Local System Voting Feature for Machine Translation System Combination

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 467–476, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. Local System Voting Feature for Machine Translation System Combination Markus Freitag, Jan-Thorsten Peter, Stephan Peitz, Minwei Feng and Hermann Ney Human Language Technology and Pattern Recognition Group Computer Science Department RWTH Aachen University D-52056 Aachen, Germany <surname>@cs.rwth-aachen.de Abstract In this paper, we enhance the traditional confusion network system combination approach with an additional model trained by a neural network. This work is moti- vated by the fact that the commonly used binary system voting models only assign each input system a global weight which is responsible for the global impact of each input system on all translations. This prevents individual systems with low sys- tem weights from having influence on the system combination output, although in some situations this could be helpful. Fur- ther, words which have only been seen by one or few systems rarely have a chance of being present in the combined output. We train a local system voting model by a neural network which is based on the words themselves and the combinatorial occurrences of the different system out- puts. This gives system combination the option to prefer other systems at different word positions even for the same sentence. 1 Introduction Adding more linguistic informed models (e.g. language model or translation model) additionally to the standard models into system combination seems to yield no or only small improvements. The reason is that all these models should have al- ready been applied during the decoding process of the individual systems (which serve as input hy- potheses for system combination) and hence al- ready fired before system combination. To im- prove system combination with additional models, we need to define a model which can not be ap- plied by an individual system. In state-of-the-art confusion network system combination the following models are usually ap- plied: System voting (globalVote) models For each word the voting model for system i (1 ≤ i ≤ I) is 1 iff the word is from system i, otherwise 0. Binary primary system model (primary) A model that marks the primary hypothesis. Language model 3-gram language model (LM) trained on the input hypotheses. Word penalty Counts the number of words. To gain improvements with additional models, it is better to define models which are not used by an individual system. A simple model which can not be applied by any individual system is the binary system voting model (globalVote). This model is the most important one during system combination decoding as it determines the impact of each individual system. Each system i is as- signed one globalVote model which fires if the word is generated by system i. Nevertheless, this simple model is independent of the actual words and the score is only based on the global prefer- ences of the individual systems. This disadvan- tage prevents system combination from produc- ing words which have only been seen by systems with low system …

Content truncated to 500 words