CogDeBias: An LLM-Based Multilingual Framework for Cognitive Bias Detection and Mitigation in Corporate Decision-Making Texts

You’re listening to “CogDeBias: An LLM-Based Multilingual Framework for Cognitive Bias Detection and Mitigation in Corporate Decision-Making Texts,” by Y. Shen, W. Yang, and Y. Shen. Published in 2026.

Abstract.

Cognitive biases embedded in corporate strategic communications pose significant risks to investment decisions and governance quality, yet existing detection approaches rely on manual analysis that lacks scalability. This study presents CogDeBias, a multilingual framework integrating large language models (LLMs) and machine learning (ML) for automated detection and mitigation of cognitive biases in enterprise decision-making texts. A bilingual annotated corpus of 1,200 corporate annual reports (600 English, 600 Chinese) was constructed, encompassing 30,416 sentences, of which 8,082 are bias-positive (26.6%) and carry approximately 9,230 bias-category instances across six categories: confirmation bias, sunk cost fallacy, overconfidence, anchoring effect, bandwagon effect, and recency bias.

The hybrid architecture combines XLM-RoBERTa multilingual encoding, Llama-3-70B prompt-based classification, and XGBoost ensemble learning, achieving a weighted F1-score of 0.82 on the test set, surpassing rule-based (0.59), BERT-based (0.72), zero-shot GPT-4 (0.76), and few-shot Llama-3 (0.785) baselines, with the 3.5-percentage-point margin over the strongest baseline statistically significant (p = 0.008, McNemar’s test). On a fully expert-annotated subset labelled without any model assistance, the framework retained a weighted F1 of 0.81, indicating that the reported performance is not an artefact of model-to-model label agreement. Cross-lingual evaluation revealed English F1 of 0.84 versus Chinese F1 of 0.80, with zero-shot transfer from English to Chinese yielding F1 of 0.68, improving to 0.78 with minimal target-language fine-tuning.

The Mixtral-8x7B-powered mitigation generator produced actionable suggestions rated 4.1/5 by 30 financial analysts (Fleiss’ kappa = 0.71). These suggestions are decision-support outputs for human review, not automated corrections to regulated filings. Processing efficiency of 0.42 seconds for core model inference (3.0 seconds end-to-end per document) supports batch-oriented use as decision support in investor due diligence, corporate audit, and regulatory monitoring workflows, subject to domain-specific calibration. This work establishes a computational paradigm for decision science applications, demonstrating that linguistic patterns reliably surface systematic reasoning distortions across languages and corporate contexts.

Introduction.

Decision-making failure within corporations has significant economic implications. Recent history within the business

The associate editor coordinating the review of this manuscript and approving it for publication was Wai-Keung Fung.

world has highlighted examples of how cognitive biases within management discourse contributed to decision-making failure, ultimately costing companies billions of dollars in shareholder equity. These biases, or systematic deviations from rational decision-making, are often not directly observable within financial data analysis, occurring at the semantic level within annual reports, earnings announcements, and other strategic decision-making dis-course. This is in contrast to quantitatively derived metrics that indicate decision-making failure within companies, such as financial distress indicators. The increasing acknowledg-ment of cognitive factors within financial markets serves to further emphasize the need for an automated detection of cognitive biases.

Despite growing recognition of cognitive factors in finan-cial markets, three persistent gaps limit existing approaches: (i) limited scalability of manual annotation, (ii) linguistic fragmentation between English-dominant tools and bilingual corporate contexts, and (iii) methodological fragmentation across single-bias studies. A systematic review of these gaps, organized by research stream, is presented in Section II.

This research introduces three main innovations. First, this study establishes a six-category cognitive bias taxon-omy (confirmation bias, sunk cost fallacy, overconfidence, anchoring effect, bandwagon effect, and recency bias) operationalized through corpus-validated linguistic patterns, enabling automated detection at weighted F1 = 0.82 on a 240-document held-out test set—outperforming the strongest baseline (Few-shot Llama-3, F1 = 0.785) by 3.5 percent-age points (p = 0.008, McNemar’s test). The taxonomy adapts foundational heuristics-and-biases theory and recent medical-domain bias classifications to the corporate disclosure context, addressing a relatively underdeveloped area in enterprise decision-making bias research, building on prior investor-level bias modeling.

Second, this study introduces what is, to our knowledge, among the first end-to-end frameworks to jointly address cognitive-bias detection and mitigation in bilingual (English–Chinese) corporate disclosures (Table 1), integrating XLM-RoBERTa multi-lingual encoding, hybrid LLM-ML bias classification, and Mixtral-8x7B-powered mitigation generation for enterprise disclosure analysis. By moving beyond passive annotation to actionable rewriting suggestions, the framework delivers mitigation outputs rated 4.1/5 by 30 financial analysts (Fleiss’ kappa = 0.71, with 86% of suggestions judged highly rele-vant), and achieves end-to-end processing of 3.0 seconds per document (0.42-second core model inference), supporting scalable batch processing to assist investor due diligence and regulatory monitoring workflows.

This responds to the call of practitioners for useful tools, rather than purely proof-of-concept systems, in decision-making bias research. Third, this study constructs a sentence-level annotated bilin-gual (English–Chinese) cognitive-bias corpus for corporate disclosures (Table 1), comprising 1,200 documents and 8,082 bias-positive sentences (≈9,230 category-level bias instances) across six categories. Fine-grained, multi-domain annotated benchmarks with cross-platform transfer have proven valuable in adjacent text-classification tasks, a design principle this corpus follows. While benchmark datasets exist for adjacent financial NLP tasks such as opinion mining and question answering, no comparable bilingual annotated resource is available for cognitive-bias detection in corporate contexts.

The corpus enables systematic cross-lingual evaluation, demonstrating English F1 = 0.84 and Chinese F1 = 0.80, with zero-shot English-to-Chinese trans-fer at F1 = 0.68 and rapid recovery to F1 = 0.78 using only 100 target-language documents. As this few-shot adaptation effect does not survive Benjamini–Hochberg correction (p = 0.041), it is reported as suggestive rather than confirmatory.

