Cohesive constraints allow the phrase-based decoder to employ arbitrary, non-syntactic phrases, and encourage it to translate those phrases in an order that respects the source dependency tree structure. We present extensions of the cohesive constraints, such as exhaustive interruption count and rich interruption check. Furthermore, we present analyses related to the impact of cohesive constraints across language pairs with different reordering models and dependency parsers. Our experiments show that the cohesion-enhanced decoder performs statistically significant better than the standard phrasebased decoder on English→Spanish. Improvements between 0.4 and 1.8 BLEU point are also obtained on English→Iraqi, Arabic→English and Chinese→English systems. 1

Translation systems are generally trained to optimize BLEU, but many alternative metrics are available. We explore how optimizing toward various automatic evaluation metrics (BLEU, METEOR, NIST, TER) affects the resulting model. We train a state-of-the-art MT system using MERT on many parameterizations of each metric and evaluate the resulting models on the other metrics and also using human judges. In accordance with popular wisdom, we find that it’s important to train on the same metric used in testing. However, we also find that training to a newer metric is only useful to the extent that the MT model’s structure and features allow it to take advantage of the metric. Contrasting with TER’s good correlation with human judgments, we show that people tend to prefer BLEU and NIST trained models to those trained on edit distance based metrics like TER or WER. Human preferences for METEOR trained models varies depending on the source language. Since using BLEU or NIST produces models that are more robust to evaluation by other metrics and perform well in human judgments, we conclude they are still the best choice for training. 1

Syntax-based translation models should in principle be efficient with polynomially-sized search space, but in practice they are often embarassingly slow, partly due to the cost of language model integration. In this paper we borrow from phrase-based decoding the idea to generate a translation incrementally left-to-right, and show that for tree-to-string models, with a clever encoding of derivation history, this method runs in averagecase polynomial-time in theory, and lineartime with beam search in practice (whereas phrase-based decoding is exponential-time in theory and quadratic-time in practice). Experiments show that, with comparable translation quality, our tree-to-string system (in Python) can run more than 30 times faster than the phrase-based system Moses (in C++). 1

We present a new Java-based open source toolkit for phrase-based machine translation. The key innovation provided by the toolkit is to use APIs for integrating new features (/knowledge sources) into the decoding model and for extracting feature statistics from aligned bitexts. The package was used to develop a number of useful features written to these APIs including features for hierarchical reordering, discriminatively trained linear distortion, and syntax based language models. Useful utilities distributed with the toolkit include: a conditional phrase extraction system that builds a phrase table just for a specific dataset; and an implementation of MERT that allows for pluggable evaluation metrics for both training and evaluation with built in support for a variety of metrics (e.g., TERp, BLEU, METEOR). 1

We propose a novel source-side dependency tree reordering model for statistical machine translation, in which subtree movements and constraints are represented as reordering events associated with the widely used lexicalized reordering models. This model allows us to not only efficiently capture the statistical distribution of the subtree-to-subtree transitions in training data, but also utilize it directly at the decoding time to guide the search process. Using subtree movements and constraints as features in a log-linear model, we are able to help the reordering models make better selections. It also allows the subtle importance of monolingual syntactic movements to be learned alongside other reordering features. We show improvements in translation quality in English→Spanish and English→Iraqi translation tasks. 1

In statistical machine translation, decoding without any reordering constraint is an NP-hard problem. Inversion Transduction Grammars (ITGs) exploit linguistic structure and can well balance the needed flexibility against complexity constraints. Currently, translation models with ITG constraints usually employs the cube-time CYK algorithm. In this paper, we present a shift-reduce decoding algorithm that can generate ITG-legal translation from left to right in linear time. This algorithm runs in a reduce-eager style and is suited to phrase-based models. Using the state-ofthe-art decoder Moses as the baseline, experiment results show that the shift-reduce algorithm can significantly improve both the accuracy and the speed on different test sets. 1

Abstract not found

Lexicalized reordering models play a crucial role in phrase-based translation systems. They are usually learned from the word-aligned bilingual corpus by examining the reordering relations of adjacent phrases. Instead of just checking whether there is one phrase adjacent to a given phrase, we argue that it is important to take the number of adjacent phrases into account for better estimations of reordering models. We propose to use a structure named reordering graph, which represents all phrase segmentations of a sentence pair, to learn lexicalized reordering models efficiently. Experimental results on the NIST Chinese-English test sets show that our approach significantly outperforms the baseline method. 1

The distortion cost function used in Mosesstyle machine translation systems has two flaws. First, it does not estimate the future cost of known required moves, thus increasing search errors. Second, all distortion is penalized linearly, even when appropriate reorderings are performed. Because the cost function does not effectively constrain search, translation quality decreases at higher distortion limits, which are often needed when translating between languages of different typologies such as Arabic and English. To address these problems, we introduce a method for estimating future linear distortion cost, and a new discriminative distortion model that predicts word movement during translation. In combination, these extensions give a statistically significant improvement over a baseline distortion parameterization. When we triple the distortion limit, our model achieves a +2.32 BLEU average gain over Moses. 1

In this paper we propose several novel approaches to improve phrase reordering for statistical machine translation in the framework of maximum-entropy-based modeling. A smoothed prior probability is introduced to take into account the distortion effect in the priors. In addition to that we propose multiple novel distortion features based on syntactic parsing. A new metric is also introduced to measure the effect of distortion in the translation hypotheses. We show that both smoothed priors and syntax-based features help to significantly improve the reordering and hence the translation performance on a large-scale Chinese-to-English machine translation task. 1