A project for detecting whether an Italian text was written by a human or generated by a language model.
This repository contains the code for a Machine-Generated Text Detection project developed for the course Linguistica Computazionale II — A.A. 2025/2026.
The goal of the project is to build and compare different text classification systems able to distinguish between:
| Label | Meaning |
|---|---|
0 |
👤 Human-written text |
1 |
🤖 Machine-generated text |
The project is based on the DESEGMA-IT dataset, proposed as part of EVALITA 2026, and focuses only on the first sub-task: binary document-level classification.
The aim is to compare multiple approaches to Italian text classification, ranging from traditional machine learning models to neural language models.
The project includes:
- 🧩 Linear SVM models based on n-grams;
- 🧠 Linear SVM models based on non-lexical linguistic features;
- 🟪 Linear SVM models based on word embeddings;
- 🚀 Fine-tuning of a neural language model.
The final objective is not only to identify the best-performing model, but also to analyze how different textual representations affect classification performance.
The dataset used in this project is DESEGMA-IT — Detection and Segmentation of Machine Generated Text in Italian.
Official task page:
https://desegma.github.io/