Fin_Agent at FinEval 2026: Task-Aware Methods for Multilingual Financial NLP

Mohotarema Rashid; Tanzina Hossain; Xiaoying Song; Lingzi Hong

cs.CLOct 8, 2026

Fin_Agent at FinEval 2026: Task-Aware Methods for Multilingual Financial NLP

Authors: Mohotarema Rashid, Tanzina Hossain, Xiaoying Song, Lingzi Hong

Organizations: University of North Texas

Abstract

We present Fin_Agent, our submission to the FinEval 2026 shared task on multilingual fi nancial language understanding. Our sys tems address Greek financial named-entity recognition, Japanese implicit communication recognition, and multilingual financial ques tion answering. For Greek NER, we com bine annotation-aware numeric rules with zero shot Qwen2.5-32B classification, achieving 0.9521 public F1 and 0.8884 private F1. For Japanese implicit communication, an inter pretable cue-based ordinal scorer achieves 98.0% test accuracy. For multilingual financial QA, BGE-M3retrieval with Qwen3-32B gen eration achieves 0.3023 and 0.3419 ROUGE-1 F1 on the public and private splits. Key limita tions include dataset-specific rules, handcrafted cues, and limited supervision. Implementa tion code is available https://github.com/ Rashidnira/finagent-fineval-2026.

Figures & tables

Evaluation	System	F1	Rank
Public test	Hybrid system	0.9521	4
Private test	Hybrid system	0.8884	6
Validation	XLM-R-LARGE	0.9287	–
Validation	Rules (ours)	0.9862	–

Table 1: Subtask 1 official results and numeric-NER ablation.

Task	Public	Private	Metric
Subtask 2 Japanese ICR	100.0%	97.14%	Accuracy

Table 2: Official Subtask 2 results.

Split	F1	Precision	Recall
Public (24)	0.3023	0.3040	0.3528
Private (52)	0.3419	0.3227	0.3899

Table 3: Official Subtask 3 results.

Configuration	Words	F1	P	R
Grouped, no retrieval	37.6	.2618	.344	.253
Per-question + BGE-M3†	91.8	.2987	.274	.386
Per-question + 55-word	61.2	.3022	.321	.329
Per-question + 72-word	71.8	.2944	.293	.342
Rewrite + normalization	69.4	.3023	.304	.353

Table 4: Public-set submission variants.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Split	Precision	Recall	Micro-F1
Training (320)	0.9877	0.9683	0.9779
Validation (80)	0.9882	0.9843	0.9862

Appendix

Table 5: Subtask 1 numeric-rule performance.

Family	Submitted Japanese cues (regular-expression alternatives)
DECIDED	決議し/いた/済, 決定し/いた/済, 承認し/いた/済, 締結し/済/いた, 確定して/しま/済, 正式に決, 可決, 合意に達し, 署名し, 決まりました, 取締役会で決議/承認/確認/決定
DONE	完了します/しました/いたします, 開始します, 実施します, 移行します, 適用します, 充当します, 返済します, 継続します, 実行します, 導入します, 払い戻します, 上場します, 着工します
NEG_HARD	いいえ, 予定/計画/こと/事はありません・ございません, する考えはありません, 意思/方針はありません, しません, いたしません, 行いません, 実施しません, 提供しない, 検討していません, 中止, 終了します, 撤回, 除外して, 白紙, 社内規程, お答えできません, 回答を差し控え, 開示する考えはない, ノーコメント
NEG_SOFT	難しい, 困難と, 見送る/見送り, 時期尚早, 優先順位は低い, 考えていません, 踏み切る考えは, 慎重に, 控えます, 差し控え
REOPEN	ただし/但し, 一方で, 改めて検討/判断/見直し, 再検討, 見直します, 余地は/がある, 状況が変われば, 環境が変わる, 整った段階, 数年後, 将来的には検討, 選択肢の一つ, 高まれば, 柔軟に
VOLIT	たいと考え, たいと思う/です, 努めて, 目指し/目指す, 注力して, 進めてまいり, 拡大したい, 引き上げたい, 検討してまいり, 検討を進め, 前向き, 想定している, 判断します, 排除しておらず

Appendix

Table 6: Compact rendering of the exact Subtask 2 regex inventory. Slash-separated forms denote alternatives; the released source preserves the precise grouping and suffixes.

Decision rule	Accuracy	Macro-recall
Raw rule-score argmax	0.3846	0.4000
Sinkhorn, strength 1.0	0.2692	0.3000

Appendix

Table 7: Subtask 2 results on the fixed 26-item development split.

Component	Role	Parameters
Qwen3-32B	Generation/rewriting	32,762,123,264
BGE-M3	Passage retrieval	567,754,752
Deterministic code	Chunking/validation	0
Total		33,329,878,016

Appendix

Table 8: Learned components and total parameter count for Subtask 3.

Label	Generation	Retrieval
G	Grouped (4 questions/call)	None
G+R	Grouped	Per-question evidence
Q	One call per question	None
Q+R	One call per question	Evidence before question

Appendix

Table 9: Controlled-analysis configurations.

Configuration	Default F1	Multiscript F1
G	$0.4547\pm 0.0151$	$0.4396\pm 0.0191$
G+R	$0.4525\pm 0.0103$	$0.4391\pm 0.0097$
Q	$0.4564\pm 0.0054$	$0.4383\pm 0.0087$
Q+R	$0.4559\pm 0.0080$	$0.4406\pm 0.0092$

Appendix

Table 10: Development ROUGE-1 F1, mean and standard deviation over five seeds.

