This repository contains the code for the paper "IndicMT Eval: A Dataset to Meta-Evaluate Machine Translation Metrics for Indian Languages" to appear at ACL 2023
We contribute a Multidimensional Quality Metric (MQM) dataset for Indian languages created by taking outputs generated by 7 popular MT systems and asking human annotators to judge the quality of the translations using the MQM style guidelines. Using this rich set of annotated data, we show the performance of 16 metrics of various types on evaluating en-xx translations for 5 Indian languages. We provide an updated metric called Indic-COMET which not only shows stronger correlations with human judgement on Indian languages, but is also more robust to perturbations.
Please find more details of this work in our paper (link coming soon).
The MQM annotated dataset collected with the help of language experts for the 5 Indian lamguages (Hindi, Tamil, Marathi, Malayalam, Gujarati) can be downloaded from here (link coming soon).
An example of an MQM annotation containing the source, reference and the translated output with error spans as demarcated by the annotator looks like the following:
More details regarding the instructions provided and the procedures followed for annotations are present in the paper.
The easiest method to access / view the data is to visit this link More details in data folder
cd data
We load the pretrained encoder and initialize it with either XLM-Roberta, COMET-DA or COME-MQM weights. During training, we divide the model parameters into two groups: the encoder parameters, that include the encoder model and the regressor parameters, that include the parameters from the top feed-forward network. We apply gradual unfreezing and discriminative learning rates, meaning that the encoder model is frozen for one epoch while the feed-forward is optimized with a learning rate. After the first epoch, the entire model is fine-tuned with a different learning rate. Since we are fine-tuning on a small dataset, we make use of early stopping with a patience of 3. The best saved checkpoint is decided using the overall Kendall-tau correlation on the test set. We use the COMET repository for training and our checkpoints are compatible with their setup.
Download the best checkpoint here
| MQM | DA |
|---|---|
| indic-comet-mqm | indic-comet-da |
| hparams.yaml | hparamas.yaml |
We followed the implementation of metrics with the help of the following repositories: For BLEU, METEOR, ROUGE-L, CIDEr, Embedding Averaging, Greedy Matching, and Vector Extrema, we use the implementation provided by Sharma et al. (2017). For chrF++, TER, BERTScore, and BLEURT, we use the repository of Castro Ferreira et al. (2020). For SMS, WMDo, and Mover-Score, we use the implementation provided by Fabbri et al. (2020). For all the remaining task-specific metrics, we use the official codes from the respective papers.
The python file code/evaluate.py runs all of these metrics on the given dataset.
If you find IndicMTEval useful in your research or work, please consider citing our paper.
@article{DBLP:journals/corr/abs-2212-10180,
author = {Ananya B. Sai and
Tanay Dixit and
Vignesh Nagarajan and
Anoop Kunchukuttan and
Pratyush Kumar and
Mitesh M. Khapra and
Raj Dabre},
title = {IndicMT Eval: {A} Dataset to Meta-Evaluate Machine Translation metrics
for Indian Languages},
journal = {CoRR},
volume = {abs/2212.10180},
year = {2022}
}
@article{singh2024good,
title={How Good is Zero-Shot MT Evaluation for Low Resource Indian Languages?},
author={Singh, Anushka and Sai, Ananya B and Dabre, Raj and Puduppully, Ratish and Kunchukuttan, Anoop and Khapra, Mitesh M},
journal={arXiv preprint arXiv:2406.03893},
year={2024}
}