Massively Translingual Compound Analysis and Translation Discovery

Massively Translingual Compound Analysis and Translation Discovery Winston Wu, David Yarowsky Department of Computer Science, Center for Language and Speech Processing Johns Hopkins University {wswu,yarowsky}@jhu.edu Abstract Word formation via compounding is a very widely observed but quite diverse phenomenon across the world’s languages, but the compositional semantics of a compound are often productively correlated between even distant languages. Using only freely available bilingual dictionaries and no annotated training data, we derive novel models for analyzing compound words and effectively generate novel foreign-language translations of English concepts using these models. In addition, we release a massively multilingual dataset of compound words along with their decompositions, covering over 21,000 instances in 329 languages, a previously unprecedented scale which should both productively support machine translation (especially in low resource languages) and also facilitate researchers in their further analysis and modeling of compounds and compound processes across the world’s languages. Keywords: compounds, multilingual, translation 1. Introduction Morphological compounding (e.g. lighthouse or airport) is one of the most common and productive methods of word formation across the world’s languages (Denning et al., 2007), and yet its derivational processes and semantics can be quite complex. Consider the semantic concept hospital, which can be real- ized via compound morphology in a remarkable diversity of semantic compositions, including: Lang. Compound Literal Semantics nl ziekenhuis sick + house no sykehus sick + house hu kórház disease + house eo malsanuelejo sick + place ms rumah sakit house + sick zh 病院 disease + institution There are clearly a wide variety of semantic associ- ations constituting this concept (e.g. sick/disease + house/place/institution), a variety of constituent orders (e.g. sick+house vs. house+sick) and potentially a variety of compounding processes beyond simple concatenation (e.g. sykehus in Norse (no) is a compound of syk and hus with the insertion of an e). The following paper presents a massively cross-linguistic computational model of both compound morphology com- positional processes and compound semantics on a scale of over 300 languages. Furthermore, the paper not only presents a derived analysis of the compounding process and semantics of compounds within a single language (e.g. German), as with much prior related work (e.g. Koehn and Knight (2003)), but does so via a joint model across essentially all the world’s lan- guages with adequate dictionary resources, an unprecedent- edly large scale for this class of research, and with signifi- cant additional synergistic multilingual power. In addition, the paper successfully applies these models and results to the valuable application of predicting novel trans- lations of compound words, both to English (e.g. kórház → disease + house → hospital) and from English (e.g. hospi- tal → disease + house, sick + place, etc. → kórház etc.), with valuable applications for translation dictionary expan- sion and out-of-vocabulary handling in machine translation, again on this uniquely large multilingual scale. Finally, in conjunction with this paper, we release a novel and uniquely large-scale 329-language, 21,000+ example dataset1 of these compound morphological analyses and their associated compositional and compound translations, a valuable resource for training …

Content truncated to 500 words