md_gender_bias

Riferimenti:

parole_di-genere

Utilizzare il comando seguente per caricare questo set di dati in TFDS:

ds = tfds.load('huggingface:md_gender_bias/gendered_words')
  • Descrizione :
Machine learning models are trained to find patterns in data.
NLP models can inadvertently learn socially undesirable patterns when training on gender biased text.
In this work, we propose a general framework that decomposes gender bias in text along several pragmatic and semantic dimensions:
bias from the gender of the person being spoken about, bias from the gender of the person being spoken to, and bias from the gender of the speaker.
Using this fine-grained framework, we automatically annotate eight large scale datasets with gender information.
In addition, we collect a novel, crowdsourced evaluation benchmark of utterance-level gender rewrites.
Distinguishing between gender bias along multiple dimensions is important, as it enables us to train finer-grained gender bias classifiers.
We show our classifiers prove valuable for a variety of important applications, such as controlling for gender bias in generative models,
detecting gender bias in arbitrary text, and shed light on offensive language in terms of genderedness.
  • Licenza : licenza MIT
  • Versione : 1.0.0
  • Divide :
Diviso Esempi
'train' 222
  • Caratteristiche :
{
    "word_masculine": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "word_feminine": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

nome_generi

Utilizzare il comando seguente per caricare questo set di dati in TFDS:

ds = tfds.load('huggingface:md_gender_bias/name_genders')
  • Descrizione :
Machine learning models are trained to find patterns in data.
NLP models can inadvertently learn socially undesirable patterns when training on gender biased text.
In this work, we propose a general framework that decomposes gender bias in text along several pragmatic and semantic dimensions:
bias from the gender of the person being spoken about, bias from the gender of the person being spoken to, and bias from the gender of the speaker.
Using this fine-grained framework, we automatically annotate eight large scale datasets with gender information.
In addition, we collect a novel, crowdsourced evaluation benchmark of utterance-level gender rewrites.
Distinguishing between gender bias along multiple dimensions is important, as it enables us to train finer-grained gender bias classifiers.
We show our classifiers prove valuable for a variety of important applications, such as controlling for gender bias in generative models,
detecting gender bias in arbitrary text, and shed light on offensive language in terms of genderedness.
  • Licenza : licenza MIT
  • Versione : 1.0.0
  • Divide :
Diviso Esempi
'yob1880' 2000
'yob1881' 1935
'yob1882' 2127
'yob1883' 2084
'yob1884' 2297
'yob1885' 2294
'yob1886' 2392
'yob1887' 2373
'yob1888' 2651
'yob1889' 2590
'yob1890' 2695
'yob1891' 2660
'yob1892' 2921
'yob1893' 2831
'yob1894' 2941
'yob1895' 3049
'yob1896' 3091
'yob1897' 3028
'yob1898' 3264
'yob1899' 3042
'yob1900' 3730
'yob1901' 3153
'yob1902' 3362
'yob1903' 3389
'yob1904' 3560
'yob1905' 3655
'yob1906' 3633
'yob1907' 3948
'yob1908' 4018
'yob1909' 4227
'yob1910' 4629
'yob1911' 4867
'yob1912' 6351
'yob1913' 6968
'yob1914' 7965
'yob1915' 9357
'yob1916' 9696
'yob1917' 9913
'yob1918' 10398
'yob1919' 10369
'yob1920' 10756
'yob1921' 10857
'yob1922' 10756
'yob1923' 10643
'yob1924' 10869
'yob1925' 10638
'yob1926' 10458
'yob1927' 10406
'yob1928' 10159
'yob1929' 9820
'yob1930' 9791
'yob1931' 9298
'yob1932' 9381
'yob1933' 9013
'yob1934' 9180
'yob1935' 9037
'yob1936' 8894
'yob1937' 8946
'yob1938' 9032
'yob1939' 8918
'yob1940' 8961
'yob1941' 9085
'yob1942' 9425
'yob1943' 9408
'yob1944' 9152
'yob1945' 9025
'yob1946' 9705
'yob1947' 10371
'yob1948' 10241
'yob1949' 10269
'yob1950' 10303
'yob1951' 10462
'yob1952' 10646
'yob1953' 10837
'yob1954' 10968
'yob1955' 11115
'yob1956' 11340
'yob1957' 11564
'yob1958' 11522
'yob1959' 11767
'yob1960' 11921
'yob1961' 12182
'yob1962' 12209
'yob1963' 12282
'yob1964' 12397
'yob1965' 11952
'yob1966' 12151
'yob1967' 12397
'yob1968' 12936
'yob1969' 13749
'yob1970' 14779
'yob1971' 15295
'yob1972' 15412
'yob1973' 15682
'yob1974' 16249
'yob1975' 16944
'yob1976' 17391
'yob1977' 18175
'yob1978' 18231
'yob1979' 19039
'yob1980' 19452
'yob1981' 19475
'yob1982' 19694
'yob1983' 19407
'yob1984' 19506
'yob1985' 20085
'yob1986' 20657
'yob1987' 21406
'yob1988' 22367
'yob1989' 23775
'yob1990' 24716
'yob1991' 25109
'yob1992' 25427
'yob1993' 25966
'yob1994' 25997
'yob1995' 26080
'yob1996' 26423
'yob1997' 26970
'yob1998' 27902
'yob1999' 28552
'yob2000' 29772
'yob2001' 30274
'yob2002' 30564
'yob2003' 31185
'yob2004' 32048
'yob2005' 32549
'yob2006' 34088
'yob2007' 34961
'yob2008' 35079
'yob2009' 34709
'yob2010' 34073
'yob2011' 33908
'yob2012' 33747
'yob2013' 33282
'yob2014' 33243
'yob2015' 33121
'yob2016' 33010
'yob2017' 32590
'yob2018' 32033
  • Caratteristiche :
{
    "name": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "assigned_gender": {
        "num_classes": 2,
        "names": [
            "M",
            "F"
        ],
        "names_file": null,
        "id": null,
        "_type": "ClassLabel"
    },
    "count": {
        "dtype": "int32",
        "id": null,
        "_type": "Value"
    }
}

