metadata

dataset_info:
  features:
    - name: pregunta
      dtype: string
    - name: respuesta
      dtype: string
  splits:
    - name: train
      num_bytes: 922215
      num_examples: 4196
  download_size: 404937
  dataset_size: 922215
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: cc-by-nc-sa-4.0
language:
  - es

Dataset Card for Es-Inclusive-Language-it

This dataset is the instructions version of the dataset Es-Inclusive-Language. In this version there are two columns of data: pregunta y respuesta. The column pregunta contains the input as a prompt giving instructions to rewrite a piece of spanish text to the inclusive version of the same text. The expected output, i.e. the text in inclusive language, is given in the column respuesta.

Find below the dataset card for es-inclusive-language dataset.

Dataset Card for Es-Inclusive-Language

Languages are powerful tools to communicate ideas, but their use is not impartial. The selection of words carries inherent biases and reflects subjective perspectives. In some cases, language is wielded to enforce ideologies, marginalize certain groups, or promote specific political agendas.

Spanish is not the exception to that. For instance, when we say “los alumnos” or “los ingenieros”, we are excluding women from those groups. Similarly, expressions such as “los gitanos” o “los musulmanes” perpetuate discrimination against these communities. In response to these linguistic challenges, this dataset offers neutral alternatives in accordance with official guidelines on inclusive language from various Spanish speaking countries. Its purpose is to provide grammatically correct and inclusive solutions to situations where our language choices might otherwise be exclusive.

This dataset consists of pairs of texts with one entry featuring exclusive language and the other one its corresponding inclusive rewrite. All pairs are tagged with the origin (source) of the data and, in order to account for completeness of inclusive translation, also with labels for translation difficulty.

This is a tool that contributes to the Sustainable Development Goals number five (Achieve gender equality and empower all women and girls) and ten (Reduce inequality within and among countries).

Dataset Details

Dataset Description

Curated by: Andrés Martínez Fernández-Salguero, Gaia Quintana Fleitas, Miguel López Pérez, Imanuel Rozenberg and Josué Sauca
Funded by: SomosNLP, HuggingFace, Argilla
Language(s) (NLP): Spanish (es-ES, es-AR, es-MX, es-CR, es-CL)
License: cc-by-nc-sa-4.0

Dataset Sources

Repository: https://github.com/Andresmfs/es-inclusive-language-dataset-creation
Video presentation: https://www.youtube.com/watch?v=7rrNGJIXEHU

Uses

Direct Use

This dataset can be used to fine-tune LLMs to perform text2text generation tasks, specifically to train models that are able to rewrite Spanish texts using inclusive language.

Out-of-Scope Use

This dataset is specifically designed for translating Spanish texts to Spanish texts in inclusive language. Using the dataset for unrelated tasks is considered out of scope. This dataset can not be used with commercial purposes, it is intended for research or educational purposes only.

Dataset Structure

The dataset has a total of 4196 rows and contains the following columns:

gender_exclusive (input): text in non inclusive language
gender_inclusive (target): text in inclusive language
difficulty: translation difficulty category. Descriptions and distribution below.
origin: data source. Descriptions and distribution below.

Difficulty tags descriptions

We used different labels, most of them gender related, and can be describe like this:

Tag	Description	Example
no_cambia	No changes are needed	"Los alumnos Carlos y Manuel son muy problemáticos" cannot be translated as "El alumnado Carlos y Manuel son muy problemáticos”
plural_complejo	Plural words for which there is not a neutral term. There are different formulas that will vary according to the context.	"Los agricultores" -> "La comunidad agrícola", "Los y las agricultoras". “Las limpiadoras” -> “El equipo de limpieza”. More: "El grupo de...", "El sector de...", "El personal de..."
plural_neutro	Change the plural for a generic noun.	"Los alumnos" -> "El alumnado"
culturas	People and cultures	"Los andaluces" -> "El pueblo andaluz", "La comunidad andaluza"
feminizar_profesiones	Professions with androcentric feminine forms	“La médico” -> "La médica". “La técnico de sonido” -> "La técnica de sonido"
nombres_propios	Proper names	"Los alumnos Carlos y Manuel son muy problemáticos" cannot be translated as "El alumnado es muy problemático
persona_generica	Reference to a generic person	"Nota al lector" -> "Nota a quien lee", "Nota a la persona que lee"
dificultades_variadas	Mix of difficulties (to tag big chunks of diverse data)
plurales	Mix of neutral and complex plurals
falsa_concordancia	Androcentric agreement errors	"Estas siete parejas van a dar lo mejor de sí mismos" -> "Estas siete parejas van a dar lo mejor de sí mismas."
omision	The subject or some pronouns are omitted, or the phrase is restructured with verboids.	"los participantes mantendrán un debate" -> "habrá un debate", "Si los científicos trabajan adecuadamente" -> "Trabajando adecuadamente, "los estudiantes" -> "estudiantes
terminologia	Correction of terms with ableist, racist, or other types of discrimination bias.
parafrasis	Avoid words with generic connotations by reformulating the phrase
otros	Difficulties that don’t fit in the other labels

