ListNet-based MT Rescoring

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 248–255, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. ListNet-based MT Rescoring †Jan Niehues, ∗Quoc Khanh Do, ∗Alexandre Allauzen and †Alex Waibel †Karlsruhe Institute of Technology, Karlsruhe, Germany ∗LIMSI-CNRS, Orsay, France †firstname.surname@kit.edu ∗surname@limsi.fr Abstract The log-linear combination of different features is an important component of SMT systems. It allows for the easy in- tegartion of models into the system and is used during decoding as well as for n- best list rescoring. With the recent suc- cess of more complex models like neural network-based translation models, n-best list rescoring attracts again more attention. In this work, we present a new technique to train the log-linear model based on the ListNet algorithm. This technique scales to many features, considers the whole list and not single entries during learning and can also be applied to more complex mod- els than a log-linear combination. Using the new learning approach, we im- prove the translation quality of a large- scale system by 0.8 BLEU points during rescoring and generate translations which are up to 0.3 BLEU points better than other learning techniques such as MERT or MIRA. 1 Introduction Nowadays, statistical machine translation is the most promising approach to translate from one natural language into another one, when sufficient training data is available. While there are several powerful approaches to model the translation pro- cess, nearly all of them rely on a log-linear com- bination of different models. This approach al- lows the system an easy integration of additional models into the translation process and therefore a great flexibility to address the various issues and the different language pairs. The log-linear model is used during decoding and for n-best list rescoring. Recently, the success of rich but computationally complex models, such as neural network based translation models (Le et al., 2012), leads to an increased interest in rescor- ing. It was shown that the n-best list rescoring is an easy and efficient way to integrate complex models. From a machine learning perspective the log- linear model is used to solve a ranking problem. Given a list of candidates associated with different features, we need to find the best ranking accord- ing to a reference ranking. In machine translation, this ranking is, for example, given by an automatic evaluation metric. One promising approach for this type of problems is the ListNet algorithm (Cao et al., 2007), which has already been applied suc- cessfully to the information retrieval task. Using this algorithm it is possible to train many features. In contrast to other algorithms, which work only on single pairs of entries, it considers the whole list during learning. Furthermore, in addition to train the weights of a linear combination, it can be used for more complex models such as neural networks. In this paper, we present an adaptation of this al- gorithm to the task of machine translation. There- fore, we investigate different methods to normal- ize the features and adapt the algorithm …

Content truncated to 500 words