nuovi_dati

Utilizzare il comando seguente per caricare questo set di dati in TFDS:

ds = tfds.load('huggingface:md_gender_bias/new_data')
  • Descrizione :
Machine learning models are trained to find patterns in data.
NLP models can inadvertently learn socially undesirable patterns when training on gender biased text.
In this work, we propose a general framework that decomposes gender bias in text along several pragmatic and semantic dimensions:
bias from the gender of the person being spoken about, bias from the gender of the person being spoken to, and bias from the gender of the speaker.
Using this fine-grained framework, we automatically annotate eight large scale datasets with gender information.
In addition, we collect a novel, crowdsourced evaluation benchmark of utterance-level gender rewrites.
Distinguishing between gender bias along multiple dimensions is important, as it enables us to train finer-grained gender bias classifiers.
We show our classifiers prove valuable for a variety of important applications, such as controlling for gender bias in generative models,
detecting gender bias in arbitrary text, and shed light on offensive language in terms of genderedness.
  • Licenza : licenza MIT
  • Versione : 1.0.0
  • Divide :
Diviso Esempi
'train' 2345
  • Caratteristiche :
{
    "text": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "original": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "labels": [
        {
            "num_classes": 6,
            "names": [
                "ABOUT:female",
                "ABOUT:male",
                "PARTNER:female",
                "PARTNER:male",
                "SELF:female",
                "SELF:male"
            ],
            "names_file": null,
            "id": null,
            "_type": "ClassLabel"
        }
    ],
    "class_type": {
        "num_classes": 3,
        "names": [
            "about",
            "partner",
            "self"
        ],
        "names_file": null,
        "id": null,
        "_type": "ClassLabel"
    },
    "turker_gender": {
        "num_classes": 5,
        "names": [
            "man",
            "woman",
            "nonbinary",
            "prefer not to say",
            "no answer"
        ],
        "names_file": null,
        "id": null,
        "_type": "ClassLabel"
    },
    "episode_done": {
        "dtype": "bool_",
        "id": null,
        "_type": "Value"
    },
    "confidence": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

funpedia

Utilizzare il comando seguente per caricare questo set di dati in TFDS:

ds = tfds.load('huggingface:md_gender_bias/funpedia')
  • Descrizione :
Machine learning models are trained to find patterns in data.
NLP models can inadvertently learn socially undesirable patterns when training on gender biased text.
In this work, we propose a general framework that decomposes gender bias in text along several pragmatic and semantic dimensions:
bias from the gender of the person being spoken about, bias from the gender of the person being spoken to, and bias from the gender of the speaker.
Using this fine-grained framework, we automatically annotate eight large scale datasets with gender information.
In addition, we collect a novel, crowdsourced evaluation benchmark of utterance-level gender rewrites.
Distinguishing between gender bias along multiple dimensions is important, as it enables us to train finer-grained gender bias classifiers.
We show our classifiers prove valuable for a variety of important applications, such as controlling for gender bias in generative models,
detecting gender bias in arbitrary text, and shed light on offensive language in terms of genderedness.
  • Licenza : licenza MIT
  • Versione : 1.0.0
  • Divide :
Diviso Esempi
'test' 2938
'train' 23897
'validation' 2984
  • Caratteristiche :
{
    "text": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "title": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "persona": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "gender": {
        "num_classes": 3,
        "names": [
            "gender-neutral",
            "female",
            "male"
        ],
        "names_file": null,
        "id": null,
        "_type": "ClassLabel"
    }
}

