Neural Automatic Evaluation Of Machine Translation Based On Multiple Information Fusion

Automatic evaluation of machine translation is an important research task in the field of machine translation,which refers to the quality evaluation of machine translation system output by computer technology.The guidance provided by automatic evaluation of machine translation is vital for optimizing machine translation systems.Currently,the mainstream approach to automatic evaluation of machine translation is based on neural network.The lastest neural automatic evaluation methods of machine translation use pre-trained contextual embeddings to extract different deep semantic features,and then simply concatenate them feed into the multi-layer neural network to predict translation quality.Simply concatenated features results in lack of deep fusion between features.And fine-grained accurate matching information tends to be lost when layer by layer abstraction is used for prediction.To address these limitations,this paper proposes a new neural automatic evaluation method for machine translation based on multiple information fusion.Specifically,we introduce middle fusion and late fusion into machine translation evaluation.We propose to use embrace fusion to interactively fuse different features in the middle stage,and to fuse sentence mover’s distance and sentence cosine similarity that are based on fine-grained accurate matching in the late stage.Moreover,current automatic evaluation methods for machine translation utilize large-scale pre-trained language models to extract the semantic representations of machine translations and reference translations,and then calculate the similarity of these representations.However,current pre-trained language models may map semantically similar sentences into a dense vector space that is far away from each other.In this paper,we propose a new neural automatic evaluation method of machine translation based on siamese similarity feature.We use the siamese network structure to fine-tune the pre-trained language model,so that it can map sentences with similar semantics into a dense vector space with a closer distance,which is more suitable for the task of machine translation evaluation,and then use the fine-tuned siamese pre-trained language model to extract features of semantic similarity,which are incorporated into the neural machine translation evaluation method to improve its performance.In order to verify the effectiveness of the proposed methods,extensive experiments are carried out on the WMT’21 Metrics Task,experimental results show that our proposed methods can effectively improve the correlation with human judgement,and achieve competitive performance with the best metrics in the evaluation campaign.