... clusteredtraining data.2 DomainAdaptation for Translation ModelsTo motivate efforts in domain adaptation, let usreview why additional training data can improve,but also decrease translation quality.Adding ... trans-lation models, which is only a subset of domain adaptation techniques in SMT.Mixture-modelling for language models is wellestablished (Foster and Kuhn, 2007). Language model adaptation ... purpose as translation model adaptation, i.e. skewing theprobability distribution in favour of in -domain translations. This means that LM adaptation mayhave similar effects as TM adaptation, ...
... the additional in -domain monolingualcorpora to adapt the out-of -domain translation mod-el for domain- specific translation task. In detail, webuild an adapted translationmodel in the followingsteps:• ... of further translation quality improve-ment.In this paper, we propose a novel adaptation method to adapt the translationmodel for domain- specific translation task by utilizing in -domain 459Philipp ... sen-tences, and Wu et al. (2008) used an in -domain translation dictionary and monolingual corpora toadapt an out-of -domain translationmodel for the in- domain text.Differing from the above-mentioned...
... lexical weighting by domain. We induceunsupervised domains from large corpora, and weincorporate soft, probabilistic domain membershipinto a translation model. Unsupervised modeling ofthe training ... training1Language modeladaptation is also prevalent but is not thefocus of this work.115data come from; and even if we do, “subcorpus” maynot be the most useful notion of domain for bettertranslations.We ... Korea, 8-14 July 2012.c2012 Association for Computational LinguisticsTopic Models for Dynamic TranslationModel Adaptation Vladimir EidelmanComputer Scienceand UMIACSUniversity of MarylandCollege...
... the domain adaptation setting, by “bootstrapping” them basedon known translations from the source domain. (3)Develop methods for integrating these mined dictio-naries into a phrase-based translation ... weights of a learned translationmodel todo well on a new domain. As expected, we shallsee that unseen words pose a major challenge foradapting translation systems to distant domains. Nomachine ... of the translation error when mov-ing to new domains. Using an extension ofa recent approach to mining translations fromcomparable corpora (Haghighi et al., 2008),we are able to find translations...
... do-main adaptation technique. Namely, we regard thepseudo-error corpus as the source domain and thereal-error corpus as the target domain, and modelsare learnt that fit the target domain. In ... for the domain adaptation, whicheliminates the need to change the learning algo-rithm. This method regards the models for thesource domain as the prior distribution and learnsthe models for ... automatically. Fur-thermore, we apply domain adaptation, thepseudo-error sentences are from the source domain, and the real-error sentences are fromthe target domain. Experiments show that sta-ble...
... word-based translationmodel into a unifiedframework, called TransLM. The experiments showthat this model gains better performance than boththe language model and the word-based translation model. ... translation model. 3 Our Approach: Phrase-Based Translation Model for QuestionRetrieval3.1 Phrase-Based Translation Model Phrase-based machine translation models (Koehnet al., 2003; D. Chiang, 2005; ... we proposea novel phrase-based translationmodel forquestion retrieval. Compared to the traditionalword-based translation models, the phrase-based translationmodel is more effective be-cause...
... interpolation of several models,with weights tuned by MERT on the developmentset (Och, 2003). The baselines consisted of the lan-guage model, two phrase translation models, twolexical models, and a ... target-given-source phrase model 1.3 Exponential phrase models with sharedfeaturesThe model used in this work is based on the familiarequation for conditional exponential models:1To avoid confusion ... gen-eration and modeling of inflected forms un-observed in training data and decoding proce-dures for a model with non-local target-sidefeature dependencies.1 Introduction Translation into...
... corpus dependon the domain from which the corpus is drawn.A change of predominant sense is often indicativeof a change in domain, as different corpora drawnfrom different domains usually give ... 89–96,Sydney, July 2006.c2006 Association for Computational LinguisticsEstimating Class Priors in Domain Adaptation for Word Sense DisambiguationYee Seng Chan and Hwee Tou NgDepartment of Computer ... to these sense priors. The workof these researchers showed that when the domain of the training data differs from the domain of thedata on which the system is applied, there will bea decrease...
