HOHilário Oliveira

Hilário Oliveira

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Hilário Oliveira has a total of 74 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • A Synthetic Dataset for Evaluating Grammar Correction in Brazilian Portuguese Using Offline LLMs

    by Luciano Cabral, Maria Correia, Bárbara Lima, Hilário Oliveira, Álvaro Sobrinho, André Araújo, Rafael Araújo, Diego Dermeval, Sebastian Munoz-Najar Galvez

  • Open Access

    From VEDs to VLMs: Assessing Fine-Tuning Efficiency and Metric Reliability in Brazilian Portuguese Image Captioning

    by Gabriel Bromonschenkel, Alessandro Lameiras Koerich, Thiago Paixão, Hilário Tomaz Alves de Oliveira

    The adaptability and robustness of Visual Language Models (VLMs) to low-resource linguistic contexts remain a critical frontier in multimodal machine learning. This work evaluates the performance of state-of-the-art architectures, including ViTucano, PaliGemma, Phi-3 V, and LLaMA 3.2 V, in the context of Brazilian Portuguese image captioning (IC). By contrasting native human-annotated captions with machine-translated equivalents from the Flickr30K dataset, we delineate the performance gain and drop of using a translated corpus for fine-tuning versus its version annotated by native speakers, along with an additional comparison to the zero-shot setting. Furthermore, we demonstrate that VLMs fine-tuned under restricted conditions significantly outperform Vision Encoder-Decoder (VED) models, with ViTucano 2B, Phi-3 V, and PaliGemma achieving the highest results across translated and native data. Finally, our metrics adequacy analysis suggests that CLIP-Score is more robust than traditional unimodal metrics in distinguishing correct captions from incorrect, randomly selected captions.

  • Open Access

    Brazilian Portuguese Image Captioning with Transformers: A Study on Cross-Native-Translated Dataset

    by Gabriel Bromonschenkel, Alessandro L. Koerich, Thiago M. Paixão, Hilário Tomaz Alves de Oliveira

    Image captioning (IC) refers to the automatic generation of natural language descriptions for images, with applications ranging from social media content generation to assisting individuals with visual impairments. While most research has been focused on English-based models, low-resource languages such as Brazilian Portuguese face significant challenges due to the lack of specialized datasets and models. Several studies create datasets by automatically translating existing ones to mitigate resource scarcity. This work addresses this gap by proposing a cross-native-translated evaluation of Transformer-based vision and language models for Brazilian Portuguese IC. We use a version of Flickr30K comprised of captions manually created by native Brazilian Portuguese speakers and compare it to a version with captions automatically translated from English to Portuguese. The experiments include a cross-context approach, where models trained on one dataset are tested on the other to assess the translation impact. Additionally, we incorporate attention maps for model inference interpretation and use the CLIP-Score metric to evaluate the image-description alignment. Our findings show that Swin-DistilBERTimbau consistently outperforms other models, demonstrating strong generalization across datasets. ViTucano, a Brazilian Portuguese pre-trained VLM, surpasses larger multilingual models (GPT-4o, LLaMa 3.2 Vision) in traditional text-based evaluation metrics, while GPT-4 models achieve the highest CLIP-Score, highlighting improved image-text alignment. Attention analysis reveals systematic biases, including gender misclassification, object enumeration errors, and spatial inconsistencies. The datasets and the models generated and analyzed during the current study are available in: https://github.com/laicsiifes/transformer-caption-ptbr.

  • Open Access

    Brazilian Portuguese Image Captioning with Transformers: A Study on Cross-Native-Translated Dataset

    by Gabriel Bromonschenkel, Alessandro L. Koerich, Thiago M. Paixão, Hilário Tomaz Alves De Oliveira

    Image captioning (IC) refers to the automatic generation of natural language descriptions for images, with applications ranging from social media content generation to assisting individuals with visual impairments. While most research has been focused on English-based models, low-resource languages such as Brazilian Portuguese face significant challenges due to the lack of specialized datasets and models. Several studies create datasets by automatically translating existing ones to mitigate resource scarcity. This work addresses this gap by proposing a cross-native-translated evaluation of Transformer-based vision and language models for Brazilian Portuguese IC. We use a version of Flickr30K comprised of captions manually created by native Brazilian Portuguese speakers and compare it to a version with captions automatically translated from English to Portuguese. The experiments include a cross-context approach, where models trained on one dataset are tested on the other to assess the translation impact. Additionally, we incorporate attention maps for model inference interpretation and use the CLIP-Score metric to evaluate the image-description alignment. Our findings show that Swin-DistilBERTimbau consistently outperforms other models, demonstrating strong generalization across datasets. ViTucano, a Brazilian Portuguese pre-trained VLM, surpasses larger multilingual models (GPT-4o, LLaMa 3.2 Vision) in traditional text-based evaluation metrics, while GPT-4 models achieve the highest CLIP-Score, highlighting improved image-text alignment. Attention analysis reveals systematic biases, including gender misclassification, object enumeration errors, and spatial inconsistencies.

