Research on the discrimination of translation difficulty level based on spoken language signal processing technology

Abstract

Computer translation systems require the ability to use external dictionaries in order to quickly improve the accuracy of professional field translations, and establishing a more accurate grading and evaluation system has important research value. This article selects spoken signals as the pre training processing object, establishes a pre training ELMo model based on optimized similar word vectors based on spoken signals, and applies this model and method to Chinese logistics outbound call data. Subsequently, compared with the traditional industry best translation mechanism, this paper proposes an improved machine translation intervention method based on dictionary guided decoding. This method, given a Transformer baseline model, supervises additional attention heads during the training process, which can achieve better intervention success rate and translation quality after intervention. Finally, using a tree shaped multi-layer combination model, a tree shaped two-layer combination model is constructed. Several BP neural networks are used as the lowest level classifiers, and the LVQ neural network is used as the upper level classification combiner to train the DuIE2.0 dataset. Experimental results have shown that compared with the BERT multi head model, the F1 value of the proposed model in this chapter has increased by 1.85%.