... active learning for domain adaptation, fol-lowed by count-merging. Next, we describe an EM-based algorithm to estimate the sense priors in thenew domain. Performance of domainadaptation us-ing ... to a new domain. Domain adaptation is necessary when the train-ing and target domains are different. In this paper,49did not deal with the porting of a WSD system toa different domain. Escudero ... (%)Percentage of adaptation examples added (%)a-cara-truePriorFigure 2: Adaptation process for all 21 nouns.of the BC training examples. At each adaptation iter-ation, WSJ adaptation examples...
... domainsof discourse makes it an ideal candidate for domain adaptation. This work addressed two importantquestions of domain adaptation. First, we showedthat for a given source and target domain, ... model, but all of these newtechniques also make use of lexical features. Thuswe believe that our adaptation methods could be alsoapplied to those more refined models.While work on domainadaptation ... in -domain gold standard is 80.4%. We say that the adaptation loss for the baseline model is 7.6% and the adapta-tion loss for the SCL-MI model is 0.7%. The relativereduction in error due to adaptation...
... unintuitive word-for-word models; at thesame time, seemingly capable models often have se-rious hidden problems — intuition is no substitutefor experimentation. With translation ideas growingmore ... visualize other syntax-based MT mod-els, other tree-to-tree or tree-to-string MT models, ormodels for paraphrasing.2 Translation FrameworkIt is useful at this point to give a brief descrip-tion ... new approaches to statistical ma-chine translation, and more ideas are being sug-gested all the time. However, it is difficult to deter-mine how well a model will actually perform. Ex-perienced...
... implement other translation models easily: 1) STSG-based model when1d =; 2) SCFG-based model when1d = and2h =; 3) phrase-based translationmodel only (no reordering model) when 0c = ... We conducted Chinese-to-English translation ex-periments. We trained the translationmodel on the FBIS corpus (7.2M+9.2M words) and trained a 4-gram language model on the Xinhua portion of ... sequence alignment-based translation model. In this subsection, we first re-port the rule distributions and compare our model with the three baseline systems. Then we study the model s expressive...
... on feature-levelmulti-view adaptation. 3 Our Model Intuitively, source-specific and target-specific fea-tures can be drawn together by mining theirco-occurrence with domain- independent (common)features, ... deemed as differentdomains. The task is defined as top category classifi-cation. For example, the dataset denoted as DA-ECconsists of source domain: DA 1(+), EC 1(-); andtarget domain: DA 2(+), ... Co-Training for Domain Adaptation. In Pro-ceedings of NIPS, pages 1-9.Wenyuan Dai, Gui-Rong Xue, Qiang Yang and YongYu. 2007. Co-clustering Based Classification for Out-of -domain Documents....
... hi-erarchical rules simultaneously.2 Proposed Model Given an original translation model, T M, our goalis to find the optimally reduced translation model, T M∗, which minimizes the degradation ... This restriction guarantees thatthe translational coverage of the reduced model isas high as those of the original translation model. In addition, our model does not prune the phrasesand the ... sentence s, given the translation model T M. Consistency measures the similarity betweenthe two groups of decoded target sentences producedby two different translation models. There are num-ber...
... and target models in both N-grams andME models. It seems that with the adapted MEmodels, the same recognition accuracy for the tar-get evaluation data can be obtained with 50% less adaptation ... smaller,by penalizing weights with big absolute values.3 DomainAdaptation of MaximumEntropy ModelsRecently, a hierarchical Bayesian adaptation method was proposed that can be applied to a largefamily ... the domain- specific model is applied. Intu-itively, this approach can be described as follows:the domain- specific parameters are largely deter-mined by global data, unless there is good domain- specific...