ATOGAN:Adaptive Training Objective Generative Adversarial Network for Cross-lingual Word Alignment in Non-Isomorphic Embedding Spaces

Anonymous

Abstract: Cross-lingual word alignment is a task for word translation from monolingual word embedding spaces of two languages. Recent works are mostly based on supervised approaches, which need specific bilingual seed dictionaries. The unsupervised adversarial approaches, which utilize the generative adversarial networks to map the whole monolingual space, do not need any aligned data. However these approaches pay no attention to the problem of mode collapse and gradient disappearance in generative adversarial networks(GAN). We proposed an adaptive training objective generative adversarial network(ATOGAN). We combined particle swarm optimization(PSO) with GAN to select the training objective in GAN's training, which alleviates the problem of mode collapse and gradient disappearance. Moreover, we improved the word alignment by bi-directional mapping and consistency loss. Experimental results demonstrate that our approach is better than several state-of-the-art approaches in distant language pairs(non-isomorphic embedding spaces).