immagine_chat

Utilizzare il comando seguente per caricare questo set di dati in TFDS:

ds = tfds.load('huggingface:md_gender_bias/image_chat')
  • Descrizione :
Machine learning models are trained to find patterns in data.
NLP models can inadvertently learn socially undesirable patterns when training on gender biased text.
In this work, we propose a general framework that decomposes gender bias in text along several pragmatic and semantic dimensions:
bias from the gender of the person being spoken about, bias from the gender of the person being spoken to, and bias from the gender of the speaker.
Using this fine-grained framework, we automatically annotate eight large scale datasets with gender information.
In addition, we collect a novel, crowdsourced evaluation benchmark of utterance-level gender rewrites.
Distinguishing between gender bias along multiple dimensions is important, as it enables us to train finer-grained gender bias classifiers.
We show our classifiers prove valuable for a variety of important applications, such as controlling for gender bias in generative models,
detecting gender bias in arbitrary text, and shed light on offensive language in terms of genderedness.
  • Licenza : licenza MIT
  • Versione : 1.0.0
  • Divide :
Diviso Esempi
'test' 5000
'train' 9997
'validation' 338180
  • Caratteristiche :
{
    "caption": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "id": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "male": {
        "dtype": "bool_",
        "id": null,
        "_type": "Value"
    },
    "female": {
        "dtype": "bool_",
        "id": null,
        "_type": "Value"
    }
}

procedura guidata

Utilizzare il comando seguente per caricare questo set di dati in TFDS:

ds = tfds.load('huggingface:md_gender_bias/wizard')
  • Descrizione :
Machine learning models are trained to find patterns in data.
NLP models can inadvertently learn socially undesirable patterns when training on gender biased text.
In this work, we propose a general framework that decomposes gender bias in text along several pragmatic and semantic dimensions:
bias from the gender of the person being spoken about, bias from the gender of the person being spoken to, and bias from the gender of the speaker.
Using this fine-grained framework, we automatically annotate eight large scale datasets with gender information.
In addition, we collect a novel, crowdsourced evaluation benchmark of utterance-level gender rewrites.
Distinguishing between gender bias along multiple dimensions is important, as it enables us to train finer-grained gender bias classifiers.
We show our classifiers prove valuable for a variety of important applications, such as controlling for gender bias in generative models,
detecting gender bias in arbitrary text, and shed light on offensive language in terms of genderedness.
  • Licenza : licenza MIT
  • Versione : 1.0.0
  • Divide :
Diviso Esempi
'test' 470
'train' 10449
'validation' 537
  • Caratteristiche :
{
    "text": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "chosen_topic": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "gender": {
        "num_classes": 3,
        "names": [
            "gender-neutral",
            "female",
            "male"
        ],
        "names_file": null,
        "id": null,
        "_type": "ClassLabel"
    }
}

convai2_dedotto

Utilizzare il comando seguente per caricare questo set di dati in TFDS:

ds = tfds.load('huggingface:md_gender_bias/convai2_inferred')
  • Descrizione :
Machine learning models are trained to find patterns in data.
NLP models can inadvertently learn socially undesirable patterns when training on gender biased text.
In this work, we propose a general framework that decomposes gender bias in text along several pragmatic and semantic dimensions:
bias from the gender of the person being spoken about, bias from the gender of the person being spoken to, and bias from the gender of the speaker.
Using this fine-grained framework, we automatically annotate eight large scale datasets with gender information.
In addition, we collect a novel, crowdsourced evaluation benchmark of utterance-level gender rewrites.
Distinguishing between gender bias along multiple dimensions is important, as it enables us to train finer-grained gender bias classifiers.
We show our classifiers prove valuable for a variety of important applications, such as controlling for gender bias in generative models,
detecting gender bias in arbitrary text, and shed light on offensive language in terms of genderedness.
  • Licenza : licenza MIT
  • Versione : 1.0.0
  • Divide :
Diviso Esempi
'test' 7801
'train' 131438
'validation' 7801
  • Caratteristiche :
{
    "text": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "binary_label": {
        "num_classes": 2,
        "names": [
            "ABOUT:female",
            "ABOUT:male"
        ],
        "names_file": null,
        "id": null,
        "_type": "ClassLabel"
    },
    "binary_score": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "ternary_label": {
        "num_classes": 3,
        "names": [
            "ABOUT:female",
            "ABOUT:male",
            "ABOUT:gender-neutral"
        ],
        "names_file": null,
        "id": null,
        "_type": "ClassLabel"
    },
    "ternary_score": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    }
}

light_inferited

Utilizzare il comando seguente per caricare questo set di dati in TFDS:

