Elamur Research⌕
☰
home / papers / 2026 / Extending an Ensemble Baseline with Corpus-Based Graph Features for Portuguese Pun Detection

Extending an Ensemble Baseline with Corpus-Based Graph Features for Portuguese Pun Detection

Paper · 2026 · Avelar Rodrigues de Sousa, Camilla Soares Sousa, Carlos Henrique Santos Barros et al.

Summary

The paper asks whether relational information drawn from a corpus can strengthen a lightweight, interpretable Portuguese pun detector. The starting point is the soft-voting ensemble of Leal et al. (2025), which uses TF-IDF unigram and bigram features. The authors build three graphs from the Puntuguese corpus:

  • a lexical Co-occurrence graph,
  • a PPMI-weighted graph,
  • a Pun-Context graph built only from pun-labelled texts.

Each graph is factorised with TruncatedSVD into node embeddings. These are pooled per document and appended to the TF-IDF vectors before classification.

The graph features do not reliably beat the purely lexical baseline. Their main effect is to change the balance between recall on puns and specificity on non-puns. Pun-Context and PPMI give the best graph-augmented results, while merging all three graphs hurts performance. Error analysis shows that puns using neither homography nor homophony remain the hardest category, although Pun-Context reduces these misses most.

The work matters as an honest negative-leaning result. It suggests that the way graph signals are selected and compressed to document level matters more than simply adding structural information.

Key points

  • Data: the public subset of Puntuguese contains 2,850 humorous one-liners (2,053 PT-BR, 797 PT-PT). Each is paired with a manually micro-edited non-humorous version. Training uses the train split only and evaluation the test split only; the validation file is unused.
  • Baseline: an ensemble of RF, LR and SVM combined through soft voting. Its TF-IDF features use lowercasing, accent normalisation and Portuguese stopword removal.
  • Agreement on the test set (1,140 items): 767 items are correct under every graph variant and 175 are wrong under every variant. Pun-Context has 13 uniquely correct cases, while Co-occurrence and PPMI have 3 each.
  • All Graphs configuration: it recovers 51 items that every single-graph model misses, but fails on 85 items that all of them get right, which explains its weaker overall performance.
  • Limitations named by the authors: very compact pooling of graph features (mean embeddings, coverage, hit counts), slight setup differences from the original baseline, and a single train/test protocol.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

Cite

Cite as Avelar Rodrigues de Sousa, Camilla Soares Sousa, Carlos Henrique Santos Barros and Rafael Torres Anchiêta. 2026. Extending an Ensemble Baseline with Corpus-Based Graph Features for Portuguese Pun Detection. Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{sousa2026extending,
  title = {{Extending an Ensemble Baseline with Corpus-Based Graph Features for Portuguese Pun Detection}},
  author = {Avelar Rodrigues {de Sousa}, Camilla Soares Sousa, Carlos Henrique Santos Barros and Rafael Torres Anchiêta},
  booktitle = {Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1},
  year = {2026},
  url = {https://aclanthology.org/2026.propor-1.75/}
}

References

7 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Kenneth Ward Church and Patrick Hanks. 1990. Word association norms, mutual information, and lexicography. Computational Linguistics, 16(1):22–29.
  2. Avelar de Sousa, Camilla Sousa, and Carlos Barros. 2025. Usabilidade e engajamento em jogos casuais diários: Estudo de caso do guess of dreams. In Anais da XIII Escola Regional de Informática de Goiás, pages 1–10, Porto Alegre, RS, Brasil. SBC. find
  3. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,…
  4. Marcio Lima Inacio, Gabriela Wick-Pedro, Renata Ramisch, Luís Espírito Santo, Xiomara S. Q. Chacon, Roney Santos, Rogério Sousa, Rafael Anchiêta, and Hugo Goncalo Oliveira. 2024. Puntuguese: A corpus of puns in Portuguese with micro-edits. In Proceedings of the 2024 Joint International Conference on… find
  5. Daniel Jurafsky and James H. Martin. 2025. Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition, with Language Models, 3rd edition. Online manuscript released August 24, 2025. find
  6. Antonios Kalloniatis and Panagiotis Adamidis. 2024. Computational humor recognition: a systematic literature review. Artificial Intelligence Review, 58(2):43. find
  7. Justine T. Kao, Roger Levy, and Noah D. Goodman. 2016. A computational model of linguistic humor in puns. Cognitive Science, 40(5):1270–1285. find

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sousa-2026-extending-ensemble-baseline-corpus-based/