Skip to content
Multi-Task Neural Models for Translating Between Styles Within and Across Languages
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
scripts
system_output
LICENSE
README.md
setup.sh

README.md

multitask-ft-fsmt

Multi-Task Neural Models for Translating Between Styles Within and Across Languages -- Formality Transfer (FT) and Formality-Sensitive Machine Translation (FSMT).

This repository contains implementations for

@InProceedings{niu-rao-carpuat:2018:COLING2018,
  author    = {Niu, Xing  and  Rao, Sudha  and  Carpuat, Marine},
  title     = {Multi-Task Neural Models for Translating Between Styles Within and Across Languages},
  booktitle = {{COLING}},
  year      = {2018}
}

Usage Instructions

  1. Set-up -- Follow the instructions in setup.sh to obtain necessary software and data.
  2. Training
Bi-FT-ensemble (Bi-directional FT + domain combination + ensemble decoding)
> bash scripts/main.sh -e
MultiTask-tag-style
> bash scripts/main.sh -m tag-style -k 12 -e -f
MultiTask-style
> bash scripts/main.sh -m style -k 12 -e -f
MultiTask-random
> bash scripts/main.sh -m random -k 12 -e -f
  1. Evaluation -- Adjust parameters in scripts/evaluate.sh
> bash scripts/evaluate.sh

System Output

We provide the system output for GYAFC/OpenSubtitles2016 test sets (see setup.sh).

  • Formality Transfer (GYAFC)

    • Bi-FT-ensemble (Bi-directional FT + domain combination + ensemble decoding)
    • MultiTask-tag-style
  • Formality-Sensitive Machine Translation (OpenSubtitles2016)

    • NMT-constraint
    • MultiTask-tag-style
    • MultiTask-style
    • MultiTask-random
    • PBMT-random
You can’t perform that action at this time.