Non-autoregressive Mongolian-Chinese machine translation method based on round-robin decoding and vocabulary attention

Abstract

The invention provides a non-autoregressive Mongolian-Chinese machine translation method based on round-robin decoding and vocabulary attention, and the possibility of repeated generation is reduced by using round-robin generation; according to the method, vocabulary attention is used, each layer of a decoder is operated, each position pays attention to the whole word list, characters which are possibly generated are obtained, through the vocabulary attention, the decoder can enable the characters which are expected to be generated at all the positions to be interacted, and therefore the possibility of repeated generation is reduced, and meanwhile extra parameters do not need to be introduced; according to the method, the aligned cross entropy is used as a loss function of the non-autoregressive translation model, so that the problem that Mongolian word orders are difficult to model due to lack of autoregressive factors in the non-autoregressive translation model is solved; and duringdecoding, dynamic bidirectional decoding is adopted, so that a sentence-level better sampling result is obtained. According to the method, the decoding speed is increased in the Mongolian-Chinese translation process, and meanwhile the translation quality is improved.

Description

technical field [0001] The invention belongs to the technical field of machine translation, in particular to a non-autoregressive Mongolian-Chinese machine translation method based on look-around decoding and vocabulary attention. Background technique [0002] As one of the fundamental components of artificial intelligence, machine translation provides a method for solving language translation problems. Machine translation is the process of using computers to realize rapid conversion between two natural languages. The development of machine translation technology has been closely related to computer technology, information theory, linguistics The development of other disciplines closely followed. From the 1949 translation memorandum to the present, during this period, machine translation has gone through many different stages of development, and many methods have emerged. To sum up, there are three main categories, which are rule-based methods at first, and then develop into...

AI Technical Summary

Problems solved by technology

[0024] Moreover, existing autoregressive and non-autoregressive models are usually trained using cross-entropy loss. Cross-entropy is a strict loss function that predicts out-of-position words are penalized, even for output sequences with small edit distances.

Images