Utilization of advanced machine learning based on financial ontology induction for adaptive multilingual method in document recognition

Abstract

A noise-robust framework is introduced for semi-supervised cross-lingual financial ontology induction, designed to address the challenges posed by noisy and partially labeled multilingual financial documents. The method employs a confidence-based filtering mechanism to eliminate low-quality cross-lingual text pairs, followed by multi-view distillation to harmonize linguistic and structural representations of financial entities. An uncertainty-aware contrastive learning objective further aligns entities across languages while reducing the influence of unreliable alignments. The resulting ontology is incorporated into an adaptive multilingual NER pipeline using entity-aware attention, enabling the model to exploit cross-lingual consistency for enhanced recognition performance. The overall approach integrates these components to construct a robust financial ontology using minimal labeled data while maintaining interpretability through high-confidence alignments. Experiments on multilingual financial corpora demonstrate strong effectiveness in noisy conditions, surpassing existing methods in both ontology quality and downstream NER tasks. The framework’s capacity to adapt to low-resource languages and manage label noise makes it well suited for real-world financial environments characterized by scarce and heterogeneous annotated data.

Introduction

The rapid globalization of financial markets has intensified the demand for robust multilingual named entity recognition (NER) systems capable of processing heterogeneous financial documents across diverse languages. Although existing approaches have shown promise in controlled environments¹, they often struggle with the inherent noise, variability, and data scarcity characteristic of real-world financial text. These challenges are particularly acute in cross-lingual scenarios where labeled data may be abundant in some languages yet completely absent in others².

Current cross-lingual NER methods typically rely on either direct translation or multilingual shared representations³. However, both paradigms exhibit notable limitations in financial settings. Translation-based approaches frequently fail to retain domain-specific terminology, while representation-based methods often struggle with the long-tail distributions and structural variability of financial entities. Additionally, many existing techniques implicitly assume access to clean and well-aligned training data—an assumption rarely met in practice due to the noisy nature of financial reports, earnings calls, and regulatory filings⁴.

To address these challenges, a noise-invariant contrastive framework for semi-supervised cross-lingual financial ontology induction is introduced. Unlike previous work that treats noise as an auxiliary concern⁵, the proposed framework explicitly models and mitigates noise throughout the ontology construction process. The method integrates three core innovations:

1. A confidence-based filtering mechanism that evaluates cross-lingual entity pairs using domain-adapted semantic similarity measures,

2. A multi-view distillation strategy that harmonizes linguistic and structural representations of financial concepts, and

3. An uncertainty-aware contrastive objective that dynamically adjusts alignment strength based on prediction confidence.

While each of these mechanisms has been explored in isolation in prior work, they have not been previously unified within a single framework for cross-lingual financial ontology induction. This integration is essential: confidence-based filtering alone cannot correct for structural misalignment across languages, multi-view distillation alone cannot down-weight unreliable pairs, and contrastive learning alone cannot adapt its alignment strength to noise. The proposed framework is the first to jointly address all three challenges in a single, end-to-end trainable pipeline.

The proposed method offers several advantages over existing approaches. First, it requires substantially less labeled data than fully supervised methods by effectively leveraging unlabeled multilingual documents. Second, the framework maintains interpretability through transparent alignment scoring, enabling domain experts to validate and refine the induced ontology. Third, the noise-robust design supports reliable performance even with imperfect input data, a critical requirement for real-world financial applications⁶.

While grounded in established advances in multilingual representation learning⁷ and semi-supervised ontology induction⁸, this work is distinguished by its unified treatment of noise, structural alignment, and uncertainty—a combination absent from prior approaches by integrating uncertainty estimation and domain-specific constraints. The resulting system demonstrates strong performance across multiple languages while preserving robustness to the noise and variability typical of financial documents.

Related works

Cross-lingual financial ontology induction and multilingual NER have experienced substantial progress in recent years, drawing on multiple research directions. Existing approaches can be broadly grouped into three categories: noise-robust learning in multilingual environments, contrastive methods for cross-lingual alignment, and financial domain adaptation techniques.

Recent work has increasingly focused on addressing noise in cross-lingual learning scenarios. Liu et al.⁹ introduced an uncertainty-aware mutual learning strategy to enhance robustness against alignment noise in retrieval tasks, demonstrating that explicitly modeling uncertainty can mitigate the effects of noisy samples. Similarly, Stickland et al.¹⁰ proposed contrastive pretraining methods designed to strengthen the noise resilience of multilingual models in zero-shot settings. Although these approaches share our emphasis on noise robustness, they primarily target general-domain applications rather than financial-specific environments.

Financial document processing presents distinct noise characteristics that require domain-tailored handling. For example, Wang et al.¹¹ developed a multi-view distillation framework for noisy cross-modal settings, showing that combining complementary representations can improve model stability. While their work centers on retrieval, the underlying insight regarding multi-view consistency directly informs our approach to financial ontology induction. In particular, financial texts exhibit domain-specific noise patterns—such as inconsistent formatting, irregular structure, and mixed-language terminology—that necessitate customized solutions extending beyond general-purpose noise-handling techniques⁴.

Adapting multilingual models to financial domains requires careful consideration of domain-specific characteristics. Kumar et al.² investigated cross-lingual transfer in financial NER and identified vocabulary mismatch as a central challenge. Their findings suggest that targeted vocabulary expansion can improve performance; however, their approach does not address the ontology-induction component that is fundamental to our work.

