The question
Can rephrasing questions and aggregating candidates improve verbatim extraction of causes and effects from financial texts?
We study English and Spanish text spans from FinCausal 2026. Answers must occur verbatim in the context. The task extracts what the text states; it does not establish economic causality or forecast prices.
Method and baseline
The system uses five sequential stages: causal analysis, question rephrasing, candidate generation, iterative aggregation and verbatim validation with word-boundary alignment. The generative agents use Gemini 3 Flash Preview, with candidates also supplied by a reading-comprehension model. The baseline extracts in one pass using the same language model (PDF pages 2–4, printed pages 140–142).
Development results
Table 2 declares 20 questions per language, drawn from the released training set. These are development samples, not an independent test. F1 measures the overlap between the extracted span and the reference answer. Only the baseline and full pipeline are shown.
| Language and system | F1 (%) |
|---|---|
| English · baseline | 82.0 |
| English · VERSA | 87.7 |
| Spanish · baseline | 75.5 |
| Spanish · VERSA | 79.4 |
Some component tests in the paper are preliminary estimates rather than complete runs. See the paper for the remaining results.
Official evaluation
The paper also describes an official blind evaluation, using an adequacy score that differs from F1. The improvements over the baseline shown here come from development samples and do not establish an equivalent improvement in that evaluation.
Official results: Table 3, PDF page 5, printed page 143.
| Language | Score (1–5) | Rank |
|---|---|---|
| English | 4.404 | 173 |
| Spanish | 4.336 | 152 |
Cost and limitations
The paper reports approximately 39.1 million tokens and 56,000 API calls, with an estimated cost of USD 27.3, across the full evaluation, including development and both official submissions (PDF page 6, printed page 144). These are figures reported in the paper, without independent measurement. They are separate from the costs of trading in markets.
- Development samples are small and some component tests are preliminary estimates.
- Official adequacy scores and development F1 measure different things and cannot be compared directly.
- The system depends on the language model and requires many calls. Verbatim output alone does not guarantee causal accuracy.
- The study does not examine cryptocurrency or portfolio decisions. A complete reproduction has not been verified.
Paper and citation
Aldan Jay; Rafael Berlanga; Yoelvis Moreno; Vicent Santamarta. (2026). VERSA: Verbatim Extraction via Rephrasing and Self-Aggregation for Financial Causality. The 7th Financial Narrative Processing Workshop @ LREC 2026 · ELRA · 139–145.
DOI: 10.63317/523w58xpfhbz · CC BY-NC 4.0
BibTeX
@inproceedings{jay2026versa,
title = {{VERSA}: Verbatim Extraction via Rephrasing and Self-Aggregation for Financial Causality},
author = {Jay, Aldan and Berlanga, Rafael and Moreno, Yoelvis and Santamarta, Vicent},
booktitle = {The 7th Financial Narrative Processing Workshop},
publisher = {European Language Resources Association (ELRA)},
year = {2026},
month = may,
pages = {139--145},
doi = {10.63317/523w58xpfhbz},
url = {https://lrec.elra.info/lrec2026-ws-fnp-14},
note = {CC BY-NC 4.0}
}