Origin tags descriptions

Data quality can depend on their origin, so data are tagged with origin labels according to this table:

Tag	Description	Link to origin
neutral_es	Curated and refined version of neutral-es dataset	https://huggingface.co/datasets/hackathon-pln-es/neutral-es
GPT-3.5_fewshot	Chat GPT-3.5 generated with few shot technique
GPT-3.5_CaDi ^*	Data created based on the dataset used for developing CaDi project^*	https://lenguaje-incluyente.ibero.mx/
GPT-3.5_fs_multiplication	Data multiplicated from GPT-3.5_fewshot using GPT-3.5
guia_CCGG	Examples from Spanish General Courts language inclusive Guide	https://www.congreso.es/docu/igualdad/Recomendaciones_uso_no_sexista_lenguajeCC.GG..pdf
guia_TAI	Examples from Trenes Argentinos' Guide to the use of inclusive language	https://www.argentina.gob.ar/sites/default/files/guia_para_uso_de_lenguaje_inclusivo_v1.pdf
guia_CONICET	Examples from Guide to inclusive, non-sexist language (CONICET)	https://cenpat.conicet.gov.ar/wp-content/uploads/sites/91/2020/08/Guia-lenguaje-inclusivo-no-sexista-CENPAT_final-1.pdf
guia_INAES	Examples of Guidelines for Inclusive Language Recommendations (INAES)	https://www.argentina.gob.ar/sites/default/files/2020/10/lenguaje_inclusivo_inaes_2021.pdf
guia_CHRYSALLIS	Examples from Practical Guide to Inclusive Language (Chrysallis)	https://www.lgbtqiahealtheducation.org/wp-content/uploads/2020/04/Guia-practica-de-lenguaje-inclusivo-Chrysallis.pdf
guia_ONU	Examples from Guidance for the use of gender-inclusive language (UN)	https://www.unwomen.org/sites/default/files/Headquarters/Attachments/Sections/Library/Gender-inclusive%20language/Guidelines-on-gender-inclusive-language-es.pdf
guia_MX	Examples from Manual for the use of inclusive and gender-sensitive language (MX)	https://www.gob.mx/cms/uploads/attachment/file/183695/Manual_Lenguaje_Incluyente_con_perspectiva_de_g_nero-octubre-2016.pdf
guia_CL	Examples from Gender Inclusive Language Guide of the Government of Chile	https://www.cultura.gob.cl/wp-content/uploads/2023/01/guia-de-lenguaje-inclusivo-de-genero.pdf
guia_IEM	Examples from Uso del Lenguaje Inclusivo de Género	https://secretariagenero.poder-judicial.go.cr/images/Documentos/LenguajeInclusivo/Documentos/Uso-de-lenguaje-inclusivo-de-Genero-IEM-UNA.pdf
human_combinatory	Combinatorics of text fragments generated with GPT3.5
GPT-4_human	Chat GPT-4 generated and human revised
human	Human created data based on information from inclusive language guidelines

^*©Universidad Iberoamericana, A.C. , Ciudad de México, México

^*©Capitolina Díaz Martínez, Elvia María Guadalupe González del Pliego Dorantes, Marco Antonio López Hernández, Alberto López Medina, Héctor Celallos Avalos, Laura Mejía Hernández

Dataset Creation

Curation Rationale

The selection of words carries inherent biases and reflects subjective perspectives. In some cases, language is wielded to enforce ideologies, marginalize certain groups, or promote specific political agendas. In response to these linguistic challenges, this dataset offers neutral alternatives in accordance with official guidelines on inclusive language from various Spanish-speaking countries. Its purpose is to provide grammatically correct and inclusive solutions to situations where our language choices might otherwise be exclusive.