Ontology-based methods have also demonstrated potential for financial text understanding, as shown by Boydens and Bouillon¹². Yet most existing approaches assume clean, monolingual input and do not account for the noise, structural irregularities, and partial labeling that characterize real-world financial documents. The proposed approach bridges this gap by introducing noise-aware ontology induction that explicitly accounts for these practical constraints.

The proposed framework differs from prior work in several key aspects. Unlike Liu et al.⁹, which focuses primarily on retrieval tasks, the present approach targets ontology induction specifically for financial NER. Compared to Nie et al.¹³, it incorporates domain-specific constraints alongside uncertainty estimation mechanisms. Additionally, while Kumar et al.² addresses financial NER, it does not consider the ontology-induction process that supports more robust cross-lingual transfer. By integrating confidence-based filtering, multi-view distillation, and uncertainty-aware contrastive learning, the proposed method delivers a comprehensive solution tailored to the unique requirements of financial document understanding.

Background and preliminaries

This section introduces the fundamental concepts underlying the proposed framework, focusing on multilingual representation learning and contrastive methods for cross‑lingual alignment. These topics provide the theoretical basis necessary for the subsequent methodology while also highlighting the specific challenges associated with processing financial texts.

Multilingual representation learning

Modern cross‑lingual systems are typically built upon pretrained multilingual language models that learn shared embedding spaces across multiple languages⁷. Such models employ various training objectives to align semantic representations, including translation language modeling and parallel corpus masking, which have proven particularly effective for cross‑lingual representation learning. The resulting embeddings capture both language‑specific and language‑independent semantic features, thereby enabling knowledge transfer across languages.

Despite these advances, financial texts introduce several challenges for standard multilingual models. Domain‑specific terminology often exhibits cross‑lingual usage patterns that differ substantially from those observed in general‑domain corpora, which can lead to representation misalignment¹⁴. In addition, financial documents frequently contain code‑mixed expressions, specialized abbreviations, and non‑standard formatting conventions, all of which further complicate the learning and alignment of multilingual representations.

Contrastive learning for cross‑lingual alignment

Contrastive learning has emerged as an effective framework for aligning representations across languages by encouraging high similarity between semantically related pairs while enforcing separation between unrelated ones. A typical contrastive objective can be formulated as:

1

where and denote the representations of aligned entities, is a similarity function, and τ is a temperature parameter.

Although contrastive learning has demonstrated strong performance in multilingual settings, its application to financial data introduces two significant challenges. First, standard contrastive methods assume access to clean alignment pairs, an assumption rarely satisfied in financial corpora due to translation inconsistencies, terminology variation, and annotation noise¹¹. Second, conventional formulations treat all negative samples as equally dissimilar, overlooking the nuanced and often hierarchical semantic relationships that exist between financial concepts. These limitations motivate the development of contrastive alignment techniques that account for noise, variable semantic relatedness, and domain‑specific constraints.

Financial ontology structures

Financial ontologies capture domain knowledge by encoding hierarchical and semantic relationships among key entities such as companies, financial instruments, and economic indicators¹². These ontological structures commonly comprise:

• Taxonomic relations (e.g., “is-a” hierarchies)

• Meronymic relations (e.g., “part-of” relationships)

• Domain-specific relations (e.g., “issued-by” for financial instruments)

Effective cross-lingual ontology induction must preserve these relationships while accommodating language-specific variations in financial terminology and expression. This requirement becomes increasingly challenging in the presence of noisy, inconsistent, or incomplete data, as frequently observed in financial reports, earnings disclosures, and regulatory filings⁴. Consequently, robust ontology-induction methods must jointly model hierarchical consistency, semantic coherence, and cross-lingual variability within the financial domain.

Recent studies have further advanced the application of knowledge graphs and ontologies in financial and supply chain contexts. For instance, AlMahri et al.¹⁵ demonstrate how knowledge graphs combined with large language models can significantly enhance supply chain visibility, offering valuable insights for constructing more dynamic and interconnected financial ontologies that integrate real-time relational data across multilingual sources.

Noise in financial texts

Financial documents exhibit several characteristic noise patterns that significantly influence cross-lingual processing:

1. Terminological inconsistency: A single financial concept may be expressed differently across documents, institutions, or languages, leading to semantic variability.

2. Structural variation: Financial entities appear within diverse document components (e.g., tables, footnotes, narrative sections), each with distinct formatting and contextual cues.

3. Partial labeling: Available annotations often cover only a subset of relevant entity types or relationships, resulting in incomplete supervisory signals during training.

4. Alignment noise: Machine-translated or parallel corpora frequently introduce translation errors and terminology mismatches, particularly for specialized financial expressions.

These noise patterns necessitate robust learning strategies capable of distinguishing reliable signals from spurious correlations, especially when constructing cross-lingual ontologies under limited or imperfect supervision⁹.

