A Pilot Study on the Semantic Classification of Two German Prepositions: Combining Monolingual and Multilingual Evidence

Proceedings of Recent Advances in Natural Language Processing, pages 148–155, Hissar, Bulgaria, 7-13 September 2013. A Pilot Study on the Semantic Classification of Two German Prepositions: Combining Monolingual and Multilingual Evidence Simon Clematide Computational Linguistics University of Zurich simon.clematide@cl.uzh.ch Manfred Klenner Computational Linguistics University of Zurich klenner@cl.uzh.ch Abstract This paper reports on the annotation and maximum-entropy modeling of the se- mantics of two German prepositions, mit (‘with’) and auf (‘on’). 500 occurrences of each preposition were sampled from a treebank and annotated with syntacto- semantic classes by two annotators. The classification is guided by a perspective of information extraction, relies on linguis- tic tests and aims at the separation of se- mantically transparent and opaque mean- ings (that is of collocational construc- tions). Apart from descriptive statistical material, we present results of experiments using monolingual and multilingual evi- dence (the latter from informative English and Spanish translations) in order to pre- dict the semantic classes. 1 Introduction In linguistics, scientific grammars (Zifonun et al., 1997) as well as grammars for language learners (Helbig and Buscha, 2001) follow a long-standing tradition of semantic classification of prepositional phrases. However, it is less well-known which classification schemes can be used for automatic sense disambiguation, supporting for instance ap- plications of information extraction and knowl- edge discovery. In this pilot study, we want to gain experience of how to classify the semantic contributions of var- ious prepositions from a multilingual perspective. Our main goal is to distinguish between seman- tically transparent contributions that prepositions can provide in a general or productive manner and the less transparent contributions in collocational constructions. Many prepositions are subcatego- rized by verbs (or adjectives) and the semantic contribution of a selected preposition is weak or unspecific—a fact that is often revealed by cross- lingual comparisons of subcategorization frames. In this study we want to assess the influence of syntactic dependencies and subcategorization on semantic classification. Therefore, we chose to take our material from a syntactically annotated treebank. The rest of this paper is organized as follows. Section 2 presents related work and approaches. In Section 3, we describe our syntacto-semantic clas- sification system used in the annotation of prepo- sitions sampled from a German treebank. We also present the types of evidence used in the machine learning experiments for the automatic prediction of the classes. Section 4 contains a systematic evaluation of the performance of the different evi- dence that we have integrated in our approach. 2 Related Work As Baldwin et al. (2009, p.134) have put it in their introduction to a special issue on that topic in the Computational Linguistics Journal: ”Infor- mation extraction is one application where prepo- sitions are uncontroversially crucial to system ac- curacy”. The underlying task can be cast as prepo- sition (word) sense disambiguation (WSD). It also has been recognized in the machine translation community that ”prepositions are hard to trans- late” (Shilon et al., 2012, p.106). Although seman- tic information helps to tackle the translation task, the semantic class of a …

Content truncated to 500 words