Zero parallel corpus multi-modal neural machine translation method

Abstract

The invention discloses a zero parallel corpus multi-modal neural machine translation method, which comprises the following steps: pre-training a pre-constructed neural machine translation model by utilizing a source language with corresponding picture information and a target language monolingual corpus; translating sentences in the source language into sentences in a target language by using a pre-trained neural machine translation model; calculating a reward value of a sentence level according to an internal relation between the sentence obtained by translation and a corresponding picture, and updating pre-trained neural machine translation model parameters by adopting a strategy gradient reinforcement learning method by taking the maximum expected total reward value as an optimization target so as to obtain a trained neural machine translation model; and translating the given source language sentence by using the trained neural machine translation model. According to the zero parallel corpus multi-modal neural machine translation method, a neural machine translation model of a zero resource language pair can be established by utilizing the internal relation between pictures and characters.

Description

technical field [0001] The invention relates to the technical field of machine translation, in particular to a zero-parallel corpus multi-modal neural machine translation method. Background technique [0002] End-to-end neural machine translation (hereinafter referred to as neural machine translation) has achieved rapid development both at home and abroad in recent years. Neural machine translation abandons the cumbersome structure and complex feature design in statistical machine translation, and directly sends parallel corpus to the neural network to complete the training of a complete translation system. [0003] In neural machine translation, even the simplest single-layer model has a large number of parameters (usually tens of millions of parameters) to optimize, while training a complex model with a large number of parameters in machine learning requires a lot of training data. As the model design of neural machine translation becomes more and more complex, and the d...

AI Technical Summary

Problems solved by technology

Although the above method can realize the training of the neural machine translation model by constructing parallel corpus, due to the limited effect of the picture description model, the wrong description generated by it will lead to the formation of low-quality parallel corpus training data, thereby affecting the trained neural machine translation model quality, resulting in less accurate translation results

Images