Reproduce “PublicHearingBR: A Brazilian Portuguese Dataset of Public Hearing Transcripts for Summarization of Long Documents” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Leandro Carísio Fernandes, Guilherme Zeferino Rodrigues Dobins, Roberto Lotufo, Jayr Alencar Pereira

Researchers from the Brazilian Chamber of Deputies, Unicamp, and UFCA introduce PublicHearingBR, the first open dataset for summarizing long public hearing transcripts in Brazilian Portuguese. The work establishes a baseline for opinion extraction using ChatGPT and develops a robust LLM-based hallucination detection methodology, identifying 11.89% possible hallucinations and achieving up to 90.87% true positive detection.

Key insights

  • The creation of PublicHearingBR directly addresses a critical gap by providing the first open dataset for long, multi-speaker document summarization in Brazilian Portuguese.
  • Hallucination detection using LLMs is significantly influenced by prompt design, with more specific prompts leading to higher true positive rates but also increased false positives.
  • LLM-generated summaries, even from general-purpose models, can contain a notable percentage of factual inaccuracies (estimated 11.89% in this study), underscoring verification challenges.
  • The manual annotation for hallucination detection resulted in a new, valuable Natural Language Inference (NLI) dataset for Brazilian Portuguese.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.