Bidirectional Reasoning Supervision for Multilingual Financial Decision Making

✅ Accepted at EMNLP 2025 (Industry Track) · Pages 1576–1587, Suzhou, China

📄 Paper: ACL Anthology · PDF
💻 Code: github.com/milab-nsu/FinR-M

Abstract

We explore the potential of generative Large Language Models (LLMs) for classifying financial sustainability across four diverse languages: English, Hindi, Bengali, and Telugu — representing low, medium, and high-resource languages. We introduce bidirectional reasoning supervision that enhances classification by supervising the model with three signals simultaneously: the classification label (Y), positive reason (R+), and negative reason (R−). This contrastive-style reasoning framework improves interpretability, robustness, and cross-lingual generalization in multilingual financial NLP. Smaller models fine-tuned with this approach achieve competitive or superior results compared to significantly larger models trained with conventional label-only methods.

Introduction

Financial sustainability classification requires nuanced reasoning about corporate disclosures. Standard fine-tuning trains only on labels, ignoring the rationale behind decisions. Unidirectional reasoning adds only positive rationales — why a statement IS sustainable — but fails to teach the model what makes a statement NOT sustainable. Our key insight: training with both positive and negative rationales gives a contrastive learning signal that better aligns model behavior with expert financial reasoning.

Datasets used:

• English: MultiFin shared task (Kang & El Maarouf, 2022) — sustainable vs. unsustainable classification
• Hindi, Bengali, Telugu: IndicFinNLP Task 2 (Ghosh et al., 2024)
• Train/test split: 90:10 for all languages

🎯 Key Contributions

• Bidirectional Reasoning Supervision — Novel training objective: [label + R+ + R−] jointly, training model to distinguish why a statement belongs AND does not belong to a class.
• Multilingual Financial NLP — Evaluated on English, Hindi, Bengali, Telugu covering low-, medium-, and high-resource settings.
• Small-Model Efficiency — Qwen-2.5 (1.5B) with bidirectional reasoning achieves 95.59% (EN) — outperforming Qwen-2.5 (72B) and LLaMA-3.1 (70B) trained with labels only.
• Cross-domain Generalizability — Method also outperforms baselines on hate speech (ETHOS) and ethics (DFAR) classification.

Methodology

Figure: Three fine-tuning approaches compared. (a) Labels only: LLM trained to output label ŷ. (b) Unidirectional: label + positive reason R+. (c) Bidirectional (Ours): label + positive reason R+ + negative reason R−. The contrastive supervision in (c) enables the model to learn why a statement is sustainable AND why it is not unsustainable.

Reason Generation

Positive reason prompt:

###Statement: s_i
###Label: {y_i}
###Explain why the statement is labeled as {y_i}.
Provide a brief explanation in no more than two sentences.

Negative reason prompt:

###Statement: s_i
###Label: {y_i}
###Explain why the statement is **not** labeled as {¬y_i}.
Provide a brief explanation in no more than two sentences.

Reasons generated by GPT-4o (temperature 0.1). Used only during training — not at inference, so inference cost is unchanged.

Training Details

Models: LLaMA-3.2 (3B), LLaMA-3.1 (8B & 70B), Qwen-2.5 (1.5B, 7B & 72B) — fine-tuned with QLoRA (parameter-efficient fine-tuning)

📊 Table 1 — Main Results: Multilingual Financial Sustainability Classification

(♢ = Labels only, ♡ = Labels + Unidirectional Reason, ♠ = Labels + Bidirectional Reason [Ours]; gray rows = our method)

English:

Bengali:

Hindi:

Telugu:

📊 Table 2 — Comparison with Existing Benchmarks

📊 Table 3 — Reason Quality Evaluation (GPT-4o reasons, Claude Opus-4.1 as judge, score out of 5)

GPT-4o-generated rationales achieve avg. scores > 4.8/5 across all languages and metrics, confirming high-quality supervision signals.

📊 Cross-Domain Results (Table 7 in paper)

Our bidirectional method also outperforms label-only and unidirectional on:

Hate Speech (ETHOS dataset):

Ethics (DFAR dataset):

📚 Citation

@inproceedings{kabir2025bidirectional,
  title={Bidirectional Reasoning Supervision for Multilingual Financial Decision Making},
  author={Kabir, Muhammad Rafsan and Ahad, Jawad Ibn and Krambroeckers, Robin and Ahmed, Silvia and Elahi, M M Lutfe and Mohammed, Nabeel and Rahman, Shafin},
  booktitle={Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track},
  pages={1576--1587},
  year={2025}
}