wmt20_mlqe_task1

مراجع:

أون دي

استخدم الأمر التالي لتحميل مجموعة البيانات هذه في TFDS:

ds = tfds.load('huggingface:wmt20_mlqe_task1/en-de')
  • وصف :
This shared task (part of WMT20) will build on its previous editions
to further examine automatic methods for estimating the quality
of neural machine translation output at run-time, without relying
on reference translations. As in previous years, we cover estimation
at various levels. Important elements introduced this year include: a new
task where sentences are annotated with Direct Assessment (DA)
scores instead of labels based on post-editing; a new multilingual
sentence-level dataset mainly from Wikipedia articles, where the
source articles can be retrieved for document-wide context; the
availability of NMT models to explore system-internal information for the task.

Task 1 uses Wikipedia data for 6 language pairs that includes high-resource
English--German (En-De) and English--Chinese (En-Zh), medium-resource
Romanian--English (Ro-En) and Estonian--English (Et-En), and low-resource
Sinhalese--English (Si-En) and Nepalese--English (Ne-En), as well as a
dataset with a combination of Wikipedia articles and Reddit articles
for Russian-English (En-Ru). The datasets were collected by translating
sentences sampled from source language articles using state-of-the-art NMT
models built using the fairseq toolkit and annotated with Direct Assessment (DA)
scores by professional translators. Each sentence was annotated following the
FLORES setup, which presents a form of DA, where at least three professional
translators rate each sentence from 0-100 according to the perceived translation
quality. DA scores are standardised using the z-score by rater. Participating systems
are required to score sentences according to z-standardised DA scores.
  • الترخيص : غير معروف
  • الإصدار : 1.1.0
  • الإنشقاقات :
ينقسم أمثلة
'test' 1000
'train' 7000
'validation' 1000
  • سمات :
{
    "segid": {
        "dtype": "int32",
        "id": null,
        "_type": "Value"
    },
    "translation": {
        "languages": [
            "en",
            "de"
        ],
        "id": null,
        "_type": "Translation"
    },
    "scores": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "mean": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "z_scores": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "z_mean": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "model_score": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "doc_id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "nmt_output": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "word_probas": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

أون ز

استخدم الأمر التالي لتحميل مجموعة البيانات هذه في TFDS:

ds = tfds.load('huggingface:wmt20_mlqe_task1/en-zh')
  • وصف :
This shared task (part of WMT20) will build on its previous editions
to further examine automatic methods for estimating the quality
of neural machine translation output at run-time, without relying
on reference translations. As in previous years, we cover estimation
at various levels. Important elements introduced this year include: a new
task where sentences are annotated with Direct Assessment (DA)
scores instead of labels based on post-editing; a new multilingual
sentence-level dataset mainly from Wikipedia articles, where the
source articles can be retrieved for document-wide context; the
availability of NMT models to explore system-internal information for the task.

Task 1 uses Wikipedia data for 6 language pairs that includes high-resource
English--German (En-De) and English--Chinese (En-Zh), medium-resource
Romanian--English (Ro-En) and Estonian--English (Et-En), and low-resource
Sinhalese--English (Si-En) and Nepalese--English (Ne-En), as well as a
dataset with a combination of Wikipedia articles and Reddit articles
for Russian-English (En-Ru). The datasets were collected by translating
sentences sampled from source language articles using state-of-the-art NMT
models built using the fairseq toolkit and annotated with Direct Assessment (DA)
scores by professional translators. Each sentence was annotated following the
FLORES setup, which presents a form of DA, where at least three professional
translators rate each sentence from 0-100 according to the perceived translation
quality. DA scores are standardised using the z-score by rater. Participating systems
are required to score sentences according to z-standardised DA scores.
  • الترخيص : غير معروف
  • الإصدار : 1.1.0
  • الإنشقاقات :
ينقسم أمثلة
'test' 1000
'train' 7000
'validation' 1000
  • سمات :
{
    "segid": {
        "dtype": "int32",
        "id": null,
        "_type": "Value"
    },
    "translation": {
        "languages": [
            "en",
            "zh"
        ],
        "id": null,
        "_type": "Translation"
    },
    "scores": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "mean": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "z_scores": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "z_mean": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "model_score": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "doc_id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "nmt_output": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "word_probas": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

وآخرون أون

استخدم الأمر التالي لتحميل مجموعة البيانات هذه في TFDS:

ds = tfds.load('huggingface:wmt20_mlqe_task1/et-en')
  • وصف :
This shared task (part of WMT20) will build on its previous editions
to further examine automatic methods for estimating the quality
of neural machine translation output at run-time, without relying
on reference translations. As in previous years, we cover estimation
at various levels. Important elements introduced this year include: a new
task where sentences are annotated with Direct Assessment (DA)
scores instead of labels based on post-editing; a new multilingual
sentence-level dataset mainly from Wikipedia articles, where the
source articles can be retrieved for document-wide context; the
availability of NMT models to explore system-internal information for the task.

