UPF-Cobalt is a MT evaluation metric that exploits alignment and syntactic context to assess MT quality.
- Install Stanford Parser (http://nlp.stanford.edu/software/corenlp.shtml)
- Download dependency-based word vectors from https://levyomer.wordpress.com/2014/04/25/dependency-based-word-embeddings/ (optional)
- Download upf-cobalt
git clone https://github.com/mfomicheva/upf-cobalt.git
To use the metric, run evaluate.py with the following parameters:
Parameter | Description |
---|---|
-r | parsed reference file |
-t | parsed system output file |
-v | file containing word vectors - optional |
-a | output alignments - optional |
-o | specify output directory - optional (default is "./Data") |
If no parameters are specified the metric will process the example files stored in Data folder.
@inproceedings{fomicheva2016cobaltf,
title={CobaltF: a fluent metric for MT evaluation},
author={Fomicheva, Marina and Bel Rafecas, N{\'u}ria and Specia, Lucia and da Cunha Fanego, Iria and Malinovsiy, Anton},
booktitle={The 54th Annual Meeting of the Association for Computational Linguistics. Proceedings of the First Conference on Machine Translation (WMT); 2016 Aug 7-12; Berlin, Germany},
year={2016},
pages={483--490},
organization={ACL (Association for Computational Linguistics)}
}