ds = tfds.load('huggingface:md_gender_bias/light_inferred')
  • Descrizione :
Machine learning models are trained to find patterns in data.
NLP models can inadvertently learn socially undesirable patterns when training on gender biased text.
In this work, we propose a general framework that decomposes gender bias in text along several pragmatic and semantic dimensions:
bias from the gender of the person being spoken about, bias from the gender of the person being spoken to, and bias from the gender of the speaker.
Using this fine-grained framework, we automatically annotate eight large scale datasets with gender information.
In addition, we collect a novel, crowdsourced evaluation benchmark of utterance-level gender rewrites.
Distinguishing between gender bias along multiple dimensions is important, as it enables us to train finer-grained gender bias classifiers.
We show our classifiers prove valuable for a variety of important applications, such as controlling for gender bias in generative models,
detecting gender bias in arbitrary text, and shed light on offensive language in terms of genderedness.
  • Licenza : licenza MIT
  • Versione : 1.0.0
  • Divide :
Diviso Esempi
'test' 12765
'train' 106122
'validation' 6362
  • Caratteristiche :
{
    "text": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "binary_label": {
        "num_classes": 2,
        "names": [
            "ABOUT:female",
            "ABOUT:male"
        ],
        "names_file": null,
        "id": null,
        "_type": "ClassLabel"
    },
    "binary_score": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "ternary_label": {
        "num_classes": 3,
        "names": [
            "ABOUT:female",
            "ABOUT:male",
            "ABOUT:gender-neutral"
        ],
        "names_file": null,
        "id": null,
        "_type": "ClassLabel"
    },
    "ternary_score": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    }
}

opensubtitles_inferred

Utilizzare il comando seguente per caricare questo set di dati in TFDS:

ds = tfds.load('huggingface:md_gender_bias/opensubtitles_inferred')
  • Descrizione :
Machine learning models are trained to find patterns in data.
NLP models can inadvertently learn socially undesirable patterns when training on gender biased text.
In this work, we propose a general framework that decomposes gender bias in text along several pragmatic and semantic dimensions:
bias from the gender of the person being spoken about, bias from the gender of the person being spoken to, and bias from the gender of the speaker.
Using this fine-grained framework, we automatically annotate eight large scale datasets with gender information.
In addition, we collect a novel, crowdsourced evaluation benchmark of utterance-level gender rewrites.
Distinguishing between gender bias along multiple dimensions is important, as it enables us to train finer-grained gender bias classifiers.
We show our classifiers prove valuable for a variety of important applications, such as controlling for gender bias in generative models,
detecting gender bias in arbitrary text, and shed light on offensive language in terms of genderedness.
  • Licenza : licenza MIT
  • Versione : 1.0.0
  • Divide :
Diviso Esempi
'test' 49108
'train' 351036
'validation' 41957
  • Caratteristiche :
{
    "text": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "binary_label": {
        "num_classes": 2,
        "names": [
            "ABOUT:female",
            "ABOUT:male"
        ],
        "names_file": null,
        "id": null,
        "_type": "ClassLabel"
    },
    "binary_score": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    },
    "ternary_label": {
        "num_classes": 3,
        "names": [
            "ABOUT:female",
            "ABOUT:male",
            "ABOUT:gender-neutral"
        ],
        "names_file": null,
        "id": null,
        "_type": "ClassLabel"
    },
    "ternary_score": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    }
}

yelp_inferred

Utilizzare il comando seguente per caricare questo set di dati in TFDS:

ds = tfds.load('huggingface:md_gender_bias/yelp_inferred')
  • Descrizione :
Machine learning models are trained to find patterns in data.
NLP models can inadvertently learn socially undesirable patterns when training on gender biased text.
In this work, we propose a general framework that decomposes gender bias in text along several pragmatic and semantic dimensions:
bias from the gender of the person being spoken about, bias from the gender of the person being spoken to, and bias from the gender of the speaker.
Using this fine-grained framework, we automatically annotate eight large scale datasets with gender information.
In addition, we collect a novel, crowdsourced evaluation benchmark of utterance-level gender rewrites.
Distinguishing between gender bias along multiple dimensions is important, as it enables us to train finer-grained gender bias classifiers.
We show our classifiers prove valuable for a variety of important applications, such as controlling for gender bias in generative models,
detecting gender bias in arbitrary text, and shed light on offensive language in terms of genderedness.
  • Licenza : licenza MIT
  • Versione : 1.0.0
  • Divide :
Diviso Esempi
'test' 534460
'train' 2577862
'validation' 4492
  • Caratteristiche :
{
    "text": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "binary_label": {
        "num_classes": 2,
        "names": [
            "ABOUT:female",
            "ABOUT:male"
        ],
        "names_file": null,
        "id": null,
        "_type": "ClassLabel"
    },
    "binary_score": {
        "dtype": "float32",
        "id": null,
        "_type": "Value"
    }
}