What good are ‘Nominalkomposita’ for ‘noun compounds’: Multilingual Extraction and Structure Analysis of Nominal Compositions using Linguistic Restrictors

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1047–1058, Dublin, Ireland, August 23-29 2014. What good are ‘Nominalkomposita’ for ‘noun compounds’: Multilingual Extraction and Structure Analysis of Nominal Compositions using Linguistic Restrictors Patrick Ziering Lonneke van der Plas Institute for Natural Language Processing University of Stuttgart, Germany {Patrick.Ziering, Lonneke.vanderPlas}@ims.uni-stuttgart.de Abstract Finding a definition of compoundhood that is cross-lingually valid is a non-trivial task as shown by linguistic literature. We present an iterative method for defining and extracting English noun compounds in a multilingual setting. We show how linguistic criteria can be used to extract compounds automatically and vice versa how the results of this extraction can shed new lights on linguistic theories about compounding. The extracted compound nouns and their multilingual contexts are a rich source that serves several purposes. In an additional case study we show how the database serves to predict the internal structure of tripartite noun compounds using spelling variations across languages, which leads to a precision of over 91%. 1 Introduction Compounding is a phenomenon that is studied extensively in linguistic literature. Also in computational linguistics, compounds are enjoying more and more attention ( ´O S´eaghdha, 2008; Hendrickx et al., 2013). Compounding is a very productive word formation. Already 2-year-olds are able to form new words by using compounds consisting of two morphemes (Clark, 1981). As a consequence, compounds are a very common word type but many occur with a very low token count. In an analysis of the German APA corpus, Baroni et al. (2002) found that almost half (47%) of the word types were compounds. At the same time, the compounds accounted for a small portion of the overall token count (7%), which suggests that many of them are rare (83% of the compounds had a corpus frequency of 5 or lower). For English, more than half of the two-noun compounds (e.g., car park) in the BNC occur exactly once (Kim and Baldwin, 2006). The high productivity of compounds makes compositional approaches to automatic processing indispensable: listing all possible compounds in a dictionary would be as infeasible as listing all possible adjective-noun combinations. Even for compound nouns that occur 10 times or more in the BNC, static English dictionaries provide only 27% coverage (Tanaka and Baldwin, 2003). Being abundant as a phenomenon but scarce in terms of individual examples (the combination of high type frequency and low token frequency) makes the analysis of these compound nouns particularly problematic for statistical techniques that need high token frequencies to make accurate predictions. Data sparsity is expected to lead to low performance. However, the correct analysis of compound nouns is important for a number of NLP tasks, for example in machine translation (Bouillon et al., 1992; Rackow et al., 1992; Johnston and Busa, 1999; Navigli et al., 2003). The accurate translation of compounds is non-trivial, because we find a large amount of variation in the way languages deal with compounding. Some languages such as German use closed compounding (i.e., they create …

Content truncated to 500 words