Brunovncs/distilbert-swales-rhetorical-moves
DistilBERT Academic Move Classifier
Model Description
This model is a fine-tuned version of DistilBERT (distilled BERTimbau, Portuguese) trained for Rhetorical Move Classification in academic texts, grounded in John Swales' Genre Analysis theory.
Its goal is to identify the rhetorical structure of scientific text sections (such as Introduction, Methodology, Results and Discussion), supporting the analysis of academic writing in Portuguese.
The project was developed as an undergraduate thesis at UNESP — Universidade Estadual Paulista (Ibilce, São José do Rio Preto campus), and its results were published at the Brazilian Symposium in Information and Human Language Technology (STIL 2025), reaching 0.94 weighted F1 / 93.97% accuracy across 6 rhetorical classes — the best known result for this task in Portuguese.
- Author: Bruno Vinicius Veronez de Jesus — brunovinicius.dev.br · GitHub · LinkedIn
- Advisor: Prof. Dr. Arnaldo Cândido Júnior
- Institution: UNESP — Ibilce
- Publication: STIL 2025 — Paper · DOI: 10.5753/stil.2025.37824
Intended Use
The model classifies individual sentences from Portuguese academic texts, assigning categories based on Swales' rhetorical moves. It is intended for natural language processing tools focused on scientific writing and discourse analysis.
You can see it in action in [Swales — Scientific Writing](https://swales.com.br) (source code), a web app that gives students real-time feedback on the structure of scientific abstracts using this model.
Training Data
The model was trained on a corpus of Portuguese academic abstracts annotated with rhetorical moves (CorpusDT), focusing on the argumentative structure typical of scientific writing. A meaning-preserving data-augmentation pipeline (back-translation, synonym substitution, insertion/swap/deletion) was used to expand 52 annotated abstracts into 1,409 balanced training examples.
Citation
If you use this model, please cite:
@inproceedings{veronez2025aprendizado,
author = {Bruno Vinicius {Veronez de Jesus} and Arnaldo {Candido Junior}},
title = {Aprendizado Profundo para Detecção de Movimentos Retóricos},
booktitle = {Simpósio em Tecnologia da Informação e da Linguagem Humana},
year = {2025},
address = {Fortaleza, CE, Brazil},
publisher = {Sociedade Brasileira de Computação},
}