Stripping Adjectives: Integration Techniques for Selective Stemming in SMT Systems

Stripping Adjectives: Integration Techniques for Selective Stemming in SMT Systems Isabel Slawik Jan Niehues Institute for Anthropomatics and Robotics KIT - Karlsruhe Institute of Technology, Germany firstname.lastname@kit.edu Alex Waibel Abstract In this paper we present an approach to re- duce data sparsity problems when translat- ing from morphologically rich languages into less inflected languages by selectively stemming certain word types. We de- velop and compare three different integra- tion strategies: replacing words with their stemmed form, combined input using al- ternative lattice paths for the stemmed and surface forms and a novel hidden combina- tion strategy, where we replace the stems in the stemmed phrase table by the observed surface forms in the test data. This allows us to apply advanced models trained on the surface forms of the words. We evaluate our approach by stem- ming German adjectives in two German→English translation scenar- ios: a low-resource condition as well as a large-scale state-of-the-art translation sys- tem. We are able to improve between 0.2 and 0.4 BLEU points over our baseline and reduce the number of out-of-vocabulary words by up to 16.5%. 1 Introduction Statistical machine translation (SMT) is currently the most promising approach to automatically translate text from one natural language into an- other. While it has been successfully used for a lot of languages and applications, many chal- lenges still remain. Translating from a morpholog- ically rich language is one such challenge where c⃝ 2015 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. the translation quality of modern systems is often still not sufficient for many applications. Traditional SMT approaches work on a lexical level, that is every surface form of a word is treated as its own distinct token. This can create data spar- sity problems for morphologically rich languages, since the occurrences of a word are distributed over all its different surface forms. This problem be- comes even more apparent when translating from an under-resourced language, where parallel train- ing data is scarce. When we translate from a highly inflected lan- guage into a less morphologically rich language, not all syntactic information encoded in the surface forms may be needed to produce an accurate trans- lation. For example, verbs in French must agree with the noun in case and gender. When we trans- late these verbs into English, case and gender in- formation may be safely discarded. We therefore propose an approach to overcome these sparsity problems by stemming different morphological variants of a word prior to transla- tion. This allows us to not only estimate transla- tion probabilities more reliably, but also to trans- late previously unseen morphological variants of a word, thus leading to a better generalization of our models. To fully maximize the potential of our SMT system, we looked at three different integra- tion strategies. We evaluated hard decision stem- ming, where all adjectives are replaced by their stem, as well as soft integration strategies, where we consider the words and their …

Content truncated to 500 words