Gabriel Lino Garcia
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Gabriel Lino Garcia has a total of 21 co-authors. Here are the top co-authors by number of shared publications:
- João Paulo Papa
São Paulo State University (UNESP)
8 shared publications - Pedro Henrique Paiola
Sao Paulo State University
8 shared publications - João Renato Ribeiro Manesco
São Paulo State University (UNESP)
5 shared publications - Ana Lara Alves Garcia
Sao Paulo State University
1 shared publication - André da Fonseca Schuck
Sao Paulo State University
1 shared publication - Aparecido Nilceu Marana
São Paulo State University (UNESP)
1 shared publication
Latest Publications
- Open Access
- 17 January 2025
- Published by Sociedade Brasileira de Computação in Journal of the Brazilian Computer Society
- Vol. 31 (1) , 1031-1049
Abstractive Summarization with LLMs for Texts in Brazilian Portuguese
This study aims to compare large language models (LLMs) in the task of text summarization for Portuguese-language texts. A dataset of 8,116 samples was used, containing the original texts and their corresponding reference summaries. Initially, an experiment was conducted comparing three different prompts using zero-shot, one-shot, and few-shot techniques, processing 100 samples for four out of the six models (those that accept instructions as part of their input). The goal of this preliminary experiment was to determine an optimal prompt for conducting the full-scale experiment. After selecting the prompt, a second experiment was performed, running all six models on the 8,116 samples and evaluating summarization quality using metrics such as BLEU and ROUGE, as well as Compression Rate and Inference Time for the generated summaries. Finally, an experiment was conducted to analyze the impact of 4-bit and 8-bit quantization, assessing how these different configurations affect the generated summaries, evaluation metrics, Compression Rate, and Inference Time.
- 17 January 2025
- Published by Sociedade Brasileira de Computação in Journal of the Brazilian Computer Society
- Vol. 31 (1) , 1031-1049
Open AccessAbstractive Summarization with LLMs for Texts in Brazilian Portuguese
This study aims to compare large language models (LLMs) in the task of text summarization for Portuguese-language texts. A dataset of 8,116 samples was used, containing the original texts and their corresponding reference summaries. Initially, an experiment was conducted comparing three different prompts using zero-shot, one-shot, and few-shot techniques, processing 100 samples for four out of the six models (those that accept instructions as part of their input). The goal of this preliminary experiment was to determine an optimal prompt for conducting the full-scale experiment. After selecting the prompt, a second experiment was performed, running all six models on the 8,116 samples and evaluating summarization quality using metrics such as BLEU and ROUGE, as well as Compression Rate and Inference Time for the generated summaries. Finally, an experiment was conducted to analyze the impact of 4-bit and 8-bit quantization, assessing how these different configurations affect the generated summaries, evaluation metrics, Compression Rate, and Inference Time.
- Open Access
- 17 January 2025
- Published by Sociedade Brasileira de Computação in Journal of the Brazilian Computer Society
- Vol. 31 (1) , 918-939
The Bode Family of Large Language Models: Investigating the Frontiers of LLMs in Brazilian Portuguese
The rapid advancement of Large Language Models (LLMs) has significantly impacted Natural Language Processing, yet their effectiveness remains uneven across languages. Most state-of-the-art models are trained predominantly in English, leading to performance disparities in lower-resource languages such as Brazilian Portuguese (BP). This paper explores fine-tuning strategies for adapting open-weight LLMs to BP, focusing on dataset translation techniques, linguistic adaptation challenges, and parameter-efficient fine-tuning methods, such as LoRA and Q-LoRA. We present a benchmark analysis evaluating multiple fine-tuning approaches across various open models, establishing a guiding framework for future BP-specific adaptations. Our results showcase the importance of specialized fine-tuning in improving cross-lingual transfer and NLP performance in BP, contributing to the broader goal of enhancing multilingual language model accessibility.
