Training a Swedish Constituency Parser on Six Incompatible Treebanks

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 5219–5224 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 5219 Training a Swedish Constituency Parser on Six Incompatible Treebanks Richard Johansson♣, Yvonne Adesam♠ ♣Department of Computer Science and Engineering, University of Gothenburg, Sweden ♠Spr˚akbanken, Department of Swedish, University of Gothenburg, Sweden richard.johansson@gu.se, yvonne.adesam@gu.se Abstract We investigate a transition-based parser that uses Eukalyptus, a function-tagged constituent treebank for Swedish which includes discontinuous constituents. In addition, we show that the accuracy of this parser can be improved by using a multitask learning architecture that makes it possible to train the parser on additional treebanks that use other annotation models. Keywords: constituency parsing, Swedish, multitask learning 1. Introduction Syntactic parsing is a widely used intermediate step in sev- eral natural language processing (NLP) tasks. For the last couple of decades, syntactic parsing has largely been based on machine learning systems, trained in a supervised fash- ion using large collections of hand-annotated sentences – treebanks. While the accuracy of automatic parsers had largely reached a plateau a few years ago, the introduction of deep learning techniques has led to recent improvements in accuracy. Hand-annotated treebanks are based on linguistic models that can vary drastically. The most widely noted differ- ence among annotation models for syntax is probably the divergence between constituency models that conceive of the sentence structure as consisting of hierarchically orga- nized phrases or constituents, and dependency models that represent the sentence structure as a graph where the to- kens (words) are the edges. To exemplify this divergence, Figure 1 shows the Swedish sentence S˚ant tror jag inte p˚a. ‘I don’t believe in that kind of stuff.’ annotated according to constituency and dependency models. However, even within one class of models, there are many theoretical de- sign choices that can vary between different treebanks. For syntactic parsing, as for supervised learning in general, the availability of a substantial number of annotated exam- ples is crucial. Manual annotation is time-consuming and re-using previously annotated treebanks would therefore be beneficial. However, the variability of treebanks is a ma- jor nuisance: if we want to develop a parser that uses some particular annotation model, it is not evident that any other available treebanks can be used. This is an obvious prob- lem for a language such as Swedish, for which the largest treebank (Nilsson et al., 2005) is significantly smaller than treebanks for e.g. English, Chinese, and German. At the same time multiple treebanks with different types of an- notation – different types of encoded linguistic knowledge – are available. We are aware of at least five unique tree- banks for Swedish, most of which are annotated according to their own respective models (some of them more than one model). Is there a remedy: can we develop an approach that can utilize additional treebanks? For non-neural, feature- based dependency parsers, Johansson (2013) proposed two Sånt tror jag inte på . OO HD PP OA HD SB MD …

Content truncated to 500 words