Dataset Card for [author_profiling]

Dataset Summary

The corpus for the author profiling analysis contains texts in Russian-language which labeled for 5 tasks:

1. gender -- 13448 texts with the labels, who wrote this: text female or male;

2. age -- 13448 texts with the labels, how old the person who wrote the text. This is a number from 12 to 80. In addition, for the classification task we added 5 age groups: 0-19; 20-29; 30-39; 40-49; 50+;

3. age imitation -- 8460 texts, where crowdsource authors is asked to write three texts: a) in their natural manner, b) imitating the style of someone younger, c) imitating the style of someone older;

4. gender imitation -- 4988 texts, where the crowdsource authors is asked to write texts: in their origin gender and pretending to be the opposite gender;

5. style imitation -- 4988 texts, where crowdsource authors is asked to write a text on behalf of another person of your own gender, with a distortion of the authors usual style.

Dataset is collected sing the Yandex.Toloka service link.

You can read the data using the following python code:

def load_jsonl(input_path: str) -> list:
    """
    Read list of objects from a JSON lines file.
    """
    data = []
    with open(input_path, 'r', encoding='utf-8') as f:
        for line in f:
            data.append(json.loads(line.rstrip('\n|\r')))
    print('Loaded {} records from {}/n'.format(len(data), input_path))
    
    return data

path_to_file = "./data/train.jsonl"
data = load_jsonl(path_to_file)

or you can use HuggingFace style:

from datasets import load_dataset

train_df = load_dataset('sagteam/author_profiling', split='train')
valid_df = load_dataset('sagteam/author_profiling', split='validation')
test_df = load_dataset('sagteam/author_profiling', split='test')

Here are some statistics:

1. For Train file:

• No. of documents -- 9564;
• No. of unique texts -- 9553;
• Text length in characters -- min: 197, max: 2984, mean: 500.5;
• No. of documents written -- by men: 4704, by women: 4860;
• No. of unique authors -- 2344; men: 1172, women: 1172;
• Age of the authors -- min: 13, max: 80, mean: 31.2;
• No. of documents by age group -- 0-19: 813, 20-29: 4188, 30-39: 2697, 40-49: 1194, 50+: 672;
• No. of documents with gender imitation: 1215; without gender imitation: 2430; not applicable: 5919;
• No. of documents with age imitation -- younger: 1973; older: 1973; without age imitation: 1973; not applicable: 3645;
• No. of documents with style imitation: 1215; without style imitation: 2430; not applicable: 5919.

2. For Valid file:

• No. of documents -- 1320;
• No. of unique texts -- 1316;
• Text length in characters -- min: 200, max: 2809, mean: 520.8;
• No. of documents written -- by men: 633, by women: 687;
• No. of unique authors -- 336; men: 168, women: 168;
• Age of the authors -- min: 15, max: 79, mean: 32.2;
• No. of documents by age group -- 1-19: 117, 20-29: 570, 30-39: 339, 40-49: 362, 50+: 132;
• No. of documents with gender imitation: 156; without gender imitation: 312; not applicable: 852;
• No. of documents with age imitation -- younger: 284; older: 284; without age imitation: 284; not applicable: 468;
• No. of documents with style imitation: 156; without style imitation: 312; not applicable: 852.

3. For Test file:

• No. of documents -- 2564;
• No. of unique texts -- 2561;
• Text length in characters -- min: 199, max: 3981, mean: 515.6;
• No. of documents written -- by men: 1290, by women: 1274;
• No. of unique authors -- 672; men: 336, women: 336;
• Age of the authors -- min: 12, max: 67, mean: 31.8;
• No. of documents by age group -- 1-19: 195, 20-29: 1131, 30-39: 683, 40-49: 351, 50+: 204;
• No. of documents with gender imitation: 292; without gender imitation: 583; not applicable: 1689;
• No. of documents with age imitation -- younger: 563; older: 563; without age imitation: 563; not applicable: 875;
• No. of documents with style imitation: 292; without style imitation: 583; not applicable: 1689.

Supported Tasks and Leaderboards

This dataset is intended for multi-class and multi-label text classification.

The baseline models currently achieve the following F1-weighted metrics scores (table):

Model name	gender	age_group	gender_imitation	age_imitation	style_imitation	no_imitation	average
Dummy-stratified	0.49	0.29	0.56	0.32	0.57	0.55	0.46
Dummy-uniform	0.49	0.23	0.51	0.32	0.51	0.51	0.43
Dummy-most_frequent	0.34	0.27	0.53	0.17	0.53	0.53	0.40
LinearSVC + TF-IDF	0.67	0.37	0.62	0.72	0.71	0.71	0.63

Languages

The text in the dataset is in Russian.

Dataset Structure

Data Instances

Each instance is a text in Russian with some author profiling annotations.