- 17 January 2025
- Published by Sociedade Brasileira de Computação in Journal of the Brazilian Computer Society
- Vol. 31 (1) , 918-939
Open AccessThe Bode Family of Large Language Models: Investigating the Frontiers of LLMs in Brazilian Portuguese
The rapid advancement of Large Language Models (LLMs) has significantly impacted Natural Language Processing, yet their effectiveness remains uneven across languages. Most state-of-the-art models are trained predominantly in English, leading to performance disparities in lower-resource languages such as Brazilian Portuguese (BP). This paper explores fine-tuning strategies for adapting open-weight LLMs to BP, focusing on dataset translation techniques, linguistic adaptation challenges, and parameter-efficient fine-tuning methods, such as LoRA and Q-LoRA. We present a benchmark analysis evaluating multiple fine-tuning approaches across various open models, establishing a guiding framework for future BP-specific adaptations. Our results showcase the importance of specialized fine-tuning in improving cross-lingual transfer and NLP performance in BP, contributing to the broader goal of enhancing multilingual language model accessibility.
- Open Access
- 17 January 2025
- Published by Sociedade Brasileira de Computação in Journal of the Brazilian Computer Society
- Vol. 31 (1) , 885-917
Evaluating Large Language Models for Brazilian Portuguese Sentiment Analysis: A Comparative Study of Multilingual State-of-the-Art vs. Brazilian Portuguese Fine-Tuned LLMs
This study presents an extensive comparative analysis of Large Language Models (LLMs) for sentiment analysis in Brazilian Portuguese texts. We evaluated 23 LLMs—comprising 13 state-of-the-art multilingual models and 10 models specifically fine-tuned for Portuguese—across 12 public annotated datasets from diverse domains, employing the in-context learning paradigm. Our findings demonstrate that large-scale models such as Claude-$3.5$-Sonnet, GPT-4o, DeepSeek-V3, and Sabiá-3 delivered superior results with accuracies exceeding 92%, while smaller models (7-13B parameters) also showed compelling performance with top performers achieving accuracies above 90%. Notably, linguistic specialization through fine-tuning demonstrated mixed results—significantly reducing hallucination rates for some models but not consistently yielding performance improvements across all model types. We also observed that newer model generations frequently outperformed their predecessors, and in the one dataset where traditional machine learning methods were employed by the original authors for sentiment classification, all evaluated LLMs substantially surpassed these traditional approaches. Moreover, smaller-scale models exhibited a tendency toward overgeneration despite explicit instructions. These findings contribute valuable insights to the discourse on language-specific model optimization and establish empirical benchmarks for both multilingual and Portuguese-specialized LLMs in sentiment analysis tasks.
- 17 January 2025
- Published by Sociedade Brasileira de Computação in Journal of the Brazilian Computer Society
- Vol. 31 (1) , 885-917
Open AccessEvaluating Large Language Models for Brazilian Portuguese Sentiment Analysis: A Comparative Study of Multilingual State-of-the-Art vs. Brazilian Portuguese Fine-Tuned LLMs
This study presents an extensive comparative analysis of Large Language Models (LLMs) for sentiment analysis in Brazilian Portuguese texts. We evaluated 23 LLMs—comprising 13 state-of-the-art multilingual models and 10 models specifically fine-tuned for Portuguese—across 12 public annotated datasets from diverse domains, employing the in-context learning paradigm. Our findings demonstrate that large-scale models such as Claude-$3.5$-Sonnet, GPT-4o, DeepSeek-V3, and Sabiá-3 delivered superior results with accuracies exceeding 92%, while smaller models (7-13B parameters) also showed compelling performance with top performers achieving accuracies above 90%. Notably, linguistic specialization through fine-tuning demonstrated mixed results—significantly reducing hallucination rates for some models but not consistently yielding performance improvements across all model types. We also observed that newer model generations frequently outperformed their predecessors, and in the one dataset where traditional machine learning methods were employed by the original authors for sentiment classification, all evaluated LLMs substantially surpassed these traditional approaches. Moreover, smaller-scale models exhibited a tendency toward overgeneration despite explicit instructions. These findings contribute valuable insights to the discourse on language-specific model optimization and establish empirical benchmarks for both multilingual and Portuguese-specialized LLMs in sentiment analysis tasks.
- Open Access
- 17 November 2024
- Published by Sociedade Brasileira de Computação
- p. 707-718
Comparison of Relation Extraction Models for Generating Knowledge Graphs in the Oil Drilling Domain
Grafos de Conhecimento são estruturas que oferecem conhecimento explícito, raciocínio simbólico e resultados interpretáveis, além de poderem ser aprimoradas com o tempo. Por isso, a extração automática de relações entre entidades a partir de textos não estruturados é uma das atuais áreas de pesquisa do Processamento de Linguagem Natural. Neste artigo, selecionamos três modelos de extração de relações do estado da arte treinados em domínio geral e comparamos seus resultados obtidos em um dataset da área de perfuração de poços de petróleo. Os resultados demonstraram a baixa eficácia dos modelos gerais quando aplicados à linguagem técnica deste domínio.