Objectives of the research revolve around the development of an automated, scalable, and multilingual bias detection system that is deployable in a real-world setting. This includes the development of classification standards and annotation methods that have been approved by domain experts, the creation of a hybrid detection system that combines the strengths of neural and traditional ML tech-niques, as well as the empirical validation of the accuracy of the detection system and the utility of the mitigation system through human evaluation. The rest of the paper is organized as follows: Section II reviews related work across cognitive-bias detection, financial-text NLP, and multilingual NLP, and identifies the gaps this study addresses.

Section III describes data acquisition, bias taxonomy, system architecture, and evaluation; Section IV presents the results in terms of detection metrics, cross-lingual analysis, and mitigation quality assessment; Section V discusses the key findings, advantages, and limitations; and Section VI concludes the paper by highlighting the contributions and prospects.

II. RELATED WORK.

There are currently three major limitations associated with existing research: First, scalability is a major concern, and individual analysts examining corporate disclosures have severe throughput limits, making it difficult to scale up sample sizes and generalize findings on a larger scale. The study on corporate foresight revealed that detecting biases and fallacies within corporate planning documents demands specialized human expertise. Such expertise does not scale to thousands of annual reports and disclosures. The gap addressed here is that such manual, expert-dependent analysis does not scale to large bilingual corpora; CogDe-Bias replaces per-analyst throughput limits with automated sentence-level detection.

Second, linguistic fragmentation remains a major concern, and sentiment analysis tools have advanced for English-language financial disclosures, while corporate settings, especially Chinese A-share markets and Hong Kong markets, which use a mix of languages, are still not addressed. The detection of biases across languages, especially for Chinese and Hong Kong markets, remains underexplored, given that gaps have already been identified for multilingual NLP systems for non-English languages. The gap addressed here is the absence of a bias-detection system and annotated resource spanning English and Chinese corporate disclosures, which the present bilingual corpus and cross-lingual evaluation directly fill.

Third, methodological fragmentation is another concern, and prior research on AI and its application in behavioral finance have addressed individual biases, while a comprehensive framework remains absent, and applications remain at an early stage of practical deployment. Notably, the recent global review by Şeker et al. catalogues how cognitive biases have been operationalized across heuristic-based, machine-learning, and LLM-based approaches in investor decision-making, demonstrating that this research stream is well-established at the investor level but identifying a persistent gap in detection at the corporate-disclosure level. The gap addressed here is the lack of a unified framework that covers multiple bias categories and couples detection with actionable mitigation, rather than treating single biases in isolation—the gap the present work fills.

Positioned against the three literature streams, CogDeBias’s contribution is the specific combination none of them provides. Deep-learning and transformer financial-NLP work (e.g., lexicon-to-transformer sentiment, risk-disclosure topic model-ing ) targets sentiment or topic rather than reasoning-level cognitive bias, and is English-only. Cognitive-bias work in LLMs,,, probes biases inside models on medical or general benchmarks, not bias expressed in corporate disclosures, and stops at detection without mitigation. Multilingual NLP reviews, document the English–non-English gap but provide no bias-detection system or annotated resource.

CogDeBias is distinguished by the combination of four features: (i) operating on corporate disclosures, (ii) covering English and Chinese, (iii) integrating hybrid LLM–ML detection with generation-based mitigation, and (iv) supplying a sentence-level bilin-gual bias benchmark. No prior work in Table 1 combines all four; it is this combination, rather than any single element, that is novel.

Significant gaps in technical, cognitive, and behavioral capabilities of language-based systems have long been identified. Large language models (LLMs) hold tremendous promise in bridging these gaps. Transformer-based architec-tures, which are trained on multilingual data, have shown unprecedented capabilities in cross-lingual transfer learning. For example, the XLM-RoBERTa model has shown robust semantic understanding in 100 different languages using unsupervised learning of representations. In addition to technical capabilities, empirical research demonstrates that large language models are subject to human-like biases in their decision-making.

For example, in one of the first studies, it was shown that pre-ChatGPT models were subject to human-like framing effects, anchoring effects, and System 1 thinking, as observed in cognitive reflection and semantic illusion tests, which are subject to human irrationality. However, reinforcement learning from human feedback significantly reduces these effects. This demonstrates that, even as alignment training decreases the likelihood of human-like biases in decision-making, it also decreases the ability of the model to detect bias in input data. New methodologies in medical domain bias assessment are now being developed and are of significant importance to corporate applications. For example, in one recent study, datasets were created to assess six different decision-making biases in medical domains, including confirmation bias, availability heuristic, and overconfidence.

The results showed that state-of-the-art proprietary models are robust to bias-inducing prompts, whereas open-source models showed significant performance degradation. This framework of using adversarial prompts to detect latent patterns of bias can be applied to the present study. A second study, which surveyed the phenomenon of cognitive biases in large lan-guage models, validates template-based approaches to corporate text analysis. Table 1 summarizes ten representative works across these three streams and positions CogDeBias against the identified gaps.

III. MATERIALS AND METHODS.

A. DATASET CONSTRUCTION AND PREPROCESSING

The dataset was collected from four major sources, which cover various regulatory environments. English-language documents, totaling 600, were collected from the SEC EDGAR system. The authors specifically filtered the database to include the annual reports filed by publicly traded U.S. corporations, denoted as 10-K reports, from January 2023 to August 2025. Chinese-language documents, also totaling 600, were collected from the Shanghai Stock Exchange disclosure platform. The authors filtered the database to include annual reports filed by A-share listed companies during the same period. Supporting materials were collected from the Hong Kong Stock Exchange and OpenCorporates database, which included dual-language documents and disclosures filed by smaller enterprises. The final dataset included 1,200 corporate documents, with an equal number of English and Chinese documents.

The study focused on recent events and strategic narratives. These were shaped by the post-pandemic recovery and by various technology disruptions, both of which are documented to enhance managerial cognitive biases. A text mining study of financial disclosures has found that the narrative section of financial reports reveals behavioral patterns that are distinct from numerical statements.

To ensure the inclusion of the documents, the following criteria were satisfied: Firstly, the texts were required to have substantive management discussion sections, such as ‘‘Management Discussion and Analysis’’ for U.S. filings or the equivalent sections for Chinese reports. These sections carry forward-looking strategic discourse, where cognitive biases are most likely to occur. Secondly, the completeness thresholds for the documents were set such that those with less than 95% structural integrity, based on the absence of critical sections or corruption in the encoding, were excluded from the dataset. Lastly, tables, figures, and chart captions were excluded, as they are not part of the running text. The remaining text was filtered for bias-indicative linguistic patterns.

