SCIENCE@home

Towards English-centric Zero-shot Neural Machine Translation: The Analysis and Solution

Shuai Wang

Published 2023 in 2023 IEEE International Conference on Sensors, Electronics and Computer Engineering (ICSECE)

ABSTRACT

The proposal of multilingual Neural Machine Translation (NMT) models enables zero-shot translation without parallel data of the desired language pairs. However, in the real world, the data used to train multi- lingual NMT is usually English-centric (with en as the source or target). This sparsity of language pairs could affect the training of multilingual NMT for zero-shot translation. In this paper, we first give a quantitative analysis of the zero-shot NMT, and find the bias caused by the English-centric data leads to a dramatic drop of zero-shot translation performance. The reason is the model sometimes generates words in the wrong languages, often consecutively. To reduce the word language errors, we propose three new architectures to incorporate language-specific information into the model. For the error continuity problem, we propose a task-selection based curriculum learning method to efficiently enrich the valid word patterns of diversified language pairs during training. Experiments show that our methods effectively alleviate the English-centric problem, and significantly improve the performance of zero-shot NMT.

PUBLICATION RECORD

CITATION MAP

EXTRACTION MAP

CLAIMS

  • No claims are published for this paper.

CONCEPTS

  • No concepts are published for this paper.

REFERENCES

Showing 1-38 of 38 references · Page 1 of 1

CITED BY

  • No citing papers are available for this paper.

Showing 0-0 of 0 citing papers · Page 1 of 1