Using POS Information for SMT into Morphologically Rich Languages

Using POS Information for Statistical Machine Translation into Morphologically Rich Languages Nicola Ueffing and Hermann Ney Lehrstuhl ftir Informatik VI - Computer Science Department RWTH Aachen - University of Technology tueffing,neyl@cs.rwth - aachen.de Abstract When translating from languages with hardly any inflectional morphology like English into morphologically rich lan- guages, the English word forms often do not contain enough information for producing the correct fullform in the target language. We investigate meth- ods for improving the quality of such translations by making use of part-of- speech information and maximum en- tropy modeling. Results for translations from English into Spanish and Catalan are presented on the LC-STAR corpus which consists of spontaneously spoken dialogues in the domain of appointment scheduling and travel planning. 1 Introduction In this paper, we address the question of how part- of-speech (POS) information can help improv- ing the quality of Statistical Machine Translation (SMT). One of the main problems when translat- ing from a language with hardly any inflectional morphology (which is English in our experiments) into one with richer morphology (here: Spanish and Catalan) is the production of the correct in- flected form in the target language. We introduce transformations to the English string that are based on the part-of-speech information and show how this knowledge source can help SMT. Systematic evaluations will show that the quality of the gen- erated translations is improved. The transformations we apply are the following: Treatment of verbs In Catalan and Spanish, the pronoun before a verb is often omitted and in- stead, the person is expressed via the ending of the verb. The same holds for future tense and for the modes expressed through 'would' and 'should' in English. Since this makes it hard to generate the correct translation of a given English verb, we propose a method re- sulting in English word forms containing suf- ficient information. Question inversion In English, interrogative phrases have a word order that is different from declarative sentences: Either an auxil- iary 'do' is inserted or the order of verb and pronoun is inverted. Since this is different in Spanish and Catalan, we modify the word order in English to make it more similar to the Spanish/Catalan one and to help the verb treatment mentioned above. The paper is organized as follows: Related work is treated in Section 2. In Section 3, we shortly review the statistical approach to machine transla- tion. Then, we introduce the transformations that we apply to the less inflected language of the two under consideration (namely English) in Section 4. After describing the maximum entropy approach and the training procedure we use for the statisti- cal lexicon in Section 5, we present results on the trilingual LC-STAR corpus in Section 6. Then, we conclude and present ideas about future work in Section 7. 347 2 Related Work Publications dealing with the integration of lin- guistic information into the process of statisti- cal machine translation are rather few although this had already been suggested in (Brown et al., …

Content truncated to 500 words