Danie1Arias/sentiment-analysis-catalan-reviews

CSXSC: Classificador de Sentiments de Xarxes Socials en Català

This repository contains the CSXSC (Classificador de Sentiments a Xarxes Socials en Català) dataset, a comprehensive corpus designed for sentiment analysis of Catalan-language content from social media.

The dataset contains 23,788 text entries, each classified as positive, negative, or neutral. It was specifically constructed to address the significant class imbalance often found in user-generated content, resulting in a balanced, multi-source dataset suitable for training robust classification models.

This dataset was developed as part of a Master's Thesis at Rovira i Virgili University by Daniel Arias Cámara.

Dataset Structure

Data Splits

The dataset is divided into three splits, following an 80/10/10 partition:

• —train.csv: 19,030 entries
• —validation.csv: 2,379 entries
• —test.csv: 2,379 entries

Data Fields

Each CSV file contains two columns:

• —text: (string) The Catalan text of the social media post, review, or comment.
• —label: (string) The sentiment label corresponding to the text Positive, Negative or Neutral

Dataset Creation

The CSXSC dataset is a composite corpus created by processing, combining, and curating three distinct sources. The primary goal was to create a larger, more balanced dataset than any single source could provide.

Source Data

1. 1.[GuiaCat](https://huggingface.co/datasets/projecte-aina/GuiaCat): A dataset of 5,750 Catalan restaurant reviews. The original 5-level sentiment labels were mapped to our 3-class schema (Positive, Negative, Neutral). This dataset was heavily skewed towards positive reviews.
2. 2.[CaSSA (Catalan Structured Sentiment Analysis)](https://huggingface.co/datasets/projecte-aina/CaSSA-catalan-structured-sentiment-analysis): A collection of 6,400 reviews and forum posts. The fine-grained "polar expressions" were aggregated using a majority-voting strategy to assign a single sentiment label to each text.
3. 3.[GoEmotions](https://www.kaggle.com/datasets/debarshichanda/goemotions): A large-scale dataset of 58,000 English Reddit comments with fine-grained emotion labels. This dataset was used to augment and rebalance the Catalan corpus.

Curation Process

The creation process involved the following key steps:

1. 1.Baseline Corpus Creation: The GuiaCat and CaSSA datasets were processed and combined, resulting in an initial, highly imbalanced corpus of 12,150 entries.
2. 2.Strategic Sampling: To address the imbalance, a specific number of negative and neutral instances were strategically sampled from the GoEmotions dataset to achieve a target distribution of approximately 40% positive, 30% negative, and 30% neutral.
3. 3.Machine Translation: The sampled English comments from GoEmotions were translated into Catalan using the high-quality [Aina English-Catalan Translator](https://huggingface.co/projecte-aina/aina-translator-en-ca).
4. 4.Quality Curation: The translated texts underwent a rigorous two-step quality control process:
5. 5.An automated evaluation using the [Salamandra-7B-Instruct](https://huggingface.co/BSC-LT/salamandra-7b-instruct) model to flag and remove poor-quality translations.
6. 6.A final manual review to remove remaining noise (nonsensical text, artifacts) and correct mislabeled sentiments.

Licensing Information

The CSXSC dataset is licensed under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0).

Citation Information

If you use this dataset in your research, please cite the original Master's Thesis:

@mastersthesis{arias2025csxsc,
  title={From Traditional to Large Language Models: A Novel NLP-Based Model for Sentiment Analysis in Social Media},
  author={Arias Cámara, Daniel},
  year={2025},
  school={Rovira i Virgili University}
}