zouharvi's picture
Update README.md
286affc verified
metadata
license: cc
configs:
  - config_name: ort_human
    data_files: ort_human.json
  - config_name: ort_wmt
    data_files: ort_wmt.json
    default: true
task_categories:
  - translation
language:
  - cs
  - en
tags:
  - quality
  - human_translation
  - evaluation
pretty_name: Optimal Reference Translations
size_categories:
  - 1K<n<10K

This is the dataset for two papers: Quality and Quantity of Machine Translation References for Automated Metrics [paper] - effect of reference quality and quantity on automatic metric performance, and Evaluating Optimal Reference Translations [paper] - creation of the data and human aspects of annotation and translation. Please see the original repository for more information and the raw data or contact the authors with any questions.

Please make sure that you have the latest datasets installed:

data_human = load_dataset("zouharvi/optimal-reference-translations", 'ort_human')["train"]
data_wmt = load_dataset("zouharvi/optimal-reference-translations", 'ort_wmt')["train"]

Quality and Quantity of Machine Translation References for Automated Metrics [paper]

Abstract: Automatic machine translation metrics often use human translations to determine the quality system translations. Common wisdom in the field dictates that the human references should be of very high quality. However, there are no cost-benefit analyses that could be used to guide practitioners who plan to collect references for machine translation evaluation. We find that higher-quality references lead to better metric correlations with humans at the segment-level. Having up to 7 references per segment and taking their average helps. Interestingly, the references from vendors of different qualities can be mixed together and improve metric success. Higher quality references, however, cost more to create and we frame this as an optimization problem: given a specific budget, what types of references should be collected to maximize metric success. These findings can be used by evaluators of shared tasks when references need to be created under a certain budget.

Cite this paper as:

@misc{zouhar2024quality,
      title={Quality and Quantity of Machine Translation References for Automated Metrics}, 
      author={Vilém Zouhar and Ondřej Bojar},
      year={2024},
      eprint={2401.01283},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Evaluating Optimal Reference Translations [paper]

Abstract: The overall translation quality reached by current machine translation (MT) systems for high-resourced language pairs is remarkably good. Standard methods of evaluation are not suitable nor intended to uncover the many translation errors and quality deficiencies that still persist. Furthermore, the quality of standard reference translations is commonly questioned and comparable quality levels have been reached by MT alone in several language pairs. Navigating further research in these high-resource settings is thus difficult. In this article, we propose a methodology for creating more reliable document-level human reference translations, called "optimal reference translations," with the simple aim to raise the bar of what should be deemed "human translation quality." We evaluate the obtained document-level optimal reference translations in comparison with "standard" ones, confirming a significant quality increase and also documenting the relationship between evaluation and translation editing.

This is project at ETH Zürich and ÚFAL Charles University. Paper to be published in Natural Language Engineering 2024. For now cite as:

@misc{zouhar2023evaluating,
      title={Evaluating Optimal Reference Translations}, 
      author={Vilém Zouhar and Věra Kloudová and Martin Popel and Ondřej Bojar},
      year={2023},
      eprint={2311.16787},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Collected English to Czech translation evaluation human data are in data/ort_human.json. The rest of this repository contains data preparation and evaluation code. Our data is based on WMT2020 data and can thus be also used to e.g. evaluate the quality of various translations as references. The process of the data is as follows:

  1. R1, R2, and R3 are independent translations from English to Czech. R4 is an expert translation by a translatologist.
  2. All the human translations are evaluated on document and segment level with detail (in data/ort_human.json) by different types of human annotators (laypeople, translatology students, professional translators). If the translation is not perfect, the annotators provide a post-edited version for which they would assign the highest grade (6).

