List of Accepted Research Papers

Summary: We built 462 machine translation systems for all language pairs of the Acquis Communautaire corpus. We report and analyse the performance of these system, and compare them against pivot translation and a number of system combination methods (multi-pivot, multi-source) that are possible due to the available systems.

Summary: In the formally syntax-based MT, a hierarchical tree generated by synchronous CFG rules associates the source sentence with the target sentence. In this paper, we propose a source dependency model to calculate the probability of the hierarchical tree generated in decoding. We develop this source dependency model from word-aligned corpus, without using any linguistically motivated parsing. Our experimental results display that integrating the source dependency model into the formally syntax-based machine translation significantly improves the performance on Chinese-to-English translation tasks.

Summary: We examine the contribution of reliable elements in French-- and English--Japanese alignment from comparable corpora, using transliterated elements and scientific compounds as anchor points among context-vectors of elements to align. We highlight those elements in context-vector normalisation to give them a higher priority in context-vector comparison. We carry out experiments on small comparable corpora to show that those elements can efficiently be used to improve the quality of the alignment.

Summary: We investigate the possibility of automatically detecting whether a piece of text is an original or a translation. On a large parallel English-French corpus where reference information is available, we find that this is possible with around 90% accuracy. We further study the implication this has on Machine Translation performance. By separating the corpus according to translation direction and training separate Phrase-Based MT systems, we show that reliably detecting translation direction yields improved translation performance. This suggests that paying attention to the translation direction when building a parallel corpus for the purpose of training a statistical MT system may have an effect on the output quality.

Summary: The way of mining comparable corpora and the strategy of dictionary extraction are two essential elements of bilingual dictionary extraction from comparable corpora. This paper first proposes a method, which uses the inter-language link in Wikipedia, to build robust and large-scale comparable corpora. Then, this paper presents an approach, which combines context heterogeneity similarity and dependency heterogeneity similarity, to extract bilingual dictionary from the collected comparable corpora. Experimental results show that because of combining the advantages of context heterogeneity similarity and dependency heterogeneity similarity appropriately, the proposed approach outperforms both the two individual approaches significantly.

Summary: This paper reports on a preliminary study testing the use of eye tracking as a method for evaluating machine translation output. 50 French machine translated sentences, 25 rated as excellent and 25 rated as poor in an earlier human evaluation, were selected. 10 native speakers of French were instructed to read the MT sentences for comprehensibility. Their eye gaze data were recorded non-invasively using a Tobii 1750 eye tracker. They were also asked to record retrospective protocols while watching a replay of their eye gaze reading data.

Summary: The paper explores a way to learn post-editing fixes of raw MT outputs automatically by combining two different types of statistical machine translation (SMT) systems in a linear fashion. Our proposed system (which we call a chained system) consists of two SMT systems: (i) a syntax-based SMT and (ii) a phrase-based SMT. We investigated the impact from the chained system on the initial SMT system using BLEU. The results of our experiments strongly indicate that the chained system can compensate the weaknesses of the initial SMT system in a robust way by providing human-like fixes.

Summary: The contribution discusses variants of architectures of hybrid MT systems. The three main types of architectures are: coupling of systems (serial or parallel), architecture adaptations (integrating novel components into SMT or RMT architectures, either by pre/post-editing, or by system core modifications), and genuine hybrid systems, combining components of different paradigms. The interest is to investigate which resources are required for which types of systems, and to which extent the proposals contribute to an overall increase in MT quality.

Summary: We describe an approach for filtering phrase tables in a Statistical Machine Translation system, which relies on a statistical independence measure called Noise, first introduced in (Moore 2004). While previous work by (Johnson et al., 2007) also addressed the question of phrase table filtering, it relied on a simpler independence measure, the p-value, which is theoretically less satisfying than the Noise in this context. In this paper, we use Noise as the filtering criterion, and show that when we partition the bi-phrase tables in several sub-classes according to their complexity, using Noise leads to improvements in BLEU score that are unreachable using p-value, while allowing a similar amount of pruning of the phrase tables.

Summary: This paper summarises the results of a pilot project conducted to investigate the correlation between automatic evaluation metric scores and post-editing time on a segment by segment basis. Firstly, the results from the comparison of various automatic metrics and post-editing time will be reported. Secondly, further analysis of the relationship between the two is carried out by taking into consideration other relevant variables, such as text length and structures, and by means of multiple linear regression. It has been found that different automatic metrics achieve different levels and types of correlation with post-editing time. We suggest that some of the source text characteristics and machine translation errors may be able to further account for the relationship between the two.

