Corpus vs. Lexicon Supervision in Morphosyntactic Tagging: the Case of Slovene

Corpus vs. Lexicon Supervision in Morphosyntactic Tagging: The Case of Slovene Nikola Ljubeˇsi´c,∗† Tomaˇz Erjavec† ∗ Dept. of Information and Communication Sciences, University of Zagreb Ivana Luˇci´ca 3, HR-10000 Zagreb, Croatia nikola.ljubesic@ffzg.hr † Dept. of Knowledge Technologies, Joˇzef Stefan Institute, Jamova cesta 3, SI-1000 Ljubljana, Slovenia tomaz.erjavec@ijs.si Abstract In this paper we present a tagger developed for inflectionally rich languages for which both a training corpus and a lexicon are available. We do not constrain the tagger by the lexicon entries, allowing both for lexicon incompleteness and noisiness. By using the lexicon indirectly through features we allow for known and unknown words to be tagged in the same manner. We test our tagger on Slovene data, obtaining a 25% error reduction of the best previous results both on known and unknown words. Given that Slovene is, in comparison to some other Slavic languages, a well-resourced language, we perform experiments on the impact of token (corpus) vs. type (lexicon) supervision, obtaining useful insights in how to balance the effort of extending resources to yield better tagging results. Keywords: Part-of-Speech tagging, evaluation, Slavic languages 1. Introduction Part-of-speech or, better, morphosyntactic tagging is still an interesting topic of research, esp. for highly inflected languages, such as Czech (Strakov´a et al., 2014), Polish (Radziszewski, 2013) or Slovene (Grˇcar et al., 2012). Such languages with their large tagsets of morphosyntactic de- scriptions (MSDs) and often limited training data still of- fer significant room for improvement in tagging accuracy. A related research question is how to best split the effort needed to compile larger training corpora against extending the tagger background lexicon, a problem already investi- gated for French (Denis and Sagot, 2012). In this paper we address both questions. We develop a new tagger, esp. optimised for tagging un- known (and partially unknown) words, useful for cases where the background lexicon is small or inflectionally in- complete, and test it on Slovene data. We build on the approach by Grˇcar et al. (2012) as it is conceptually simple and is, in addition to Denis and Sagot (2012), one of the few proposals that treats tagging of seen and unseen tokens as an identical problem, utilising the knowledge from a morphosyntactic lexicon indirectly in form of classification features. It therefore does not con- sider the lexicon supplied MSDs as the only possible MSDs for the word, i.e. it uses the lexicon-as-features rather than the lexicon-as-constraint approach. In contrast to Grˇcar et al. (2012) we replace an instance classifier (SVM) with a sequential one (CRF) and test ad- ditional features, which significantly improves their results, with an error reduction of ∼25% on both known and un- known words. Finally, we repeat the experiment performed for French (Denis and Sagot, 2012) on the impact of the amount of token (corpus) and type (lexicon) supervision for Slovene. 2. Related Work For Czech, Strakov´a et al. (2014) describe the open-source tagger MorphoDiTa, based on the averaged perceptron, yielding 95.75% accuracy on MSD tagging and 97.8% on lemmatisation when training and …

Content truncated to 500 words