Adaptive Model Weighting and Transductive Regression for Predicting Best System Combinations

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 276–281, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Adaptive Model Weighting and Transductive Regression for Predicting Best System Combinations Ergun Bi¸cici Ko¸c University 34450 Sariyer, Istanbul, Turkey ebicici@ku.edu.tr S. Serdar Kozat Ko¸c University 34450 Sariyer, Istanbul, Turkey skozat@ku.edu.tr Abstract We analyze adaptive model weight- ing techniques for reranking using in- stance scores obtained by L1 regular- ized transductive regression. Compet- itive statistical machine translation is an on-line learning technique for se- quential translation tasks where we try to select the best among com- peting statistical machine translators. The competitive predictor assigns a probability per model weighted by the sequential performance. We de- fine additive, multiplicative, and loss- based weight updates with exponential loss functions for competitive statisti- cal machine translation. Without any pre-knowledge of the performance of the translation models, we succeed in achieving the performance of the best model in all systems and surpass their performance in most of the language pairs we considered. 1 Introduction When seen as independent instances, system combination task can be solved with a sequen- tial learning algorithm. Online learning algo- rithms enable us to benefit from previous good model choices to estimate the next best model. We use transductive regression based machine translation model to estimate the scores for each sentence. We analyze adaptive model weighting tech- niques for system combination when the com- peting translators are SMT models. We use separate model weights weighted by the se- quential performance. We use additive, mul- tiplicative, or loss based weight updates to update model weights. Without any pre- knowledge of the performance of the transla- tion models, we are able to achieve the per- formance of the best model in all systems and we can surpass its performance as well as the regression based machine translation’s perfor- mance. The next section reviews the transductive regression approach for machine translation, which we use to obtain instance scores. In sec- tion 3 we present competitive statistical ma- chine translation model for solving sequential translation tasks with competing translation models. Section 4 presents our results and ex- periments and the last section gives a sum- mary of our contributions. 2 Transductive Regression Based Machine Translation Transduction uses test instances, which can sometimes be accessible at training time, to learn specific models tailored towards the test set. Transduction has computational advan- tages since we are not using the full train- ing set and a smaller set of constraints exist to satisfy. Transductive regression based ma- chine translation (TRegMT) aims to reduce the computational burden of the regression ap- proach by reducing the dimensionality of the training set and the feature set and also im- prove the translation quality by using trans- duction. Regression Based Machine Translation: Let n training instances be represented as (x1, y1), . . . , (xn, yn) ∈ X∗×Y ∗, where (xi, yi) corresponds to a pair of source and target lan- guage token sequences. Our goal is …

Content truncated to 500 words