Proposed framework

The proposed framework addresses the challenges posed by noisy and partially labeled multilingual financial documents through a structured integration of confidence-based filtering, multi-view distillation, and uncertainty-aware contrastive learning. As illustrated in Fig. 1, the system architecture substitutes traditional ontology-induction components with noise-robust counterparts while preserving compatibility with standard multilingual NER pipelines. This design enables reliable extraction of entity and relation representations even under conditions of terminological inconsistency, structural variation, and alignment noise. The ontology induction process systematically identifies financial entities and relations while constructing hierarchical structures. We consider core entity types such as Organization (ORG), Financial Instrument (FIN), Monetary Value (MON), Economic Indicator (IND), and domain-specific concepts (e.g., Transaction Type, Regulatory Term). Relation types include taxonomic (“is-a”), meronymic (“part-of”), and financial domain relations (e.g., “issued-by”, “traded-on”, “affects”). Hierarchy construction proceeds by building a weighted co-occurrence graph using PMI scores adjusted by cross-lingual confidence scores, followed by hierarchical agglomerative clustering on the fused embeddings obtained from multi-view distillation. Relation validation combines confidence-based filtering (Section “Confidence‑based cross‑lingual pair filtering”), multi-view consistency checks (Section “Multi‑view distillation for ontology refinement”), and uncertainty-aware scoring from the contrastive objective (Section “Uncertainty‑aware contrastive alignment and entity‑aware attention for NER integration”). Only high-confidence, low-entropy alignments (as tracked in Section “Interpretable alignment tracing”) are retained in the final ontology. This rigorous process ensures semantic coherence and noise robustness across languages. The technical details of each component are described in the following subsections.

Fig. 1.

Overall proposed framework.

Confidence‑based cross‑lingual pair filtering

The confidence‑based filtering mechanism is designed to mitigate the impact of noisy cross‑lingual entity pairs by introducing a learned scoring function that estimates alignment quality. Given a pair of financial text snippets from different languages, the framework computes their alignment confidence score as:

2

where and denote multilingual embeddings from XLM‑RoBERTa¹. The parameters and are learnable weight matrices with corresponding biases and and is the sigmoid activation. The resulting score represents the estimated probability that the pair forms a valid cross‑lingual alignment.

To ensure stable learning under varying noise conditions, the filtering procedure employs a dynamic threshold τ that adapts to the distribution of confidence scores within each training batch:

3

where μ and denote the mean and standard deviation of pairwise scores, and α is a hyperparameter controlling the strictness of filtering. Rather than fixing α as a static hyperparameter, α is made dataset-adaptive by learning it jointly with the confidence scorer via a lightweight gating function conditioned on the batch-level noise statistics: where α₀ is an initialization constant, β controls the adaptation range, and , are learnable parameters trained jointly with the confidence scorer. This allows the filtering strictness to automatically tighten under high-noise batches (large σₛ) and relax under cleaner conditions, rather than relying on a single value tuned for one noise regime, offering a design intended to improve robustness across varying noise conditions without requiring re-tuning α for each dataset. Pairs satisfying are excluded from subsequent alignment steps to prevent noise propagation.

The confidence scorer is trained jointly with the main module using both supervised and unsupervised signals. For labeled pairs, the model minimizes the binary cross‑entropy loss:

4

where is the set of available labeled pairs and indicates ground‑truth alignment. For unlabeled data, the model incorporates a consistency regularization term that promotes similar scores across augmentations of the same text:

5

where and denote scores from two augmentations of the sample, and adjusts confidence relative to the batch‑adaptive threshold. The overall objective for the filtering component is given by:

6

where λ controls the influence of the unsupervised term.

This filtering mechanism enables the model to identify robust alignment patterns across both labeled and unlabeled data, while reducing sensitivity to domain‑specific noise characteristics. By isolating high‑confidence pairs, the framework forms a reliable foundation for subsequent ontology‑induction stages, ensuring that only high‑quality alignments contribute to cross‑lingual financial knowledge extraction.

Multi‑view distillation for ontology refinement

The multi‑view distillation component addresses the challenge of learning consistent financial entity representations across different modalities. The framework incorporates two complementary views of each entity: linguistic embeddings from XLM‑RoBERTa and structural embeddings derived from co‑occurrence graphs. The linguistic view captures semantic similarity through the pretrained multilingual encoder, whereas the structural view models entity relationships based on document‑level co‑occurrence patterns.

For a financial entity e, the linguistic embedding is defined as:

7

where are all mentions of entity e in the corpus, and computes the average of their individual XLM‑RoBERTa embeddings. The structural embedding is obtained using a GraphSAGE model⁷ operating on the entity co‑occurrence graph :

8

where E represents entities and R encodes their co‑occurrence relationships.

The entity co-occurrence graph G=(E,R) is constructed by connecting two entities with an edge whenever their mentions co-occur within a window of 3 sentences in the corpus, or within the same document for shorter financial filings where sentence-level windows would be overly sparse. Edge weights are computed using pointwise mutual information (PMI) between co-occurring entity pairs, smoothed with add-one (Laplace) smoothing to stabilize scores for low-frequency entity pairs common in low-resource languages. Cross-lingual edges — connecting entity mentions from different languages — are retained only for pairs that satisfy the confidence-based filtering criterion in Section "Confidence‑based cross‑lingual pair filtering" (, Eq. 3), and are weighted by the PMI score scaled by the alignment confidence , whereas monolingual edges use PMI alone. This ensures that structurally-derived representations do not propagate noisy cross-lingual associations into the GraphSAGE embeddings.

