Sentence-Level Machine Translation Quality Estimation Based On Neural Network Features

The quality estimation of machine translation plays an important role in the development of machine translation,which is a new research direction.Unlike the traditional automatic evaluation method of machine translation,the quality estimation method does not need to use human reference translation to evaluate the quality of machine translation.This feature makes the quality estimation method more flexible and practical than the traditional automatic evaluation method.At present,most of the common sentence-level translation quality estimation methods rely on the linguistic analysis of sentences to extract features.These methods cannot be used in all languages.In order to solve this problem,some researchers began to propose using neural network to extract features for sentence-level quality estimation tasks.Shah et al.proposed to extract sentence language model probability,sentence embedding feature and sentence cross entropy feature using continuous space language model for sentence-level quality estimation task.The continuous space language model is a feed-forward neural network language model,the input is a fixed-length of word sequence,with the increase of the number of hidden layers and hidden units,the computational complexity of the model will increase sharply.Based on the research work of Shah et al.,in this paper we first extract sentence embedding features by using continuous space language model,context word prediction model and word cooccurrence model respectively,and then compare them with those they extracted.Secondly,we use the recurrent neural network language model to extract sentence cross entropy features,which are used in sentence-level translation quality estimation task together with sentence embedding features.Considering that both sentence embedding features and sentence cross entropy features are monolingual features,they are suitable for reflecting translation difficulty and fluency,but have little effect on fidelity information in translation.In order to further extract the features that reflect the fidelity of translation,we train a neural machine translation system,and use this system to extract the alignment features of translation words to measure the fidelity of translation.Finally,on the basis of benchmark features,we combine the neural network features extracted in this paper,and find that the evaluation results have a great improvement compared with the baseline system,and our results are superior to shah et al.,which shows that the neural network features extracted in this paper can capture more translation quality information.