Modeling Inflection and Word-Formation in SMT

Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 664–674, Avignon, France, April 23 - 27 2012. c⃝2012 Association for Computational Linguistics Modeling Inflection and Word-Formation in SMT Alexander Fraser∗ Marion Weller∗ Aoife Cahill† Fabienne Cap∗ ∗Institut f¨ur Maschinelle Sprachverarbeitung †Educational Testing Service Universit¨at Stuttgart Princeton, NJ 08541 D–70174 Stuttgart, Germany USA {fraser,wellermn,cap}@ims.uni-stuttgart.de acahill@ets.org Abstract The current state-of-the-art in statistical machine translation (SMT) suffers from is- sues of sparsity and inadequate modeling power when translating into morphologi- cally rich languages. We model both in- flection and word-formation for the task of translating into German. We translate from English words to an underspecified German representation and then use linear- chain CRFs to predict the fully specified German representation. We show that im- proved modeling of inflection and word- formation leads to improved SMT. 1 Introduction Phrase-based statistical machine translation (SMT) suffers from problems of data sparsity with respect to inflection and word-formation which are particularly strong when translating to a morphologically rich target language, such as German. We address the problem of inflection by first translating to a stem-based representation, and then using a second process to inflect these stems. We study several models for doing this, including: strongly lexicalized models, unlexicalized models using linguistic features, and models combining the strengths of both of these approaches. We address the problem of word-formation for compounds in German, by translating from English into German word parts, and then determining whether to merge these parts to form compounds. We make the following new contributions: (i) we introduce the first SMT system combining inflection prediction with synthesis of portman- teaus and compounds. (ii) For inflection, we com- pare the mostly unlexicalized prediction of lin- guistic features (with a subsequent surface form generation step) versus the direct prediction of surface forms, and show that both approaches have complementary strengths. (iii) We com- bine the advantages of the prediction of linguis- tic features with the prediction of surface forms. We implement this in a CRF framework which improves on a standard phrase-based SMT base- line. (iv) We develop separate (but related) pro- cedures for inflection prediction and dealing with word-formation (compounds and portmanteaus), in contrast with most previous work which usu- ally either approaches both problems as inflec- tional problems, or approaches both problems as word-formation problems. We evaluate on the end-to-end SMT task of translating from English to German of the 2009 ACL workshop on SMT. We achieve BLEU score increases on both the test set and the blind test set. 2 Overview of the translation process for inflection prediction The work we describe is focused on generaliz- ing phrase-based statistical machine translation to better model German NPs and PPs. We particu- larly want to ensure that we can generate novel German NPs, where what we mean by novel is that the (inflected) realization is not present in the parallel German training data used to build the SMT system, and hence cannot be produced by our baseline (a standard …

Content truncated to 500 words