This step also addressed common NLP preprocessing problems, such as character-encoding errors in cross-platform aggregation, which are especially prevalent in multilingual and low-resource text.

Raw documents were subjected to systematic transforma-tion through a series of four stages. In the text cleaning stage, HTML residues, XML tags from regulatory templates, and non-standard Unicode characters were removed, which are known to interfere with tokenization. In the next stage of sentence segmentation, the text was divided into discrete analytical units, while language classification was performed to partition the text into English and Chinese categories using the langdetect library. Finally, text vectorization was done to transform the text into a compatible format for the neural network model, which was achieved through the sentence-transformers library. Through the text preprocessing stage, 30,416 annotated sentences were obtained, as indicated in the dataset composition and partitioning statistics as presented below in Table 2.

The composition of the dataset is evenly distributed across the training, validation, and test sets, while the bias levels were consistent at 26.6% throughout the dataset.

Each document was annotated with filing timestamp and industry sector (Global Industry Classification Standard (GICS)). Industry-specific stress events (e.g., semiconductor export controls in Oct 2023/July 2024, identified through Bloomberg/Reuters archives) were catalogued. Documents filed within 90 days of stress events were labeled for post-hoc temporal analysis (Section V-A). Baseline comparison group (n = 780, 65%) comprised documents from non-disruption periods.

Documents were sampled to reflect sectoral diversity: English corpus included Technology (22%), Financials (18%), Healthcare (15%), Consumer (12%), Industrials (10%), and other sectors (23%); Chinese corpus comprised Manufacturing (28%), IT (19%), Financials (14%), Materials (11%), Consumer (10%), and others (18%). Industry classi-fications followed GICS standards for consistency.

Market capitalization distribution: Large-cap >$10B (35%, n = 420), Mid-cap $2-10B (45%, n = 540), Small-cap

<$2B (20%, n = 240). The mid-cap oversampling (45% vs ∼30% typical market weight) reflects institutional research coverage patterns and enhances generalizability across firm sizes, though limits direct applicability to cap-weighted portfolio analysis.

B. COGNITIVE BIAS TAXONOMY AND GROUND TRUTH LABELING

The cognitive bias taxonomy was developed based on the foundational research by Kahneman and Tversky on ’heuristics and biases in judgment under uncertainty’. Six cognitive biases were identified based on their prevalence within corporate-level strategic discourse and relevance to the behavioral finance literature on investor decision-making processes. Cognitive biases are systematic deviations from rational economic decision-making processes that often permeate management discourse. Confirmation bias occurs when managers selectively use supporting evidence for predetermined corporate strategies while overlooking disconfirming evidence. Sunk cost fallacy occurs when managers justify continued investments in failing projects based on prior invested capital rather than expected return on investment.

Overconfidence bias, prevalent in equity markets and linked to suboptimal trading and portfolio performance, manifests as management discourse that uses absolute language in discussing market outlook or competitive advantage. Anchoring effect occurs through con-tinued reliance on initial targets or price anchors regardless of changing market conditions. Bandwagon effect occurs through the adoption of industry trends characterized as ‘‘best practices’’ without independent strategic analysis. Recency bias occurs through the use of short-term metrics as a basis for long-term planning decisions.

The theoretical origins of the six categories can be traced as follows: anchoring effect and recency bias derive directly from the original heuristics-and-biases framework; confirmation bias and overconfidence bias appear both in and in the recent medical-domain bias evaluation taxonomy; sunk cost fallacy and bandwagon effect were adapted from the broader behavioral-economics literature and operationalized specifically for the corporate-disclosure context in this work. The six categories were selected by two criteria: prevalence in management discussion sections (jointly about 86 percent of expert-flagged instances) and grounding in the established taxonomies cited above. Biases common in other settings but sparse in disclosure text were deliberately excluded; for example, status-quo bias, central to habitual decision-making, rarely surfaces in forward-looking strategic narratives.

This non-exhaustiveness is revisited in Section V-C.

Operational definitions specific to the context of corporate disclosures were derived through an iterative process involv-ing domain experts. Confirmation bias in annual reports tends to be characterized by the asymmetric presentation of data, where market research is prominently featured while risk factors are relegated to the fine print. Sunk cost fallacy tends to be exemplified by the use of the following words or phrases: ‘‘given our substantial R&D investment to date’’ or ‘‘considering resources already committed.’’ Overconfidence bias tends to be exemplified by the use of words such as ‘‘will achieve,’’ ‘‘guaranteed to succeed,’’ or claims to superiority without supporting quantitative evidence. Anchoring bias tends to be exemplified by references to the initial guidance or competitor levels established over the course of the preceding periods.

Bandwagon bias tends to be exemplified by the use of words such as ‘‘as leaders across our sector recognize’’ or ‘‘following proven models’’ to justify decisions. Recency bias tends to be exemplified by the use of words such as ‘‘based on last quarter’s momentum’’ to justify decisions over long periods. The prevalence and characteristic manifestations of the aforementioned biases are shown in Table 3, where it is evident that confirmation bias is the most dominant bias (28.3% of the training set), while recency bias is the least prevalent (6.7%), which may be due to the subtlety of its linguistic manifestations requiring temporal reasoning to detect.

The annotation process followed a two-stage hybrid approach. The first stage involved three business psychol-ogy experts with doctoral training in the field. All three experts independently annotated the same 300 randomly selected documents (150 English and 150 Chinese) using a standardized codebook, providing a complete three-rater intersection on which Fleiss’ kappa was computed. Eval-uations followed a single-blind protocol—analysts were unaware of each output’s source — with presentation order randomized per analyst.

The codebook contains, for each of the six bias categories: (i) a working definition; (ii) 6– 10 lexical and syntactic indicators (e.g., absolutist modal verbs, investment-anchor phrases) with positive and negative examples drawn from a 60-document pilot pool; (iii) decision rules for boundary cases; (iv) a five-point severity rubric; and (v) language-specific notes covering English and Chi-nese rhetorical conventions. The de-identified codebook is available from the corresponding author upon reasonable academic request, consistent with the Data Availability statement. The experts annotated the documents at the sentence level and provided one or more bias categories. Each expert also rated the bias severity using a five-point Likert scale.

