README.md

---
annotations_creators:
- Barcelona Supercomputing Center
language_creators:
- Twitter
language:
- ca
license: cc-by-4.0
multilinguality:
- monolingual
pretty_name: CaSET
size_categories:
- 'null': null
task_categories:
- text-classification
task_ids: []
---

# Dataset Card for CaSET, the Catalan Stance and Emotions Dataset from Twitter

## Table of Contents
- [Table of Contents](#table-of-contents)
- [Dataset Description](#dataset-description)
  - [Dataset Summary](#dataset-summary)
  - [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
  - [Languages](#languages)
- [Dataset Structure](#dataset-structure)
  - [Data Instances](#data-instances)
  - [Data Fields](#data-fields)
  - [Data Splits](#data-splits)
- [Dataset Creation](#dataset-creation)
  - [Curation Rationale](#curation-rationale)
  - [Annotations](#annotations)
  - [Personal and Sensitive Information](#personal-and-sensitive-information)
- [Considerations for Using the Data](#considerations-for-using-the-data)
  - [Social Impact of Dataset](#social-impact-of-dataset)
  - [Discussion of Biases](#discussion-of-biases)
  - [Other Known Limitations](#other-known-limitations)
- [Additional Information](#additional-information)
  - [Dataset Curators](#dataset-curators)
  - [Licensing Information](#licensing-information)
  - [Citation Information](#citation-information)
  - [Contributions](#contributions)

## Dataset Description

- **Point of Contact:** [Blanca Calvo](aina@bsc.es)

### Dataset Summary

The CaSET dataset is a Catalan corpus of Tweets annotated with Emotions, Static Stance, and Dynamic Stance. The dataset contains 11k unique sentence on five polemical topics, grouped in 6k pairs of sentences, paired as original messages and answers to these messages.

### Supported Tasks and Leaderboards

This dataset can be used to train models for emotion detection, static stance detection, and dynamic stance detection. 

### Languages

The dataset is in Catalan (`ca-CA`).

## Dataset Structure

Each instance in the dataset is a pair of original-answer messages, annotated with the relation between the two messages (the dynamic stance) and the topic of the messages. For each message there is the id to retrieve it with the Twitter API, the emotions identified in the message, and the relation between the message and the topic (static stance). The text fields have to be retrieved using the Twitter API.

### Data Instances

```
{
"id_original": "1413960970066710533", 
"id_answer": "1413968453690658816", 
  "original_text": "", 
  "answer_text": "", 
    "topic": "vaccines", 
    "dynamic_stance": "Disagree", 
  "original_stance": "FAVOUR", 
  "answer_stance": "AGAINST", 
  "original_emotion": ["distrust", "joy", "disgust"], 
  "answer_emotion": ["distrust"]
}
```

### Data Splits

The dataset does not contain splits.

## Dataset Creation

### Curation Rationale

We created this corpus to contribute to the development of language models in Catalan, a low-resource language. 

### Source Data

The data was collected using the Twitter API by the Barcelona Supercomputing Center.

#### Initial Data Collection and Normalization

The data was collected based on a list of keywords related to the five topics included in the dataset: vaccines, rent regulation, surrogate pregnancy, airport expansion, and a TV show rigging. Specific periods in which the topic was under discussion were also selected.

#### Who are the source language producers?

The source language producers are users of Twitter.

### Annotations

- Emotions are annotated in a multi-label fashion. The labels can be: Anger, Anticipation, Disgust, Fear, Joy, Sadness, Surprise, Distrust, and No emotion.

- Static stance is annotated per message. The labels can be: FAVOUR, AGAINST, NEUTRAL, NA.

- Dynamic stance is annotated per pair. The labels can be: Agree, Disagree, Elaborate, Query, Neutral, Unrelated, NA.

#### Annotation process

- For emotions there were 3 annotators. The gold labels are an aggregation of all the labels annotated by the 3. The IAA calculated with Fleiss' Kappa per label was, on average, 45.38.

- For static stance there were 2 annotators, in the cases of disagreement a third annotated chose the gold label. The overall Fleiss' Kappa between the 2 annotators is 82.71.

- For dynamic stance there were 4 annotators. If at least 3 of the annotators disagreed, a fifth annotator chose the gold label. The overall Fleiss' Kappa between the 4 annotators was 56.51, and the average Fleiss' Kappa of the annotators with the gold labels is 85.17. 

#### Who are the annotators?

All the annotators are native speakers of Catalan. 

### Personal and Sensitive Information

## Considerations for Using the Data

### Social Impact of Dataset

We hope this corpus contributes to the development of language models in Catalan, a low-resource language.

### Discussion of Biases

We are aware that, since the data comes from social media, this will contain biases, hate speech and toxic content. We have not applied any steps to reduce their impact.

### Other Known Limitations

The dataset has to be downloaded using the Twitter API, therefore some instances might be lost. 

## Additional Information

### Dataset Curators

Language Technologies Unit (LangTech) at the Barcelona Supercomputing Center.

This work was funded by the [Departament de la Vicepresidència i de Polítiques Digitals i Territori de la Generalitat de Catalunya](https://politiquesdigitals.gencat.cat/ca/inici/index.html#googtrans(ca|en) within the framework of [Projecte AINA](https://politiquesdigitals.gencat.cat/ca/economia/catalonia-ai/aina).


### Licensing Information

[Creative Commons Attribution 4.0](https://creativecommons.org/licenses/by/4.0/).

### Citation Information

```

```

### Contributions