English-Hungarian NP Alignment in MetaMorpho™

English-Hungarian NP Alignment in MetaMorpho TM Gábor Pohl Faculty of Information Technology Péter Pázmány Catholic University H-1083 Budapest, Práter utca 50/A, Hungary pohl@itk.ppke.hu Abstract. In this paper, a fast automatic NP alignment technique developed for MetaMorpho TM is presented. MetaMorpho TM is an EBMT-based translation memory that stores not only full sentence pairs but also NP pairs in its database of translations. In order to fulfill speed requirements of a translation memory (segments have to be stored quickly), in the proposed NP alignment algorithm time consuming statistical data collection is substituted for stemmed lexical matching using a bilingual dictionary, cognate matching and POS matching. A simple heuristic means of extracting Hungarian NP candidates with- out a deep parser is also presented in this paper. Parsed NPs of an English sentence are mapped to the words of the Hungarian translation and the shortest span containing all matched words is expanded to a full Hungarian NP using simple rules. The first experiment shows that high precision can be reached using the two algorithms discussed. 1 Introduction Noun phrase (NP) alignment is the process of identifying corresponding NP pairs in human translations. In this paper, automatic NP align- ment techniques developed for our EBMT- based translation memory (TM), MetaMorpho TM (Grőbler et al., 2004; Hodász & Pohl, 2005), will be presented and analyzed. MetaMorpho TM differs from the mostly language independent commercially available TM products in the following features. (1) In order to find the stored segment most similar to the one searched, a morpho-syntactic similarity measure is applied instead of simple character- based similarity measures. (2) MetaMorpho TM is a simple EBMT1 system (Nagao, 1984; Somers, 2003). Unlike traditional TM products not only whole sentences are searched in the memory. NPs, and the sentence skeleton (de- rived from the sentence by substituting NPs with symbolic NP slots) are also searched in the database, and their most probable translations 1 Example Based Machine Translation are morphologically altered and combined to form a possible translation sentence. Hence the recall of the translation memory is increased by suggesting translations built up from NP and sentence skeleton translations looked up sepa- rately. If the translation of words or phrases in the NPs or the sentence skeleton depend on segments looked up separately, this approach can produce incorrect translations, but even in such cases, the suggested translation might con- tain well-translated parts that can help the hu- man translator. We think the achieved higher recall justifies the slightly lower precision. In the MetaMorpho TM system, not only full sentence translations but also NP and sentence skeleton translations have to be stored in the database of translation equivalents. NPs of the stored sentence pair have to be aligned either by the translator or by an automatic means. Leav- ing the tedious task of NP alignment to the translator would decrease productivity, thus an automatic means of NP alignment was devel- oped for our TM. NP alignment is considered to be a major error source in our translation mem- …

Content truncated to 500 words