Because three annotators participated, inter-annotator agreement was measured using Fleiss’ kappa, which is designed for three or more raters, and yielded κ = 0.74 (substantial agreement). For reference, the three pairwise Cohen’s kappa values were 0.78, 0.76, and 0.74 (mean = 0.76); these are reported only as supplementary information and not used as the primary aggregate metric. Disagreements were resolved by majority voting across the three annotators. The remaining 900 documents were annotated using a two-step LLM-assisted approach. GPT-4 (gpt-4-0613) first generated candidate sentence-level bias labels using six-shot prompts.

The prompt for each input sentence concatenated: (i) a system instruction describing the six-bias taxonomy and the JSON output schema ({biastype, span, confidence}); (ii) six in-context examples — one per bias category — randomly sampled from the stage-one expert-annotated pool, each pairing a positive sentence with its expert label and a one-line rationale; and (iii) the target sentence to be classified. Examples were re-sampled per document to reduce ordering bias, and decoding used temperature = 0 with a 256-token output cap. Two trained graduate research assistants (each having completed a 4-hour calibration session against the stage-one expert annotations and reaching ≥90% agreement on a 50-sentence calibration set before commencing verification) independently reviewed each GPT-4 candidate label, classifying it as accept, modify (with corrected bias category or span), or reject.

Disagreements between the two assistants were resolved in two tiers: cases where both assistants agreed were finalized directly; cases of inter-assistant disagreement, or any case flagged by either assistant as ambiguous, were escalated to the three stage-one doctoral experts, whose majority vote constituted the final label. Roughly 12% of candidate labels were escalated under this protocol. Of the candidate labels processed, 76.4% were accepted as-is, 15.1% had their bias category modified, 6.7% had their span boundaries adjusted, and 1.8% were rejected as false positives; verifiers additionally added bias instances missed by GPT-4, primarily for recency bias and Chinese-language sunk-cost cases.

Cost savings of approximately 60% were realized relative to full expert annotation, and a post-hoc re-evaluation of 100 randomly sampled stage-two documents against independent expert relabeling yielded Cohen’s kappa = 0.74, comparable to stage-one expert agree-ment (Fleiss’ kappa = 0.74; pairwise Cohen’s mean = 0.76).

The annotated corpus was split into training (720 doc-uments, 60%), validation (240 documents, 20%), and test (240 documents, 20%) sets using a stratified sampling approach. This split is an appropriate balance between the need for a large training set to cover linguistic diversity and the need for a sufficient number of documents in the other two sets for performance evaluation without overfitting. Document-level stratification used language (English/Chinese), industry sector (GICS), and bias-positive flag as joint stratification keys; per-category bias proportions across the three splits were 28.2/28.5/28.1% (confirma-tion), 18.7/18.5/18.9% (sunk cost), 22.1/22.3/21.8% (over-confidence), 12.4/12.2/12.6% (anchoring), 11.8/11.9/11.7% (bandwagon), and 6.7/6.6/6.9% (recency), confirming that proportions were preserved within ±0.4 percentage points.

Because the split was performed at the document level over a multi-year window (2023 to 2025), we additionally ran a company-identifier leakage check. Of the firms in the corpus, roughly 90 contributed documents to more than one split; the 24 test-set documents (10 percent of the test set) whose issuing firm also appeared in the training set were removed and the model re-evaluated. Test weighted F1 changed only marginally, from 0.82 to 0.815 (a drop of 0.5 percentage points), indicating that cross-split firm overlap does not materially inflate the reported performance.

C. COGDEBIAS FRAMEWORK ARCHITECTURE

Before describing the architecture, we formalize the detection task. The cognitive bias detection problem is formulated as a ‘‘multi-label sentence-level classification’’task. Let S = {s1, s2..., sM} denote the sentence collection across the corpus (M = 30,416). Each sentence s is mapped via the multilingual encoder into a 768-dimensional representation x ∈ R768, and is associated with a binary label vector y = [y1..., y6]T ∈ {0,1}6 indicating the presence of each of the six bias categories: confirmation bias, sunk cost fallacy, overconfidence, anchoring effect, bandwagon effect, and recency bias. A sentence may carry zero, one, or multiple bias labels simultaneously, reflecting the empirical finding that 14.2% of bias-positive sentences exhibit more than one bias type. The detection task seeks a hypothesis h:R768 →6 producing the per-category bias probability for each sentence.

In the realized system this hypothesis is instantiated as the weighted fusion of the embedding-based ML branch and the text-based LLM branch (Section III-C, Eq. ); the encoder representation therefore enters detection through the ML branch rather than as a standalone classifier. Encoder fine-tuning minimizes a combined objective L = LBCE + αLSupCon (α = 0.3), where LBCE is the binary cross-entropy averaged over the six categories and LSupCon is the supervised contrastive term encouraging category-consistent embedding clustering. With l2-normalized embeddings zi and temperature τ = 0.07, it is defined as

zi · zp/τ exp −1 X X log LSupCon = exp (zi · za/τ) P |P(i)| i∈B p∈P(i) a∈B\{i} where the positive set P(i) contains in-batch sentences sharing at least one bias label with i, the rest are negatives, and bias-negative sentences form positives only among themselves.

The proposed framework rests on four explicit method-ological assumptions. First, cognitive biases in corporate disclosures manifest through detectable linguistic pat-terns, supported by behavioral-finance evidence that bias-laden discourse exhibits distinctive lexical and syntactic markers (e.g., absolutist modal verbs for overconfi-dence; investment-anchor phrases for sunk cost). This is consistent with evidence that managerial language style in corporate disclosures systematically shapes investor response. Second, sentence-level granularity is sufficient as the unit of analysis, given that the targeted biases typically surface within single statements rather than across paragraphs; multi-sentence reasoning is reserved for future extensions.

Third, the six bias categories are non-exhaustive but cover the dominant patterns observed in management discussion sections, accounting for approximately 86% of bias instances flagged by domain experts during pilot annotation. Fourth, expert annotators can reliably identify these biases from text alone, supported by the Fleiss’ kappa = 0.74 inter-annotator agreement (Section III-B) and confirmed by the multi-label re-annotation procedure used to address overlap between overconfidence and confirmation bias (Section IV-B).

