text
stringlengths
2
146
binary_label
class label
2 classes
binary_score
float32
0.5
1
ternary_label
class label
3 classes
ternary_score
float32
0.33
0.99
hi , how are you doing ? i'm getting ready to do some cheetah chasing to stay in shape .
1ABOUT:male
0.6522
2ABOUT:gender-neutral
0.4496
i am ! for my hobby i like to do canning or some whittling .
1ABOUT:male
0.6505
2ABOUT:gender-neutral
0.8413
that's neat . when i was in high school i placed 6th in 100m dash !
1ABOUT:male
0.8615
1ABOUT:male
0.529
i do not . but i do have a favorite meat since that is all i eat exclusively .
0ABOUT:female
0.6487
2ABOUT:gender-neutral
0.5111
i would have to say its prime rib . do you have any favorite foods ?
0ABOUT:female
0.5697
2ABOUT:gender-neutral
0.5346
do you have anything planned for today ? i think i am going to do some canning .
0ABOUT:female
0.7657
2ABOUT:gender-neutral
0.5458
i think i will can some jam . do you also play footfall for fun ?
1ABOUT:male
0.8343
2ABOUT:gender-neutral
0.5547
hi , how are you doing today ?
1ABOUT:male
0.5223
1ABOUT:male
0.362
wow , four sisters . just watching game of thrones .
0ABOUT:female
0.7858
0ABOUT:female
0.6612
i agree . what do you do for a living ?
1ABOUT:male
0.6433
2ABOUT:gender-neutral
0.5219
interesting . i'm a website designer . pretty much spend all my time on the computer .
0ABOUT:female
0.6294
2ABOUT:gender-neutral
0.4271
that's awesome . i have always had a love for technology .
1ABOUT:male
0.8084
2ABOUT:gender-neutral
0.5211
i really enjoy free diving , how about you , have any hobbies ?
1ABOUT:male
0.7435
2ABOUT:gender-neutral
0.595
that's nice . moms are pretty cool too .
0ABOUT:female
0.8143
0ABOUT:female
0.7246
we all live in a yellow submarine , a yellow submarine . morning !
1ABOUT:male
0.5129
2ABOUT:gender-neutral
0.6538
lol . i am shy , anything to break the ice , and i am a beatles fan .
1ABOUT:male
0.5769
0ABOUT:female
0.4154
really ? what shows ? i like tv , it makes me forget i do not like my family
0ABOUT:female
0.5684
0ABOUT:female
0.4639
i did too . i do not get along with mine . they have no class .
0ABOUT:female
0.5439
0ABOUT:female
0.391
put the lime in the coconut as well . . .
0ABOUT:female
0.5324
2ABOUT:gender-neutral
0.7012
i prefer mojitos . watermelon or cucumber .
1ABOUT:male
0.7555
2ABOUT:gender-neutral
0.4616
hi ! i work as a gourmet cook .
0ABOUT:female
0.645
0ABOUT:female
0.5549
really . but , i can sing pitch perfect .
1ABOUT:male
0.657
1ABOUT:male
0.5627
great ! i had won an award for spelling bee .
0ABOUT:female
0.7777
0ABOUT:female
0.3997
okay but i was published in new yorker once
0ABOUT:female
0.626
0ABOUT:female
0.5699
i have not . i can cook any word you want me to
0ABOUT:female
0.6717
0ABOUT:female
0.4296
i'm asian and have no hair .
0ABOUT:female
0.7287
0ABOUT:female
0.6674
i love carrots . i eat carrots like a horse .
0ABOUT:female
0.7075
2ABOUT:gender-neutral
0.6913
i work as a gourmet cook who also has a pitch perfect voice .
0ABOUT:female
0.7664
0ABOUT:female
0.536
how are you doing today
1ABOUT:male
0.5017
0ABOUT:female
0.3829
