Application of Big Data Visualization Technology in Quality Assessment of English Translation of Ancient Books

Abstract

The development and application of machine translation cannot be separated from the evaluation of its quality. Correct evaluation is a correct lead to its development direction, and the translation direction respected by mainstream evaluation can be said to be the development direction of machine translation in research. In this paper, the Transformer model of the multi-head attention mechanism is constructed, and the optimization and improvement schemes in both model architecture and training data are proposed for the mBERT and XLM models to form the XLM-R model. Based on the predictor-evaluator QE model and the pre-trained model-evaluator QE two models, the XLM-R translation quality assessment baseline is proposed. A CLS pooling method is proposed for the baseline system, which ensures similarity at a finer-grained subword level. Meanwhile, the subword similarity scoring index is embodied using a simple splicing method, taking into account the global semantic quality. Simulation experiments are set up to evaluate the quality of translation tasks at the sentence and system levels, respectively. There was a 7.92% improvement in the Pearson correlation coefficient of the model with similarity characterization when compared to the one without the model in the sentence-level QE task development set, which was 0.0379 higher. The model’s translation fluency also showed an increasing trend in the interval of 0.15 to 0.5 with the growth of the manual score, indicating that the model proposed in this paper is effective in baseline system evaluation.