The CogDeBias system is based on an architectural design that incorporates a pipeline of text understanding, bias detection, and mitigation strategy generation. Each of these blocks is independent, allowing for optimization of each module separately, while the overall system is differentiable, making it possible to perform joint optimiza-tion when required. The overall architecture is based on three layers of functionality: the multilingual layer, which processes the input text into an encoded representation, the bias identification layer, which uses hybrid classification techniques, and the strategy generation layer, which generates recommendations for mitigation strategies. Figure 1 presents the overall architectural schematic of the system, showing the flow of information and interactions between the different blocks of the system.

The multilingual encoding module is based on XLM-RoBERTa-large (HuggingFace snapshot xlm-roberta-large, 550M parameters), selected for its demonstrated +14.6% XNLI accuracy gain over multilingual BERT and strong English–Chinese alignment, which is based on the transformer architecture; transformer models have become central to decision-making tasks beyond classifi-cation, motivating their use here for disclosure-level reasoning analysis. Pre-trained on 2.5 terabytes of Common-Crawl text in 100 languages, with masked language modeling as the pre-training task. This model is then fine-tuned on the annotated dataset using supervised contrastive learn-ing, where the sentence embeddings of bias-positive and bias-negative examples were optimized to maximize the inter-class differences in the 768-dimensional latent space.

Fine-tuning used the AdamW optimizer (learning rate 2 × 10×−5, weight decay 0.01, linear warmup over the first 10% of steps), batch size 32 with gradient accumulation to an effective size of 64, 5 epochs, contrastive temperature τ = 0.07, and supervised-contrastive loss weight α = 0.3 (combined with binary cross-entropy as defined in Section III-C); the best checkpoint was selected by validation F1. This task-specific adaptation of the pre-trained model achieved an F1 score improvement of 4.2% when using frozen pre-trained weights on the validation set. The sentence embeddings produced by this model have the additional advantage of providing efficient similarity-based retrieval of similar bias examples from historical documents, which is an important feature of the overall system.

Distributed document representations have shown significant advantages over bag-of-words representations in capturing subtle seman-tic relationships between text, which is important in bias detection. One example is distinguishing hedging that signals overconfidence from hedging that reflects legitimate uncertainty.

The bias detection module utilizes the combination of different classification approaches through an ensemble voting mechanism. In the LLM pathway, Llama-3-70B-Instruct (HuggingFace snapshot meta-llama/Meta-Llama-3-70B-Instruct) was selected as the open-weights frontier model with the strongest published instruction-following scores at the time of development. This choice bal-anced detection accuracy against on-premise deploya-bility, given the data-confidentiality constraints in the Data Availability statement. The 70B-parameter decoder-only model is used for zero-shot classification through specifically crafted prompts.

Six bias-specific templates were adapted from medical-domain bias prompts through three steps: (i) reframing — patient-history language replaced with corporate-disclosure language (e.g., ‘management statements’ replacing ’patient symptoms’); (ii) category substitution — medical-specific biases (availability heuristic, self-diagnosis) dropped and corporate-finance categories (sunk cost, bandwagon, recency) added with descrip-tors from a 60-document pilot annotation; (iii) four iterative refinement rounds against a 50-sentence cali-bration set until two-expert agreement reached Cohen’s kappa = 0.78. The final templates are formatted as follows: ‘‘Analyze if the following corporate statement exhibits BIASTYPE by CHARACTERISTICBEHAVIOR. Statement: INPUT.

Classify and provide confidence (Yes/No) and confidence score.’’ An example of the confirmation bias template is as follows: ‘‘Analyze if the following corporate statement exhibits confirmation bias by selectively presenting information supporting predetermined conclu-sions while omitting contradictory evidence. Statement: INPUT. Classify and provide confidence.’’ In contrast, the ML pathway utilizes XGBoost as the gradient boosting clas-sifier, chosen for its proven robustness on small-to-medium tabular feature sets, native support for missing values, and consistent leadership on financial-text classification benchmarks. It operates on the fine-tuned XLM-RoBERTa-large sentence embedding (768-dimensional, Section III-C) concatenated with crafted features: syntactic parse tree statistics, sentiment polarity scores via FinBERT, and TF-IDF weighted unigrams through trigrams.

FinBERT (Hugging-Face snapshot ProsusAI/finbert) is a financial-domain BERT variant. The encoder embedding thus constitutes the multilin-gual representation consumed by the ML detection branch, while the same contrastively fine-tuned space additionally supports similarity-based example retrieval. To address language-specific bias manifestation patterns observed during preliminary analysis, 28 additional Chinese-language rule patterns were incorporated into the ML branch, targeting culturally embedded sunk-cost expressions such as ’having invested substantial resources’ and ‘cannot give up halfway’. Hyperparameter optimization via Bayesian search was used to set the optimal tree depth to 6 and the optimal number of estimators to 200.

The final predictions are produced through a weighted fusion mechanism by = λbyLLM + (1 − λ)byML with λ = 0.7 determined by validation grid search (LLM weight 0.7, ML weight 0.3), and tie-breaking based on the LLM confidence margins. Table 4 reports the weighted F1 across nine fusion ratios on both the validation set (used for selection) and the held-out test set (reported as confirmation only). Validation F1 peaks at λ = 0.7 (0.823), with the test-set F1 reaching 0.821 at the same setting. Performance remains within ±0.005 across λ ∈ [0.6, 0.8], confirming robustness; pure-LLM (λ = 1.0) and pure-ML (λ = 0.0) drop to 0.789 and 0.694 respectively, both on the test set. The chosen λ = 0.7 was fixed prior to test-set evaluation to avoid leakage.

The mitigation recommendation generator translates the identified biases into actionable editorial recommendations. The generator is positioned as decision support: it produces candidate revisions for human review by disclosure, com-pliance, or communications staff, rather than performing automated correction of regulatory filings, which carry legal, regulatory, and ethical risks that require human judgment.