Summary: We introduce the first machine translation system for Yiddish-English and English-Yiddish. We discuss challenges presented by this language and their solutions, including an algorithm for cognate extraction.

Summary: In this paper we present an extension of a phrase-based decoder that dynamically chunks, reorders, and applies phrase translations in tandem. A maximum entropy classifier is trained based on the word alignments to find the best positions to chunk the source sentence. No language specific or syntactic information is used to build the chunking classifier. Words inside the chunks are moved together to enable the decoder to make long-distance reorderings to capture the word order differences between languages with different sentence structures. To keep the search space manageable, phrases inside the chunks are monotonically translated.

Summary: To address the shortage of Japanese-English parallel corpora, we developed a parallel corpus by collecting open source software manuals from the Web. The constructed corpus contains approximately 500 thousand sentence pairs that were aligned automatically by an existing method. We also conducted statistical machine translation (SMT) experiments with the corpus and confirmed that the corpus is useful for SMT.

Summary: We propose two variations on the conventional beam thresholding. The first variation is the dynamic beam thresholding, in which the beam threshold varies with the length of source sequences covered by hypotheses. The second one incorporates a language model look-ahead probability into the beam thresholding so that the interaction between a hypothesis and the contexts outside the hypothesis can be captured. Both thresholding methods achieve significant speed improvements when used separately. By combining them together, we obtain a further speedup, which is comparable to that of the cube pruning approach \cite{Chiang:07}.

Summary: We propose a generic rule induction framework that is informed by syntax from both sides of a parsed parallel corpus, as sets of structural, boundary and labeling related constraints. We then explore the issue of lexical coverage of translation models learned in different scenarios using syntax from one side vs. both sides. We specifically look at how the non-isomorphic nature of parse trees for the two languages affects coverage. We propose a novel technique for restructuring targetside parse trees, that generates alternate isomorphic target trees that preserve the syntactic boundaries of constituents that were aligned in the original parse trees. We also show that combining rules extracted by restructuring syntactic trees on both sides produces significantly better translation models.

Summary: We have developed a Web site called "Minna no Hon'yaku" ("Translation for Everyone by Everyone"), which hosts online volunteer translators. Its core features are (1) a blog-like look and feel, (2) the legal sharing of translations, (3) high quality, comprehensive language resources, and (4) the translation aid editor; QRedit. Translators who use QRedit daily reported an up to 30 per cent reduction of the overall translation time. As of 24 April 2008, there are about 300 users and 4 groups registered to MNH. The groups using it include such major NGOs as Amnesty International Japan and Democracy Now! Japan.

Summary: In this paper, we propose a hybrid spoken language translation method utilizing sentence segmentation. By portioning the sentence using the result of syntax analysis, we can utilize rule-based control of the integration of subtranslations translated by a suitable method for each segment. We also report a preliminary experiment on translation quality of our prototype Japanese-to-English translation system. We confirmed that our method achieved a 13.4\% advantage in NIST score for the individual RBMT method, and a 6.0\% advantage for the individual EBMT method.

Summary: In this paper, we present a reordering model based on Maximum Entropy. This model is extended from a hierarchical reordering model with PBSMT (Galley and Manning, 2008), which integrate syntactic information directly in decoder as features of MaxEnt model. The advantages of this model are (1) maintaining the strength of phrase based approach with a hierarchical reordering model, (2) many kinds of linguistic information integrated in PBSMT as arbitrary features of MaxEntropy model. The experiment results with English-Vietnamese pair showed that our approach achieves improvements over the system which use a lexical hierarchical reordering model (Galley and Manning, 2008).

Summary: We propose a reordering method to improve the fluency of the output of the phrase-based SMT (PBSMT) system. We parse the translation results that follow the the source language order into non-projective dependency trees, then reorder dependency trees to obtain more fluent target sentence.s. This method insures that the translation result are grammatically correct. Our method achieves great improvements over PBSMT using a dependency-based metric.

Summary: We investigate the problem of estimating the quality of the output of machine translation systems at the sentence level when reference translations are not available. The focus is on automatically identifying a threshold to map a continuous predicted score into ``good'' / ``bad'' categories for filtering out bad-quality cases in a translation post-edition task. We use the theory of Inductive Confidence Machines (ICM) to identify this threshold according to a confidence level that is expected for a given task. Experiments show that this approach gives improved estimates when compared to those based on classification or regression algorithms without ICM.

