A Joint Approach to Compound Splitting and Idiomatic Compound Detection

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4410–4417 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4410 A Joint Approach to Compound Splitting and Idiomatic Compound Detection Irina Krotova1, Sergey Aksenov2, Ekaterina Artemova3 1 MobileTeleSystems 2 Sberbank of Russia 3 National Research University Higher School of Economics Moscow, Russia ivkrotov@mts.ru, aksenov.s.an@sberbank.ru, eartemova@hse.ru Abstract Applications such as machine translation, speech recognition, and information retrieval require efficient handling of noun compounds as they are one of the possible sources for out-of-vocabulary (OOV) words. In-depth processing of noun compounds requires not only splitting them into smaller components (or even roots) but also the identification of instances that should remain unsplitted as they are of idiomatic nature. We develop a two-fold deep learning-based approach of noun compound splitting and idiomatic compound detection for the German language that we train using a newly collected corpus of annotated German compounds. Our neural noun compound splitter operates on a sub-word level and outperforms the current state of the art by about 5%. Keywords: compound splitting, idiomatic compounds detection, word embeddings, sequence models, sub-word models 1. Introduction Compounding is a common word-formation process in Germanic languages (e.g. German, Dutch, Swedish) that poses challenges for many natural language processing ap- plications, such as machine translation (Daiber et al., 2015; Fritzinger and Fraser, 2010; Popovi´c et al., 2006), speech recognition (Larson et al., 2000), information retrieval (Al- fonseca et al., 2008; Monz and De Rijke, 2001) and coref- erence resolution (Tuggener, 2016). Difficulties are primarily caused by high productivity and low corpus frequency of compounds, which increases the vocabulary size and leads to sparse data problems. Accord- ing to (Baroni et al., 2002), almost half (47%) of the word types in a 28-million German newswire corpus are com- pounds. At the same time, 83% of them are not frequent words or productively formed hapax legomena and have a corpus frequency of 5 or lower. German compounds are not orthographically separated by hyphen or whitespaces and are mostly written as a single word. For example, the equivalent of the German word Arbeitstag is written in English as two-word compound “working day”. This leads to more out-of-vocabulary (OOV) words, which can not be listed in a lexicon and translated, but at the same time, may be split and represented as at least two compo- nents or roots. The right-most component is a noun, the head of the compound. The leftmost component is the mod- ifier and can be a noun, verb, adjective, number, or a prepo- sition. The decomposition of a complex compound or com- pound splitting is a well defined, but yet not a simple task. Two parts of the compound are not always concate- nated as in Tischtennis (“table tennis”) or Wolkenkratzer (“skyscraper”). Compound parts can undergo morpho- logical modifications from the normal form, such as ad- dition (Arbeitszeit (“working time”)) or truncation of let- ters (Kirchhof (“church garden”)), umlaut (B¨ucherregal (“bookshelf”)) or a combination of modifications. All these …

Content truncated to 500 words