Note: If you you also want to use the WMT2020 system submissions, please contact Vilém Zouhar. The code is here, just not pretty yet. 🙂

Example usage

from datasets import load_dataset
data = load_dataset("zouharvi/optimal-reference-translations", 'ort_human')["train"]

# 220 annotated documents
len(data)

# 1760 annotated source lines
sum([len(doc["lines"]) for doc in data])

# 7040 annotated translations
sum([sum([len(line["translations"]) for line in doc["lines"]]) for doc in data])

# 11 annotators
len(set(doc["uid"] for doc in data))

import numpy as np
# Average document-level for R4: 5.865
np.average([doc["rating"]["4"]["overall"] for doc in data])

# Average document-level for R3: 4.810
np.average([doc["rating"]["3"]["overall"] for doc in data])

Data structure

Beginning of ort_wmt (human evaluation of multiple WMT systems):

[
  {
    "src": "The government has compulsorily retired 15 more tax officers in the fourth tranche of its crackdown on errant officials accused of corruption and other malpractices.",
    "systems": {
      "Online-Z.1630": {
        "tgt": "Vláda povinně vysloužila 15 dalších daňových úředníků ve čtvrtém tranši svého zásahu proti potulným úředníkům obviněným z korupce a dalších zanedbaných praktik.",
        "score": 0.11005665393442401
      },
      "SRPOL.522": {
        "tgt": "Vláda povinně odvolala dalších 15 daňových úředníků ve čtvrté části svého zásahu proti chybujícím úředníkům obviněným z korupce a dalších nekalých praktik.",
        "score": -0.13950326931169
      },
      "CUNI-DocTransformer.1450": {
        "tgt": "Vláda poslala do penze dalších 15 daňových úředníků ve čtvrté části svého zátahu proti zatoulaným úředníkům obviněným z korupce a dalších nekalých praktik.",
        "score": 0.768625609971165
      },
      "Online-G.1555": {
        "tgt": "Vláda nuceně poslala do důchodu dalších 15 daňových úředníků ve čtvrté části svého zákroku proti zbloudilým úředníkům obviněným z korupce a jiných pochybení.",
        "score": 0.13307070461983803
      },
      "UEDIN-CUNI.1482": {
        "tgt": "Vláda povinně odvolala dalších 15 daňových úředníků ve čtvrté tranši zátahu proti potulným úředníkům obviněným z korupce a dalších nekalých praktik.",
        "score": 0
      },
      "Online-B.1589": {
        "tgt": "Vláda povinně odešla do důchodu dalších 15 daňových úředníků ve čtvrté tranši svého zákroku proti errancujícím úředníkům obviněným z korupce a jiných nezákonných praktik.",
        "score": -0.961245691263453
      },
      "CUNI-Transformer.1080": {
        "tgt": "Vláda nuceně propustila dalších 15 daňových úředníků ve čtvrté tranši svého zátahu proti chybujícím úředníkům obviněným z korupce a dalších nekalých praktik.",
        "score": 0.33815971885602303
      },
      "Online-A.1573": {
        "tgt": "Vláda povinně odešel 15 více daňových úředníků ve čtvrté tranši svého zákroku proti potulný úředníků obviněných z korupce a dalších nekalých praktik.",
        "score": -0.46360315621142206
      },
      "ref": {
        "tgt": "Vláda přikázala odchod do důchodu dalším 15 daňovým úředníkům v rámci čtvrtého balíčku opatření proti úředníkům obviněným z korupce a dalších nezákonných praktik.",
        "score": 0.282610797572385
      },
      "CUNI-T2T-2018.1071": {
        "tgt": "Vláda ve čtvrté tranši svého zátahu proti pochybným úředníkům obviněným z korupce a dalších nekalých praktik povinně odvolala dalších 15 daňových úředníků.",
        "score": -0.28717921591702694
      },
      "eTranslation.1048": {
        "tgt": "Vláda ve čtvrté tranši svého zákroku proti chybujícím úředníkům obviněným z korupce a dalších nekalých praktik nuceně odvolala dalších 15 daňových úředníků.",
        "score": 0.444703520638052
      },
      "OPPO.1121": {
        "tgt": "Vláda ve čtvrté tranši svého zásahu proti zbloudilým úředníkům obviněným z korupce a dalších nekalých praktik nuceně propustila dalších 15 daňových úředníků.",
        "score": 1.9093879205480302
      },
      "zlabs-nlp.1151": {
        "tgt": "Vláda povinně odešel do důchodu 15 dalších daňových důstojníků ve čtvrtém tranši jeho praskání na vymazané úředníky obviněné z korupce a dalších malpraktices.",
        "score": -1.20533986295174
      }
    },
    "ref": {
      "R1": "Vláda přikázala odchod do důchodu dalším 15 daňovým úředníkům v rámci čtvrtého balíčku opatření proti úředníkům obviněným z korupce a dalších nezákonných praktik.",
      "R1_pe_student_sahara": "Indická vláda přikázala odchod do důchodu dalším patnácti daňovým úředníkům v rámci čtvrtého balíčku opatření proti úředníkům obviněným z korupce a dalších nezákonných praktik.",
      "R2": "Vláda povinně poslala do důchodu dalších 15 daňových úředníků ve čtvrté vlně svého zásahu proti špatným úředníkům obviněným z korupce a dalších profesních pochybení.",
      "R2_pe_student_sahara": "Indická vláda ve čtvrté vlně svého zásahu proti úředníkům obviněným z korupce a dalších profesních pochybení poslala dalších patnáct daňových úředníků do povinného důchodu.",
      "R3": "Vláda odvolala 15 dalších daňových úředníků ve čtvrté tranši stíhání vykyvujících úředníků obviněných z korupce a dalších nezákonných praktik.",
    ...

