather than creating new man-made resources from scratch. We also integrate this form of supervision into an off-the-shelf frame-semantic parser and allow cross-lingual transfer. Using Google’s SLING architecture, we show significant improvements on the English and Spanish CoNLL 2009 datasets, whether training on the full available datasets or small subsamples thereof. Keywords: cross-lingual frame semantic parsing, multilinguality, data augmentation 1. Introduction Shallow semantic parsing comes in many varieties, includ- ing frame-semantic parsing (T¨ackstr¨om et al., 2015; Ring- gaard et al., 2017), semantic role labeling (SRL) (Surdeanu et al., 2008; Pradhan and Xue, 2009; Weischedel et al., 2013; Zhang et al., 2019), and semantic dependency pars- ing (Oepen et al., 2014). In this paper, we present WIK- IBANK, a multilingual resource with partial semantic de- pendency structures projected from an existing knowledge base. WIKIBANK is created automatically, and used to augment pre-existing resources, or reduce the annotation effort for low-resource languages. We show how it can be used to improve an off-the-shelf frame-semantic parser, but this resource could also be equally useful for seman- tic role labeling systems and other semantic parsing frame- works. The frames in frame-semantic parsing present ex- tended predicate-argument structures that determine “who did what to whom”, “when”, “where” and “why”. For instance, in a sentence such as John gives Mary a com- puter in the morning, a frame-semantic parser may identify John, Mary, and computer as core arguments of the “giv- ing” frame, and in the morning as a temporal, optional argu- ment. This kind of information has been shown to improve multiple downstream tasks such as information extraction (Bastianelli et al., 2013), machine translation (Knight and Luk, 1994; Ueffing et al., 2007; Wu and Fung, 2009; Shi et al., 2016; Beloucif and Wu, 2018), discourse parsing (Mi- haylov and Frank, 2016), question answering (Surdeanu et al., 2003; Moschitti et al., 2003; Shen and Lapata, 2007; Berant and Liang, 2014) and classifying reason and stance in online debates (Hasan and Ng, 2014). Frame-semantic parsers, however, are normally trained on manually anno- tated resources such as the FrameNet corpus (Baker et al., 1998) or the OntoNotes corpus (Pradhan and Xue, 2009; Weischedel et al., 2013). However, such annotations only exist for a small subset of the world’s languages. Some recent work (Aminian et al., 2019) solves this problem us- ing annotation projection. Moreover, in contrast with pre- vious methods, the …

Content truncated to 500 words