license: cc-by-4.0
task_categories:
- text-classification
language:
- es
pretty_name: ClimID
size_categories:
- 1K<n<10K
Dataset Card for BERTIN-ClimID: BERTIN-Base Climate-related text Identification
README Spanish Version: README_ES
Dataset for BERTIN-ClimID was developed as the fusion of different sources (open-source).
Dataset Details
Dataset Description
- Curated by: Gerardo Huerta Gabriela Zuñiga
- Funded by: SomosNLP, HuggingFace
- Language(s): es-ES, es-PE
- License: cc-by-nc-sa-4.0
Dataset Sources
- Repository: somosnlp/spa_climate_detection
- Paper: [WIP]
- Video presentation: Proyecto BERTIN-ClimID
Uses
Direct Use
- News classification: With this model it is possible to classify news headlines related to the areas of climate change.
- Paper classification: The identification of scientific texts that disclose solutions and/or effects of climate change. For this use, the abstract of each paper can be used for identification. -Social Media posts Classification: Classify social media posts (short texts) related or not to climate areas
Out-of-Scope Use
- For the creation of information repositories regarding climate issues.
- This model can serve as a basis for creating new classification systems for climate solutions to disseminate new efforts to combat climate change in different sectors.
- Creation of new datasets that address the issue.
Dataset Structure
- question: Text
- answer: binary label, if the text is related to climate change or sustainability (1) if the text is not related (0)
- domain: Identifies what topic the text is related to, in our case we have 3 types "climate_change_reports", "miscellaneous_press", "climate_change". Climate change reports refers to the paragraphs that talk about climate change but were extracted from corporate annual reports. Miscellaneous press are paragraphs on various topics extracted from the press. Climate change, all paragraphs that talk about this topic and do not have any special source of information.
- Country of origin: Where this data comes from geographically. We include 3 categories: "global", "Spain", "USA". Global is data that was taken from sources that do not indicate its specific origin but we know that it was taken from data repositories with sources from any country of origin.
- Language: Geographic variety of Spanish used. In this case we used 2 types "es_pe", "es_esp", this is because many of the data had to be translated from English to Spanish, annotations were made using the regional language of the team that collaborated with the translation.
- Registration: Functional variety of language. Within this dataset, 3 types are identified: "cult", "medium", "colloquial" depending on the origin of the data.
- Task: Identifies the purpose for which the input data is intended.
- Period: In what era the language used is located. This dataset uses actual language.
[More Information Needed]
Dataset Creation
Curation Rationale
The motivation of the dataset creation was to create a repository in Spanish on information or resources on topics such as: climate change, sustainability, global warming, energy, etc; this because we didn't found a dataset like this one. Climate change and global warming are current main problems globally so it's important to fight this harm in all places with all languages, also to bring solutions and share information accesible for everyone
Source Data
We used several sources of data to make a varied dataset that could work with different types of texts, from articles, news, social media posts and other texts. we included:
- Spanish translation of the dataset: [climate Bert] (https://huggingface.co/datasets/climatebert/climate_detection)
- News in Spanish on topics not related to climate change:Spanish news headers
- Translation of opinions related to climate change: Opinions
- Translation of news tweets not related to climate change: Posts
Data Collection and Processing
- Spanish translation of the dataset: climate Bert
- News in Spanish on topics not related to climate change:Spanish news headers For this dataset, the column with news and the topics Macroeconomics, Innovation, Regulations, Alliances, Reputation have been discriminated, which have been labeled with (0) The dataset also contained the topic Sustainability as a topic but it was removed (we only required unrelated texts).
- Translation of opinions related to climate change: Opinions In this dataset all opinions are related to climate change, which is why they were labeled with (1). Data cleaning has been carried out by removing harshtags, usernames and emogis to use only the textual content of the tweets.
- Translation of news tweets not related to climate change: Posts In this dataset the news is categorized and has short length (like opinions) all text is not related to climate change so they were labeled with (0). Data cleaning has been carried out by removing harshtags, usernames and emogis to use only the textual content of the tweets. This dataset has been chosen to balance the amount of related text and to include short texts not related to training.
Who are the source data producers?
- Climate bert dataset: Large Companies listed on paper(original dataset).
- Spanish News: Web scrapping from Bank news sites
- Opinions from climate change: Tweets extraction
- Opinions not related to climate change: Tweets of around 2 month of Los Angeles News from twitter.
Annotations
Annotation process
All the records had the corresponding annotation (related or not related to climate and global warming), but we only changed the text values to binary values (1 / 0)
Who are the annotators?
[More Information Needed]
Personal and Sensitive Information
In this case it was not necessary to have an anonymization process
Bias, Risks, and Limitations
At this point, no specific studies have been carried out on biases and limitations, however we make the following notes based on previous experience and model testing:
It inherits the biases and limitations of the base model with which it was trained.
Direct biases such as the majority use of high-level language in the dataset due to the use of texts extracted from news, legal documentation of companies that can complicate the identification of texts with low-level languages (example: colloquial). To mitigate these biases, diverse opinions on climate change topics extracted from sources such as social networks were included in the dataset, and the labels were additionally rebalanced (see tables below).
The dataset inherits other limitations such as: the model loses performance in short texts, this is because most of the texts used in the dataset have a long length of between 200 - 500 words. Once again, an attempt was made to mitigate these limitations with the inclusion of short texts.
train:
| Número | Label | % |
|---|---|---|
| 1600 | 1 | 55% |
| 1300 | 0 | 45% |
- test:
| Número | Label | % |
|---|---|---|
| 480 | 1 | 62% |
| 300 | 0 | 38% |
Recommendations
Our recommendation is to continue adding more samples of spanish text in both large and short length.
[More Information Needed]
License
cc-by-nc-sa-4.0 Due to inheritance of the data used in the dataset.
Citation
BibTeX:
@misc{BERTIN-ClimID,
author = {Gerardo Huerta, Gabriela Zuñiga},
title = {Dataset for BERTIN-ClimID: BERTIN-Base Climate-related text Identification},
month = Abril,
year = 2024,
url = {https://huggingface.co/datasets/somosnlp/spa_climate_detectiona}
}
More Information
This project was developed during the Hackathon #Somos600M organized by SomosNLP. We thank all event organizers and sponsors for their support during the event.
Team: