Neural machine translation method and device based on word vector connection technology

Abstract

The invention discloses a neural machine translation method based on a word vector connection technology. The method comprises the steps that at the encoding stage, an encoder acquires the word vectorsequence of a source statement; the hidden layer vector sequence corresponding to the source statement is determined according to the determined forward vector sequence and the opposite vector sequence; the vector expression, containing contextual information, corresponding to each source word includes the forward hidden layer state, the opposite hidden layer state and the word vector corresponding to the source word; the contextual vector can be acquireed; at the decoding stage, a decoder forecasts the target word of the corresponding source word, so that the target statement of the source statement is generated. After the technical scheme provided by the embodiment is applied, an information channel between the source end word vector and the target end word vector is shortened; the connection and mapping among the word vectors are enhanced; the translation system performance is enhanced; the translation quality is improved. The invention also discloses a neural machine translation device based on the word vector connection technology, and the corresponding technical effect is achieved.

Description

technical field [0001] The present invention relates to the technical field of neural machine translation (NMT), in particular to a neural machine translation method and device based on word vector connection technology. Background technique [0002] With the rapid development of computer technology, computer computing power has been continuously improved, big data has been widely used, and deep learning has also been further applied. The NMT (Neural Machine Translation) technology based on deep learning has received more and more attention. [0003] In the NMT field, the more commonly used translation model is the encoder-decoder model with an attention-based mechanism. The main idea is to encode the sentence to be translated, that is, the source sentence, through the encoder, and use a vector representation, and then use the decoder to decode the vector representation of the source sentence, and translate it into the corresponding translation, that is, the target sentence...

AI Technical Summary

Problems solved by technology

However, because the learned word vector of the source sentence and the word vector of the target sentence are located at both ends of the encoder-decoder framework (source end and target end), a very complicated information conversion channel (encoder and decoder) is required in the middle , so that there is a lack of direct connection and mapping between the word vectors of the source sentence and the word vectors of the target sentence, which can easily cause the NMT system to generate wrong word alignment information, thereby reducing the translation quality

Images