Contrast	Default	Multiscript
G+R $-$ G	$-0.0022\pm 0.0239$	$-0.0005\pm 0.0269$
Q+R $-$ Q	$-0.0005\pm 0.0072$	$+0.0023\pm 0.0141$
Q $-$ G	$+0.0017\pm 0.0195$	$-0.0013\pm 0.0231$
(Q+R) $-$ (G+R)	$+0.0034\pm 0.0113$	$+0.0015\pm 0.0110$
Interaction	$+0.0017\pm 0.0250$	$+0.0028\pm 0.0273$

Appendix

Table 11: Within-seed paired effects on development ROUGE-1 F1 (mean

\pm

s.d. over five seeds).

Configuration	Mean words	Ratio to reference
Reference	47.9	1.00
G	33.3	0.70
G+R	35.3	0.74
Q	51.6	1.08
Q+R	55.8	1.16

Appendix

Table 12: Mean answer length in the controlled development analysis.

Configuration	Revenue	Balance sheet	Cash flow	R&D
G	.422	.460	.312	.625
G+R	.420	.464	.309	.617
Q	.424	.483	.340	.578
Q+R	.442	.482	.308	.591

Appendix

Table 13: Development F1 by question type, default scorer, averaged over five seeds. Each type contains 19 instances.

Flagged rows	Raw	+Rewr.	+Len.	+Lang.
Untraceable numbers	5	2	2	2
Non-verbatim quotes	4	4	5	5
Missing evidence label	0	0	6	6
Rows with any flag	19	8	11	11

Appendix

Table 14: Validator flags after each stage. Values are numbers of rows. The final row reports the union of all flagged rows, including checks not listed separately, chiefly answers longer than 110 words at raw generation. Because one answer may carry multiple flags, this value need not equal the sum of the three displayed classes. The label check tests for an exact string and is reliable; the number and quotation checks are heuristic and may flag derived figures or normalised wording.

Stage	Mean words	Max words	Over 100
Raw generation	91.8	153	23
+ Rewrite	73.7	126	7
+ Length guard	69.4	99	0
+ Language guard	69.4	99	0

Appendix

Table 15: Answer length after each stage of the submitted post-processing pipeline. The two compression variants in Table 4 were produced by separate scripts and are not sequential stages of this pipeline.

Financial QA

Financial question answering develops systems that answer questions about reports, markets, regulations, and other financial data with reliable evidence and reasoning.

See all 51 papers

Overview of FinMMEval 2026 Task 2: Multilingual Financial Short-Answer Question Answering
Jul 22, 2026Zhuohan Xie, Xueqing Peng, Georgi Georgiev +18Multilingual Language Model Evaluation Financial QA
Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, United Arab Emirates · The Fin AI, United States · FMI, Sofia University “St. Kliment Ohridski”, Sofia, Bulgaria +7
Language-Routed RAG and Direct Option Scoring for Multilingual Financial QA: DS@GT at FinMMEval
Jul 24, 2026Justice Ayela, Kabir SahniFinance Multilingual Language Model Evaluation
Georgia Institute of Technology, North Ave NW, Atlanta, GA 30332
Overview of FinMMEval 2026 Task 1: Multilingual Financial Multiple-Choice Question Answering
Jul 22, 2026Zhuohan Xie, Yuyang Dai, Rania Elbadry +18Multilingual Language Model Evaluation Multiple-Choice Question Answering
Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, United Arab Emirates · INSAIT, Sofia University “St. Kliment Ohridski”, Sofia, Bulgaria · University of Arizona, Tucson, United States +6

Jul 22, 2026J/K move · Enter open · S save

Overview of FinMMEval 2026 Task 2: Multilingual Financial Short-Answer Question Answering

Zhuohan Xie, Xueqing Peng, Georgi Georgiev, Dimitar Dimitrov+17

Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, United Arab Emirates · The Fin AI, United States · FMI, Sofia University “St. Kliment Ohridski”, Sofia, Bulgaria+7

Multilingual Language Model Evaluation Financial QA Multilingual QA Finance Financial Language Model Evaluation

Open paper AskPDF arXiv

FinMMEval 2026 Task 2 evaluates short-answer financial question answering over multilingual evidence. Each final-test item pairs an English question with financial statements and news in English, Chinese, Japanese, Spanish, and Greek. Participating systems submit one concise answer per item in JSONL format. The final-test set contains 256 items, split evenly between easy and expert tiers; each tier contains four question templates instantiated over 32 company-report groups. Gold answers were withheld during submission, and systems were ranked by macro-averaged item-level ROUGE-1 F1 against organizer-held reference answers. The final leaderboard includes 12 ranked submissions. The strongest systems are closely clustered, with the top four separated by less than one percentage point in ROUGE-1 F1. The submitted system papers document retrieval-augmented generation, cross-lingual evidence handling, structured prompting, answer compression, and validation strategies.

Abstract

Financial QA

Explore similar work

Overview of FinMMEval 2026 Task 2: Multilingual Financial Short-Answer Question Answering

Language-Routed RAG and Direct Option Scoring for Multilingual Financial QA: DS@GT at FinMMEval

Overview of FinMMEval 2026 Task 1: Multilingual Financial Multiple-Choice Question Answering