metadata

license: apache-2.0
task_categories:
  - question-answering
language:
  - es
tags:
  - computational linguistics
  - spanish
  - NLP
  - json
size_categories:
  - 1K<n<10K

Dataset Card for LingComp_QA, un corpus educativo de lingüística computacional en español

Dataset Details

Dataset Description

Curated by: Jorge Zamora Rey, Isabel Moyano Moreno, Mario Crespo Miguel
Funded by: SomosNLP, HuggingFace, Argilla, Instituto de Lingüística Aplicada de la Universidad de Cádiz
Language(s) (NLP): es-ES
License: apache-2.0

Dataset Sources

Repository: https://github.com/reddrex/lingcomp_QA/tree/main
Paper: Comming soon!

Uses

This dataset is intended for educational purposes. When we further develop this resource, we would like it to serve as learning resource for NLP and Computational Linguistics beginners - either for tests, looking up the answers to common questions or studying key concepts and methodologies of CL.

Direct Use

There is no specific use case intended for this dataset. However, we would like to develop a conversational language model that answers questions on Computational Linguistics. The dataset could be used to develop other educational tools or resources too, such as interactive quizzes, tutorials, and study materials, to help students learn about computational linguistics concepts, methodologies, and applications.

Out-of-Scope Use

The dataset is specifically designed for tasks related to computational linguistics, language processing, and natural language understanding. Therefore, using the dataset for unrelated tasks, such as image processing or numerical analysis, would be considered out of scope. Other out of scope uses would be using this dataset for product development or marketing, any commercial use really, as it is intended for research and educational purposes.

Dataset Structure

The dataset's structure looks like this:

[
  {
    "pregunta": "¿Qué implica la lingüística computacional teórica?",
    "respuesta": "La lingüística computacional teórica incluye el desarrollo de teorías formales de gramática y semántica, basadas en lógicas formales o enfoques simbólicos. Las áreas de estudio teórico en este ámbito incluyen la complejidad computacional y la semántica computacional."
  },
  {
    "pregunta": "¿Qué es una gramática libre de contexto?",
    "respuesta": "Una gramática libre de contexto es una gramática formal en la que cada regla de producción es de la forma V → w, donde V es un símbolo no terminal y w es una cadena de terminales y/o no terminales."
  },
  {
    "pregunta": "¿Qué es el algoritmo CYK y cuál es su propósito?",
    "respuesta": "El algoritmo de Cocke-Younger-Kasami (CYK) es un algoritmo de análisis sintáctico ascendente que determina si una cadena puede ser generada por una gramática libre de contexto y, en caso afirmativo, cómo puede ser generada. Su propósito es realizar un análisis sintáctico de la cadena para determinar su estructura gramatical."
  },
  {...}
]

We have a "pregunta" or question column, and a "respuesta" or answer column, where each question has an answer associated.

Dataset Creation

Curation Rationale

The lack of NLP educational resources meant for linguists, especially in Spanish, drove us to make a first attempt of collecting information on this topic from open internet sources. We aim to grow the corpus and create a a foundational resource for teaching linguists (and other beginners) about the principles, techniques, and applications of computational linguistics and NLP.

Source Data

Blogs, wikipedia articles and our Computational Linguistics and Language Engineering course materials at the University of Cádiz comprise the source data for this dataset.

Data Collection and Processing

First, we collected information on different aspects of Computational Linguistics (statistics, computer science, linguistics, corpus linguistics, etc.) from open blogs and webpages with Bootcat. After this, we manually extracted information and created questions for each information segment. Then all three revised the whole document, deleted duplicated questions from each member's corpus portion and checked expressions. We also tried to make some explanations easier to comprehend for a broader audience, as the purpose of this project is mainly educational.

Here we link the scripts used in the creation of the corpus, which are the following:

Who are the source data producers?

Our team members manually checked and organized the information into sets of questions and answers, while rewriting some of the info in a more suitable style for learners.

Annotation process

We manually sorted the information into question-answer pairs. However, we did use the following Colaboratory notebook to create the JSON file:

https://github.com/reddrex/lingcomp_QA/blob/main/dataset/Creación_archivo_JSON_a_partir_de_txt.ipynb

Who are the annotators?

The annotators are the members of our team: Jorge Zamora, Isabel Moyano and Mario Crespo.

Personal and Sensitive Information

There are no personal, sensitive or private data that should not be shown in the dataset. The only names and dates that appear in it are those of the scientists, programmes and core dates in the development of the Artificial Intelligence area NLP.

Bias, Risks, and Limitations

Main bias might belong to the sources from which we extracted the information. Some blogs or wikipedia articles might employ different terminology for the same concept (and while we have tried to correct this, some terms could have escaped our supervisors). Also, the low availability of information on Computational Linguistics and NLP on Spanish on the Internet may have created an imbalance on topics tackled by the dataset. For example, there could be more information on Python usage than NLTK, or more on NLTK than Spacy, as it happens. Among our future plans there is balancing the topics out by translating from English sources. Plus, we would like to add QA pairs that might not appear in any relevant open info source and that we believe would be good for learners - mostly from our experience in the Linguistics and Applied Languages bachelor, although we are open to requests.

The limitations we found while building the dataset are mostly time-related, as such a broad topic can be difficult to cover in such a limited amount of time. Furthermore, we found ourselves unable to fully balance the coverage of all the involved themes, as there were not enough information sources on the internet - plus, open to the public - that we could use in order to document our QA pairs.

Recommendations

Users should be made aware of the risks, biases and limitations of the dataset. We recommend checking the dataset from time to time or checking our social media/contacting us via email. There we will be announcing whether a new version with more information and a broader range of sources will be launched and when.

License

Apache 2.0

Citation

BibTeX:

@software{LingComp_QA,
  author = {Zamora Rey, Jorge and Crespo Miguel, Mario and Moyano Moreno, Isabel},
  title = {LingComp_QA, un corpus educativo de lingüística computacional en español},
  month = March,
  year = 2024,
  url = {https://huggingface.co/datasets/somosnlp/LingComp_QA}
}

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:

Contact

mario.crespo@uca.es isabel.moyano@uca.es