Larger feature set approach for machine translation in IWSLT 2007

Larger Feature Set Approach for Machine Translation in IWSLT 2007 Taro Watanabe, Jun Suzuki, Katsuhito Sudoh, Hajime Tsukada, Hideki Isozaki NTT Communication Science Laboratories 2-4 Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0237 Japan {taro,jun,sudoh,tsukada,isozaki}@cslab.kecl.ntt.co.jp 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 Interna- tional Workshop on Spoken Language Translation (IWSLT) 2007. 1. Introduction This paper presents NTT Statistical Machine Translation System evaluated in the evaluation campaign of International Workshop on Spoken Language Translation (IWSLT) 2007. Our system is composed of two steps: First, k-best transla- tion candidates are generated using an efficient decoder for hierarchical phrase-based translation [1]. Next, the large k- best translation is reordered using a reranking voted percep- tron [2]. Both systems employ a large number of sparse fea- tures — of the order of millions — to achieve a state of the art performance [3]. The large number of parameters are trained using an ef- ficient online training algorithm: The decoder employs an online large-margin training method [4] that has been suc- cessfully applied in dependency parsing [5] or joint label- ing/chunking [6]. The reranker uses a reranking voted per- ceptron which gave significant improvement in the last year’s IWSLT 2006 evaluation [2]. Both systems are tuned using approximated BLEU as an objective function that scales the sentence-wise BLEU to a document-wise BLEU. Domain mismatch is handled by a simple task adaptation scheme by selecting training data that resembles a test set [7]. In order to handle the ASR’s error prone input, we decoded all the n- best translations and let the reranker choose the right trans- lation by treating the individually translated list as a single k-best list combined with the ASR’s n-best list’s confidence measures. This paper is organized as follows: The overview of our decoder is presented in Section 2. We will describe the fea- ture functions experimented in [3] together with additional features. The reranking system is described in Section 3. The reranker is biased to use a slightly different feature set to avoid over training. Both systems share the same online training algorithm, but differ in that the decoder’s parame- ters are updated based on the dynamically generated candi- date list, whereby the reranking training is based on a fixed translation candidate list. Section 4 presents the results for the evaluation campaign of IWSLT 2007. 2. Machine Translation System We use a linear feature combination approach [8] in which a foreign language sentence f is translated into another lan- guage, for example English, e, by seeking a maximum solu- tion: ˆe = argmax e w⊤ · h(f, e) (1) where h(f, e) is a large-dimension feature vector. w …

Content truncated to 500 words