Publications

Discriminative translation models utilizing source context have been shown to help statistical machine translation performance. We propose a novel extension of this work using target context information. Surprisingly, we show that this model can be efficiently integrated directly in the decoding process. Our approach scales to large training data sizes and results in consistent improvements in translation quality on four language pairs. We also provide an analysis comparing the strengths of the baseline source-context model with our extended source-context and target-context model and we show that our extension allows us to better capture morphological coherence. Our work is freely available as part of Moses.

The paper presents the results of the first large-scale human evaluation of automatic grammatical error correction (GEC) systems. Twelve participating systems and the unchanged input of the CoNLL-2014 shared task have been reassessed in a WMT-inspired human evaluation procedure. Methods introduced for the Workshop of Machine Translation evaluation campaigns have been adapted to GEC and extended where necessary. The produced rankings are used to evaluate standard metrics for grammatical error correction in terms of correlation with human judgment.

In this paper, we described the process of building a large-scale speech-to-text pipeline. Two target domains, daily conversations and travel-related conversations between two agents, for the English-German language pair (both directions) are examined. The SMT component is built from out-of-domain but freely-available bilingual and monolingual data. We make use of most of the known available resources to examine the effects of unrestricted data and large scale models. A naive baseline delivers solid results in terms of MT-quality.Extending the baseline with context-based translation model features like operations sequence models, higher-order class-based language models, and additional web-scale word-based language models leads to a system that significantly outperforms the baseline. Domain adaption is performed by separately weighting the influence of the out-of-domain subcorpora. This is explored for translation models and language models yielding significant improvements in both cases. Automatic and manual evaluation results are provided for raw MT-quality and ASR+MT-quality.

This paper presents a machine translation tool -- based on Moses -- developed for the International Maritime Organization (IMO) for the automatic translation of documents from Spanish, French, Russian and Arabic to/from English. The main challenge lies in the insufficient size of in-house corpora (especially for Russian and Arabic). The United Nations (UN) granted IMO the right to use UN resources and we describe experiments and results we obtained with different translation model combination techniques. While BLEU results remain inconclusive for combinations, we also analyze user preferences for certain models (when choosing between IMO only or combined with UN). The combined models are perceived by translators as being much better for general texts while IMO only models seem better for technical texts.

This paper introduces the freely available WikEd Error Corpus. We describe the data mining process from Wikipedia revision histories, corpus content and format. The corpus consists of more than 12 million sentences with a total of 14 million edits of various types.

As one possible application, we show that WikEd can be successfully adapted to improve a strong baseline in an ESL grammatical error correction task by 2.63%. Used together with an ESL error corpus, a composed system gains 1.64% when compared to the ESL-trained system.

Statistical machine translation toolkits like Moses have not been designed with grammatical error correction in mind. In order to achieve competitive results in this area, it is not enough to simply add more data. Optimization procedures need to be customized, task-specific features should be introduced. Only then can the decoder take advantage of relevant data.

We demonstrate the validity of the above claims by combining web-scale language models and large-scale error-corrected texts with parameter tuning according to the task metric and correction-specific features. Our system achieves a result of 35.0% F0.5 on the blind CoNLL-2014 test set, ranking on third place. A similar system, equipped with identical models but without tuned parameters and specialized features, stagnates at 25.4%.

We describe fragments of the SMT pipeline at WIPO for German as a source language. Two subsystems are discussed in detail: word decompounding and verb structure pre-reordering. Apart from automatic evaluation results for both subsystems, for the pre-reordering mechanism manual evaluation results are reported.

Described is a large-scale implementation of a Moses-based machine translation system in the United Nations aiming at accelerating the work of translators. The system (called TAPTA4UN) has been trained on extensive parallel corpora in 6 languages. Both automatic and human evaluations are provided. The system is now used in production by professional translators. The technical challenges of scalability and the final evaluation by users are also described.

We describe Phrasal Rank-Encoding (PR-Enc), a novel method for the compression of word-aligned target language data in phrase tables as used in phrase-based SMT. This method reduces the redundancy in phrase tables which is a direct effect of the phrase-based approach. A combination of PR-Enc with Huffman coding allows to reduce the size of an aggressively compressed phrase table by another 39 percent. Using this and other methods for space reduction in a new binary phrase table implementation, a size reduction by an order of magnitude is achieved when comparing to the Moses on-disk phrase table implementation. Concerning decoding speed, all variants of the new phrase table are faster than the Moses binary phrase table implementation while the PR-Enc encoded variant outperforms all other methods.

This paper describes the first steps towards a minimum-size phrase table implementation to be used for phrase-based statistical machine translation. The focus lies on the size reduction of target language data in a phrase table. Rank Encoding (R-Enc), a novel method for the compression of word-aligned target language in phrase tables is presented. Combined with Huffman coding a relative size reduction of 56 percent for target phrase words and alignment data is achieved when compared to bare Huffman coding without R-Enc. In the context of the complete phrase table the size reduction is 22 percent.