i like to watch kids
0ABOUT:female
0.5749
0ABOUT:female
0.5675
what do you weld ? houses ?
1ABOUT:male
0.5266
2ABOUT:gender-neutral
0.6765
what is your secret that you have
0ABOUT:female
0.6869
0ABOUT:female
0.3739
how does that feel for you
1ABOUT:male
0.5496
0ABOUT:female
0.3996
i bet that it does
0ABOUT:female
0.508
2ABOUT:gender-neutral
0.3995
i watch kids for a living
0ABOUT:female
0.559
0ABOUT:female
0.5725
hello friend , how is it going
1ABOUT:male
0.5177
1ABOUT:male
0.4051
i'm great enjoying the football season
1ABOUT:male
0.9578
1ABOUT:male
0.7083
you work for a funeral home ?
1ABOUT:male
0.5689
2ABOUT:gender-neutral
0.3811
lol , i can imagine . i'll be reading a lot when football is over
1ABOUT:male
0.8982
1ABOUT:male
0.67
ok i see , that's your halloween costume
0ABOUT:female
0.5795
0ABOUT:female
0.5574
i like anything to do with mystery
0ABOUT:female
0.6198
2ABOUT:gender-neutral
0.3699
lol , oh i see , taught you own it
0ABOUT:female
0.5139
2ABOUT:gender-neutral
0.3603
well i like sherlock holmes and others
1ABOUT:male
0.8187
0ABOUT:female
0.4647
rock on , i'm listening to my favorite band guns and roses .
1ABOUT:male
0.8896
1ABOUT:male
0.494
of course . i love to listen to rock .
1ABOUT:male
0.9076
1ABOUT:male
0.6035
well i'm into black everything . so at least it wouldn't show on my black carpet .
0ABOUT:female
0.5876
2ABOUT:gender-neutral
0.5125
i've a black car , purse , wear all black .
0ABOUT:female
0.579
2ABOUT:gender-neutral
0.424
wow , does he live there or work ?
1ABOUT:male
0.844
1ABOUT:male
0.8106
have you visited him there before ?
1ABOUT:male
0.7982
1ABOUT:male
0.7072
sounds a bit scary . i ve never been .
1ABOUT:male
0.6001
0ABOUT:female
0.3486
hi . how are you doing today ?
1ABOUT:male
0.5364
1ABOUT:male
0.3631
i'm alright . i just got done writing .
1ABOUT:male
0.5033
1ABOUT:male
0.367
it is my living . i like culture .
1ABOUT:male
0.5299
2ABOUT:gender-neutral
0.4439
what are you going to school for
1ABOUT:male
0.5511
1ABOUT:male
0.3771
do you own your own company
0ABOUT:female
0.5228
2ABOUT:gender-neutral
0.3412
do you get free pizza
1ABOUT:male
0.6728
2ABOUT:gender-neutral
0.6281
that is a good start
1ABOUT:male
0.6914
1ABOUT:male
0.5087
i used to party a lot
0ABOUT:female
0.6665
2ABOUT:gender-neutral
0.3879
hi how are you today
1ABOUT:male
0.5082
0ABOUT:female
0.4059
i only worked half a day i work at the bank
1ABOUT:male
0.6216
2ABOUT:gender-neutral
0.4325
what do you do ?
1ABOUT:male
0.603
1ABOUT:male
0.3692
oh ok i'm a teller was the best job i could find with no college
1ABOUT:male
0.5895
2ABOUT:gender-neutral
0.3995
yes it does pay the bills
1ABOUT:male
0.593
1ABOUT:male
0.4167
i hike at the park
1ABOUT:male
0.5198
2ABOUT:gender-neutral
0.5008
i love nature and it keeps my mind off things . do you travel
1ABOUT:male
0.6744
2ABOUT:gender-neutral
0.5514
i want to go to another country at least once . i have never left the states
0ABOUT:female
0.5231
2ABOUT:gender-neutral
0.4825
hello , how are you today ?
1ABOUT:male
0.5005
0ABOUT:female
0.4172
i'm very good . did you watch the football games today ?
