alibaba / translation-suggestion-psgd

PSGD

The code for ACL 2023 paper Easy Guided Decoding in Providing Suggestions for Interactive Machine Translation.

The code is forked from this commit of fairseq.

PSGD is short for "Prefix-Suffix Guided Decoding". It's a constrained decoding algorithm that generate a span in a sequence with given prefix and suffix. We propose and apply this algorithm for Translation Suggestion, improving both suggestion quality and computation efficiency.

Getting Started

An example script of running PSGD and compare it with the Lexically Constrained Decoding on a Translation Suggestion dataset can be found in examples/translation_suggestion/example.sh.

Requirements

See the README of fairseq for details

License

PSGD is MIT-licensed.

Citation

Please cite as:

@inproceedings{wang2023easy,
    title={Easy Guided Decoding in Providing Suggestions for Interactive Machine Translation},
    author={Ke Wang and Xin Ge and Jiayi Wang and Yu Zhao and Yuqi Zhang},
    year={2023},
    booktitle={Proceedings of the 61th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)}
}