Research On Neural Machine Translation Combining Lexicology And Syntax

Translation model based on Neural Machine Learning(NMT)has gradually become the most commonly used method in the field of machine translation.However,the complexity of natural languages and the randomness of neural network lead to rare word problem and uncontrollable translation process etc.Meanwhile,the current attention-based neural translating model usually conducts translation via serialized encoding and decoding,while ignoring the dependency parsing information and hierarchical feature of languages,which provides an inspiration for NMT system optimization.In this paper,lexical and syntactical information from preprocessed language data is introduced as priori knowledge to optimize the current NMT model.1)Named entity recognition is imported to take some entities out from sentences for direct translation.Tagging-replacing strategy is used to make up for the negative effect from the randomness of neural network.Word alignment is improved by the optimized attention mechanism and multi-task learning.2)Dependency parsing is imported to obtain depending relationships of words in a sentence,which is rebuilt as matrix or array according to the model structures.Incorporating with multi-head selfattention mechanism and gating,dependency relationships provide syntactic information for both encoder and decoder of the NMT system,enhancing the ability to extract hidden features.Several experiments are designed and implemented to evaluate effects of the lexical and syntactical information.Recall rate of named entities and BLEU scores are calculated in lexicology optimization and BLEU improvements are calculated in syntax optimization.Results show that most modifications improve the performance of the basic model,even better than some recent novel works.Compared with respective baseline,the BLEU score increases 2.32 points with lexicology and 1.86 with syntax at maximum.The combination of the two methods results in an integrated optimization.Experiments prove that this work has positive effect on the NMT system.