Task 1 uses Wikipedia data for 6 language pairs that includes high-resource
English--German (En-De) and English--Chinese (En-Zh), medium-resource
Romanian--English (Ro-En) and Estonian--English (Et-En), and low-resource
Sinhalese--English (Si-En) and Nepalese--English (Ne-En), as well as a
dataset with a combination of Wikipedia articles and Reddit articles
for Russian-English (En-Ru). The datasets were collected by translating
sentences sampled from source language articles using state-of-the-art NMT
models built using the fairseq toolkit and annotated with Direct Assessment (DA)
scores by professional translators. Each sentence was annotated following the
FLORES setup, which presents a form of DA, where at least three professional
translators rate each sentence from 0-100 according to the perceived translation
quality. DA scores are standardised using the z-score by rater. Participating systems
are required to score sentences according to z-standardised DA scores.
  • الترخيص : غير معروف
  • الإصدار : 1.1.0
  • الإنشقاقات :
ينقسم أمثلة
'test' 1000
'train' 7000
'validation' 1000
  • سمات :
{
    "segid": {
        "dtype": "int32",
        "id": null,
        "_type": "Value"
    },
    "translation": {
        "languages": [
            "et",
            "en"
        ],
        "id": null,
        "_type": "Translation"
    },
    "scores": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "mean": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "z_scores": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "z_mean": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "model_score": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "doc_id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "nmt_output": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "word_probas": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

شمال شرق أون

استخدم الأمر التالي لتحميل مجموعة البيانات هذه في TFDS:

ds = tfds.load('huggingface:wmt20_mlqe_task1/ne-en')
  • وصف :
This shared task (part of WMT20) will build on its previous editions
to further examine automatic methods for estimating the quality
of neural machine translation output at run-time, without relying
on reference translations. As in previous years, we cover estimation
at various levels. Important elements introduced this year include: a new
task where sentences are annotated with Direct Assessment (DA)
scores instead of labels based on post-editing; a new multilingual
sentence-level dataset mainly from Wikipedia articles, where the
source articles can be retrieved for document-wide context; the
availability of NMT models to explore system-internal information for the task.

Task 1 uses Wikipedia data for 6 language pairs that includes high-resource
English--German (En-De) and English--Chinese (En-Zh), medium-resource
Romanian--English (Ro-En) and Estonian--English (Et-En), and low-resource
Sinhalese--English (Si-En) and Nepalese--English (Ne-En), as well as a
dataset with a combination of Wikipedia articles and Reddit articles
for Russian-English (En-Ru). The datasets were collected by translating
sentences sampled from source language articles using state-of-the-art NMT
models built using the fairseq toolkit and annotated with Direct Assessment (DA)
scores by professional translators. Each sentence was annotated following the
FLORES setup, which presents a form of DA, where at least three professional
translators rate each sentence from 0-100 according to the perceived translation
quality. DA scores are standardised using the z-score by rater. Participating systems
are required to score sentences according to z-standardised DA scores.
  • الترخيص : غير معروف
  • الإصدار : 1.1.0
  • الإنشقاقات :
ينقسم أمثلة
'test' 1000
'train' 7000
'validation' 1000
  • سمات :
{
    "segid": {
        "dtype": "int32",
        "id": null,
        "_type": "Value"
    },
    "translation": {
        "languages": [
            "ne",
            "en"
        ],
        "id": null,
        "_type": "Translation"
    },
    "scores": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "mean": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "z_scores": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "z_mean": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "model_score": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "doc_id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "nmt_output": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "word_probas": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

رو-إن

استخدم الأمر التالي لتحميل مجموعة البيانات هذه في TFDS:

ds = tfds.load('huggingface:wmt20_mlqe_task1/ro-en')
  • وصف :
This shared task (part of WMT20) will build on its previous editions
to further examine automatic methods for estimating the quality
of neural machine translation output at run-time, without relying
on reference translations. As in previous years, we cover estimation
at various levels. Important elements introduced this year include: a new
task where sentences are annotated with Direct Assessment (DA)
scores instead of labels based on post-editing; a new multilingual
sentence-level dataset mainly from Wikipedia articles, where the
source articles can be retrieved for document-wide context; the
availability of NMT models to explore system-internal information for the task.

Task 1 uses Wikipedia data for 6 language pairs that includes high-resource
English--German (En-De) and English--Chinese (En-Zh), medium-resource
Romanian--English (Ro-En) and Estonian--English (Et-En), and low-resource
Sinhalese--English (Si-En) and Nepalese--English (Ne-En), as well as a
dataset with a combination of Wikipedia articles and Reddit articles
for Russian-English (En-Ru). The datasets were collected by translating
sentences sampled from source language articles using state-of-the-art NMT
models built using the fairseq toolkit and annotated with Direct Assessment (DA)
scores by professional translators. Each sentence was annotated following the
FLORES setup, which presents a form of DA, where at least three professional
translators rate each sentence from 0-100 according to the perceived translation
quality. DA scores are standardised using the z-score by rater. Participating systems
are required to score sentences according to z-standardised DA scores.
  • الترخيص : غير معروف
  • الإصدار : 1.1.0
  • الإنشقاقات :
ينقسم أمثلة
'test' 1000
'train' 7000
'validation' 1000
  • سمات :
{
    "segid": {
        "dtype": "int32",
        "id": null,
        "_type": "Value"
    },
    "translation": {
        "languages": [
            "ro",
            "en"
        ],
        "id": null,
        "_type": "Translation"
    },
    "scores": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "mean": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "z_scores": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "z_mean": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "model_score": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "doc_id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "nmt_output": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "word_probas": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