Summary: We present empirical results in which parameters tuned by MERT on a certain metric (e.g. Bleu) may not lead to optimal scores on that metric. The score can be improved significantly by tuning on an entirely different metric (e.g. Meteor, by 0.82 Bleu points or 3.38% relative improvement on WMT08 English--French data). We analyse this phenomenon in MERT and further propose three combination strategies of different metrics to reduce the bias of a single metric and obtain parameters that receive better scores (0.99 Bleu points or 4.08%) on evaluation metrics than those tuned on the standalone metric itself.

Summary: A method is presented which integrates target language automatic speech recognition (ASR) models with source language statistical machine translation (SMT) and named entity recognition (NER) information at the phonetic level. Information extracted from a source language document including translation model probabilities and translated named entities are combined with acoustic-phonetic information obtained from phone lattices produced by the ASR system. Phone-level integration allows the combined MAHT system to correctly decode words that are either not in the ASR vocabulary or would have been incorrectly decoded by the ASR system. It is shown that the combined MAHT system results in a decrease in word error rate on the dictated translations of up to 32% relative to a stand alone baseline ASR system.

Summary: Mining officially published web pages can be an invaluable undertaking for translators in government departments who are producing the translations, and for machine translation researchers who are studying how those translations are produced. In this paper, we present the StatCan Daily Translation Extraction System (SDTES) and demonstrate how it is used to induce transla-tions from officially published bilingual mate-rials from government websites in Canada. New evaluation results show that SDTES is a very effective system for identifying and ex-tracting sentences that are translation pairs from most of the federal government web pages which are currently under the CLF2 (Common Look and Feel for the Internet 2.0) framework.

Summary: We investigate novel types of assistance for human translators, based on statistical machine translation methods. We developed a tool that makes suggestions for sentence completion, shows word and phrase translation options, and allows post-editing of machine translation output. A user study validates the types of assistance and provides insight into the human translation process.

Summary: We presents an introduction to the CWMT2008 evaluation and focus on its two new metrics: BLEU-SBP (Chiang et al., 2008) and linguistic check-point method (Zhou et al., 2008). Our experiments validated BLEU-SBP’s effectivity in resolving the nondecomposability problem of both NIST-BLEU and IBM-BLEU at sentence level. Our evaluation indicates linguistic check-point method is a valid metric to evaluate the capability of an MT system in translating various linguistic phenomena. With the aid of these metrics, we disclosed some performance differences between statistical MT systems and rule-based MT systems. We suggests the high BLEU score doesn’t necessarily mean high translation adequacy.

Summary: Statistical MT is limited by reliance on large parallel corpora. We propose Lemmatic MT, a new paradigm that extends MT to a far broader set of languages, but requires manual encoding effort. The author encodes each sentence as a sequence of words drawn from a translation dictionary. We report on an experimental investigation of LEMUEL, a prototype Lemmatic MT system that outperforms Google Translate and also has high translation adequacy on language pairs not handled by Google Translate.

Summary: We propose to mine parallel texts from mixed-language Web pages. We define a mixed-language Web page as a Web page consisting of (at least) two languages. We mined Japanese-English parallel texts from mixed-language Web pages. We presented the statistics for extracted parallel texts and conducted machine translation experiments. These statistics and experiments showed that mixed-language Web pages are rich sources of parallel texts.

Summary: Evidence is presented to support the hypothesis that variation and inflection in Arabic has a negative impact on scores from automated measures of speech translation (e.g., WER, BLEU). Normalization operations improve correlation between BLEU scores and Likert-type judgments of semantic adequacy — as well as between BLEU scores and human judgments of successful transfer of the meaning of individual content words from English to Arabic.

Summary: In this paper, we propose to enhance the phrase translation model with association measures as new feature functions. These features are estimated on counts of phrase pair co-occurrence and their marginal counts. Four feature functions, namely, Dice coefficient, log-likelihood-ratio, hyper-geometric distribution and link probability are exploited and compared. Experimental results demonstrate that the performance of the phrase translation model can be improved by enhancing it with these association based feature functions. Moreover, we study the correlation between the features to predict the usefulness of a new association feature given the existing features.

Summary: We explore the problem of integrating a phrase-base MT system within a computer assisted translation (CAT) environment. We argue that one way of achieving successful integration is to design a MT system that behaves more like the translation memory (TM) component of CAT systems. This implies producing MT output that is consistent with that of a TM when high-similarity material exists in the training data; it also implies providing the MT system with a component to filter out machine translations that are less likely to be useful. Our results indicate that the proposed approach leads to systems that produce better output than a TM, for a larger portion of the source text.

