As these demands increase exponentially in the size of cross-sentence text translation, the currently available methods of neural machine translation fail to meet the requirements of discourse-level semantic coherence, which is primarily reflected in jerry-shaky logical relations among the translated sentences, the loss of referential information readily, and discourse-level semantic drift. To solve this, the current paper suggests a profound learning translation framework with the incorporation of semantic understanding. It provides the integration of multi-layered semantic representation, cross-sentence dependency model, and semantic coherence constraint optimization, which enables the cooperatively transferring semantic information at the word, sentence, and discourse levels of the translated text. In particular, the model offers a concise model of word, sentence and discourse semantics at the encoding phase and offers semantic consistency constraints and context sensitive attention mechanism at the decoding phase to ensure semantic continuity and logical consistency among sentence segments. The experimental verification carried out in this paper showed a significant increase in the translation semantic coherence index Coh of the baseline model from 0.796 to 0.834 on standard parallel corpora and also a significant increase in the referential consistency rate of 83.4% to 87.9% as well as in the BLEU value 28.91 to 29.47, which is a significant improvement in the ability to control details on the discourse level and guarantee the translation accuracy level.
Diao et al. (Thu,) studied this question.