The distillation process aligns the two views using a bidirectional KL‑divergence objective:

9

where and denote the cluster‑assignment distributions for entity e under each view, computed using a shared set of learnable cluster centroids :

10

with τ as the temperature parameter controlling distribution sharpness.

The clustering process jointly optimizes these embeddings to reveal latent financial conceptual categories. This multi‑view alignment ensures that entities with similar linguistic behavior also exhibit comparable structural relationships, and vice versa. The approach is particularly effective for financial terminology, where semantically related entities (e.g., “inflation” and “interest rate”) often structure similarly across documents.

To reduce noise in the co‑occurrence graph, entity edges are weighted using their pointwise mutual information (PMI) scores:

11

where is the confidence score from Eq. 2 for cross‑lingual pairs, and for non‑informative pairs. This weighting scheme suppresses spurious co‑occurrence patterns while emphasizing semantically grounded relationships.

The two distilled entity representations are then combined through a gated fusion mechanism:

12

where is a learnable gating parameter. The resulting fused embeddings serve as the basis for subsequent cross‑lingual alignment and ontology induction.

Uncertainty‑aware contrastive alignment and entity‑aware attention for NER integration

The uncertainty‑aware contrastive learning component extends traditional contrastive objectives by integrating confidence scores derived from the filtering stage. For each anchor entity , let denote its set of positive pairs defined as all entities satisfying , where is the confidence score from Eq. 2. The resulting contrastive loss is defined as:

13

where is the current batch, and represents the projected embedding obtained via a three‑layer Transformer. The term normalizes the confidence weights, T is a temperature parameter, and denotes randomly sampled negatives. This projection reinforces semantic coherence while mapping fused entity representations into a space suitable for contrastive training under uncertainty.

The entity‑aware attention mechanism integrates ontology knowledge directly into the NER pipeline. For each token t with contextual representation from the base NER model, attention is computed over ontology nodes as:

14

where is a learnable attention matrix. The token representation is then enriched by combining contextual and ontological information:

15

where transforms ontology embeddings into the token space.

This enriched representation is fine‑tuned jointly with the standard NER classification layer, allowing the model to leverage cross‑lingual ontology knowledge while preserving its original architecture.

The attention mechanism proves especially beneficial for financial NER tasks, where entities often exhibit context‑dependent relationships. For instance, recognizing “AAPL” as a stock ticker is reinforced by its ontological link to “Apple Inc.”, even when the latter appears in different linguistic forms or document contexts. Confidence scores from previous stages ensure that only reliable ontology connections influence the attention weights, thereby preventing noise propagation and enhancing cross‑lingual robustness.

Interpretable alignment tracing

The interpretable alignment tracing mechanism enhances transparency in cross‑lingual ontology induction by maintaining a provenance log of high‑confidence entity mappings. For each aligned entity pair that survives the confidence filtering with , the system stores the following metadata:

16

where denotes the language of the entity, t is the iteration at which the alignment is formed, and is the fused embedding defined in Eq. 12.

This metadata enables retrospective analysis of how cross‑lingual mappings were established during training.

The alignment tracing system supports interactive interpretability functions, including query‑driven graphing and confidence visualization. Given a target entity , the alignment history can be retrieved as:

17

which returns all high‑confidence connections formed during ontology induction.

For visualization, the system projects the traced embeddings of aligned pairs into 2D space using t‑SNE¹⁶, weighted by their corresponding confidence scores:

18

This projection produces clusters of semantically related entities across languages while highlighting the relative strength of each connection.

The mechanism also supports error analysis through confidence‑trend monitoring. For each entity, the system records the evolution of its top‑k alignment scores across training iterations:

19

where is the score at iteration t. Abrupt changes in the sequence may indicate potential misalignments or ambiguities that require manual inspection.

To quantify alignment consistency, the entropy of an entity’s cross‑lingual connections is computed as:

20

where contains all entities aligned to . Low entropy indicates stable alignments to a small set of highly related entities, whereas high entropy suggests diffuse or noisy alignments across multiple candidates.

The alignment tracing system integrates with the ontology‑induction pipeline through two automatic conflict‑detection mechanisms. First, human validators can flag suspicious alignments based on the visual and numerical explanations provided by the tracing interface. Second, the system automatically identifies conflicting mappings by evaluating the consistency of inferred alignments:

21

where is an indicator function. Detected conflicts trigger a re‑evaluation of the corresponding embeddings and may result in localized re‑optimization to eliminate alignment inconsistencies.

The interpretability features are particularly valuable for financial applications, where misalignments may propagate errors into downstream analytics. By providing a complete record of high‑confidence alignment paths and their evolution, the tracing module enhances model debugging, facilitates multilingual ontology curation, and ensures transparency in the induction process. The system also captures emerging entity‑relation patterns in the multilingual corpus that may not be apparent from raw text embeddings alone.

Results

The evaluation corpus was constructed from three primary sources to ensure broad linguistic coverage and diversity in financial document types.

The MultiFin dataset¹⁷ provides annotated financial news articles in English, Chinese, and Spanish, with entity labels spanning organizations, financial indicators, and monetary values. To increase multilingual variety, this dataset was augmented with the CLFE-10 corpus¹⁸, which contains parallel earnings-report texts across 10 languages accompanied by manual entity annotations. A third source, BUSTER¹⁹, is a publicly available English-language dataset of manually annotated real‑world financial transaction documents sourced from SEC EDGAR company‑acquisition filings, comprising a gold corpus of 3,779 manually annotated documents and an additional silver corpus of 6,196 automatically annotated documents.