Summary: We propose to estimate the probability that a target word appears in the translation of a given source sentence using a multilayer perceptron. At the expense of ignoring words order and repetition, our model does not assume word alignments and consider all source words jointly when evaluating the probability of a target word.

We compared our model against \ibmone which does not consider word order neither. Our model was competitive with \ibmone when predicting the target words that should be in the translation of a source sentence. When our model was extended to include alignment information, it surpassed \ibmone on all the metrics we evaluated.

Summary: Translation spotting consists in automatically identifying the existing translations of a user query inside a bitext. This task, which mainly relies on statistical word alignment algorithms, is a valuable tool for Computer Assisted Translation but fails to achieve excellent results. In this paper, we propose to borrow the idea of relevance feedback from the information retrieval domain to enhance these results. We show that the translations of a query identified during a first translation spotting stage provide relevant information that can be used in a second stage to improve the results.

Summary: In this paper we study in detail the relation between word alignment and phrase extraction. First, we analyze word alignment according to several characteristics and compared them to hand-aligned data. Secondly, we analyzed the phrase-pairs generated by these alignments. We observed the unaligned words in the extracted phrase pairs follow the distribution of unaligned words in the alignment from where they were extracted. A manual evaluation of phrase pair quality showed that the more unaligned words in our phrase-pairs the lower the quality. Finally, we presented translation results from using different phrase-tables build upon the different alignments.

Summary: The aim of this paper is to further compare two versions of our spoken language translator MedSLT which differ in terms of grammatical coverage. In this paper this comparison is extended to study in more detail the task performance with respect to the number of successful interactions and to explain why some utterances could not be correctly recognized; additionally this paper will explore how these results correlate with the quality of the translation produced by both versions of the systems.

Summary: In this work we have deal with the reordering problem in Spanish-Basque statistical machine translation, comparing three different approaches and analyzing their strength and weakness. Tested approaches cover the more usual techniques: lexicalized reordering implemented on Moses, preprocessing based on hand defined rules over the syntactic analysis of the source and statistical translation.

According with the obtained results, the three reordering techniques improves the results of the baseline. We observe different behaviour at combining techniques. While the use of the Syntax-Based reordered corpus together with the lexicalized reordering get the best results, training the lexicalized reordering on the statistically reordered source does not improve the performance of the single methods.

Summary: In this work, we show how an existing rule-based, linguistically motivated machine translation system may be improved and adapted automatically to a given domain, whenever parallel corpora are available. We perform this adaptation by extracting dictionary entries from the parallel data. From this initial set, the application of these rules is tested against the baseline performance. Rules are then pruned depending on sentence-level improvements and regressions, as evaluated by an automatic string-based metric. Experiments using the Europarl dataset show a +3% absolute improvement in BLEU over the original rule-based system.

Summary: In this paper, a source-side context-informed (SSCI) hypothesis alignment method is proposed to carry out the word alignment and word reordering issues. First of all, the source--target word alignment links are produced as the hidden variables by exporting source phrase spans during the translation decoding process. Secondly, a mapping strategy and normalisation model are employed to acquire the 1-to-1 alignment links and build the confusion network (CN). The source-side context-based method outperforms the state-of-the-art TER-based alignment model in our experiments on the WMT09 English-to-French and NIST Chinese-to-English data sets respectively. Experimental results demonstrate that our proposed approach scores consistently among the best results across different data and language pair conditions.

Summary: 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 efficiently capture the subtree-to-subtree transitions observed not only in the source of word-aligned training data but also at decoding time. Using subtree movements and constraints as features in a log-linear model, we are able to help the reordering models make better selection. It also allows the subtle importance of monolingual syntactic movements to be learned alongside other reordering features. We show improvements in translation quality on English-Spanish and English-Iraqi translation tasks.

Summary: In most statistical machine translation (SMT) systems, bilingual segments are extracted via word alignment. In this paper we compare alignments tuned directly according to alignment F-score and BLEU score in order to investigate the alignment characteristics that are helpful in translation. We report results for two different SMT systems (a phrase-based and an n-gram-based system) on Chinese to English IWSLT data, and Spanish to English European Parliament data. We give alignment hints to improve BLEU score, depending on the SMT system used and the type of corpus.