An example for an instance from the dataset is shown below:

{
  'id': 'crowdsource_4916',
  'text': 'Ты очень симпатичный, Я давно не с кем не встречалась. Ты мне сильно понравился, ты умный интересный и удивительный, приходи ко мне в гости , у меня есть вкусное вино , и приготовлю вкусный ужин, посидим пообщаемся, узнаем друг друга поближе.',
  'account_id': 'account_#1239',
  'author_id': 411,
  'age': 22,
  'age_group': '20-29',
  'gender': 'male',
  'no_imitation': 'with_any_imitation',
  'age_imitation': 'None',
  'gender_imitation': 'with_gender_imitation',
  'style_imitation': 'no_style_imitation'
}

Data Fields

Data Fields includes:

• id -- unique identifier of the sample;

• text -- authors text written by a crowdsourcing user;

• author_id -- unique identifier of the user;

• account_id -- unique identifier of the crowdsource account;

• age -- age annotations;

• age_group -- age group annotations;

• no_imitation -- imitation annotations. Label codes:

  • 'with_any_imitation' -- there is some imitation in the text;
  • 'no_any_imitation' -- the text is written without any imitation

• age_imitation -- age imitation annotations. Label codes:

  • 'younger' -- someone younger than the author is imitated in the text;
  • 'older' -- someone older than the author is imitated in the text;
  • 'no_age_imitation' -- the text is written without age imitation;
  • 'None' -- not supported (the text was not written for this task)

• gender_imitation -- gender imitation annotations. Label codes:

  • 'no_gender_imitation' -- the text is written without gender imitation;
  • 'with_gender_imitation' -- the text is written with a gender imitation;
  • 'None' -- not supported (the text was not written for this task)

• style_imitation -- style imitation annotations. Label codes:

  • 'no_style_imitation' -- the text is written without style imitation;
  • 'with_style_imitation' -- the text is written with a style imitation;
  • 'None' -- not supported (the text was not written for this task).

Data Splits

The dataset includes a set of train/valid/test splits with 9564, 1320 and 2564 texts respectively. The unique authors do not overlap between the splits.

Dataset Creation

Curation Rationale

The formed dataset of examples consists of texts in Russian using a crowdsourcing platform. The created dataset can be used to improve the accuracy of supervised classifiers in author profiling tasks.

Source Data

Initial Data Collection and Normalization

Data was collected from crowdsource platform. Each text was written by the author specifically for the task provided.

Who are the source language producers?

Russian-speaking Yandex.Toloka users.

Annotations

Annotation process

We used a crowdsourcing platform to collect texts. Each respondent is asked to fill a questionnaire including their gender, age and native language.

For age imitation task the respondents are to choose a topic out of a few suggested, and write three texts on it:

1. Text in their natural manner;
2. Text imitating the style of someone younger;
3. Text imitating the style of someone older.

For gender and style imitation task each author wrote three texts in certain different styles:

1. Text in the authors natural style;
2. Text imitating other gender style;
3. Text in a different style but without gender imitation.

The topics to choose from are the following.

• An attempt to persuade some arbitrary listener to meet the respondent at their place;
• A story about some memorable event/acquisition/rumour or whatever else the imaginary listener is supposed to enjoy;
• A story about oneself or about someone else, aiming to please the listener and win their favour;
• A description of oneself and one’s potential partner for a dating site;
• An attempt to persuade an unfamiliar person to come;
• A negative tour review.

The task does not pass checking and is considered improper work if it contains:

• Irrelevant answers to the questionnaire;
• Incoherent jumble of words;
• Chunks of text borrowed from somewhere else;
• Texts not conforming to the above list of topics.

Texts checking is performed firstly by automated search for borrowings (by an anti-plagiarism website), and then by manual review of compliance to the task.

Who are the annotators?

Russian-speaking Yandex.Toloka users.

Personal and Sensitive Information

All personal data was anonymized. Each author has been assigned an impersonal, unique identifier.

Considerations for Using the Data

Social Impact of Dataset

[Needs More Information]

Discussion of Biases

[Needs More Information]

Other Known Limitations

[Needs More Information]

Additional Information

Dataset Curators

Researchers at AI technology lab at NRC "Kurchatov Institute". See the website.

Licensing Information

Apache License 2.0.

Citation Information

If you have found our results helpful in your work, feel free to cite our publication.

@article{сбоев2022сравнение,
  title={СРАВНЕНИЕ ТОЧНОСТЕЙ МЕТОДОВ НА ОСНОВЕ ЯЗЫКОВЫХ И ГРАФОВЫХ НЕЙРОСЕТЕВЫХ МОДЕЛЕЙ ДЛЯ ОПРЕДЕЛЕНИЯ ПРИЗНАКОВ АВТОРСКОГО ПРОФИЛЯ ПО ТЕКСТАМ НА РУССКОМ ЯЗЫКЕ},
  author={Сбоев, АГ and Молошников, ИА and Рыбка, РБ and Наумов, АВ and Селиванов, АА},
  journal={Вестник Национального исследовательского ядерного университета МИФИ},
  volume={10},
  number={6},
  pages={529--539},
  year={2021},
  publisher={Общество с ограниченной ответственностью МАИК "Наука/Интерпериодика"}
}

Contributions

Thanks to @naumov-al for adding this dataset.