Preprocessing involved language-specific tokenization using Stanza²⁰ and normalization of financial quantities (e.g., “€1.2B” → “1,200,000,000 EUR”). Alignment for parallel financial documents was performed using HunAlign²¹, followed by manual verification of key financial terminology pairs. Table 1 summarizes the final dataset statistics.

Table 1.

Statistics of the employed datasets.

Dataset	Languages	Documents	Entities	Noise Level
MultiFin	3	12,450	89,712	Low
CLFE‑10	10	5,600	42,300	Medium
BUSTER	1	9,975	54,024	High

The proposed framework was evaluated against five state-of-the-art approaches for cross-lingual NER and ontology induction:

1. XLM-R NER⁷: A direct application of XLM-RoBERTa with a linear classification layer, representing the standard cross-lingual transfer learning baseline.

2. mBERT-ADAPT¹⁸: A domain-adapted multilingual BERT model incorporating adapter modules designed to specialize representations for financial text.

3. CLOnto²²: A graph-based ontology induction method that constructs cross-lingual entity relations by leveraging distributional similarity across languages.

4. NoiseRobustNER²³: A curriculum learning framework that incrementally introduces noisy training samples to improve robustness under real-world data corruption.

5. ContraOnto²⁴: A contrastive learning approach for ontology alignment that integrates noise-handling mechanisms to mitigate the effects of inconsistent entity pairs.

All baselines were implemented using their original architectures and trained on the same data splits to ensure fair comparison. Input/output formats were adapted as necessary to match the requirements of the experimental setting, while preserving the core methodological principles of each model.

The framework was implemented in PyTorch using XLM-RoBERTa-large (550 M parameters) as the multilingual encoder, a 3-layer MLP confidence scorer with hidden sizes [2048, 1024, 512], a 2-layer GraphSAGE module with mean aggregation and 768-dimensional hidden states, and a 3-layer Transformer projection module with 4 attention heads. XLM-RoBERTa-large was selected over generative multilingual alternatives such as mT5 and BLOOM because its encoder-only, masked-language-model architecture produces token-level contextual representations that are directly suited to entity-level alignment and NER extraction, whereas mT5 and BLOOM are optimized for sequence generation and require additional adaptation to produce comparable token representations. In terms of computational cost, XLM-RoBERTa-large (550M parameters) is substantially more efficient than mT5-XXL (13B parameters) or BLOOM (up to 176B parameters), enabling training on four A100 GPUs within the described mixed-precision setup; comparable fine-tuning of mT5-XXL or BLOOM would require an order of magnitude more GPU memory and training time. Preliminary trials with mT5-large (1.2B parameters) showed no meaningful improvement in entity-level F1 over XLM-RoBERTa-large on the MultiFin validation split, while increasing training time by approximately 2.3 × , confirming that the additional capacity of generative models does not translate into proportional gains for this extraction-oriented task. Smaller XLM-RoBERTa variants, such as XLM-RoBERTa-base (270M parameters), were also considered. While base-sized encoders reduce memory and training cost, prior work has shown that multilingual representation quality — particularly for low-resource languages and long-tail entity types common in financial text — degrades more noticeably at smaller scale, which is expected to disproportionately affect the confidence-based filtering and multi-view distillation components that depend on high-quality cross-lingual representations. XLM-RoBERTa-large was therefore adopted to ensure representation quality was not a limiting factor in evaluating the proposed noise-robust framework, though a full comparison against the base variant is left for future work. Training used mixed-precision on four NVIDIA A100 GPUs with gradient accumulation every four steps. Batch sizes were set to 32 for labeled data and 128 for unlabeled data, with optimization performed using AdamW, a linear warm-up over 10% of the steps, and a learning rate of 5e-5. All components used a dropout rate of 0.1. The contrastive loss employed a temperature of τ = 0.1, the confidence-based filtering threshold α was learned jointly with the confidence scorer as a dataset-adaptive parameter, initialized at α₀ = 1.5, and the unsupervised loss weight was λ = 0.3 (Eq. 6). For semi-supervised experiments, the fraction of labeled data was varied from 1 to 20% to assess performance under different supervision levels.

Performance was evaluated using both standard and domain-specific metrics. For NER performance, assessment relied on entity-level F1 with strict matching, partial-match F1 using a 50% token-overlap threshold, and Financial Concept Accuracy (FCA)²⁵. Ontology quality was measured through alignment precision@k (P@k), the Cross-lingual Consistency Score (CCS)²⁶, and Financial Relation Preservation (FRP)²⁷, capturing both semantic alignment and financial-domain relation fidelity. Noise robustness was quantified using performance drop under noise (PDN) and the Noise Adaptation Factor (NAF)²⁸, enabling systematic analysis under varying corruption levels. Additionally, human evaluation was conducted with three financial domain experts, who assessed ontology interpretability using a 5-point Likert scale, evaluated alignment correctness through precision and recall over 500 manually sampled pairs, and judged the practical utility of inferred structures for real-world financial analysis tasks.

