nicolauduran45/multilingual-research-classification

Multilingual Scientific Text Classification Dataset (MAG FoS L1)

Overview

This dataset contains multilingual scientific text samples (Catalan, Spanish, and English) extracted from scientific publications.
Each sample is labeled using Microsoft Academic Graph (MAG) Field of Study — Level 1 categories.

For each publication, the text field is a random selection of:

• the title
• the abstract
• the title followed by the abstract (title + ". " + abstract)

This introduces natural variation and improves model robustness for text classification tasks.

---

Dataset Structure

Features

Feature	Type	Description
text	string	Randomly selected title, abstract, or title + abstract
label	string	MAG Field of Study (FoS) Level 1 category
language	string	ISO code of publication language (ca, es, en)

Splits

Split	Samples
Train	20,059
Validation	2,507
Test	2,508
Total	25,074

Splits follow an 80/10/10 ratio.

---

Languages

The dataset includes scientific publications written in:

• Catalan (ca)
• Spanish (es)
• English (en)

---

Task

Multiclass Scientific Text Classification

Your model should predict the Field of Study (FoS) Level 1 category from a scientific text snippet.
This dataset is suitable for:

• multilingual text classification
• scientific-domain NLP
• domain adaptation
• benchmarking multilingual LLMs (mBERT, XLM-R, LLaMA, etc.)
• zero-shot or few-shot evaluation

---

Source

The labels correspond to Level 1 Fields of Study from the Microsoft Academic Graph (MAG) ontology.

Typical categories include (examples):

• Chemistry
• Physics
• Biology
• Computer Science
• Mathematics
• Medicine
• Social Sciences
• Engineering
• Earth Sciences
• Environmental Science

The exact label set matches the categories present in the processed data.

---

Creation Process

1. Load publication metadata (title, abstract, language, FoS).
2. Clean and normalize text fields.
3. Randomly choose one of:
   • title
   • abstract
   • title + abstract
4. Assign the MAG FoS L1 label.
5. Perform an 80/10/10 train-validation-test split using HuggingFace datasets.

---

Usage

Load the dataset

from datasets import load_dataset

dataset = load_dataset("YOUR_USERNAME/YOUR_DATASET_NAME")
print(dataset["train"][0])
Example record
json
Copy code
{
  "text": "Reactividad de CHI3 con radicales O... Las vías de abstracción...",
  "label": "Physical chemistry",
  "language": "es"
}