سي أون

استخدم الأمر التالي لتحميل مجموعة البيانات هذه في TFDS:

ds = tfds.load('huggingface:wmt20_mlqe_task1/si-en')
  • وصف :
This shared task (part of WMT20) will build on its previous editions
to further examine automatic methods for estimating the quality
of neural machine translation output at run-time, without relying
on reference translations. As in previous years, we cover estimation
at various levels. Important elements introduced this year include: a new
task where sentences are annotated with Direct Assessment (DA)
scores instead of labels based on post-editing; a new multilingual
sentence-level dataset mainly from Wikipedia articles, where the
source articles can be retrieved for document-wide context; the
availability of NMT models to explore system-internal information for the task.

Task 1 uses Wikipedia data for 6 language pairs that includes high-resource
English--German (En-De) and English--Chinese (En-Zh), medium-resource
Romanian--English (Ro-En) and Estonian--English (Et-En), and low-resource
Sinhalese--English (Si-En) and Nepalese--English (Ne-En), as well as a
dataset with a combination of Wikipedia articles and Reddit articles
for Russian-English (En-Ru). The datasets were collected by translating
sentences sampled from source language articles using state-of-the-art NMT
models built using the fairseq toolkit and annotated with Direct Assessment (DA)
scores by professional translators. Each sentence was annotated following the
FLORES setup, which presents a form of DA, where at least three professional
translators rate each sentence from 0-100 according to the perceived translation
quality. DA scores are standardised using the z-score by rater. Participating systems
are required to score sentences according to z-standardised DA scores.
  • الترخيص : غير معروف
  • الإصدار : 1.1.0
  • الإنشقاقات :
ينقسم أمثلة
'test' 1000
'train' 7000
'validation' 1000
  • سمات :
{
    "segid": {
        "dtype": "int32",
        "id": null,
        "_type": "Value"
    },
    "translation": {
        "languages": [
            "si",
            "en"
        ],
        "id": null,
        "_type": "Translation"
    },
    "scores": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "mean": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "z_scores": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "z_mean": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "model_score": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "doc_id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "nmt_output": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "word_probas": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

رو-إن

استخدم الأمر التالي لتحميل مجموعة البيانات هذه في TFDS:

ds = tfds.load('huggingface:wmt20_mlqe_task1/ru-en')
  • وصف :
This shared task (part of WMT20) will build on its previous editions
to further examine automatic methods for estimating the quality
of neural machine translation output at run-time, without relying
on reference translations. As in previous years, we cover estimation
at various levels. Important elements introduced this year include: a new
task where sentences are annotated with Direct Assessment (DA)
scores instead of labels based on post-editing; a new multilingual
sentence-level dataset mainly from Wikipedia articles, where the
source articles can be retrieved for document-wide context; the
availability of NMT models to explore system-internal information for the task.

Task 1 uses Wikipedia data for 6 language pairs that includes high-resource
English--German (En-De) and English--Chinese (En-Zh), medium-resource
Romanian--English (Ro-En) and Estonian--English (Et-En), and low-resource
Sinhalese--English (Si-En) and Nepalese--English (Ne-En), as well as a
dataset with a combination of Wikipedia articles and Reddit articles
for Russian-English (En-Ru). The datasets were collected by translating
sentences sampled from source language articles using state-of-the-art NMT
models built using the fairseq toolkit and annotated with Direct Assessment (DA)
scores by professional translators. Each sentence was annotated following the
FLORES setup, which presents a form of DA, where at least three professional
translators rate each sentence from 0-100 according to the perceived translation
quality. DA scores are standardised using the z-score by rater. Participating systems
are required to score sentences according to z-standardised DA scores.
  • الترخيص : غير معروف
  • الإصدار : 1.1.0
  • الإنشقاقات :
ينقسم أمثلة
'test' 1000
'train' 7000
'validation' 1000
  • سمات :
{
    "segid": {
        "dtype": "int32",
        "id": null,
        "_type": "Value"
    },
    "translation": {
        "languages": [
            "ru",
            "en"
        ],
        "id": null,
        "_type": "Translation"
    },
    "scores": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "mean": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "z_scores": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    },
    "z_mean": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "model_score": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "doc_id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "nmt_output": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "word_probas": {
        "feature": {
            "dtype": "float32",
            "id": null,
            "_type": "Value"
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}