Towards Unsupervised and Language-independent Compound Splitting using Inflectional Morphological Transformations

Proceedings of NAACL-HLT 2016, pages 644–653, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Towards Unsupervised and Language-independent Compound Splitting using Inflectional Morphological Transformations Patrick Ziering Institute for Natural Language Processing University of Stuttgart, Germany Patrick.Ziering@ ims.uni-stuttgart.de Lonneke van der Plas Institute of Linguistics University of Malta, Malta Lonneke.vanderPlas@um.edu.mt Abstract In this paper, we address the task of language- independent, knowledge-lean and unsuper- vised compound splitting, which is an essen- tial component for many natural language pro- cessing tasks such as machine translation. Pre- vious methods on statistical compound split- ting either include language-specific knowl- edge (e.g., linking elements) or rely on parallel data, which results in limited applicability. We aim to overcome these limitations by learning compounding morphology from inflectional information derived from lemmatized mono- lingual corpora. In experiments for Germanic languages, we show that our approach signifi- cantly outperforms language-dependent state- of-the-art methods in finding the correct split point and that word inflection is a good ap- proximation for compounding morphology. 1 Introduction Compounding represents one of the most productive word formation types in many languages. In partic- ular, Germanic languages (e.g., German or Dutch) show high productivity in closed compounding, i.e., in creating one-word compounds such as the Ger- man Armutsbek¨ampfungsprogramm ‘poverty elimi- nation program’. Previous studies on German cor- pora reveal that almost half of the corpus types are compounds, whereas individual compounds are very infrequent (Baroni et al., 2002). Therefore, an automatic compound analysis is indispensable and represents an essential component in many natural language processing (NLP) tasks such as machine translation (MT) or information retrieval (IR). Besides determining the concatenated constituent forms, i.e., the correct split points (e.g., Armuts | bek¨ampfungs | programm), a compound split- ter needs to normalize each part (e.g., Armut + Bek¨ampfung + Programm), because down-stream applications such as MT systems expect lemma- tized words as input. However, normalization of constituent forms is non-trivial and usually re- quires language-specific knowledge (e.g., linking el- ements). State-of-the-art lemmatizers, designed for regular word inflection, would fail, because con- stituent forms often contain linking elements lead- ing to a non-paradigmatic word form of the corre- sponding lexeme (e.g., Armuts ‘poverty + s’ never occurs as an isolated token in German corpora, since the s-suffix, often used for genitive or pluralization, is not used with Armut). Moreover, morphologi- cal operations during compounding vary a lot across languages and lexemes: we find cases that start from the lemma and have additions (e.g., linking elements), truncations (e.g., reductions to a verbal stem), word-internal operations (e.g., Umlautung) and combinations thereof (e.g., the first constituent of the German Weihnachts|baum ‘Christmas tree’, Weihnachten, undergoes both the en-truncation and the s-suffixation). In this paper, we present a language-independent, unsupervised compound splitter that normalizes constituent forms by tolerantly retrieving candi- date lemmas using an Ngram index and weighting string differences with inflectional information de- rived from lemmatized corpora. Most previous work on compound splitting in- cludes language-specific knowledge such as large 644 lexicons and morphological analyzers (Fritzinger and Fraser, …

Content truncated to 500 words