DBDennis Giovani Balreira

Dennis Giovani Balreira

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Dennis Giovani Balreira has a total of 31 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Efficient Table Structure Recognition via Deep Learning and Lightweight Heuristics

    by João Speranza Pastorello, Gustavo Lopes Tamiosso, Gabriel Bauer de Oliveira, Albert Geiger, Evandro Dalbosco da Silveira, Dennis Giovani Balreira, Anderson Rocha Tavares

  • Time to Rethink NER Splits

    by Vítor Bernardo Silveira Pereira, Rafael Oleques Nunes, Dennis Giovani Balreira

  • Open Access

    Survey of Brazilian Open Budget Data Portals: Query Interfaces and Dashboards

    by Kaline B. F. Mesquita, Dennis G. Balreira, Andre S. Spritzer, Carla M. D. S. Freitas

    To promote transparency, the Brazilian government provides access to public data through web portals featuring query interfaces and dashboards. While query interfaces are used by more experienced users to gather data for further analyses, dashboards that include visualizations help a broader audience consult and explore data. A domain of particular complexity that benefits from the use of these interfaces is government spending and budgets. This study analyzes dashboards and query interfaces of government budget data through qualitative research based on a survey. Focusing on Brazil's budget transparency initiative, we examined 83 interfaces in total: 30 dashboards and 53 query interfaces from federal, state, and major city governments. This survey assesses these interfaces using design patterns for general-purpose dashboards and design principles for open government data dashboards. Our findings reveal a critical weakness: while most portals provide access to budget data, they largely neglect user-centered design, failing to provide the necessary context or consider the data literacy of their audience. This creates a significant "transparency gap'' that undermines genuine accountability and demonstrates the need for a fundamental shift in the design of these essential public tools.

  • Open Access

    A Música Brasileira na Ditadura Militar: uma análise de tópicos com BERTopic e GSDMM

    by Henry R. Piceni, Pedro V. Alexandre, Dennis G. Balreira

    Durante a ditadura militar no Brasil, artistas recorreram à música como forma de expressão, valendo-se de uma linguagem poética e metafórica. Nesse contexto, a análise sociopolítica dessas letras apresenta-se como um desafio interpretativo. Este trabalho busca analisar letras de músicas brasileiras lançadas durante este período, utilizando BERTopic e GSDMM para modelagem de tópicos, a fim de identificar temas-chave que refletem aspectos sociais, políticos e históricos do período. Os modelos analisados revelaram uma linguagem simples e cotidiana, com pouco uso de vocabulário erudito ou político, sugerindo que os artistas optaram por uma expressão direta e emocional, possivelmente para ampliar o diálogo com o público. Particularmente, o BERTopic destacou-se por mapear a diversidade temática com pouca redundância, enquanto o GSDMM segmentou temas dominantes em subtópicos altamente coesos. Este trabalho mostra como a música brasileira, por meio de uma linguagem poética e acessível, retratou o cotidiano da época e revelou o potencial da modelagem de tópicos como ferramenta para a análise cultural.


  • Open Access

    LegiSubjects-Br: A classification corpus for estimating subjects of Brazilian legislative bills

    by Rafael Oleques Nunes, Andre Suslik Spritzer, Carla Maria Dal Sasso Freitas, Dennis Giovani Balreira

    This paper proposes a new corpus for estimating subjects for bills proposed at the lower house of Brazil's legislature, the Chamber of Deputies. Politics and legislation inherently include a vast amount of jargon-filled documents. The legislative process, in particular, also involves several steps that bills have to go through before they become law. This necessarily produces a large volume of human-entered data that often lacks important information. Considering the Brazilian Chamber of Deputies from 1991 to 2022, around 75% of the bills do not have subject classification in their associated metadata. However, due to the sheer amount of bills, this scenario is well-suited for a solution based on machine learning and natural language processing. We present a solution that introduces and compares two BERT models adapted for the Portuguese language using bill summaries and keywords, which consist of brief descriptions or overviews of the main points of the documents. We obtained our best results using the BERTimbau model variation, achieving 81.59% of the weighted F1 score and 75.06% of the macro F1 score. To the best of our knowledge, this is the first work to propose a corpus and a model for predicting the subjects of bills proposed at the Brazilian Chamber of Deputies. Our approach encourages researchers to explore similar techniques for other legal documents and potentially aids political scientists in conducting a more robust data analysis than what was possible with the previous data due to the frequent absence of metainformation.

  • An Ensemble of LLMs Finetuned with LoRA for NER in Portuguese Legal Documents

    by Rafael Oleques Nunes, Letícia Maria Puttlitz, Antonio Oss Boll, Andre Spritzer, Carla Maria Dal Sasso Freitas, Dennis Giovani Balreira, Anderson Rocha Tavares


  • Assessing European and Brazilian Portuguese LLMs for NER in Specialised Domains

    by Rafael Oleques Nunes, Joaquim Santos, Andre Spritzer, Dennis Giovani Balreira, Carla Maria Dal Sasso Freitas, Fernanda Olival, Helena Freire Cameron, Renata Vieira


  • Open Access

    Reconhecimento de Entidades Nomeadas e Vazamento de Dados em Textos Legislativos

    by Rafael Oleques Nunes, André Susliz Spritzer, Carla Maria Dal Sasso Freitas, Dennis Giovani Balreira

    Este trabalho trata do vazamento de dados no treinamento de modelos de Reconhecimento de Entidades Nomeadas (NER) em textos legislativos em português brasileiro, resultante de duplicatas e anotações inconsistentes, o que compromete a avaliação dos modelos. Após corrigir esse vazamento no corpus UlyssesNER-Br, foi realizado um novo benchmark, comparando os resultados com estudos anteriores em um cenário mais confiável. Também foi reavaliada uma abordagem semissupervisionada utilizando autoaprendizado e amostragem ativa. No entanto, ao reutilizar um threshold fixo, escolhido a partir de uma nuvem de valores antes da correção, os resultados foram insatisfatórios. Isso indica que um threshold dinâmico, que se adapte às características dos dados pós-correção, poderá proporcionar uma avaliação mais eficiente e precisa, indicando a necessidade de futuros estudos sobre a escolha de thresholds.


  • Open Access

    Ancient Greek’s New Technological Muse: Extracting Topoi in the Anacreontea with LLMs

    by Rafael O. Nunes, João G. Zandoná, Júlia V. Maia, Andre Spritzer, Dennis G. Balreira, Carla M. D. S. Freitas

    Natural Language Processing, along with Large Language Models (LLMs), holds significant potential in the domain of literature, leveraging its computational capabilities to analyze and comprehend human language. These techniques prove to be particularly useful in a specific part of Greek literature called Anacreaontea, a collection of poems emulating the style of the 6thcentury BCE Greek poet Anacreon. This paper presents an LLM approach to automatically classify Anacreontea poems in their respective topoi. Our methodology explores two well-established autoregressive language models (LLama 2 and Mistral) and investigates the use of contextual prompting in this scenario. We also provide an annotated corpus with 21 fragments of the Anacreontea with topos for Greek and Portuguese text.