  • Large Language Models for chatbot applications handling sensitive information

    by Hilário Tomaz Alves de Oliveira, Álvaro Alvares De Carvalho César Sobrinho, Andrey dos Reis Cadima Dias, André Magno Costa de Araújo, Rafael Dias Araújo, Diego Dermeval Medeiros da Cunha Matos, Sebastian Munoz-Najar Galvez


  • Enhancing E-Commerce with a RAG-Powered Conversational Recommender System

    by Danilo Xavier Valle, Hilário Tomaz Alves de Oliveira

    Conversational recommender systems have emerged as a promising approach to enhancing user experience in e-commerce by enabling interactive and personalized product discovery. This paper proposes a conversational recommender system for e-commerce that employs retrieval-augmented generation (RAG) to improve product recommendations based on natural language queries. Experiments were conducted using the ESCI-S dataset, an enriched version of the Amazon ESCI shopping queries dataset, to evaluate embedding models and large language models. The goal of this study is to assess the effectiveness of an RAG-based conversational recommender system and to identify optimal configurations for enhanced performance in e-commerce applications.

  • Evaluating Machine Learning Algorithms for Software Effort Estimation from Textual Descriptions of Requirements

    by Rodrigo de Nadai Grigoleto, Hilário Tomaz Alves de Oliveira

    A estimativa de esforço no desenvolvimento de software é uma tarefa desafiadora, subjetiva, que exige considerável empenho, impactando diretamente o planejamento e a alocação de recursos. Este trabalho investigou a eficácia de diferentes algoritmos de Aprendizado de Máquina (AM) e arquiteturas de redes neurais profundas na estimativa automática do esforço de desenvolvimento a partir de descrições textuais de requisitos. Foram avaliados algoritmos tradicionais de AM usando TF-IDF, bem como modelos baseados no BERT, em seis conjuntos de dados públicos. Os experimentos analisaram o desempenho dos modelos em dois cenários: intra e inter conjuntos de dados, utilizando a métrica de erro absoluto médio. Os resultados experimentais indicaram que os modelos baseados no BERT obtiveram o melhor desempenho no cenário intra bases de dados, mas apresentaram um declínio mais acentuado no cenário inter bases de dados em comparação com os algoritmos tradicionais.

  • Open Access

    Benchmarking Large Language Models for Text-to-SQL in Brazilian Portuguese and English

    by Luís Felippe C. de Carvalho, Paulo Sérgio dos S. Júnior, Hilário Tomaz Alves de Oliveira

    This work assessed the performance of seventeen large language models, including open-source and proprietary models, on the Text-to-SQL task in both Brazilian Portuguese and English. Two schema representation strategies were considered: textual descriptions and representations using data definition language. Experimental results on the Spider dataset demonstrated the superior performance of proprietary models, particularly the Gemini-2.5-Flash-preview, as measured by the Execution Accuracy and Exact Match metrics. Among the open-source models, Qwen-2.5-Coder-14B achieved the highest performance. An error analysis of the best-performing model revealed strong proficiency in handling clauses such as SELECT and AND/OR, while considerable challenges persisted in generating more complex constructs, including GROUP BY with HAVING and set operators like UNION and INTERSECT.


  • Open Access

    Comparação de Modelos de Embeddings e LLMs para Geração Aumentada por Recuperação em Português

    by Luiz Sabiano Ferreira Medeiros, Hilário Tomaz Alves de Oliveira

    Os modelos de linguagem de larga escala (LLMs) representam um avanço para a área de processamento de linguagem natural, impulsionando o desempenho em tarefas como geração de texto e resposta a perguntas. No entanto, eles enfrentam desafios como alucinações e falta de acesso a informações atualizadas. A técnica de geração aumentada por recuperação (RAG) busca mitigar esses problemas ao integrar recuperação de informações externas à geração de texto, melhorando a precisão e a atualidade das respostas. Este trabalho realizou uma investigação de diversos modelos embeddings e LLMs de código aberto e proprietários aplicados à técnica RAG considerando três bases de dados contendo documentos escritos em português do Brasil. Os resultados experimentais demonstraram que os modelos Multilingual E5 large e Gemma 2 9B obtiveram o melhor desempenho dentre os modelos avaliados com base em diferentes medidas de avaliação.


  • Open Access

    A benchmark dataset of narrative student essays with multi-competency grades for automatic essay scoring in Brazilian Portuguese

    by Hilário Oliveira, Rafael Ferreira Mello, Péricles Miranda, Hyan Batista, Moésio Wenceslau da Silva Filho, Thiago Cordeiro, Ig Ibert Bittencourt, Seiji Isotani