1ABOUT:male
0.9355
1ABOUT:male
0.6745
what do you do for a living ? i proofread for hallmark .
0ABOUT:female
0.7396
0ABOUT:female
0.4973
books are my greatest pleasure , i have a nice little library i'm building .
1ABOUT:male
0.5109
2ABOUT:gender-neutral
0.5408
i agree . have you seen goodfellas ?
1ABOUT:male
0.5709
0ABOUT:female
0.3636
i catch the football and hockey games .
1ABOUT:male
0.9474
1ABOUT:male
0.6863
i do . i travel extensively to europe and south america .
1ABOUT:male
0.6272
2ABOUT:gender-neutral
0.4796
ireland is very nice , japan is my next stop .
1ABOUT:male
0.602
2ABOUT:gender-neutral
0.5585
hi i am sally , i live with my sweet dogs in taos , new mexico .
0ABOUT:female
0.8875
0ABOUT:female
0.6257
i've wonderful memories of my mother playing and singing to me .
0ABOUT:female
0.5805
0ABOUT:female
0.6206
maybe you will have to record your playing and dance to it .
1ABOUT:male
0.7146
1ABOUT:male
0.5917
do you have a day job ?
1ABOUT:male
0.6739
2ABOUT:gender-neutral
0.4569
interesting , i'm a struggling artist and part time tutor
1ABOUT:male
0.5942
1ABOUT:male
0.3844
great commitment to your beliefs . i admit i still eat meat only less .
0ABOUT:female
0.5903
2ABOUT:gender-neutral
0.4297
do you have any pets ?
0ABOUT:female
0.727
0ABOUT:female
0.4098
hello . how are you ?
1ABOUT:male
0.5766
0ABOUT:female
0.388
doing good . what are some of your favorite books ?
1ABOUT:male
0.5112
0ABOUT:female
0.4385
i enjoy reading but also garden in my spare time .
1ABOUT:male
0.5171
2ABOUT:gender-neutral
0.4001
i spend a lot of time in my garden .
0ABOUT:female
0.5374
2ABOUT:gender-neutral
0.4816
yes . lots of rabbits .
0ABOUT:female
0.6636
2ABOUT:gender-neutral
0.4296
i give away lots of my vegetables to veterans . i'm a veteran .
0ABOUT:female
0.5
2ABOUT:gender-neutral
0.516
very cool . thanks for the chat . have a great day .
1ABOUT:male
0.5053
1ABOUT:male
0.4087
hey how are you today ?
1ABOUT:male
0.51
0ABOUT:female
0.3969
i have been eating tacos and getting ready to move to school .
1ABOUT:male
0.5136
2ABOUT:gender-neutral
0.4049
yes , i am going to university of michigan . what year are you ?
1ABOUT:male
0.7485
1ABOUT:male
0.4532
i love doing anything outdoors . especially in summer . you ?
1ABOUT:male
0.6135
2ABOUT:gender-neutral
0.3814
fun . have you decided on your major for school ?
1ABOUT:male
0.5866
1ABOUT:male
0.3632
pre med . i'd love to be a doctor
1ABOUT:male
0.5568
1ABOUT:male
0.4002
hey , what are you up to ?
1ABOUT:male
0.585
2ABOUT:gender-neutral
0.3561
i prefer rock music , like led zeppelin .
1ABOUT:male
0.9223
1ABOUT:male
0.7418
i would if i could , but i have a farm to maintain .
1ABOUT:male
0.5588
2ABOUT:gender-neutral
0.5818
i prefer hiking outdoors and photography rather than crowded malls .
0ABOUT:female
0.5433
2ABOUT:gender-neutral
0.5125
work is tiring . i would love to travel the world instead .
1ABOUT:male
0.7027
2ABOUT:gender-neutral
0.4794
staying here is fine too though . my two dogs keep me company .
0ABOUT:female
0.5576
2ABOUT:gender-neutral
0.5131