When it identifies a bias, the generator uses type-specific remediation strategies from a curated knowledge base of 42 debiasing strategies, distributed across the six bias categories as 8 strategies for confirmation bias (e.g., devil’s advocate analysis, structured opposing-view templates), 7 for sunk cost fallacy (e.g., forward-looking ROI reframing, precommitment exit criteria), 8 for overconfidence (e.g., pessimistic-scenario inclusion, calibration self-checks), 6 for anchoring (e.g., reference-class forecasting, multi-anchor comparison), 7 for bandwagon effect (e.g., independent strategic justification, contrarian benchmark review), and 6 for recency bias (e.g., long-horizon trend re-weighting, base-rate reminders). The strategies were compiled from prior behavioral-economics and decision-science reviews and adapted to corporate-disclosure language.

The strategies are realized through Mixtral-8x7B (HuggingFace snapshot mistralai/Mixtral-8x7B-Instruct-v0.1), a sparse mixture of experts model that is chosen due to its excellent performance on instruction-following and creative rewriting tasks. The prompt format for the generation process is as follows: the original sentence, the type of identified bias along with the rationale, and the generation of three alternative sentences that address the identified biases and preserve the factual information. The output format is also stan-dardized to ensure consistency, and it includes the original sentence, the identified bias along with the rationale, three suggestions to improve the sentence, and the final revised sentence that reflects the recommendations.

The generation process uses nucleus sampling with p = 0.9 and tempera-ture = 0.7, selected via a small-scale grid search reported in Section IV-C.

Implementation of the system utilized Python 3.10, PyTorch 2.1 as the backend for deep learning, and the Transformers library for loading the models and performing the inference. The experiments were conducted on a cluster that contains four NVIDIA A100 graphics processing units (GPUs), each having 40 GB of video random access memory (VRAM). The four GPUs are partitioned across tasks: two GPUs are allocated to XLM-RoBERTa-large encoder fine-tuning (with gradient accumulation enabling an effective batch size of 64 within the 40 GB memory budget), while the remaining two GPUs serve Llama-3-70B inference under 4-bit quantization through vLLM, supporting a sustained throughput of approximately 75 sentences per second per GPU.

This task-level partitioning allows fine-tuning and inference workloads to run concurrently without contention, and supports end-to-end document processing within the latency budget reported below. The inference was optimized by the vLLM, which is a high-throughput serving engine that utilizes PagedAttention for efficient memory management. This resulted in a speedup of 2.3x over the native implementation of the Transformers library. The batch processing of 32 balanced throughput and memory constraints. End-to-end processing of the 240-document test set required 12 minutes (avg. 3.0s per document), comprising document loading and parsing (0.8s), text preprocessing and vectorization (1.2s), hybrid model inference (0.42s), and output formatting (0.6s).

Because the LLM and ML branches execute concurrently on the partitioned GPUs, the 0.42 s hybrid-inference figure is dominated by the LLM branch (0.40 s standalone), with the ML branch (0.08 s) hidden under concurrent execution and weighted fusion adding negligible overhead; the full system is therefore never faster than its slowest branch. The pure model inference time of 0.42s per document represents the core detection algorithm efficiency, while the additional 2.58s accounts for production-level data handling requirements. Throughout the manuscript, ‘‘0.42 seconds per document’’ refers exclusively to core model inference time, while end-to-end throughput is reported as 3.0 seconds per document. The complete end-to-end processing pipeline, including input/output specifications and intermediate representations at each stage, is summarized in Figure 2.

D. EVALUATION METRICS AND BASELINE COMPARISONS Model performance was evaluated using both quantita-tive and qualitative measures. The quantitative measures addressed detection accuracy, while the qualitative measures addressed mitigation utility. Standard classification metrics were computed at the sentence level for the six-class taxonomy. In this regard, the precision measure was used to determine the proportion of correctly classified bias examples out of all positive predictions. Recall, on the other hand, measured the proportion of all actual biases detected.

F1-score, the harmonic average of the above, was adopted as the primary measure of model performance, both macro and weighted, to account for class frequency imbalance. As the task is multi-label (Section III-C), we additionally report micro-F1, sample-averaged F1, subset (exact-match) accu-racy, and Hamming loss, computed from the six-dimensional predictions at the 0.5 threshold. In multi-class classification problems, where the semantic differences between the bias classes are harder to capture than binary classification, the Area Under the Receiver Operating Characteristic Curve (AUROC) was computed through the one-vs-rest strategy, where the AUROC was calculated for every bias class.

Confusion matrices were also adopted to gain further insight into the classification, especially where the differences between confirmation bias and overconfidence are difficult to capture, as both present similar linguistic features, such as the emphasis on certain information.

Quality evaluation of mitigation results followed a set of protocols based on automated measures of text simi-larity. The Bilingual Evaluation Understudy (BLEU-4) and Recall-Oriented Understudy for Gisting Evaluation–Longest Common Subsequence (ROUGE-L) measures were used to compare lexical similarity between machine-generated corrections and human-authored reference corrections.

Ref-erence corrections were authored by two of the three doctoral-level business-psychology experts who participated in stage-one annotation (Section III-B), each independently rewriting the same 200 bias-flagged sentences (100 English, 100 Chinese) under a written guideline requiring (i) factual preservation, (ii) removal of the identified bias indicator, (iii) adoption of corporate-disclosure register, and (iv) length within ±20% of the original sentence; the two reference versions were used as parallel references in the BLEU-4/ROUGE-L computation. BLEU-4 is based on n-gram precision, while ROUGE-L is based on longest subsequence common to two references. However, these measures are based on lexical similarity rather than content similarity, which is more important in the context of this problem.

A human evaluation study was conducted involving 30 busi-ness analysts with a minimum of five years of experience in corporate disclosure review. The analysts rated 20 examples of mitigation results (600 total evaluations) based on three criteria: relevance, actionability, and professionalism, each rated on a five-point Likert scale. The same 200 sentences were also rated under three added controls: a no-mitigation control (original sentence), a GPT-4 rewrite, and an expert human rewrite (upper bound). Factual preservation was measured by analyst-judged retention of all original claims and bidirectional NLI entailment. Inter-rater reliability was measured by Fleiss’ kappa statistic to ensure consistency in the evaluation process.

