Open-Source Machine Translation Quality Estimation in PyTorch
Quality estimation (QE) is one of the missing pieces of machine translation: its goal is to evaluate a translation system’s quality without access to reference translations. We present OpenKiwi, a Pytorch-based open-source framework that implements the best QE systems from WMT 2015-18 shared tasks, making it easy to experiment with these models under the same framework. Using OpenKiwi and a stacked combination of these models we have achieved state-of-the-art results on word-level QE on the WMT 2018 English-German dataset.
Following our nomination in early July, we are happy to announce we won the Best Demo Paper at ACL 2019! Congratulations to the whole team and huge thanks for supporters and issue reporters.
Check out the published paper.
We are going to release the web interface we had put in place for the live demo presentation at ACL.
We have released the OpenKiwi tutorial we presented at MT Marathon 2019.
- Framework for training QE models and using pre-trained models for evaluating MT.
- Supports both word and sentence-level Quality estimation.
- Implementation of five QE systems in Pytorch: QUETCH [1], NuQE [2, 3], predictor-estimator [4, 5], APE-QE [3], and a stacked ensemble with a linear system [2, 3].
- Easy to use API. Import it as a package in other projects or run from the command line.
- Provides scripts to run pre-trained QE models on data from the WMT 2018 campaign.
- Easy to track and reproduce experiments via yaml configuration files.
Results for the WMT18 Quality Estimation shared task, for word level and sentence level on the test set.
| Model | En-De SMT | En-De NMT | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| MT | gaps | source | r | ⍴ | MT | gaps | source | r | ⍴ | |
| OpenKiwi | 62.70 | 52.14 | 48.88 | 71.08 | 72.70 | 44.77 | 22.89 | 36.53 | 46.72 | 58.51 |
| Wang2018 | 62.46 | 49.99 | -- | 73.97 | 75.43 | 43.61 | -- | -- | 50.12 | 60.49 |
| UNQE | -- | -- | -- | 70.00 | 72.44 | -- | -- | -- | 51.29 | 60.52 |
| deepQUEST | 42.98 | 28.24 | 33.97 | 48.72 | 50.97 | 30.31 | 11.93 | 28.59 | 38.08 | 48.00 |
To install OpenKiwi as a package, simply run
pip install openkiwiYou can now
import kiwiinside your project or run in the command line
kiwiOptionally, if you'd like to take advantage of our MLflow integration, simply install it in the same virtualenv as OpenKiwi:
pip install mlflowDetailed usage examples and instructions can be found in the Full Documentation.
We provide pre-trained models with the corresponding pre-processed datasets and configuration files. You can easily reproduce our numbers in the WMT 2018 word- and sentence-level tasks by following the reproduce instructions in the documentation.
We welcome contributions to improve OpenKiwi. Please refer to CONTRIBUTING.md for quick instructions or to contributing instructions for more detailed instructions on how to set up your development environment.
OpenKiwi is Affero GPL licensed. You can see the details of this license in LICENSE.
If you use OpenKiwi, please cite the following paper: OpenKiwi: An Open Source Framework for Quality Estimation.
@inproceedings{openkiwi,
author = {Fábio Kepler and
Jonay Trénous and
Marcos Treviso and
Miguel Vera and
André F. T. Martins},
title = {Open{K}iwi: An Open Source Framework for Quality Estimation},
year = {2019},
booktitle = {Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics--System Demonstrations},
pages = {117--122},
month = {July},
address = {Florence, Italy},
url = {https://www.aclweb.org/anthology/P19-3020},
organization = {Association for Computational Linguistics},
}