FinQA-Table-random-DeBERTa-Reranker

A passage reranker for the HiREC framework, fine-tuned from naver/trecdl22-crossencoder-debertav3 on table data from the FinQA training set. General-purpose rerankers often fail to capture table-specific cues (titles, periods, indicators) that matter more than raw numerical values; this model is adapted to address that gap.

• 📄 Paper: ACL 2025 Findings
• 💻 Code: LOFin-bench-HiREC

Training Data

Constructed from the FinQA training set, where each question is paired with an evidence page containing the gold table.

• Positive passages: tables located on the evidence page of each question.
• Negative passages: tables sampled from pages other than the evidence page within the same document (random negative sampling).
• For each positive, n_neg = 8 negatives are drawn.

Training Setup

• Base model: naver/trecdl22-crossencoder-debertav3
• Objective: Binary cross-entropy on (query, passage) pairs; the cross-encoder applies an internal sigmoid, producing relevance scores in [0, 1].
• Batch size: 128 / Epochs: 3 / Learning rate: 2e-7
• Hardware: 1× NVIDIA GeForce RTX 4090

Citation

@inproceedings{choe-etal-2025-hierarchical,
    title     = {Hierarchical Retrieval with Evidence Curation for Open-Domain Financial Question Answering on Standardized Documents},
    author    = {Choe, Jaeyoung and Kim, Jihoon and Jung, Woohwan},
    booktitle = {Findings of the Association for Computational Linguistics: ACL 2025},
    year      = {2025},
    url       = {https://aclanthology.org/2025.findings-acl.855/}
}