Translation method applied to low-resource machine translation

• Main Authors: ZHANG QIULIN, XU PING
• Format: Patent
• Language: Chinese; English
• Published: 03.06.2022
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; ELECTRIC DIGITAL DATA PROCESSING; HANDLING RECORD CARRIERS; PHYSICS; PRESENTATION OF DATA; RECOGNITION OF DATA; RECORD CARRIERS

The invention discloses a translation method applied to low-resource machine translation, which comprises the following steps of: preprocessing a source language input text, and performing word embedding operation to obtain a source language input text x belonging to RL * D of a model; in the machine translation model, sampling the processed source language input text on the basis of the multi-head self-attention model so as to obtain multi-scale data representation; taking the sampled data representation as Query and Key to carry out dot product operation to obtain n attention weight matrixes; and fusing the n attention weight matrixes obtained in the above process, performing dot product operation on the fused attention weight matrixes and a source language input text x to obtain a fused output vector, and using the fused output vector as decoder input to realize a subsequent translation process. The method has good universality, the model performance can be improved under the condition that extra parameter