Dataset Card for Multi-Dimensional Gender Bias Classification

Dataset Summary

The Multi-Dimensional Gender Bias Classification dataset is based on 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. It contains seven large scale datasets automatically annotated for gender information (there are eight in the original project but the Wikipedia set is not included in the HuggingFace distribution), one crowdsourced evaluation benchmark of utterance-level gender rewrites, a list of gendered names, and a list of gendered words in English.

Supported Tasks and Leaderboards

  • text-classification-other-gender-bias: The dataset can be used to train a model for classification of various kinds of gender bias. The model performance is evaluated based on the accuracy of the predicted labels as compared to the given labels in the dataset. Dinan et al's (2020) Transformer model achieved an average of 67.13% accuracy in binary gender prediction across the ABOUT, TO, and AS tasks. See the paper for more results.

Languages

The data is in English as spoken on the various sites where the data was collected. The associated BCP-47 code en.

Dataset Structure

Data Instances

The following are examples of data instances from the various configs in the dataset. See the MD Gender Bias dataset viewer to explore more examples.

An example from the new_data config:

{'class_type': 0, 
 'confidence': 'certain', 
 'episode_done': True, 
 'labels': [1], 
 'original': 'She designed monumental Loviisa war cemetery in 1920', 
 'text': 'He designed monumental Lovissa War Cemetery in 1920.', 
 'turker_gender': 4}

An example from the funpedia config:

{'gender': 2, 
 'persona': 'Humorous', 
 'text': 'Max Landis is a comic book writer who wrote Chronicle, American Ultra, and Victor Frankestein.', 
 'title': 'Max Landis'}

An example from the image_chat config:

{'caption': '<start> a young girl is holding a pink umbrella in her hand <eos>', 
 'female': True, 
 'id': '2923e28b6f588aff2d469ab2cccfac57', 
 'male': False}

An example from the wizard config:

{'chosen_topic': 'Krav Maga', 
 'gender': 2, 
 'text': 'Hello. I hope you might enjoy or know something about Krav Maga?'}

An example from the convai2_inferred config (the other _inferred configs have the same fields, with the exception of yelp_inferred, which does not have the ternary_label or ternary_score fields):

{'binary_label': 1, 
'binary_score': 0.6521999835968018, 
'ternary_label': 2, 
'ternary_score': 0.4496000111103058, 
'text': "hi , how are you doing ? i'm getting ready to do some cheetah chasing to stay in shape ."}

An example from the gendered_words config:

{'word_feminine': 'countrywoman',
 'word_masculine': 'countryman'}

An example from the name_genders config:

{'assigned_gender': 1,
 'count': 7065,
 'name': 'Mary'}

Data Fields

The following are the features for each of the configs.

For the new_data config:

  • text: the text to be classified
  • original: the text before reformulation
  • labels: a list of classification labels, with possible values including ABOUT:female, ABOUT:male, PARTNER:female, PARTNER:male, SELF:female.
  • class_type: a classification label, with possible values including about (0), partner (1), self (2).
  • turker_gender: a classification label, with possible values including man (0), woman (1), nonbinary (2), prefer not to say (3), no answer (4).
  • episode_done: a boolean indicating whether the conversation was completed.
  • confidence: a string indicating the confidence of the annotator in response to the instance label being ABOUT/TO/AS a man or woman. Possible values are certain, pretty sure, and unsure.

For the funpedia config:

  • text: the text to be classified.
  • gender: a classification label, with possible values including gender-neutral (0), female (1), male (2), indicating the gender of the person being talked about.
  • persona: a string describing the persona assigned to the user when talking about the entity.
  • title: a string naming the entity the text is about.

For the image_chat config:

  • caption: a string description of the contents of the original image.
  • female: a boolean indicating whether the gender of the person being talked about is female, if the image contains a person.
  • id: a string indicating the id of the image.
  • male: a boolean indicating whether the gender of the person being talked about is male, if the image contains a person.

For the wizard config:

  • text: the text to be classified.
  • chosen_topic: a string indicating the topic of the text.
  • gender: a classification label, with possible values including gender-neutral (0), female (1), male (2), indicating the gender of the person being talked about.

For the _inferred configurations (again, except the yelp_inferred split, which does not have the ternary_label or ternary_score fields):

  • text: the text to be classified.
  • binary_label: a classification label, with possible values including ABOUT:female, ABOUT:male.
  • binary_score: a float indicating a score between 0 and 1.
  • ternary_label: a classification label, with possible values including ABOUT:female, ABOUT:male, ABOUT:gender-neutral.
  • ternary_score: a float indicating a score between 0 and 1.