- 17 November 2024
- Published by Sociedade Brasileira de Computação
- p. 707-718
Open AccessComparison of Relation Extraction Models for Generating Knowledge Graphs in the Oil Drilling Domain
Grafos de Conhecimento são estruturas que oferecem conhecimento explícito, raciocínio simbólico e resultados interpretáveis, além de poderem ser aprimoradas com o tempo. Por isso, a extração automática de relações entre entidades a partir de textos não estruturados é uma das atuais áreas de pesquisa do Processamento de Linguagem Natural. Neste artigo, selecionamos três modelos de extração de relações do estado da arte treinados em domínio geral e comparamos seus resultados obtidos em um dataset da área de perfuração de poços de petróleo. Os resultados demonstraram a baixa eficácia dos modelos gerais quando aplicados à linguagem técnica deste domínio.
- 17 November 2024
- Published by Springer Nature
- p. 228-243
GemBode and PhiBode: Adapting Small Language Models to Brazilian Portuguese
- 17 November 2024
- Published by Springer Nature
- p. 228-243
GemBode and PhiBode: Adapting Small Language Models to Brazilian Portuguese
- 17 November 2024
- Published by Springer Nature
- p. 213-227
Impact of Quantization on Large Language Models for Portuguese Classification Tasks
- 17 November 2024
- Published by Springer Nature
- p. 213-227
Impact of Quantization on Large Language Models for Portuguese Classification Tasks
- 17 November 2024
- Published by Springer Nature
- p. 244-255
Hate Speech Detection in Portuguese Using BERTimbau
- 17 November 2024
- Published by Springer Nature
- p. 244-255
Hate Speech Detection in Portuguese Using BERTimbau
- Open Access
Adapting LLMs for the Medical Domain in Portuguese: A Study on Fine-Tuning and Model Evaluation
This study evaluates the performance of large language models (LLMs) as medical agents in Portuguese, aiming to develop a reliable and relevant virtual assistant for healthcare professionals. The HealthCareMagic-100k-en and MedQuAD datasets, translated from English using GPT-3.5, were used to fine-tune the ChatBode-7B model using the PEFT-QLoRA method. The InternLM2 model, with initial training on medical data, presented the best overall performance, with high precision and adequacy in metrics such as accuracy, completeness and safety. However, DrBode models, derived from ChatBode, exhibited a phenomenon of catastrophic forgetting of acquired medical knowledge. Despite this, these models performed frequently or even better in aspects such as grammaticality and coherence. A significant challenge was low inter-rater agreement, highlighting the need for more robust assessment protocols. This work paves the way for future research, such as evaluating multilingual models specific to the medical field, improving the quality of training data, and developing more consistent evaluation methodologies for the medical field.
Open AccessAdapting LLMs for the Medical Domain in Portuguese: A Study on Fine-Tuning and Model Evaluation
This study evaluates the performance of large language models (LLMs) as medical agents in Portuguese, aiming to develop a reliable and relevant virtual assistant for healthcare professionals. The HealthCareMagic-100k-en and MedQuAD datasets, translated from English using GPT-3.5, were used to fine-tune the ChatBode-7B model using the PEFT-QLoRA method. The InternLM2 model, with initial training on medical data, presented the best overall performance, with high precision and adequacy in metrics such as accuracy, completeness and safety. However, DrBode models, derived from ChatBode, exhibited a phenomenon of catastrophic forgetting of acquired medical knowledge. Despite this, these models performed frequently or even better in aspects such as grammaticality and coherence. A significant challenge was low inter-rater agreement, highlighting the need for more robust assessment protocols. This work paves the way for future research, such as evaluating multilingual models specific to the medical field, improving the quality of training data, and developing more consistent evaluation methodologies for the medical field.
- 16 March 2022
- Published by Springer Nature
- p. 57-67
FakeRecogna: A New Brazilian Corpus for Fake News Detection
- 16 March 2022
- Published by Springer Nature
- p. 57-67
FakeRecogna: A New Brazilian Corpus for Fake News Detection