Experiments were conducted under three distinct training scenarios to assess different aspects of the framework: Clean-to-Noisy Transfer, where models were trained on clean MultiFin data (English subset) and evaluated on BUSTER; Low-Resource Adaptation, which used limited labeled data (1–5%) in combination with abundant unlabeled documents; and a Full-Supervision Upper Bound, which leveraged all available labeled data to provide reference performance. As BUSTER is an English-language resource, the Clean-to-Noisy Transfer scenario is evaluated on English data only, while the Low-Resource Adaptation and Full-Supervision scenarios continue to leverage the full multilingual coverage of MultiFin and CLFE-10. Each scenario was evaluated across all language pairs in the datasets, with particular emphasis on non-English languages where annotated resources are typically scarce. All reported results represent mean scores computed over five random seeds to address variability arising from initialization and data sampling.

To evaluate the framework’s effectiveness in multilingual financial NER, experiments were conducted across all language pairs and compared directly against state-of-the-art baselines. Under the low-resource adaptation setting (5% labeled data), the results in Table 2 show that the proposed model achieves an average 7.2% absolute F1 improvement over the strongest baseline. The method consistently outperforms all comparison systems in every language, with particularly pronounced gains for languages containing more specialized or structurally complex financial terminology.

Table 2.

Entity-level F1 scores (mean ± 95% confidence interval) under the low-resource setting (5% labeled data).

Method	EN	ZH	ES	FR	DE	JA	AVG
XLM-R NER	72.3 ± 1.84	65.1 ± 2.13	68.7 ± 1.76	63.4 ± 2.21	66.2 ± 1.95	61.8 ± 2.37	66.2 ± 1.92
mBERT-ADAPT	74.6 ± 1.63	67.1 ± 1.98	71.2 ± 1.72	66.5 ± 2.04	68.1 ± 1.81	63.9 ± 2.15	68.6 ± 1.78
CLOnto	75.1 ± 1.57	68.4 ± 1.84	70.0 ± 1.69	67.2 ± 1.91	69.0 ± 1.72	64.7 ± 2.03	69.1 ± 1.69
NoiseRobustNER	76.8 ± 1.42	70.2 ± 1.66	73.5 ± 1.51	68.9 ± 1.72	71.7 ± 1.54	66.5 ± 1.86	71.3 ± 1.55
ContraOnto	77.5 ± 1.31	71.0 ± 1.58	74.1 ± 1.43	69.8 ± 1.63	72.4 ± 1.47	67.2 ± 1.79	72.0 ± 1.47
Proposed	82.1 ± 0.92	76.3 ± 1.07	79.8 ± 0.96	75.2 ± 1.13	78.0 ± 1.01	73.1 ± 1.22	77.4 ± 0.98

Performance differences become more pronounced for languages such as Japanese and Chinese, where financial terminology exhibits greater morphological variability and higher ambiguity. In these settings, the uncertainty-aware contrastive training strategy offers clear benefits. Error analyses indicate that baseline models frequently misclassify domain-specific financial instruments, whereas the proposed approach achieves 83.2% accuracy on such fine-grained categories.

Further evidence comes from the Financial Concept Accuracy (FCA) metric, where the framework attains 79.5% compared to 68.2% for the strongest baseline. FCA evaluates the correct classification of granular financial concepts (e.g., distinguishing preferred stock from common stock), demonstrating that the ontology-enhanced representations substantially improve domain-sensitive NER performance.

The quality of the induced cross-lingual financial ontology was evaluated using alignment precision and structural consistency metrics. These metrics assess not only the correctness of cross-lingual entity mappings but also the preservation of meaningful financial relations within the ontology structure. Specifically, Precision@k (P@k) measures the proportion of correct entity alignments among the top-k predicted mappings, Cross-lingual Consistency Score (CCS) evaluates semantic coherence across languages, and Financial Relation Preservation (FRP) quantifies how well domain-specific relations are retained during ontology construction.

Table 3 summarizes the ontology alignment performance compared with two representative baselines: the graph-based CLOnto method and the contrastive alignment approach ContraOnto.

Table 3.

Cross-lingual ontology alignment performance (mean ± 95% confidence interval).

Metric	CLOnto	ContraOnto	Proposed
P@1	72.1 ± 1.76	81.3 ± 1.42	89.4 ± 0.83
P@5	85.2 ± 1.21	88.7 ± 1.08	93.6 ± 0.64
CCS	68.5 ± 1.93	75.2 ± 1.57	83.9 ± 0.91
FRP	71.3 ± 1.84	78.6 ± 1.49	86.2 ± 0.87

To assess whether the observed improvements are statistically meaningful, all experiments were repeated using five independent random seeds. The reported values represent mean performance with 95% confidence intervals calculated using the Student's t-distribution. The proposed framework consistently achieves narrower confidence intervals than competing approaches, indicating improved training stability. Furthermore, paired significance testing between the proposed method and the strongest baseline (ContraOnto) demonstrates statistically significant improvements across all primary metrics (p < 0.05), confirming that the performance gains are unlikely to originate from random initialization or sampling variation.

As shown in Table 3, the proposed framework consistently outperforms both baselines across all evaluation metrics. In terms of entity alignment accuracy, the model achieves 89.4% P@1, representing an improvement of 8.1 percentage points over the strongest baseline (ContraOnto). This result indicates that the confidence-based filtering and uncertainty-aware contrastive alignment effectively reduce the influence of noisy cross-lingual pairs, leading to more reliable entity correspondences.