For the word list:

  • word_masculine: a string indicating the masculine version of the word.
  • word_feminine: a string indicating the feminine version of the word.

For the gendered name list:

  • assigned_gender: an integer, 1 for female, 0 for male.
  • count: an integer.
  • name: a string of the name.

Data Splits

The different parts of the data can be accessed through the different configurations:

  • gendered_words: A list of common nouns with a masculine and feminine variant.
  • new_data: Sentences reformulated and annotated along all three axes.
  • funpedia, wizard: Sentences from Funpedia and Wizards of Wikipedia annotated with ABOUT gender with entity gender information.
  • image_chat: sentences about images annotated with ABOUT gender based on gender information from the entities in the image
  • convai2_inferred, light_inferred, opensubtitles_inferred, yelp_inferred: Data from several source datasets with ABOUT annotations inferred by a trined classifier.
Split M F N U Dimension
Image Chat 39K 15K 154K - ABOUT
Funpedia 19K 3K 1K - ABOUT
Wizard 6K 1K 1K - ABOUT
Yelp 1M 1M - - AS
ConvAI2 22K 22K - 86K AS
ConvAI2 22K 22K - 86K TO
OpenSub 149K 69K - 131K AS
OpenSub 95K 45K - 209K TO
LIGHT 13K 8K - 83K AS
LIGHT 13K 8K - 83K TO
---------- ---- --- ---- ---- ---------
MDGender 384 401 - - ABOUT
MDGender 396 371 - - AS
MDGender 411 382 - - TO

Dataset Creation

Curation Rationale

The curators chose to annotate the existing corpora to make their classifiers reliable on all dimensions (ABOUT/TO/AS) and across multiple domains. However, none of the existing datasets cover all three dimensions at the same time, and many of the gender labels are noisy. To enable reliable evaluation, the curators collected a specialized corpus, found in the new_data config, which acts as a gold-labeled dataset for the masculine and feminine classes.

Source Data

Initial Data Collection and Normalization

For the new_data config, the curators collected conversations between two speakers. Each speaker was provided with a persona description containing gender information, then tasked with adopting that persona and having a conversation. They were also provided with small sections of a biography from Wikipedia as the conversation topic in order to encourage crowdworkers to discuss ABOUT/TO/AS gender information. To ensure there is ABOUT/TO/AS gender information contained in each utterance, the curators asked a second set of annotators to rewrite each utterance to make it very clear that they are speaking ABOUT a man or a woman, speaking AS a man or a woman, and speaking TO a man or a woman.

Who are the source language producers?

This dataset was collected from crowdworkers from Amazon鈥檚 Mechanical Turk. All workers are English-speaking and located in the United States.

Reported Gender Percent of Total
Man 67.38
Woman 18.34
Non-binary 0.21
Prefer not to say 14.07

Annotations

Annotation process

For the new_data config, annotators were asked to label how confident they are that someone else could predict the given gender label, allowing for flexibility between explicit genderedness (like the use of "he" or "she") and statistical genderedness.

Many of the annotated datasets contain cases where the ABOUT, AS, TO labels are not provided (i.e. unknown). In such instances, the curators apply one of two strategies. They apply the imputation strategy for data for which the ABOUT label is unknown using a classifier trained only on other Wikipedia data for which this label is provided. Data without a TO or AS label was assigned one at random, choosing between masculine and feminine with equal probability. Details of how each of the eight training datasets was annotated are as follows:

  1. Wikipedia- to annotate ABOUT, the curators used a Wikipedia dump and extract biography pages using named entity recognition. They labeled pages with a gender based on the number of gendered pronouns (he vs. she vs. they) and labeled each paragraph in the page with this label for the ABOUT dimension.

  2. Funpedia- Funpedia (Miller et al., 2017) contains rephrased Wikipedia sentences in a more conversational way. The curators retained only biography related sentences and annotate similar to Wikipedia, to give ABOUT labels.

  3. Wizard of Wikipedia- Wizard of Wikipedia contains two people discussing a topic in Wikipedia. The curators retain only the conversations on Wikipedia biographies and annotate to create ABOUT labels.

