Low-resource neural machine translation method fusing shared semantic space and bidirectional iteration generation optimization

Abstract

The invention belongs to the technical field of natural language processing and neural machine translation, and discloses a shared semantic space and bidirectional iteration generation optimization fused low-resource neural machine translation method, which comprises the following steps of: constructing a shared semantic space module and a bidirectional iteration generation optimization module (BIGO); according to the method, firstly, singular value decomposition and entropy regularization optimal transmission are utilized, a source language and a target low-resource language are mapped into a unified semantic space, high consistency of the two languages is achieved on the representation level, and therefore the cross-language generalization ability of a model under the condition of insufficient data is remarkably improved. Subsequently, through alternate training of forward and reverse Transformer models, translation quality is continuously improved in a cycle of pseudo bilingual corpus generation, reverse reconstruction and forward and reverse joint optimization, so that the model can correct errors generated in the previous round and enhance semantic consistency in each round of iteration. According to the method, dual convergence of the model parameters and the semantic alignment matrix can be stably realized, the accuracy, robustness and long sentence consistency of low-resource language translation are remarkably improved, and the phenomena of training instability and over-fitting are effectively reduced.