Latent space knowledge representation enhancement for low resource machine translation

Abstract

Low-resource machine translation usually uses data augmentation or constrained hidden space representations to improve translation quality, which ignores the knowledge representation divergence between different languages in latent space. We propose a latent space knowledge representation enhancement to improve the translation effect by reducing the divergence of knowledge representation between the source language and target language. Firstly, this method splits knowledge and feature representation from text representation in each language. Then, it promotes the knowledge representation of mutual learning between the two languages to reduce divergence. Next, the optimized knowledge representation and feature representation are re-combined to obtain enhanced text representation. Finally, the enhanced text representation is translated and reconstructed to reduce the differences between knowledge representations further. Through extensive experiments on public low-resource datasets 'English-German' and 'English-Turkish,' The method can achieve better performance on the test set. The results show that the method can effectively improve the ability of low-resource machine translation by reducing the divergence in knowledge representation between languages.