This dataset has been created to train the model "Traductor Inclusivo" which aims to tackle this problem.

Source Data

Data Collection and Processing

The data used for training the model has been sourced from various origins. The first and more important source was a curated and refined version of es_neutral (link to the dataset curation notebook here )

In addition, we manually generated data based on Official Guidelines from different Spanish speaking countries. Finally, we augmented this data by experimenting with various prompts and Few-Shot learning techniques. We needed to be as explicit as possible, otherwise we wouldn’t get good results. You can see some examples below.

Example 1:

Example 2:

Example 3:

We tried to be as inclusive as possible, paying close attention to the classification of difficulties that one could encounter in texts like these.

Moreover, we took care to incorporate numerous counterexamples, recognizing that there are instances where neutrality is not required in a sentence. For instance, “Las arquitectas María Nuñez y Rosa Loria presentaron el proyecto” should not be rewritten as “El equipo de arquitectura María Nuñez y Rosa Loria presentó el proyecto”.

It’s important to highlight that the Traductor Inclusivo not only promotes gender inclusivity but also addresses other forms of discrimination such as ableism, racism, xenophobia, and more.

You can access scripts used during data creation here

Who are the source data producers?

Data have been produced by different producers:

Official Spanish Inclusive Language Guidelines:
neutral-es dataset
ChatGPT-3.5 based examples and structures from official guidelines and dataset used for developing CaDi project
ChatGPT-4

Annotations

Annotation process

Most data were created using ChatGPT-3.5 based on examples and structures form Official Inclusive Language Guidelines and then supervised and corrected by the annotators always following the indications from the Official Guidelines. You can find more information about this process in the Data Collection and Processing section. Other data were created using pairs of words taken from CaDi project. Data taken from neutral-es dataset were not modified beyond correcting spelling or grammar mistakes.

Regarding difficulty labels, we stablished different groups of data using the information from Official Spanish Inclusive Language Guidelines. We created a list of difficulties that could cover all kind of examples. The difficulties list can be found above together with their descriptions. Data from Official Guidelines were manually matched to the different difficulties according to the descriptions provided in the list and then we used those examples as a base to create more data using ChatGPT difficulty by difficulty, so new data would belong to a specific difficulty and therefore tagged with that difficulty.

Who are the annotators?

Personal and Sensitive Information

This dataset does not contain any personal or sensitive information.

Bias, Risks, and Limitations

Data are based on the Spanish Inclusive Language Guidelines mentioned in the source section and they may inherit any bias comming from these guidelines and institutions behind them. These are official and updated guidelines that should not contain strong biases. Data from neutral-es are based as well on Official Spanish Inclusive Language Guidelines
Data based on CaDi may inherit biases from the institution behind it. It has been created by a group of specialists in the field from Universidad Iberoamericana
Dataset contains a significantly large amount of data related to workspace compared to other areas.
This dataset covers only difficulties that appear on the difficulties list
Dataset is composed mostly by sentences where the terms to be modified are at the beginning of the sentence.
Dataset does not contain long-complex texts.

Recommendations

Users should be made aware of the risks, biases and limitations of the dataset.

License

Creative Commons (cc-by-nc-sa-4.0) This kind of licensed is inherited from CaDi project dataset.

Citation

BibTeX:

@software{GAMIJ2024es-inclusive-language,
  author = {Gaia Quintana Fleitas, Andrés Martínez Fernández-Salguero, Miguel López Pérez, Imanuel Rozenberg, Josué Sauca},
  title = {Traductor Inclusivo},
  month = April,
  year = 2024,
  url = {https://huggingface.co/datasets/somosnlp/es-inclusive-language}
}

More Information

This project was developed during the Hackathon #Somos600M organized by SomosNLP. The dataset was created using distilabel by Argilla and endpoints sponsored by HuggingFace.

Team:

Datasets:

somosnlp
/

es-inclusive-language-it