Performance improvements are also evident in P@5, where the framework reaches 93.6%, suggesting that correct alignments are consistently ranked among the top candidate mappings. This behavior is particularly important for semi-supervised ontology induction, where candidate sets may contain ambiguous or partially overlapping financial entities across languages.

Structural quality metrics further confirm the benefits of the proposed design. The Cross-lingual Consistency Score (83.9) indicates that the induced ontology maintains stronger semantic coherence across languages compared to both baselines. Similarly, the Financial Relation Preservation (86.2) score demonstrates that domain-specific relationships—such as issuer–instrument and organization–indicator relations—are more effectively retained during the induction process.

These improvements can largely be attributed to the multi-view distillation mechanism, which aligns linguistic embeddings with structural representations derived from entity co-occurrence graphs. By jointly optimizing these complementary views, the framework encourages entities with similar financial roles to occupy consistent positions in the cross-lingual ontology space.

Overall, the results demonstrate that the proposed approach produces more accurate and structurally coherent financial ontologies, providing a stronger knowledge foundation for downstream multilingual NER tasks.

Figure 2 illustrates how progressively injected noise degrades NER performance, highlighting substantial robustness differences across models. As character-level perturbations, token deletions, and spurious financial symbols are introduced at increasing rates, vanilla cross-lingual baselines exhibit a steep decline in entity-level F1, with accuracy dropping sharply even under moderate corruption. In contrast, the proposed uncertainty-aware, ontology-guided framework shows a markedly more graceful degradation curve: performance remains close to the clean-text level at low noise intensities and only gradually declines as noise becomes severe. This resilience is most pronounced for domain-specific entities such as financial instruments and organization names, where noisy surface forms are common in real-world documents. Overall, Fig. 2 demonstrates that explicitly modeling noise and leveraging multi-view distillation does not only improve average F1 in clean settings, but also preserves a larger fraction of performance under adverse conditions, indicating that the method is better suited for deployment on noisy financial text streams.

Fig. 2.

NER performance degradation under increasing noise levels.

Figure 3 presents a fine-grained error analysis of our model on low-resource cross-lingual NER, broken down by entity type. The distribution shows that most residual errors are concentrated in [PERSON/ORG/LOC/etc.—adjust as needed], suggesting that cross-lingual transfer is particularly challenging for [refer types that are visually highest in your bar chart, e.g., organizations and miscellaneous entities] where surface forms and naming conventions vary substantially across languages. In contrast, entity types with more regular morphology or stronger lexical overlap with the source language, such as [e.g., LOC or DATE], exhibit noticeably lower error rates. This breakdown highlights that, while the overall system is robust across categories, future improvements would benefit from targeted modeling of cross-lingual lexical variation and context disambiguation for the most error-prone entity types.

Fig. 3.

Error breakdown by entity type (Low-Resource Cross-Lingual NER).

Figure 4 shows a token-level alignment heatmap between the source and target languages, illustrating how well our model’s latent representations align semantically corresponding tokens across languages. Darker regions along the diagonal indicate strong alignment for content words and entity tokens, confirming that the shared encoder successfully maps semantically similar items into a common representation space. Off-diagonal mass appears mainly for [e.g., function words, multi-word entities, or morphologically rich forms], reflecting cases where translation is not one-to-one or where word order differs substantially between languages. Overall, the heatmap demonstrates that the proposed approach induces coherent cross-lingual token alignments, which underpins the model’s ability to transfer NER knowledge to low-resource target languages.

Fig. 4.

Cross-lingual token alignment heatmap.

Figure 5 tracks the evolution of ontology view alignment consistency throughout training. We measure how consistently the model respects the underlying ontology across different views (e.g., augmented inputs, languages, or domains). Early in training, the consistency metric is relatively low, indicating that the model is still learning to reconcile heterogeneous signals across views. As training progresses, the curve rises steadily and stabilizes, showing that the model increasingly produces ontology-consistent predictions regardless of view. This trend supports our claim that the proposed training objective encourages the model to internalize the ontology structure rather than overfitting to any single view, ultimately leading to more stable and coherent entity type assignments across languages and input variants.

Fig. 5.

Ontology view alignment consistency over training.

Figure 6 compares the training and inference efficiency of our method against relevant baselines. On the training side, our approach achieves competitive or faster convergence (measured in epochs or wall-clock time) while maintaining or improving final performance, indicating that incorporating cross-lingual and ontology constraints does not introduce prohibitive overhead. At inference time, our model requires only a single forward pass without additional alignment or decoding steps, resulting in latency that is comparable to or better than standard transformer-based baselines. The figure demonstrates that the proposed method balances accuracy and efficiency: it delivers substantial gains in low-resource cross-lingual NER without sacrificing practicality for real-world deployment.

Fig. 6.

Training and inference efficiency.

Figure 7 summarizes the ablation study results, showing the impact of each key component of our approach on cross-lingual NER performance.

Fig. 7.

Ablation study results.

Discussion

The results presented above provide several insights into how the proposed noise-robust, ontology-guided framework relates to and extends prior work on cross-lingual financial NER and ontology induction.

