Deriving translation units using small additional corpora

Deriving translation units using small additional corpora. Carlos A. Henr´ıquez Q. TALP Research Centre Barcelona, Spain carlos.henriquez@upc.edu Jos´e B. Mari˜no TALP Research Centre Barcelona, Spain jose.marino@upc.edu Rafael E. Banchs Institute for Infocomm Research Singapore rembanchs@i2r.a-star.edu.sg Abstract We present a novel strategy to derive new translation units using an additional bilin- gual corpus and a previously trained SMT system. The units were used to adapt the SMT system. The derivation process can be applied when the additional corpus is very small compared with the original train corpus and it does not require to compute new word alignments using all corpora. The strategy is based in the Levenshtein Distance and its resulting path. We re- ported a statistically significant improve- ment, with a confidence level of 99%, when adapting an Ngram-based Catalan- Spanish system using an additional corpus that represents less than 0.5% of the orig- inal train corpus. The additional transla- tion units were able to solve morphologi- cal and lexical errors and added previously unknown words to the vocabulary. 1 Introduction. Statistical Machine Translation (SMT) systems are trained using parallel corpora. Therefore, once the system is trained and tuned, it is tightly coupled to the specific domain the train corpus belongs to. If later on we want to use additional bilingual corpora to improve or adapt our system, we could build additional translation models and interpolate them with the original one or we could join all the addi- tional data with the original corpus and train a new system from scratch. However, those strategies of- ten involve computing new word alignments con- sidering all corpora together, which is a computa- tional expensive task. c⃝ 2011 European Association for Machine Translation. This study focuses on the use of additional bilin- gual corpora to adapt a previously trained SMT system, without the need to recompute word align- ments. The proposed method utilizes the SMT sys- tem to translate the source side of the new corpus and then compares the translation output with its target side. This comparison allows the method to detect errors made during decoding and provide it at the same time with a possible solution, which is finally used to build additional translation units. We have experimented with a Ngram-based SMT system (Mari˜no et al., 2006), translating from Catalan into Spanish and we have obtained a significant improvement in translation quality, adapting a state-of-the-art system trained with a corpus of more than four million sentences with an additional corpus of only 1.6 thousand sentences. This document is organized as follows: Sec- tion 2 introduces us to the concept of Statistical Machine Translation, with an emphasis in Ngram- based SMT; Section 3 presents a description of the possible scenarios where the proposed strategy could be used, domain adaptation (subsection 3.1) and user feedback (subsection 3.2); Section 4 de- scribes the experimental set-up, it details the base- line system in subsection 4.1 and the additional corpus in subsection 4.2, it also explains the main algorithm to derive, filter and interpolate the addi- …

Content truncated to 500 words