We describe the structure of a space-efficient phrase table for phrase- based statistical machine translation with the Moses decoder. The new phrase table can be used in-memory or be partially mapped on-disk. Compared to the standard Moses on-disk phrase table implementation a size reduction by a factor of 6 is achieved. The focus of this work lies on the source phrase index which is implemented using minimal perfect hash functions. Two methods are discussed that reduce the memory consumption of a baseline implementation

This paper describes an experiment in grammar engineering for a shallow syntactic parser using Genetic Programming and a treebank. The goal of the experiment is to improve the Parseval score of a previously manually created seed grammar. We illustrate the adaptation of the Genetic Programming paradigm to the problem of grammar engineering. The used genetic operators are described. The performance of the evolved grammar after 1,000 generations on an unseen test set is improved by 2.7 points F-score (3.7 points on the training set). Despite the large number of generations no overfitting effect is observed

The paper presents the main ideas and the architecture of the open source PSI-Toolkit, a set of linguistic tools being developed within a project financed by the Polish Ministry of Science and Higher Education. The toolkit is intended for experienced language engineers as well as casual users not having any technological background. The former group of users is delivered a set of libraries that may be included in their Perl, Python or Java applications. The needs of the latter group should be satisfied by a user friendly web interface. The main feature of the toolkit is its data structure, the so-called PSI-lattice that assembles annotations delivered by all PSI tools. This cohesive architecture allows the user to invoke a series of processes with one command. The command has the form of a pipeline of instructions resembling shell command pipelines known from Linux-based system.

SyMGiza++ -- a tool that computes symmetric word align- ment models with the capability to take advantage of multi-p rocessor systems -- is presented. A series of fairly simple modifications to the original IBM/Giza++ word alignment models allows to update the symmetrized models between chosen iterations of the original training algorithms. We achieve a relative alignment quality improvem ent of more than 17 on the standard Canadian Hansards task, while maintaining the speed improvemen ts provided by the capability of parallel computations of MGiza++. Furthermore, the alignment models are evaluated in the cont ext of phrase- based statistical machine translation, where a consistent improvement measured in BLEU scores can be observed when SyMGiza++ is use d instead of Giza++ or MGiza++.

SyMGiza++ -- a tool that computes symmetric word alignment models with the capability to take advantage of multi-processor systems -- is presented. A series of fairly simple modifications to the original IBM/Giza++ word alignm ent models allows to update the symmetrized models between each iteration of the original training algorithms. We achieve a relative alignment quality improvement of more than 17 standard Canadian Hansards task, while maintaining the speed improvements provided by MGiza++’s capability of parallel computations.

In this paper we will present a maximum entropy filter for the translation rules of a statistical machine translation system based on tree transducers. This filter can be successfully used to reduce the number of translation rules by more than 70 quality as measured by BLEU. For some filter configurations, translation quality is even improved. Our investigations include a discussion of the relationship of Alignment Error Rate and Consistent Translation Rule Score with translation quality in the context of Syntactic Statistical Machine Translation.

The aim of this paper is to describe the first steps of research towards a hybrid MT system that combines the streng ths of rule-based syntactic transfer with recently developed syntax-based statistical translation methods within a unified framework. The similarities of both paradigms concerning the processing of syntactically parsed input trees serve as a basis for this reseach. We focus on the statistical part of the future system and present a syntax-based statistical machine translation system -- BONSAI -- for Polish-to-French translation. Although BONSAI is still under develepmont, it reaches a translation quality on par with that of a modern phrase-based system. We provide the theoretical background as well as some implementation deta ils and preliminary evaluation results for BONSAI. At the end of this paper we shortly discuss the benefits of a combined approach.

The influence of compound noun splitting on a German-Polish bilingual vocabulary extraction task is investigated. To accomplish this, several unsupervised methods for increasingly accurate compound noun splitting are introduced. Bilingual evidence from a parallel German-Polish corpus and co-occurrence counts from the web are used to disambiguate compound noun analyses directly. These collected splits serve as training data for a probabilistic model that abstracts away from the errors made by the direct methods and reaches an f-measure of 95.10 evaluated in terms of word alignment quality and extraction accuracy where linguistically accurate methods are found to outperform the corpus-based methods proposed in the literature. A comparison of alignment quality achieved with the best splitting method and the baseline implies that the effort to build super- vised splitting methods might result in minimal or no performance gains.

The aim of this paper is to investigate whether a treebank grammar can be used to automatically classify and annotate German phrases contained in a MT lexicon. Phrases from the lexicon appear in their citation form and may differ structurally from the phr ase tokens found in the corpus. We describe the grammar extraction proc ess for a formalism called Tree-Generating Binary Grammar a nd evaluate the performance of subsets of the obtained grammar on a set of four types of lexical phrases.