A Method to Restrict the Blow-up of Hypotheses of a Non-disambiguated

International Conference RANLP 2009 - Borovets, Bulgaria, pages 460–464 A method to restrict the blow-up of hypotheses of a non-disambiguated shallow machine translation system Jernej Viˇciˇc University of Primorska Koper, Slovenia jernej.vicic@upr.si Petr Homola Charles university Prague, Czech republic homola@ufal.mff.cuni.cz Vladislav Kuboˇn Charles university Prague, Czech republic vk@ufal.mff.cuni.cz Abstract The article presents two automatic methods that reduce the complexity of the ambiguous space in- troduced by the omission of the part of speech tagger from the architecture of a shallow machine translation system. The methods were imple- mented in a fully functional translation system for related languages. The language pair cho- sen for the experiments was Slovenian-Serbian as these languages are highly inflectional with morphologically ambiguous forms. The empir- ical evaluations show an improvement over the original system. Keywords MT, RBMT, SMT, related languages, ranking, Apertium. 1 Introduction The article presents two automatic methods that re- duce the complexity of the ambiguous space intro- duced by the omission of the part of speech tagger from the architecture of a shallow transfer machine translation system. The shallow parsing machine translation architecture is suitable for the translation systems for related lan- guages as shown in [5, 15]. Authors [12] and the authors of translation systems for related languages Apertium [5] and ˇCes´ılko [11] suggest using an archi- tecture similar to the one presented in figure 1. This architecture employs statistical part of speech (POS) tagger for disambiguation of the morphological anal- ysis of the source language. The quality level of the tagging process of today’s state-of-the-art POS taggers for highly inflectional languages like [10] and [8] is rel- atively low, comparing to the quality of POS taggers for the analytical languages like the English language, and also comparing to the overall quality of the trans- lation systems for related languages. [14] proposes a change of the basic architecture, the omission of the statistical POS tagger from the early stages of the translation process and the introduction of a ranking mechanism in the post-processing phase. The proposed architecture is presented in the figure 2. The ranking mechanism is based on the statistical trigram target language model and models the most probable sentence in the target language. The ranker selects the best translation among all available trans- lation candidates. The rest of the article is organized as follows: the section 2 presents the current accomplishments in the field of the machine translation for related languages, section 3 presents the basic motives that led the au- thors to this set of experiments. The section 4 shows the motivation that led to the experiment presented in the paper. the section 5 presents the main method and the section presents the evaluation methodology with the results. The article concludes with a discussion. 2 State of the art The majority of the translation systems for related lan- guages uses the shallow parsing machine translation architecture as shown in [15] and in the overview of the translation systems presented in 2.1. The figure 1 shows the architecture of the …

Content truncated to 500 words