Research on the Optimization Model of Semantic Coherence and Fluency in Language Translation

Abstract

Traditional Transformer machine translation cannot effectively deal with translation errors due to the lack of an explicit evaluation mechanism for the correctness of the target language text, which not only hinders the improvement of the translation quality of the target language but also makes the decoding process highly complex. This paper constructs a model for language translation based on decoding optimization. The problem of choosing the number of decoding rounds is transformed into a Markov decision-making process, and a reinforcement learning algorithm is used, which is based on the translation quality of the target-language translation and the number of decoding rounds obtained from the translation. The appropriate number of decoding rounds is selected according to the different complexity of the source language text to be translated, so that the translation model can achieve a balance between the translation quality of the target language translation and its corresponding time consumption. The optimization of the traditional Transformer machine translation algorithm has finally been achieved. The optimization effect is examined on the dataset. It is found that when the dimension is 1024, the number of parameters of the optimized model decreases by 16.03%, the average reasoning time is under 55 seconds, and the decoding time is always lower than that of the Transformer baseline model. The convergence trend of its Loss is similar to that of the Transformer model but with a faster and better convergence speed. On the six Chinese-English bilingual datasets, the average BLEU score of this paper’s model is 64.586. On the five Chinese-German bilingual datasets, compared with Transformer’s average BLEU value of 44.078, the score of this paper’s model is still the first one, which is 18.29% higher than that of the traditional Transformer model’s BLEU value, and it has a good language translation performance. The language translation model based on decoding optimization in this paper provides a useful exploration for machine translation optimization in terms of improving translation quality, enhancing Loss convergence speed, reducing the number of parameters, and improving decoding speed.