We describe a new approach to SMT adaptation that weights out-of-domain phrase pairs according to their relevance to the target domain, determined by both how similar to it they appear to be, and whether they belong to general language or not. This extends previous work on discriminative weighting by using a finer granularity, focusing on the properties of instances rather than corpus components, and using a simpler training procedure. We incorporate instance weighting into a mixture-model framework, and find that it yields consistent improvements over a wide range of baselines. 1

Randomised techniques allow very big language models to be represented succinctly. However, being batch-based they are unsuitable for modelling an unbounded stream of language whilst maintaining a constant error rate. We present a novel randomised language model which uses an online perfect hash function to efficiently deal with unbounded text streams. Translation experiments over a text stream show that our online randomised model matches the performance of batch-based LMs without incurring the computational overhead associated with full retraining. This opens up the possibility of randomised language models which continuously adapt to the massive volumes of texts published on the Web each day. 1

Domain adaptation has recently gained interest in statistical machine translation to cope with the performance drop observed when testing conditions deviate from training conditions. The basic idea is that in-domain training data can be exploited to adapt all components of an already developed system. Previous work showed small performance gains by adapting from limited in-domain bilingual data. Here, we aim instead at significant performance gains by exploiting large but cheap monolingual in-domain data, either in the source or in the target language. We propose to synthesize a bilingual corpus by translating the monolingual adaptation data into the counterpart language. Investigations were conducted on a stateof-the-art phrase-based system trained on the Spanish–English part of the UN corpus, and adapted on the corresponding Europarl data. Translation, re-ordering, and language models were estimated after translating in-domain texts with the baseline. By optimizing the interpolation of these models on a development set the BLEU score was improved from 22.60% to 28.10 % on a test set. 1

We describe the experiments of the UC Berkeley team on improving English-Spanish machine translation of news text, as part of the WMT’08 Shared Translation Task. We experiment with domain adaptation, combining a small in-domain news bi-text and a large out-of-domain one from the Europarl corpus, building two separate phrase translation models and two separate language models. We further add a third phrase translation model trained on a version of the news bi-text augmented with monolingual sentencelevel syntactic paraphrases on the sourcelanguage side, and we combine all models in a log-linear model using minimum error rate training. Finally, we experiment with different tokenization and recasing rules, achieving 35.09 % Bleu score on the WMT’07 news test data when translating from English to Spanish, which is a sizable improvement over the highest Bleu score achieved on that dataset at WMT’07: 33.10 % (in fact, by our system). On the WMT’08 English to Spanish news translation, we achieve 21.92%, which makes our team the second best on Bleu score. 1

We report on efforts to build large-scale translation systems for eight European language pairs. We achieve most gains from the use of larger training corpora and basic modeling, but also show promising results from integrating more linguistic annotation.

Most of the freely available parallel data to train the translation model of a statistical machine translation system comes from very specific sources (European parliament, United Nations, etc). Therefore, there is increasing interest in methods to perform an adaptation of the translation model. A popular approach is based on unsupervised training, also called self-enhancing. Both only use monolingual data to adapt the translation model. In this paper we extend the previous work and provide new insight in the existing methods. We report results on the translation between French and English. Improvements of up to 0.5 BLEU were observed with respect to a very competitive baseline trained on more than 280M words of human translated parallel data. 1

Statistical machine translation is often faced with the problem of combining training data from many diverse sources into a single translation model which then has to translate sentences in a new domain. We propose a novel approach, ensemble decoding, which combines a number of translation systems dynamically at the decoding step. In this paper, we evaluate performance on a domain adaptation setting where we translate sentences from the medical domain. Our experimental results show that ensemble decoding outperforms various strong baselines including mixture models, the current state-of-the-art for domain adaptation in machine translation. 1

We present a simple method for generating translations with the Moses toolkit (Koehn et al., 2007) from existing hypotheses produced by other translation engines. As the structures underlying these translation engines are not known, an evaluationbased strategy is applied to select systems for combination. The experiments show promising improvements in terms of BLEU. 1

Abstract. Statistical Machine Translation (SMT) systems are usually trained on large amounts of bilingual text and monolingual target language text. If a significant amount of out-of-domain data is added to the training data, the quality of translation can drop. On the other hand, training an SMT system on a small amount of training material for given indomain data leads to narrow lexical coverage which again results in a low translation quality. In this paper, (i) we explore domain-adaptation techniques to combine large out-of-domain training data with small-scale in-domain training data for English—Hindi statistical machine translation and (ii) we cluster large out-of-domain training data to extract sentences similar to in-domain sentences and apply adaptation techniques to combine clustered sub-corpora with in-domain training data into a unified framework, achieving a 0.44 absolute corresponding to a 4.03 % relative improvement in terms of BLEU over the baseline.

This paper describes the statistical machine translation systems submitted to the ACL-WMT 2008 shared translation task. Systems were submitted for two translation directions: English→Spanish and Spanish→English. Using sentence pair confidence scores estimated with source and target language models, improvements are observed on the News-Commentary test sets. Genre-dependent sentence pair confidence score and integration of sentence pair confidence score into phrase table are also investigated. 1