A Multi-domain Corpus of Swedish Word Sense Annotation

A Multi-domain Corpus of Swedish Word Sense Annotation Richard Johansson∗†, Yvonne Adesam†, Gerlof Bouma†, Karin Hedberg† ∗Department of Computer Science and Engineering, University of Gothenburg, Sweden †Spr˚akbanken, University of Gothenburg, Sweden richard.johansson@gu.se, yvonne.adesam@gu.se, gerlof.bouma@gu.se, kkmhedberg@gmail.com Abstract We describe the word sense annotation layer in Eukalyptus, a freely available five-domain corpus of contemporary Swedish with several annotation layers. The annotation uses the SALDO lexicon to define the sense inventory, and allows word sense annotation of compound segments and multiword units. We give an overview of the new annotation tool developed for this project, and finally present an analysis of the inter-annotator agreement between two annotators. Keywords: word sense, annotation, Swedish 1. Introduction In an ongoing project on natural language processing for Swedish, Koala, we are investigating the possibilities of combining different levels of annotation so as to improve overall quality of automatic processing. As part of this, we are constructing a manually annotated reference cor- pus, Eukalyptus, with annotation of morphological struc- ture, lexical sense and syntactic structure. In previous pa- pers, we have presented the morphological and syntactic annotation in this corpus (Adesam et al., 2015a), and dis- cussed the methodological choices underlying our annota- tion of multiword units (Adesam et al., 2015b). In this paper, we describe the word sense annotation in the Eukalyptus corpus, which will eventually be used to evalu- ate tools for automatic word sense disambiguation (WSD). We first introduce the corpus itself and its composition, and then describe SALDO, the Swedish semantic lexicon that provides the sense inventory we use for this annotation. We then describe the annotation process, including the new an- notation tool that we have developed, and discuss the relia- bility of the annotation by making an inter-annotator agree- ment study. 2. The Eukalyptus Corpus The Eukalyptus corpus consists of around 100,000 tokens of contemporary Swedish text. Since a goal of the project is to measure the domain sensitivity of Swedish NLP tools, we have collected text from five different genres, each of which contains about 20,000 tokens: • Fiction: the first chapters from four novels, • Encyclopedic: full articles from the Swedish Wikipedia, 100 to 3,000 tokens per article, • Social media: blog entries from the SIC corpus ( ¨Ostling, 2013), • Political debates: proceedings from the European par- liament (Koehn, 2005), • Professional prose: a mix of different types of in- formation from the goverment, and articles from Arbetaren, a Swedish weekly. For all genres, we have made sure that the text can be dis- tributed freeely, and the corpus and all its annotation will eventually be released under a Creative Commons license. 3. The SALDO Lexicon The word sense annotation described in this paper uses SALDO (Borin et al., 2013) to define its sense inventory. While there are alternative Swedish sense inventories such as the Swedish WordNet (Viberg et al., 2003), SALDO has the advantage of being licensed under a Creative Commons license, and it is also the largest resource of this kind for Swedish: as of October 2015, …

Content truncated to 500 words