Exploiting Word Transformation in Statistical Machine Translation from Spanish to English

Exploiting Word Transformation in Statistical Machine Translation from Spanish to English Deepa Gupta and Marcello Federico ITC-irst-Centro per la Ricerca Scientifica e Tecnologica Via Sommarive 18, 38050 Povo (Trento) Italy {gupta,federico}@itc.it Abstract This paper investigates the use of morpho- syntactic information to reduce data- sparseness in statistical machine transla- tion from Spanish to English. In particular, word-alignment training is performed by applying different word transformations using lemmas and stems. It has been ob- served that stem-based training is better than lemma-based training when up to 1 million running words of data are used. In this paper a new word-alignment training technique is proposed by exploiting syn- tactically motivated constraints to the par- allel data. Preliminary experimental re- sults show that stem-based training with syntactically motivated constraints gives significant improvement in translation per- formance. Finally, a technique to reduce the impact of out-of-vocabulary words is discussed. The considered task is the translation of Plenary Sessions of the Eu- ropean Parliament. 1 Introduction In this work we investigate the use of morpho- syntactic information to improve performance of a phrase-based statistical machine translation (SMT) system. The considered task is the trans- lation of Plenary Sessions of the European Parlia- ment (EPPS) from Spanish to English. Recent results on this topic are reported in Popovi`c and Ney (2004), Popovi`c and Ney (2005), and Popovi`c et. al. (2005) where stem-suffix and lemma-Part-of-Speech (POS) information were used for translating German, Serbian, Spanish into English and vice versa. Moreover, in de Gispert Input sentence: ¿ hay alguna observaci´on ? Original Lemma/base POS tag+ word form morpho-attributes ¿ ¿ Fia hay haber VAIP3S0 alguna alguno DI0FS0 observaci´on observaci´on NCFS000 ? ? Fit Figure 1: Example of FreeLing output on Spanish. The complex tag ”VAIP3S0” for word hay means: verb(V), auxiliary verb(A), indicative mode(I), present(P), third person (3), singular number (S) and no gender(0) (2005) lemma-POS information restricted to verbs was used in a translation task from Spanish to En- glish. As a difference with respect to previous work, we investigate word transformations which are blindly applied to all word categories and un- der different data sparseness conditions. More- over, morpho-syntactic knowledge is only applied during training of the SMT system, and is not di- rectly exploited during translation. As an excep- tion, stems of Spanish words have been also used for translating out-of-vocabulary (OOV) words in the test set. The organization of the paper is as follows. Section 2 introduces different lexicon transforma- tions. Section 3 gives details about the data and the SMT system. Sections 4 and 5 describe techniques to reduce data-sparseness in SMT and discuss ex- perimental results. Section 6 explains a technique to reduce the impact of OOV words. Finally, Sec- tion 7 contains concluding remarks. Spanish sentence Original : ¿ hay alguna observaci´on ? Stemmed: ¿ hay algun observ ? English Sentence Original: are there any comments ? Stemmed: ar there ani comment ? Figure 2: Example of stemmed texts. 2 Lexicon Transformations We applied two kinds of lexicon transformations, lemmatization and …

Content truncated to 500 words