Organizations: FinancialNLPLab, MODULABS Shinhan Securities Seoul, Republic of Korea · FinancialNLPLab, MODULABS KT Seoul, Republic of Korea · FinancialNLPLab, MODULABS EMRO Seoul, Republic of Korea · FinancialNLPLab, MODULABS Samsung Fire & Marine Insurance Seoul, Republic of Korea · FinancialNLPLab, MODULABS KB Securities Seoul, Republic of Korea · FinancialNLPLab, MODULABS Seoul, Republic of Korea · Seoul University Seoul, Republic of Korea
Financial text embeddings must distinguish changes in event status, perspective, and obligations even when passages share similar wording. NMIXX adapts existing encoders through 18.8k source-linked triplets: paraphrases and Korean-English translations preserve meaning, while targeted financial rewrites introduce semantic contrasts. We examine this recipe across seven backbones on English and Korean financial and general-domain semantic textual similarity (STS), and analyze the composition and passage lengths of KorFinSTS. BGE-M3 attains the highest adapted financial correlations in this comparison, improving from 0.1969 to 0.2967 on FinSTS and from 0.0512 to 0.2732 on KorFinSTS. Its general English and Korean correlations decrease by 0.0391 and 0.0463. Across the seven models, five improve their mean financial correlation, but all reduce their mean general-domain correlation. Per-language comparisons and benchmark-weight sensitivity analysis reveal differences obscured by a single aggregate score. The study contributes a finance-specific supervision design and evidence for evaluating adaptation jointly with retained general semantic capability; direct cross-language retrieval remains outside its evaluation scope.
Figures & tables
Figure 1. NMIXX at a glance. (A) Financial document types motivate semantic-shift negatives; the sentence contrast is illustrative. (B) Triplet construction and adaptation of existing encoders. (C) BGE-M3’s changes on all four benchmarks connect the two research questions: financial improvements coexist with general-domain losses. Three panels show financial semantic distinctions, the NMIXX triplet pipeline, and BGE-M3 changes of plus 0.0998, plus 0.2220, minus 0.0391, and minus 0.0463 on the four benchmarks.
Figure 2. Positioning by supervision construction. TSDAE ( Wang et al., 2021 ) , GPL ( Wang et al., 2022 ) , and Fin-E5 ( Tang and Yang, 2025 ) illustrate distinct adaptation routes. NMIXX emphasizes financial meaning changes within source-linked triplets, alongside paraphrase/translation positives. These are design emphases, not exclusive capabilities or a performance ranking. The sentence contrast is illustrative. Four horizontal routes compare denoising reconstruction in TSDAE, generated queries and teacher labels in GPL, persona-based financial task synthesis in Fin-E5, and source-linked financial semantic-shift triplets in NMIXX.
Dataset
Lang.
Rows (k)
License
sujet-finance-instruct
ko
178
Apache-2.0
finance-legal-mrc
ko
359
CC-BY-SA 4.0
KorFin-ASC
ko
8.8
Apache-2.0
finance-embeddings-investopedia
en
206
CC-BY-NC 4.0
fingpt-sentiment-train
en
76.8
MIT
FNSPID
en
1,630
CC-BY-NC 4.0
Table 1. Public corpora collected before filtering.
Figure 3. NMIXX training-data construction. (A) Public corpora and synthetic augmentation pass through filtering, balancing, and expert review. (B) Semantic-shift negatives and meaning-preserving positives are generated and screened to form triplets. Stage counts use different units and do not imply a conversion rate. The sentence triplet illustrates the intended relation between anchor, positive, and negative. A two-stage pipeline shows 2.46 million raw records plus 25.9 thousand synthetic documents, a filtered pool of 46.1 thousand sentences, expert review, semantic-shift negative and paraphrase generation, judge thresholds of 8 and 9, and 18.8 thousand training triplets.
Figure 4. KorFinSTS dataset analysis. (A) Pair counts by source. (B) Median and interquartile range of text length, pooling both sentence columns within each source. Characters include stored whitespace; lengths are not model-token counts. Counts are 355 news, 500 disclosure, 421 research, and 715 legal pairs. Median character lengths are 240, 977, 371, and 834.5 respectively.
Model
License
Language Support
bge-en-icl
Apache-2.0
Mainly English
gte-Qwen2-1.5B-instruct
Apache-2.0
English & Chinese
e5-mistral-7b-instruct
MIT
Mainly English
bge-large-en-v1.5
MIT
Mainly English
all-MiniLM-L12-v2
Apache-2.0
Mainly English
instructor-base
Apache-2.0
Mainly English
Table 2. Baseline embedding models, licenses, and language support.
FinSTS
KorFinSTS
STS
KorSTS
Model
before
after
before
after
before
after
before
after
bge-en-icl
0.1668
0.2574
0.0511
−0.0745
0.8058
0.5965
0.7078
0.2487
gte-Qwen2-1.5B-inst
0.2858
0.2518
0.0094
0.2204
0.8592
0.7556
0.3742
0.4727
e5-mistral-7b-inst
0.1476
0.2641
0.1099
−0.1738
0.8768
0.6092
0.7495
0.1492
bge-large-en-v1.5
0.1675
0.1626
−0.2119
−0.1586
0.8752
0.8835
0.3320
0.2473
all-MiniLM-L12-v2
0.1909
0.2626
−0.1837
−0.1590
0.8309
0.7109
0.3858
0.1262
Table 3. Spearman’s ρ on four STS benchmarks before and after domain adaptation. The higher value of each before/after pair is bolded . Boldface denotes a numerical comparison, not statistical significance.
Figure 5. Adaptation changes across the four evaluation settings. Financial gains vary across languages and encoders (RQ1), while general-domain losses are widespread (RQ2). Numerical labels show differences between the four-decimal scores in Table 3 on a common color scale. A seven by four heatmap labels each adaptation change. BGE-M3 improves both finance scores and loses both general scores.
Figure 6. Trade-off and sensitivity analysis of Table 3 . (A) Each point averages two financial and two general changes; the dashed segment connects the two nondominated adaptations. Higher and farther right is preferable under this summary. (B) A hypothetical weight w interpolates between the general and financial means. This measures sensitivity to benchmark weights rather than deployment benefit. A scatter plot compares financial gain with general retention for seven models. A second plot varies the finance weight from zero to one. BGE-M3 crosses zero at about 0.210.