Showing papers by "Taro Watanabe published in 2007"

Online Large-Margin Training for Statistical Machine Translation

Abstract: We achieved a state of the art performance in statistical machine translation by using a large number of features with an online large-margin training algorithm. The millions of parameters were tuned only on a small development set consisting of less than 1K sentences. Experiments on Arabic-toEnglish translation indicated that a model trained with sparse binary features outperformed a conventional SMT system with a small number of features.

Larger Feature Set Approach for Machine Translation in IWSLT 2007

Abstract: The NTT Statistical Machine Translation System employs a large number of feature functions. First, k-best translation candidates are generated by an efficient decoding method of hierarchical phrase-based translation. Second, the k-best translations are reranked. In both steps, sparse binary fea tures — of the order of millions — are integrated during the search. This paper gives the details of the two steps and shows the results for the Evaluation campaign of the International Workshop on Spoken Language Translation (IWSLT) 2007.

Statistical machine translation using hierarchical phrase alignment

Abstract: The following three problems are known to exist with statistical machine translation. (1) the modeling problem involved in prescribing translation relations, (2) the problem of determining parameter settings from a text corpus of translations, and (3) the search problem involved in determining the output text (the translation) given a statistical model and an input text. In this paper we find alignments of translations using phrase-based units in a hierarchical fashion with the intention of solving the above-mentioned modeling and training problems with such hierarchical phrase alignments. As an initial method we perform chunking on the corpus on the basis of these hierarchical alignments, and create translation models using these chunks as translation units. Then, as a second method we convert the translation relations expressed in the hierarchical phrase alignments into correspondences in the translation model, and perform additional training having initialized the model parameters to values obtained from these relations. The results of experiments with Japanese-to-English translation show that both methods improve performance with the second method being particularly effective resulting in an increase in translation rate from 61.3p to 70.0p. © 2007 Wiley Periodicals, Inc. Syst Comp Jpn, 38(6): 70–79, 2007; Published online in Wiley InterScience (). DOI 10.1002sscj.20271