Paper: arXiv 2507.09601

Authors: Hanwool Lee, Sara Yu, Yewon Hwang, Jonghyun Choi, Heejae Ahn, Sungbum Jung, Youngjae Yu

Abstract

General-purpose sentence embedding models often struggle to capture specialized financial semantics, especially in low-resource languages like Korean, due to domain-specific jargon, temporal meaning shifts, and misaligned bilingual vocabularies. To address these gaps, we introduce NMIXX (Neural eMbeddings for Cross-lingual eXploration of Finance), a suite of cross-lingual embedding models fine-tuned with 18.8K high-confidence triplets that pair in-domain paraphrases, hard negatives derived from a semantic-shift typology, and exact Korean-English translations. Concurrently, we release KorFinSTS, a 1,921-pair Korean financial STS benchmark spanning news, disclosures, research reports, and regulations, designed to expose nuances that general benchmarks miss. When evaluated against seven open-license baselines, NMIXX’s multilingual bge-m3 variant achieves Spearman’s rho gains of +0.10 on English FinSTS and +0.22 on KorFinSTS, outperforming its pre-adaptation checkpoint and surpassing other models by the largest margin, while revealing a modest trade-off in general STS performance. Our analysis further shows that models with richer Korean token coverage adapt more effectively, underscoring the importance of tokenizer design in low-resource, cross-lingual settings. By making both models and the benchmark publicly available, we provide the community with robust tools for domain-adapted, multilingual representation learning in finance.

Complexity vs Empirical Score

  • Math Complexity: 2.0/10
  • Empirical Rigor: 7.5/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper primarily focuses on empirical engineering, utilizing existing architectures (bge-m3), creating a new dataset (KorFinSTS), and reporting performance metrics (Spearman’s rho) without novel mathematical derivations. It is heavily data- and implementation-focused, with the release of models and benchmarks, making it actionable for practitioners.

Research Flowchart

  flowchart TD
  A["Research Goal: Domain-Adapted Cross-Lingual<br>Financial Embeddings"] --> B["Data Collection & Construction<br>18.8K Triplet Training Set<br>1,921-Pair KorFinSTS Benchmark"]
  B --> C["Methodology: NMIXX Model Suite<br>Fine-tuned via Cross-Lingual<br>Hard Negatives & Semantic Shifts"]
  C --> D["Training & Evaluation<br>Evaluated against 7 Baselines<br>Spearman's rho Correlation"]
  D --> E{"Key Findings & Outcomes"}
  E --> F["Scores: +0.22 on KorFinSTS<br>+0.10 on English FinSTS"]
  E --> G["Analysis: Richer Korean Token<br>Coverage improves Adaptation"]
  E --> H["Deliverables: Publicly Available<br>Models & Benchmark"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes NMIXX, a suite of cross‑lingual financial embedding models and the KorFinSTS benchmark for Korean financial semantic textual similarity.

Source sentence
We introduce NMIXX, a novel suite of cross-lingual financial embedding models informed by semantic shift patterns extracted from a large-scale multilingual financial corpus.

Assumes

No explicit assumptions found in the text.

Data

  • KorFinSTS benchmark public
    Source sentence
    We also developed and released KorFinSTS, a comprehensive new benchmark for Korean financial semantic textual similarity.

Costs

Not applicable The paper reports no trading or execution result.

Finds 4

  • The multilingual bge‑m3 variant of NMIXX achieves a Spearman’s ρ gain of +0.0998 on FinSTS and +0.2220 on KorFinSTS compared to pre‑adaptation baselines. performance
    Source sentence
    Notably, the multilingual bge-m3 variant achieves a Spearman’s ρ gain of +0.0998 on FinSTS and +0.2220 on KorFinSTS compared to the pre-adaptation baselines.
  • Models with richer Korean token coverage adapt more effectively, highlighting tokenizer design as crucial for low‑resource cross‑lingual settings. other
    Source sentence
    Our analysis further shows that models with richer Korean token coverage adapt more effectively, underscoring the importance of tokenizer design in low-resource, cross-lingual settings.
  • Domain specialization improves financial‑domain STS performance but reduces general‑domain STS performance, indicating a trade‑off. negative
    Source sentence
    this domain specialization comes at the cost of reduced general‑domain STS performance, underscoring a persistent challenge in balancing domain relevance and cross-domain generalizability.
  • Only the bge‑m3 model achieves a positive average gain (¯Δ = +0.1609), confirming that sufficient Korean token coverage is key to successful bilingual domain adaptation. performance
    Source sentence
    Only bge-m3 achieves a positive average gain (¯Δ = +0.1609), underscoring that sufficient full-syllable Korean coverage is key to successful bilingual domain adaptation.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 18 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (20)

  1. TWICE: What Advantages Can Low-Resource Domain-Specific Embedding Model Bring? – A Case Study on Korea Financial Texts — → 2502.07131
  2. Beyond Surface Similarity: Detecting Subtle Semantic Shifts in Financial Narratives (2024) — link
  3. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks (2019) — link
  4. SimCSE : Simple Contrastive Learning of Sentence Embeddings (2021) — link
  5. MTEB: Massive Text Embedding Benchmark (2023) — link
  6. Towards General Text Embeddings with Multi-stage Contrastive Learning (2023) — link
  7. Text Embeddings by Weakly-Supervised Contrastive Pre-training (2024) — link
  8. FinMTEB : Finance Massive Text Embedding Benchmark (2025) — link
and 12 more
  1. Domain Adaptation for Japanese Sentence Embeddings with Contrastive Learning based on Synthetic Sentence Generation (2025) — link
  2. Prometheus: Inducing Fine-grained Evaluation Capability in Language Models (2024) — link
  3. Improving Text Embeddings with Large Language Models (2024) — link
  4. MMTEB: Massive Multilingual Text Embedding Benchmark (2025) — link
  5. WHEN FLUE MEETS FLANG: Benchmarks and Large Pre-trained Language Model for Financial Domain (2022) — link
  6. Are All Languages Created Equal in Multilingual BERT? (2020) — link
  7. Tokenization is Sensitive to Language Variation (2025) — link
  8. TSDAE : Using Transformer-based Sequential Denoising Auto-Encoder for Unsupervised Sentence Embedding Learning (2021) — link
  9. GPL : Generative Pseudo Labeling for Unsupervised Domain Adaptation of Dense Retrieval (2022) — link
  10. Multilingual E5 Text Embeddings: A Technical Report (2024) — link
  11. BGE M3-Embedding : Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation (2024) — link
  12. NV-Retriever : Improving Text Embedding Models with Effective Hard-Negative Mining (2024) — link

1 of 20 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.