Statistical machine translation (SMT) treats the translation of natural language as a machine learning problem. By examining many samples of human-produced translation, SMT algorithms automatically learn how to translate. SMT has made tremendous strides in less than two decades, and many popular techniques have only emerged within the last few years. This survey presents a tutorial overview of state-of-the-art SMT at the beginning of 2007. We begin with the context of the current research, and then move to a formal problem description and an overview of the four main subproblems: translational equivalence modeling, mathematical modeling, parameter estimation, and decoding. Along the way, we present a taxonomy of some different approaches within these areas. We conclude with an overview of evaluation and notes on future directions.

We introduce a simple method to pack words for statistical word alignment. Our goal is to simplify the task of automatic word alignment by packing several consecutive words together when we believe they correspond to a single word in the opposite language. This is done using the word aligner itself, i.e. by bootstrapping on its output. We evaluate the performance of our approach on a Chinese-to-English machine translation task, and report a 12.2 % relative increase in BLEU score over a state-of-the art phrasebased SMT system. 1

This paper proposes a method to improve word alignment by combining various clues. Our method first trains a baseline statistical IBM word alignment model. Then we improve it with various clues, which are mainly based on features such as lemmatization, translation dictionary, named entities, and chunks. We incorporate these features into an unified framework. Experimental results show that our method improves word alignment quality by achieving a relative error rate reduction of 39.8%. We also conduct phrase-based machine translation based on the word alignment results. Using BLEU as an evaluation metric, our method achieves an absolute improvement of about 0.02 (about 18 % relative) over a baseline method.

Word alignments, the mappings between source and target language words for two languages, are a critical component of statistical machine translation. A long-standing issue in statistical machine translation is that the quality of word alignments does not correlate as well as would be expected with measures of translation quality. A number of recent papers have shed light on this issue by improving on existing metrics such as Alignment Error Rate and examining the importance of word alignment quality in terms of phrase alignments. In this paper, we attempt to elucidate this situation further by first presenting a new word alignment evaluation metric, Word Alignment Agreement F1 (WAAF1), which improves upon existing alignment quality metrics. We then present experiments which demonstrate that WAAF1 also correlates better with measures of translation quality than do previous metrics.

The best systems for machine translation of natural language are based on statistical models learned from data. Conventional representation of a statistical translation model requires substantial offline computation and representation in main memory. Therefore, the principal bottlenecks to the amount of data we can exploit and the complexity of models we can use are available memory and CPU time, and current state of the art already pushes these limits. With data size and model complexity continually increasing, a scalable solution to this problem is central to future improvement. Callison-Burch et al. (2005) and Zhang and Vogel (2005) proposed a solution that we call translation by pattern matching, which we bring to fruition in this dissertation. The training data itself serves as a proxy to the model; rules and parameters are computed on demand. It achieves our desiderata of minimal offline computation and compact representation, but is dependent on fast pattern matching algorithms on text. They demonstrated its application to a common model based on the translation of contiguous substrings, but leave some open problems. Among these is a question: can this approach match the performance of conventional methods despite unavoidable differences that it induces in the model? We show how to answer this question affirmatively. The main

Modern statistical machine translation (SMT) is driven by large quantities of aligned bilingual sentence pairs (so-called bitexts), from which translation models are automatically learned. I propose to develop a framework to reduce the effort involved in using extremely large quantities of training data to develop SMT systems. This task decomposes into several sub-problems, which can be addressed independently: generation of word alignments, estimation of a translation model, estimation of a language model, decoding a tuning set with the estimated models (this requires efficient access to models of potentially very large size), optimization of model parameters according to some loss function, and decoding of an evaluation set with the estimated model. Currently, the research community deals with large data in three ways. First, some solutions for efficiently handling large amounts of training data have been developed, for example, in the domain of language model estimation and representation [1,2]. However, since cluster architectures are quite diverse, these solutions, if publicly available at all, tend to be ad-hoc and environmentdependent. Second, and more commonly, non-parallel implementations of SMT model estimators (such as GIZA++ and the Moses training suite) are applied to large data sets resulting in extremely long experiment run-times, which limits the kinds of experiments that can be run. Finally, many researches circumvent these problems entirely by using small corpora. The use of small corpora for research is so widespread that many papers draw conclusions from systems trained on orders of magnitude less training data than is actually available (e.g., [3,4]). The first phase of this project will focus on more efficient translation model estimation since this task is particularly well suited for Hadoop and because there currently is no available distributed solution to this problem. Improvements to word alignment will also be investigated. 2. Resources The basis of this project will be the Moses decoder tool suite

Abstract. In this paper, we describe an experiment whose goal is to improve the quality of machine translation. Phrase-based machine translation, which is the state-of-the-art in the field of statistical machine translation, learns its phrase tables from large parallel corpora, which have to be aligned on the word level. The most common word-alignment tool is GIZA++. It is very universal and language independent. In this text, we introduce a different approach – the tectogrammatical alignment. It works on content (autosemantic) words only, but on these words it widely outperforms GIZA++. The GIZA++ word-alignment can be therefore improved using tectogrammatical alignment and if we use this improved alignment for training phrase-based automatic translators, the translation quality also slightly increases.

We present an approach to using multiple preprocessing schemes to improve statistical word alignments. We show a relative reduction of alignment error rate of about 38%. 1