Four baseline approaches provided a comparative per-formance reference. (i) Rule-based: a keyword-and-regex matching system with 120 manually designed patterns (≈ 20 per bias category) covering lexical triggers (e.g., absolutist modal verbs, anchor phrases) and shallow syntactic templates; a sentence is labeled positive if it matches at least one pattern, with category assigned by the matching pattern. (ii) BERT-based: a fine-tuned BERT-base-multilingual-cased model (HuggingFace snapshot bert-base-multilingual-cased, 177M parameters) with a standard linear six-way classification head, trained on the 720-document training split using AdamW (lr = 3 × 10−5, weight decay = 0.01), batch size 32, 5 epochs, with the best checkpoint selected by validation F1.

An XLM-RoBERTa-base baseline (xlm-roberta-base, 270M parameters) was trained under identical hyperparameters to isolate the effect of architectural improvement from increased model scale. (iii) Zero-shot GPT-4: the gpt-4-0613 API variant queried with the same six bias-specific prompts as the LLM pathway in CogDe-Bias but with no in-context examples (temperature = 0, max tokens = 256). (iv) Few-shot Llama-3: the Llama-3-70B-Instruct (meta-llama/Meta-Llama-3-70B-Instruct) vari-ant queried with the same six bias-specific prompts plus five labeled examples per bias category drawn from the training split (30 examples in total), under identical decoding settings to the LLM pathway.

Processing costs were compared using Q3 2025 pricing: GPT-4 API at $0.03/$0.06 per 1K input/output tokens (avg. $0.09/document for 2,847-token documents) versus self-hosted infrastructure (AWS p4d.24xlarge at $32.77/hour, amortized over 50,000 documents/month typical enterprise volume, yielding $0.065/document including overhead). The 30% cost reduction represents operational savings in batch deployment at enterprise scale, excluding development costs. Low-volume scenarios (<5,000 docs/month) may favor API solutions.

Through informal consultation with 12 financial analysts (mean experience: 8 years), comprehensive bias-focused annual report review was estimated at 3-8 hours per document depending on complexity. This translates to 1-2 documents per 8-hour workday, conservatively reported as

‘‘2-3 documents per day’’ in Discussion when accounting for productivity variance.

Component-level ablation experiments evaluated four architectural choices: (i) frozen vs supervised contrastive fine-tuning of the XLM-RoBERTa encoder, (ii) the LLM-ML fusion weight λ across 0.0–1.0, (iii) individual ML feature groups (TF-IDF, syntactic parse, FinBERT sentiment) by leave-one-out removal, and (iv) bias-specific vs generic prompt templates. All ablations were trained and evaluated under identical hyperparameter and data-split protocols.

Classifier comparisons employed McNemar’s test; per-formance differences used paired t-tests; subset compar-isons used paired bootstrap (1,000 resamples). Across the manuscript, nine statistical tests were performed at α = 0.05. To control Type I error inflation, Benjamini–Hochberg false discovery rate (FDR) correction at q = 0.05 was applied to the nine reported p-values; results that remain significant under FDR correction are explicitly flagged in the surrounding text. The Bonferroni-corrected threshold (α’ = 0.00625) is also reported as a stricter reference for borderline cases.

IV. RESULTS.

A. OVERALL DETECTION PERFORMANCE

The CogDeBias framework scored an overall weighted F1-score of 0.82 on the 240-document test set, indicating strong performance in the detection of cognitive biases embedded in strategic business discourse. Bootstrap (1,000 resamples) gives a 95 percent confidence interval of [0.80, 0.84] for this F1, and the 3.5-point advantage over Few-shot Llama-3 corresponds to an odds ratio of 1.6 (McNemar), a small but consistent effect; per-language gaps and the temporal comparison report effect sizes elsewhere (English– Chinese, and rank-biserial r = 0.34 in Section V-A). As the task is multi-label (Section III-C), the primary metrics are micro-F1 = 0.83, sample-F1 = 0.81, subset accuracy = 0.74, and Hamming loss = 0.061; the weighted F1 = 0.82 is a per-category aggregate kept only for baseline comparability.

Performance varied systematically across the different bias categories, which is an inherent property of these biases, given their different levels of linguistic detectability and reasoning complexity. Confirmation bias was found to be the most accurately detected bias, scoring an F1 of 0.85, which can be attributed to the unique language features that include asymmetric information presentation and the use of hedging qualifiers for citing evidence. Overconfidence bias scored an F1 of 0.81, which can be attributed to the unique features that include absolutist language, such as unqualified future tense verb phrases and superlative comparatives without quantitative substantiation.

Sunk cost fallacy scored an F1 of 0.79, while the detection of anchoring effect scored an F1 of 0.78, which can be attributed to the inherent contextual ambiguity that arises in differentiating between historical context references and fallacious reasoning. Bandwagon effect detection scored an F1 of 0.80, which can be attributed to the unique features that include industry trend terminology and consensus-involving phrases that are relatively consistent across documents. Recency bias, on the other hand, was found to be the most challenging to detect, scoring an F1 of 0.76, which can be attributed to the inherent difficulty in differentiating between the disproportionate influence of recent events and historical trends, which requires temporal reasoning that is beyond the scope of current architectures.

A comparative evaluation against four different baseline techniques showed significant performance advantages for the hybrid CogDeBias approach, as shown in Table 5. The rule-based approach scored an F1 of 0.59, where pattern matching was brittle and unable to generalize beyond manually crafted heuristics and showed poor recall at 0.58 due to gaps in handling paraphrased expressions. BERT-base-multilingual, trained on the training set, scored an F1 of 0.72, showing that supervised training on labeled data significantly outperforms hand-crafted rules but is still bounded by its 177M parameter capacity for extracting complex contextual features.

Zero-shot GPT-4 scored an F1 of 0.76 without any training data, proving that frontier models do indeed possess innate bias detection capabilities; however, its 2.40 seconds per document evaluation time (5.7x slower than CogDeBias) makes this approach completely impractical for deployment. Few-shot Llama-3, evaluated against five examples from each category, scored an F1 of 0.785, showing the strongest baseline performance; however, CogDeBias outperformed this baseline by 3.5 percentage points (p = 0.008 using McNemar’s test), equivalent to an additional 57 correctly classified bias examples across the test set. The few-shot baseline (0.785) did not exceed the zero-shot LLM-only pathway (0.79; a 0.5-point gap within evaluation noise), as the calibrated bias-specific zero-shot prompts already approach their ceiling.