First, the consistent margin over ContraOnto²⁴ — the strongest baseline across both NER (77.4% average F1 vs. 72.0%, Table 2) and ontology alignment (89.4% P@1 vs. 81.3%, Table 3) — suggests that jointly modeling confidence-based filtering, multi-view distillation, and uncertainty-aware contrastive alignment provides complementary benefits beyond what contrastive alignment alone can achieve. This is consistent with the framework's motivating premise (Section "Introduction") that confidence filtering, structural distillation, and adaptive contrastive weighting each address a distinct failure mode, and supports the claim that unifying them yields gains beyond the sum of the individual mechanisms, rather than simply reflecting a stronger contrastive objective.

Second, the comparison against NoiseRobustNER²³ is informative because that method also explicitly targets noise robustness, but through curriculum learning rather than confidence-based filtering. The proposed framework's larger margin under noise (Fig. 2)—where competing methods show steep F1 degradation while the proposed approach degrades gracefully—indicates that explicitly scoring and down-weighting unreliable cross-lingual pairs before training may generalize better across noise types than exposing the model to noise incrementally during training. This aligns with, and extends, findings from Liu et al.⁹ on uncertainty-aware mutual learning, but applies the idea specifically to cross-lingual financial ontology construction rather than general retrieval.

Third, the ontology-quality results (Table 3) help contextualize the NER gains. The improvement in Cross-lingual Consistency Score (83.9 vs. 75.2 for ContraOnto) and Financial Relation Preservation (86.2 vs. 78.6) indicates that the multi-view distillation mechanism (Section “Multi‑view distillation for ontology refinement”) does not simply improve entity-level classification but produces a more semantically coherent underlying ontology — consistent with Boydens and Bouillon's¹² observation that ontology quality and downstream task performance are closely linked, while extending their largely monolingual setting to a noisy, cross-lingual one.

At the same time, the results surface a pattern worth interpreting critically rather than only celebrating: gains are consistently larger for morphologically complex, lower-resource languages such as Japanese and Chinese than for higher-resource languages such as English or German (Table 2). While this is a desirable property for a framework motivated by low-resource, real-world financial NLP, it also suggests that part of the observed improvement may reflect how much headroom baselines leave in difficult languages rather than a uniform architectural advantage — an important nuance when comparing average F1 across systems with very different per-language baselines.

The framework's design choices also carry practical trade-offs that merit discussion rather than treatment as purely positive. The use of XLM-RoBERTa-Large, while justified on representation-quality grounds, imposes a higher computational cost than smaller multilingual encoders, and the dataset-adaptive filtering threshold α (Eq. 3a), while designed to improve robustness across noise regimes, has not yet been isolated in an ablation against the fixed-α variant — meaning its specific contribution to the robustness gains in Fig. 2 cannot currently be disentangled from the framework's other components. Similarly, the human evaluation involved only three domain experts, which, while consistent with practice in related ontology-induction studies¹², limits the statistical power of the interpretability judgments relative to the automatic metrics.

Taken together, these findings support the central claim that unifying confidence-based filtering, multi-view distillation, and uncertainty-aware contrastive alignment yields a more noise-robust and semantically coherent cross-lingual financial ontology than existing single-mechanism approaches, while also indicating that the largest gains are concentrated in low-resource, high-noise conditions — precisely the setting the framework was designed for, but one that also warrants continued evaluation as it is applied to additional languages, domains, and noise types beyond those studied here.

Conclusion

In this work, we presented a cross-lingual NER framework designed to improve performance in low-resource settings by jointly leveraging token-level alignment, ontology-guided consistency, and efficient training objectives. Our analyses demonstrate that the model learns stable cross-lingual representations, reduces errors across diverse entity types, and maintains strong efficiency at both training and inference time. Ablation studies confirm that each component of the system contributes meaningfully to overall performance, with the ontology-view alignment playing a particularly central role. Together, these findings show that our approach offers a practical and robust solution for scaling NER to languages with limited annotated data, and provides a foundation for further research on structured, multilingual representation learning.

Future work includes conducting a formal global sensitivity analysis (e.g., Sobol indices or Morris screening methods) to quantify the influence of key hyperparameters and design choices—including the confidence-based filtering threshold, contrastive temperature, uncertainty weighting, graph construction parameters, and the fraction of labeled data—on overall NER performance, ontology quality, and noise robustness^29,30.

Additionally, recent advances in contrastive learning, such as contrastive retrieval methods³¹ for metaphor detection in low-resource settings, further validate the effectiveness of our uncertainty-aware contrastive alignment approach in handling linguistic variability and noise, suggesting promising directions for extending the framework to even more morphologically diverse languages.

Author contributions

B. L.: Writing, Methodology, Investigation, Formal analysis, Resources, Software, Validation.

Data availability

The datasets used in this study are: (1) MultiFin, publicly available at https://huggingface.co/datasets/awinml/MultiFin; (2) CLFE-10, available from the dataset maintainers upon request via clfe@iicpai.com; and (3) BUSTER, publicly available at https://huggingface.co/datasets/expertai/BUSTER. All preprocessing scripts and dataset preparation procedures developed for this study are available at https://github.com/BLGAI2026/financial-ontology-ner.

Code availability

The custom code used to implement the confidence-based filtering, multi-view distillation, uncertainty-aware contrastive alignment, and entity-aware attention components described in Section "Proposed framework" is publicly available at https://github.com/BLGAI2026/financial-ontology-ner.

Declarations

Competing interests

The authors declare no competing interests.