liaad/facta-pt-relevance-bertimbau-base

FACTA-PT Document Relevance — BERTimbau-base

This repository contains the BERTimbau-base model fine-tuned for document relevance classification in FACTA-PT.

Model Description

The model is derived from neuralmind/bert-base-portuguese-cased and was fine-tuned to determine whether a retrieved document is relevant to the verification of a given claim span.

Task

The model receives a claim span-document pair and predicts one of two classes:

• Relevant
• Irrelevant

Document relevance is treated independently from Claim-Evidence Entailment. A document classified as Relevant may subsequently contain Supported, Refuted, or NEI evidence.

Base Model

• Base model: neuralmind/bert-base-portuguese-cased
• Model: BERTimbau-base
• Base-model language: Brazilian Portuguese
• Base-model license: MIT

Training Data

The model was fine-tuned using the CLEVER external-evidence annotations developed within the associated dissertation.

Intended Use

The model is intended to filter retrieved candidate documents before passage-level Claim-Evidence Entailment in European Portuguese automated fact-checking.

It does not determine whether a claim is Supported, Refuted, or NEI.

Limitations

False negatives at this stage may remove relevant evidence from subsequent analysis.

License

This fine-tuned model is released under the MIT License, following the license of the base model neuralmind/bert-base-portuguese-cased.

FACTA-PT

This model is one of the components developed and selected for FACTA-PT.