  4. ImageChat- ImageChat contains conversations discussing the contents of an image. The curators used the Xu et al. image captioning system to identify the contents of an image and select gendered examples.

  5. Yelp- The curators used the Yelp reviewer gender predictor developed by (Subramanian et al., 2018) and retain reviews for which the classifier is very confident 鈥 this creates labels for the content creator of the review (AS). They impute ABOUT labels on this dataset using a classifier trained on the datasets 1-4.

  6. ConvAI2- ConvAI2 contains persona-based conversations. Many personas contain sentences such as 'I am a old woman' or 'My name is Bob' which allows annotators to annotate the gender of the speaker (AS) and addressee (TO) with some confidence. Many of the personas have unknown gender. The curators impute ABOUT labels on this dataset using a classifier trained on the datasets 1-4.

  7. OpenSubtitles- OpenSubtitles contains subtitles for movies in different languages. The curators retained English subtitles that contain a character name or identity. They annotated the character鈥檚 gender using gender kinship terms such as daughter and gender probability distribution calculated by counting the masculine and feminine names of baby names in the United States. Using the character鈥檚 gender, they produced labels for the AS dimension. They produced labels for the TO dimension by taking the gender of the next character to speak if there is another utterance in the conversation; otherwise, they take the gender of the last character to speak. They impute ABOUT labels on this dataset using a classifier trained on the datasets 1-4.

  8. LIGHT- LIGHT contains persona-based conversation. Similarly to ConvAI2, annotators labeled the gender of each persona, giving labels for the speaker (AS) and speaking partner (TO). The curators impute ABOUT labels on this dataset using a classifier trained on the datasets 1-4.

Who are the annotators?

This dataset was annotated by crowdworkers from Amazon鈥檚 Mechanical Turk. All workers are English-speaking and located in the United States.

Personal and Sensitive Information

For privacy reasons the curators did not associate the self-reported gender of the annotator with the labeled examples in the dataset and only report these statistics in aggregate.

Considerations for Using the Data

Social Impact of Dataset

This dataset is intended for applications such as controlling for gender bias in generative models, detecting gender bias in arbitrary text, and classifying text as offensive based on its genderedness.

Discussion of Biases

Over two thirds of annotators identified as men, which may introduce biases into the dataset.

Wikipedia is also well known to have gender bias in equity of biographical coverage and lexical bias in noun references to women (see the paper's appendix for citations).

Other Known Limitations

The limitations of the Multi-Dimensional Gender Bias Classification dataset have not yet been investigated, but the curators acknowledge that more work is required to address the intersectionality of gender identities, i.e., when gender non-additively interacts with other identity characteristics. The curators point out that negative gender stereotyping is known to be alternatively weakened or reinforced by the presence of social attributes like dialect, class and race and that these differences have been found to affect gender classification in images and sentences encoders. See the paper for references.

Additional Information

Dataset Curators

Emily Dinan, Angela Fan, Ledell Wu, Jason Weston, Douwe Kiela, and Adina Williams at Facebook AI Research. Angela Fan is also affiliated with Laboratoire Lorrain d鈥橧nformatique et Applications (LORIA).

Licensing Information

The Multi-Dimensional Gender Bias Classification dataset is licensed under the MIT License.

Citation Information

@inproceedings{dinan-etal-2020-multi,
    title = "Multi-Dimensional Gender Bias Classification",
    author = "Dinan, Emily  and
      Fan, Angela  and
      Wu, Ledell  and
      Weston, Jason  and
      Kiela, Douwe  and
      Williams, Adina",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.emnlp-main.23",
    doi = "10.18653/v1/2020.emnlp-main.23",
    pages = "314--331",
    abstract = "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 novel, 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 new, crowdsourced evaluation benchmark. Distinguishing between gender bias along multiple dimensions enables us to train better and more fine-grained gender bias classifiers. We show our classifiers are valuable for a variety of applications, like controlling for gender bias in generative models, detecting gender bias in arbitrary text, and classifying text as offensive based on its genderedness.",
}

Contributions

Thanks to @yjernite and @mcmillanmajorafor adding this dataset.

Downloads last month
13,464