Augmenting a statistical translation system with a translation memory

126 EAMT 2005 Conference Proceedings Augmenting a Statistical Translation System with a Translation Memory Sanjika Hewavitharana, Stephan Vogel, Alex Waibel Language Technologies Institute Carnegie Mellon University, Pittsburgh U.S.A. {sanjika, vogel+, ahw}@cs.cmu.edu Abstract. In this paper, we present a translation memory (TM) based system to augment a statistical translation (SMT) system. It is used for translating sentences which have close matches in the training corpus. Given a test sentence, we first extract sentence pairs from the training corpus, whose source side is similar to the test sentence. Then, the TM system modifies the translation of the sentences by a sequence of substitution, deletion and inser- tion operations, to obtain the desired result. Statistical phrase alignment model of the SMT system is used for this purpose. The system was evaluated using a corpus of Chinese- English conversational data. For close matching sentences, the translations produced by the translation memory approach were compared with the translations of the statistical decoder. 1. Introduction Spoken language translation has received more attention in recent times. Some of the notable attempts include Verbmobil (Wahlster, 2000) and Nespole (Metze et at., 2002). Many corpora have been compiled for this purpose covering various domains, including conversations in travel and medical domains. Typically, these cor- pora contain shorter sentences. For example, in the Basic Travel Expression Corpus (BTEC) cor- pus (Takezawa et al., 2002), the sentences have 6-7 words on average. Another noticeable fea- ture is that they have sentence with similar pat- terns, as shown in Figure 1 with Spanish-Eng- lish sentence pairs from the BTEC corpus. en qué tipo de trabajo estás interesado ? what kind of job are you interested in ? en qué tipo de cosas estás interesado ? what kind of things are you interested in ? en qué tipo de excursiones estás interesado ? what kind of tour are you interested in ? Figure 1: Similar patterns in sentences These three sentences differ only in one word in both Spanish sentences as well as their English translations. For a given test sentence, we often find in the training corpus, a very similar sen- tence with few mismatching words; sometimes even an exact matching sentence. Translation memory (TM) systems typically work well in these situations. In its pure form, a TM system is simply a database of past translations, stored as sentence pairs in source and target languages. Whenever an exact match is found for a new sentence to be translated, the desired translation is extracted from the translation memory. TM systems have been successfully used in Com- puter Aided Translations (CAT) as a tool for human translators. There have been attempts to combine trans- lation memory with other machine translation approaches. In (Marcu, 2001) an automatically derived TM is used along with a statistical mo- del to obtain translations of higher probability than those found using only a statistical model. Sumita (2001) describes an example-based tech- nique which extracts similar translations and modifies them using a bilingual dictionary. Wa- tanabe and Sumita (2003) proposed …

Content truncated to 500 words