Lemmatization and Morphological Tagging in German and Latin: A Comparison and a Survey of the State-of-the-art

Lemmatization and morphological tagging in German and Latin: A comparison and a survey of the state-of-the-art Steffen Eger1, R¨udiger Gleim2, Alexander Mehler2 1 Ubiquitous Knowledge Processing Lab, Technische Universit¨at Darmstadt, Germany 2Text Technology Lab, Goethe Universit¨at Frankfurt am Main, Germany eger@ukp.informatik.tu-darmstadt.de,{gleim,mehler}@em.uni-frankfurt.de Abstract This paper relates to the challenge of morphological tagging and lemmatization in morphologically rich languages by example of German and Latin. We focus on the question what a practitioner can expect when using state-of-the-art solutions out of the box. Moreover, we contrast these with old(er) methods and implementations for POS tagging. We examine to what degree recent efforts in tagger development pay out in improved accuracies — and at what cost, in terms of training and processing time. We also conduct in-domain vs. out-domain evaluation. Out-domain evaluations are particularly insightful because the distribution of the data which is being tagged by a user will typically differ from the distribution on which the tagger has been trained. Furthermore, two lemmatization techniques are evaluated. Finally, we compare pipeline tagging vs. a tagging approach that acknowledges dependencies between inflectional categories. Keywords: morphological tagging, lemmatization, morphologically rich languages 1. Introduction Lemmatization and part-of-speech (POS) tagging are criti- cal preprocessing steps for many natural language process- ing (NLP) tasks such as information retrieval, knowledge extraction, or semantic analysis. In morphologically rich languages such as German and Latin, both problems are non-trivial due to the variability of lexical forms. This re- sults both in large tagsets for POS tagging — which list such inflectional categories as case, gender, degree, etc., besides coarse-grained POS labels — and a large number of (potentially unseen) forms associated with each lemma. In this work, we survey tagging and lemmatization techniques for the two languages mentioned. Our survey includes both older, such as the TreeTagger (Schmid, 1994) and TnT (Brants, 2000), and more modern approaches to tagging and lemmatization. Although our expectation is clearly that technology steadily improves with time, it is apropri not obvious how large the gap between older and more modern approaches is, and also what the ordering of the most recent generation of systems is. We test our sys- tems under the following requirements:1 • Ideally, we would want a learned system to perform well on the distribution (a specific text genre, histori- cal language variant, etc.) on which it has been trained (in-domain (ID)) but also to perform decently on cor- pora of similar but different genres, registers, language varieties, etc. (out-domain (OD)). • Since coarse-grained POS tagging may be insufficient for linguistic applications and unsatisfactory for prac- titioners, we expect a system to perform well on fine- grained morphological tagging, not only on coarse- grained POS labels. 1An ideal system can also make use of the fact that there are strong dependencies between POS tagging and lemmatization, which should substantially improve its performance relative to ap- proaches where the tasks are treated independently (M¨uller et al., 2015). • Finally, run times of systems may be of considerable interest for practitioners. Therefore, we include …

Content truncated to 500 words