RegMT System for Machine Translation, System Combination, and Evaluation

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 323–329, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics RegMT System for Machine Translation, System Combination, and Evaluation Ergun Bic¸ici Koc¸ University 34450 Sariyer, Istanbul, Turkey ebicici@ku.edu.tr Deniz Yuret Koc¸ University 34450 Sariyer, Istanbul, Turkey dyuret@ku.edu.tr Abstract We present the results we obtain using our RegMT system, which uses transductive re- gression techniques to learn mappings be- tween source and target features of given par- allel corpora and use these mappings to gen- erate machine translation outputs. Our train- ing instance selection methods perform fea- ture decay for proper selection of training in- stances, which plays an important role to learn correct feature mappings. RegMT uses L2 regularized regression as well as L1 regular- ized regression for sparse regression estima- tion of target features. We present transla- tion results using our training instance selec- tion methods, translation results using graph decoding, system combination results with RegMT, and performance evaluation with the F1 measure over target features as a metric for evaluating translation quality. 1 Introduction Regression can be used to find mappings between the source and target feature sets derived from given parallel corpora. Transduction learning uses a sub- set of the training examples that are closely related to the test set without using the model induced by the full training set. In the context of statistical ma- chine translation, translations are performed at the sentence level and this enables us to select a small number of training instances for each test instance to guide the translation process. This also gives us a computational advantage when considering the high dimensionality of the problem as each sentence can be mapped to many features. The goal in transductive regression based ma- chine translation (RegMT) is both reducing the com- putational burden of the regression approach by re- ducing the dimensionality of the training set and the feature set and also improving the translation quality by using transduction. We present translation results using our training instance selection methods, translation results us- ing graph decoding, system combination results with RegMT, and performance evaluation with the F1 measure over target features as a metric for eval- uating translation quality. RegMT work builds on our previous regression-based machine translation results (Bicici and Yuret, 2010) especially with in- stance selection and additional graph decoding ca- pability. We present our results to this year’s chal- lenges. Outline: Section 2 gives an overview of the RegMT model. In section 3, we present our train- ing instance selection techniques and WMT’11 re- sults. In section 4, we present the graph decoding re- sults on the Haitian Creole-English translation task. Section 5 presents our system combination results using reranking with the RegMT score. Section 6 evaluates the F1 measure that we use for the auto- matic evaluation metrics challenge. The last section present our contributions. 2 Machine Translation Using Regression Let X and Y correspond to the sets of tokens that can be used in the source and …

Content truncated to 500 words