Attend the Women in ML Symposium on December 7 Register now

md_gender_bias

References:

gendered_words

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:md_gender_bias/gendered_words')
  • Description:
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.
  • License: MIT License
  • Version: 1.0.0
  • Splits:
Split Examples
'train' 222
  • Features:
{
    "word_masculine": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    },
    "word_feminine": {
        "dtype": "string",
        "id": null,
        "_type": "Value"
    }
}

name_genders

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:md_gender_bias/name_genders')
  • Description:
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.
  • License: MIT License
  • Version: 1.0.0
  • Splits:
Split Examples
'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
  • Features:
{
    "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"
    }
}

new_data

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:md_gender_bias/new_data')
  • Description:
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.
  • License: MIT License
  • Version: 1.0.0
  • Splits:
Split Examples
'train' 2345
  • Features:
{
    "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

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:md_gender_bias/funpedia')
  • Description:
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.
  • License: MIT License
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 2938
'train' 23897
'validation' 2984
  • Features:
{
    "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"
    }
}

image_chat

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:md_gender_bias/image_chat')
  • Description:
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.
  • License: MIT License
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 5000
'train' 9997
'validation' 338180
  • Features:
{
    "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"
    }
}

wizard

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:md_gender_bias/wizard')
  • Description:
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.
  • License: MIT License
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 470
'train' 10449
'validation' 537
  • Features:
{
    "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_inferred

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:md_gender_bias/convai2_inferred')
  • Description:
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.
  • License: MIT License
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 7801
'train' 131438
'validation' 7801
  • Features:
{
    "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_inferred

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:md_gender_bias/light_inferred')
  • Description:
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.
  • License: MIT License
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 12765
'train' 106122
'validation' 6362
  • Features:
{
    "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

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:md_gender_bias/opensubtitles_inferred')
  • Description:
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.
  • License: MIT License
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 49108
'train' 351036
'validation' 41957
  • Features:
{
    "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

Use the following command to load this dataset in TFDS:

ds = tfds.load('huggingface:md_gender_bias/yelp_inferred')
  • Description:
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.
  • License: MIT License
  • Version: 1.0.0
  • Splits:
Split Examples
'test' 534460
'train' 2577862
'validation' 4492
  • Features:
{
    "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"
    }
}