RicardoRei commited on
Commit
6a868f4
1 Parent(s): 9c1c871

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +6 -4
README.md CHANGED
@@ -15,7 +15,9 @@ size_categories:
15
 
16
  # Dataset Summary
17
 
18
- This dataset contains all MQM human annotations from previous [WMT Metrics shared tasks](https://wmt-metrics-task.github.io/) and the MQM annotations from [Experts, Errors, and Context](https://aclanthology.org/2021.tacl-1.87/) in a form of error spans.
 
 
19
 
20
  The data is organised into 8 columns:
21
 
@@ -25,9 +27,8 @@ The data is organised into 8 columns:
25
  - annotations: List of error spans (dictionaries with 'start', 'end', 'severity', 'text')
26
  - lp: language pair
27
 
28
- **Note that this is not an official release of the data** and the original data can be found [here](https://github.com/google/wmt-mqm-human-evaluation).
29
-
30
- Also, while `en-ru` was annotated by Unbabel, `en-de` and `zh-en` was annotated by Google. This means that for en-de and zh-en you will only find minor and major errors while for en-ru you can find a few critical errors.
31
 
32
  ## Python usage:
33
 
@@ -49,3 +50,4 @@ If you use this data please cite the following works:
49
  - [Experts, Errors, and Context: A Large-Scale Study of Human Evaluation for Machine Translation](https://aclanthology.org/2021.tacl-1.87/)
50
  - [Results of the WMT21 Metrics Shared Task: Evaluating Metrics with Expert-based Human Evaluations on TED and News Domain](https://aclanthology.org/2021.wmt-1.73/)
51
  - [Results of WMT22 Metrics Shared Task: Stop Using BLEU – Neural Metrics Are Better and More Robust](https://aclanthology.org/2022.wmt-1.2/)
 
 
15
 
16
  # Dataset Summary
17
 
18
+ This dataset contains all MQM human annotations from previous [WMT Metrics shared tasks](https://wmt-metrics-task.github.io/) and the MQM annotations from [Experts, Errors, and Context](https://aclanthology.org/2021.tacl-1.87/) in a form of error spans. Moreover, it contains some hallucinations used in the training of [XCOMET models](https://huggingface.co/Unbabel/XCOMET-XXL).
19
+
20
+ ** Please note that this is not an official release of the data** and the original data can be found [here](https://github.com/google/wmt-mqm-human-evaluation).
21
 
22
  The data is organised into 8 columns:
23
 
 
27
  - annotations: List of error spans (dictionaries with 'start', 'end', 'severity', 'text')
28
  - lp: language pair
29
 
30
+
31
+ While `en-ru` was annotated by Unbabel, `en-de` and `zh-en` was annotated by Google. This means that for en-de and zh-en you will only find minor and major errors while for en-ru you can find a few critical errors.
 
32
 
33
  ## Python usage:
34
 
 
50
  - [Experts, Errors, and Context: A Large-Scale Study of Human Evaluation for Machine Translation](https://aclanthology.org/2021.tacl-1.87/)
51
  - [Results of the WMT21 Metrics Shared Task: Evaluating Metrics with Expert-based Human Evaluations on TED and News Domain](https://aclanthology.org/2021.wmt-1.73/)
52
  - [Results of WMT22 Metrics Shared Task: Stop Using BLEU – Neural Metrics Are Better and More Robust](https://aclanthology.org/2022.wmt-1.2/)
53
+ - [xCOMET: Transparent Machine Translation Evaluation through Fine-grained Error Detection](https://arxiv.org/pdf/2310.10482.pdf)