TLSPG: Transfer learning-based semi-supervised pseudo-corpus generation approach for zero-shot translation

Amit Kumar, Rajesh Kumar Mundotiya, Ajay Pratap, Anil Kumar Singh

Abstract: Highlights•The zero-shot machine translation model aims to develop a machine translation system for those languages which do not have parallel training data.•A Transfer Learning-based Semi-supervised Pseudo-corpus Generation Approach for Zero-shot Translation (TLSPG) approach is proposed to generate the parallel data for zero-resource language pairs by exploiting the relatedness between the related language pairs of zero resource language.•Transformer and Moses-based hybrid architecture is used to train the translation model.•The proposed model consists of three modules: Transformer-based Semi-supervised learning (TSL), Moses-based Semi-supervised learning (MSL) and Transfer learning-based pseudo-corpus generation.•To validate our results on zero-shot experiments, we have analysed different parameters such as language relatedness, morphological complexity, different distribution of data size and compare with existing state-of-the-art architectures.