Summary: Rule based machine translation methods require sophisticated transfer rules which require many man-years of work to build. This cost is too high especially for less represented language pairs, such as Hungarian and Japanese. This paper proposes a simple and robust method do automatically build a transfer rule set for the Hungarian-Japanese language pair. We use a small parsed bilingual corpus and a bilingual dictionary. We concentrate on inducing the most frequent target language translation rules from all instances of a source language rule. We achieved good accuracy especially for low level rules, especially important in case of agglutinative languages.

Summary: Most of the existing, easily available parallel texts to train a statistical machine translation system are from international organizations that use a particular jargon. We consider the automatic adaptation of the translation model to the news domain. The initial system was trained on more than 200M words of UN bitexts. We then explore large amounts of in-domain monolingual texts to modify the probability distribution of the phrase-table and to learn new phrase-pairs. This procedure achieved an improvement of 3.5 BLEU points on the test set in an Arabic/French statistical machine translation system. This results compares favorably with other large state-of-the-art systems for this language pair.

Summary: In this paper we describe a six-ways parallel public-domain corpus consisting of 2100 United Nations General Assembly Resolutions with translations in the six official languages of the United Nations, with an average of around 3 million tokens per language. The corpus is available in a preprocessed, formatting-normalized TMX format with paragraphs aligned across multiple languages. We describe the background to the corpus and its content, the process of its construction, and some of its interesting properties.

Summary: A method for evaluating MT performance embedded in Cross-Language Instant Messaging (CLIM) systems is presented. A web interface that provided concurrent real-time translation for instant messaging from multiple MT services was developed and used by paid participants to collaborate on a photo identification task. The method showed a task performance benefit due to the availability of multiple translation alternatives. The method also provides a new evaluation metric for MT systems based on user’s task motivated choices. This method was used to compare two English-Japanese online translation systems, one from Google, and one from Excite/Japan.

Summary: Anecdotally, Statistical Machine Translation works much better in some language pairs than others, but methodological problems mean that it is difficult to draw hard conclusions. In this paper, we report on an experiment where a small-vocabulary multilingual interlingua-based translation system was used to generate data to train SMT models for the 9 pairs involving the source languages English, French, Japanese and the target languages English, French, Japanese, Arabic. As expected, translation between English and French in both directions performed much better than translation involving Japanese. Less obviously, translation from English and French to Arabic performed approximately as well as translation between English and French, and translation to Japanese performed better than translation from Japanese.

Summary: We evaluate the use of Machine Translation technology to repair general errors in second language (L2) authoring. This method takes into account both languages, and is thus able to model linguistic interference phenomena where the author produces an erroneous word for word translation of his L1 thought. We evaluate a simple roundtrip MT approach on a corpus of foreign-sounding errors produced in a FSL context. We find that the roundtrip approach is better at repairing linguistic interference errors than non-interference ones, and that it is better at repairing errors which only involve function words. We also show that the approach is better at inferring the author's correct L1 thought, than it is at correctly re-expressing it into L2.

Summary: There are numerous ways to extract phrases in Phrase-based Statistical Machine Translation. Previous approaches include inducing phrase pairs from parallel treebanks (both constituency and dependency). We introduce phrase pairs induced from percolated dependencies (converting constituency parses to dependency parses using head percolation tables, in the absence of a dependency parser) as a unique knowledge source in the PBSMT framework.

This paper presents results on French to English SMT by emulating previously published results and scaling up data size to 13 times. We also contrast our percolated phrase table with three other phrase extractions and present results on combining these extracted phrase pairs to build new systems.

Summary: In this paper, we describe our ongoing research project of Virtual Babel: a context-aware machine translation system for Second Life, one of the most popular virtual worlds. We augment the Second Life viewer to intercept the incoming/outgoing chat messages and reroute the message to a statistical machine translation server. The returned translations are appended to the original text message to help users to understand the foreign language. Virtual Babel provides a platform to study cross-lingual conversations facilitated by machine translation in virtual worlds and we observe interesting phenomena that are not present in document translations.

Summary: In this paper, we present 2D-Linking, a new unsupervised method for word alignment that is based on association scores between words in a bitext. 2D-Linking can align m-to-n units. It is very efficient because it requires only two passes over the data and less memory than other methods. We show that 2D-Linking is superior to competitive linking and as good as or better than symmetrized IBM Model 1 in terms of alignment quality and that it supports trading off precision against recall. 2D-Linking is easy to implement and is sufficiently flexible and modular to be combined with other linguistic or statistical methods. The m-to-n alignment produced by 2D-Linking can be used directly in many applications of word alignment.