Beginning of ort_human (human evaluation of multiple human translations):

[
    {
        "uid": "sahara",
        "expertise": "student",
        "doc": "huffingtonpost.com.19385",
        "time": 210.0,                             # self-reported in minutes
        "rating": {
            "2": {                                 # 2 = P2
                "spelling": 4.0,                   # ranges from 0 to 6
                "terminology": 5.5,
                "grammar": 5.5,
                "meaning": 5.0,
                "style": 4.5,
                "pragmatics": 6.0,
                "overall": 4.5
            },
            "4": {                                 # 4 = N1
                "spelling": 6.0,
                "terminology": 6.0,
                "grammar": 6.0,
                "meaning": 5.0,
                "style": 5.0,
                "pragmatics": 6.0,
                "overall": 5.7
            },
            "1": {                                 # 1 = P1
                "spelling": 6.0,
                "terminology": 5.9,
                "grammar": 5.4,
                "meaning": 4.7,
                "style": 4.6,
                "pragmatics": 5.8,
                "overall": 5.0
            },
            "3": {                                 # 3 = P3
                "spelling": 4.5,
                "terminology": 4.7,
                "grammar": 5.0,
                "meaning": 4.5,
                "style": 5.0,
                "pragmatics": 6.0,
                "overall": 4.6
            }
        },
        "lines": [
            {
                "source": "Sony, Disney Back To Work On Third Spider-Man Film",               # source sentence
                "comment": null,
                "translations": {
                    "2": {
                        "orig": "Sony a Disney opět pracují na třetím filmu o Spider-Manovi", # original translation
                        "done": "Sony a Disney pracují na třetím filmu o Spider-Manovi",      # post-edited translation
                        "rating": {
                            "spelling": 6.0,
                            "terminology": 6.0,
                            "grammar": 6.0,
                            "meaning": 5.0,
                            "style": 6.0,
                            "pragmatics": 6.0,
                            "overall": 5.0
                        }
                    },
                    "4": {
                        "orig": "Sony a Disney opět spolupracují na třetím filmu o Spider-Manovi",
                        "done": "Sony a Disney opět spolupracují na třetím filmu o Spider-Manovi",
                        "rating": {
                            "spelling": 6.0,
                            "terminology": 6.0,
                            "grammar": 6.0,
                            "meaning": 6.0,
                            "style": 6.0,
                            "pragmatics": 6.0,
                            "overall": 6.0
                        }
                    },
...