| The development of encoder-decoder neural network technology has led to the emer-gence of neural machine translation.Significant progress has been made in improving translation quality through neural machine translation.Due to the diversity of human language,which includes lexical diversity,syntactic diversity,and synonyms,a source sentence often corresponds to multiple valid translations.Translation is essentially a one-to-many matching relationship.Existing neural machine translation models mainly con-sider the one-to-one mapping between source and target languages,while ignoring the one-to-many mapping between source and target languages.Existing diverse machine translation models focus on capturing the uncertainty of text generation and ignore syn-tactic information in sentences,resulting in diversity only reflecting on the lexical level,and multiple generated translations having similar syntactic structures.Therefore,if the model can learn and capture the complex syntactic hierarchical structure in the sentence,it can generate multiple valid and syntax diverse target language sentences under the guid-ance of syntactic structure information.As a result,the model provide users of machine translation systems with more choices.We have separately considered the diversity at the grammatical and lexical levels,and modeled the syntax structure information of the sentences.The main content of this paper consists of three parts.Firstly,this paper proposes creative syntax-aware mixture of experts model.Based on the encoder of the mixture of experts model,a constituent attention module is introduced to learn and capture syntactic information in the source sentence.As a result,the model can better process and understand the syntactic structure in the source sentence,and use it to guide the model decoding process.Secondly,this paper proposes an adaptive label smoothing approach.We adopt the adaptive label smoothing algorithm to generate an adaptive soft target distribution,which assigns probabilities to both target and non-target words.This soft target distribution is utilized for model training,which can further enhance the diversity of the model.In the end,this paper proposes a multi-candidate optimization framework.Based on the principles of reinforcement learning,our model learns a one-to-many mapping relationship from its generated candidate translations.Firstly,we utilize a pre-trained diverse translation model to generate multiple candidate translations,and then we calculate the diversity and translation quality rewards for each candidate translation.Finally,we apply these rewards to optimize the model.Experiments on the WMT’17 English-German and WMT’14 English-French datasets show that our proposed syntax-aware mixture of experts model can effectively improve the diversity of generated translations while ensuring translation quality.This model trained with adaptive label smoothing algorithm can further enhance its performance.Building upon this,the model optimized with a multi-candidate framework has achieved a better balance between translation quality and diversity. |
