Improving machine translation between closely related Romance languages

Improving Machine Translation Between Closely Related Romance Languages Petr Homola1 and Vladislav Kuboˇn1 Institute of Formal and Applied Linguistics Malostransk´e n´am. 25 CZ-118 00 Prague Czech Republic Abstract. The paper gives an overview of the shallow-transfer MT sys- tem Apertium, describes an experiment with the language pair Portuguese- Spanish and suggests a modification of the system architecture which leads to higher translation quality. Finally, consequences of the archi- tecture improvement for the design of language resources for shallow- transfer based systems are discussed. 1 Introduction Machine translation (MT) between closely related languages has been paid at- tention in more than the last twenty years by researchers from many countries. Intuitively, this kind of machine translation should be easier than MT between distant languages since it is possible to exploit the similarity at various linguis- tic levels. On the other hand, at least one language in the language pair often lacks extensive linguistic resources (such as at least morphologically annotated monolingual corpus or aligned bilingual corpus) that are essential for statistical NLP methods. The most common strategy in MT between related languages, including the open-source system Apertium, is the use of a statistical tagger in order to obtain unambiguous input and a shallow transfer module which adapts source sentences locally so that they conform to grammatical rules of the target language. However, this strategy has severe disadvantages: 1) statistical taggers introduce errors into the morphological annotation of the translated text, and 2) the dictionary is required to contain only one translation for each lemma. We suggest a modification of the shallow-transfer architecture used in the system Apertium. According to our opinion the tagger should be removed from the system while a statistical ranker should be added after the morphological processing of the target language. Our experiments show that this modification improves the quality of the translation and a pleasant side-effect of the new architecture is that the dictionary can contain more translations for one entry and the transfer rules can be applied non-deterministically, giving possibly more than one result. This paper is organized as follows: Section 2 contains a brief description of the shallow-transfer MT system Apertium. Section 3 suggests a new system 12th EAMT conference, 22-23 September 2008, Hamburg, Germany 72 architecture that leads to higher translation quality. The consequences of the architecture modification for the design of language resources in Apertium are discussed in Section 4 and finally, we conclude in Section 5. 2 The MT System Apertium There were several experiments in the area of MT between closely related lan- guages, for example [1] for Celtic languages, [2], [3] and [4] for Scandinavian languages, [5] and [6] for Slavic languages or [7] for Turkic languages. The system Apertium was originally designed for the Romance languages of Spain and it is described in detail in [8]. The authors claim that a word-to-word translation may give an adequate translation of 75% of the text. To improve the translation quality, the architecture given in Figure 1 has been implemented. deformatter � …

Content truncated to 500 words