Research On Multi-group Parameter Tuning And Decoding In Statistical Machine Translation

As the Internet develops, people all over the world are getting closer. Machinetranslation has become a bridge across users of different languages, for it is cheapand convenient. Although Machine translation technique has improved a lot since itwas proposed, it’s still a long way to go to be as good as human translation. As forthe widely used phrase based statistical machine translation, there are many defectsto be improved.In statistical machine translation, usually, to reduce the search space,decodingalgorithms are always used with pruning. But this brings a problem called searcherror, which means the decoder can’t find the best translation according to the model.We’ll find ways to solve this.We made a plan for the problem mentioned above. We put forward a methodcalled multi-group weights based decoding. Different from traditional methods, ituses multi-group weights to decode a sentence. To train the multi-group weightsused, we propose multi-group weights training method. There is a significantdifference between the two proposed methods and previous methods, that they arebased on not sentences but derivations which are more granular. We also proposedusing forced decoding to get reference derivations in the training.Finally, we use Cubit decoder and MERT toolkit from Moses to implement themulti-group weights training and decoding method. Then with reasonableconfiguration, we designed several experiments to check the validity. The resultsshow that by using the more granular weights, search errors decrease, and thetranslation quality improves.