The 30 in-context examples add demonstration-selection and label-distribution biases that offset their benefit, especially for low-frequency categories such as recency bias. CogDeBias’s advantage thus stems from the hybrid LLM–ML fusion, not from in-context examples. This performance advantage was consistent across all types of bias, as shown in Figure 3, where CogDeBias showed 3-8% advantages across all six categories. To further isolate architectural from scale effects, an XLM-RoBERTa-base baseline (270M parameters, F1 = 0.745) was added: it surpasses BERT-base-multilingual (F1 = 0.72) by 2.5 percentage points, indicating a modest XLM-R architectural advantage at comparable scale, while CogDeBias (Full) gains a further 7.5 percentage points over XLM-R-base, attributable jointly to the encoder scale-up to large (550M) and the hybrid LLM–ML fusion.

This decomposition confirms that the principal performance gain comes from the hybrid architecture rather than encoder size alone.

Ablation studies were conducted to measure the contri-butions of the architectural components. Removing the ML ensemble branch and relying solely on LLM classification resulted in a drop in F1 from 0.82 to 0.79, suggesting that feature engineering extracts complementary information not fully captured by the learned representations—such as the use of syntactic structures like condition clauses, which are linked to anchoring bias. Training the model exclusively on English and then testing it on Chinese resulted in an F1 score of 0.74—a drop of 8 points—highlighting the necessity of fine-tuning the model despite the cross-lingual capacity of XLM-RoBERTa to recognize culturally embedded rhetorical devices.

Replacing the XLM-RoBERTa encoder that supplies the ML-branch representation with monolingual BERT encoders resulted in an F1 score of 0.76, affirming the value of multilingual pre-training beyond the obvious benefits of shared vocabularies to the alignment of semantic spaces.

Component-level ablations further quantified individual contributions. Supervised contrastive fine-tuning of XLM-RoBERTa improved F1 by 2.2 percentage points over frozen weights (validation set: 0.798 → 0.820). The 0.7:0.3 LLM:ML fusion ratio identified by validation grid search was confirmed on the held-out test set (F1 = 0.821) and remained within ±0.005 of the optimum across λ ∈ [0.6, 0.8]. Within the ML branch, TF-IDF features proved most impactful (−3.1% F1 when removed), followed by syntactic parse features (−2.4%) and sentiment scores (−1.8%). Bias-specific prompt templates outperformed generic prompts by 4.8% F1. These findings demonstrate that performance arises from architectural synergy rather than over-reliance on any single component.

Summarizing the four design components: supervised contrastive fine-tuning contributes +2.2 percentage points over frozen weights, bias-specific prompt engineering +4.8 points over generic prompts, multilingual representation (XLM-RoBERTa vs mBERT) +5.5 points...Each component yields a measurable gain, although the ML-branch contribution (−3.0 points when removed, p = 0.018) does not survive FDR correction and is therefore treated as suggestive.

Across the 1,200-document corpus, the average document contains 6.7 bias-positive sentences (standard deviation ≈ 3.6, range 0–19), consistent with Table 2. Confirmation bias and overconfidence combined account for 50.4% of all detected bias instances, while the remaining four categories collectively account for 49.6%. These descriptive patterns motivate the bias-by-bias performance analysis reported above and the temporal contextualization presented in Section V-A.

To address potential LLM-to-LLM label transfer bias, we partitioned the 240-document test set by annotation provenance: 60 documents stratified from the stage-one expert-only pool and 180 from the stage-two LLM-assisted pool. CogDeBias achieved weighted F1 = 0.81 on the expert-only-derived subset and F1 = 0.83 on the LLM-assisted-derived subset. The 2-percentage-point gap, tested via paired bootstrap (1,000 resamples), was marginally significant (p = 0.087), and widened to roughly 3 percentage points on confirmation bias and overconfidence — the two categories most exposed to GPT-4 stylistic preferences. The gap is consistent with mild but detectable annotation-source effects, though too small to alter the principal cross-baseline comparisons.

To verify performance against a model-free gold standard, the three stage-one experts independently re-annotated 120 test documents (50 percent of the test set; 60 English, 60 Chinese) from scratch, without any model-generated suggestions. On this fully expert-labelled subset

CogDeBias attained a weighted F1 of 0.81, only 1 point below the full-test value (0.82), and the strongest baseline (Few-shot Llama-3) reached 0.775, preserving the roughly 3.5-point CogDeBias advantage. This expert-only value, rather than the LLM-assisted 0.82, should be regarded as the model-independent estimate of performance. Sentence-level agreement between the model-free re-annotation and the original test labels was Cohen’s kappa of 0.79, indi-cating that GPT-4 assistance did not materially inflate the reported test performance.

Taken together, three checks bound the LLM-induced annotation effect: stage-two human verification accepted only 76.4 percent of GPT-4 candidates as-is (15.1 percent recategorized, 6.7 percent span-adjusted, 1.8 percent rejected, with missed instances added); the expert-only versus LLM-assisted subset gap was 2 percentage points (p = 0.087); and the model-free expert re-annotation differed from reported test performance by only 1 point (F1 0.81 versus 0.82, agreement Cohen’s kappa 0.79). The residual effect is small, around 1 to 2 percentage points, detectable mainly on confirmation and overconfidence, and does not alter the principal cross-baseline comparisons.

B. MULTILINGUAL PERFORMANCE ANALYSIS

Language-specific performance patterns highlighted the strengths and weaknesses of cross-lingual bias detection. English texts from the test set (n = 120) scored an F1 = 0.84, owing to greater pre-training familiarity with English linguistic structures and an increased number of English business texts within the training data set. Accuracy of detection for English texts was found to be particularly high for linguistic features of overconfidence bias—the text ‘‘We remain confident despite market headwinds’’ was correctly detected for overconfidence bias with 0.91 probability via the LLM pathway, as the text contained an inherent contradiction between expressed confidence and acknowledged adverse market conditions.

The linguistic style of English corporate discourse tends to favor direct and assertive statements that make detection of bias features easier, such as unqualified statements about future actions (‘‘will deliver 20% growth’’) and absolute comparative statements about performance (‘‘industry-leading performance’’). Chinese texts from the test set (n = 120) scored an F1 = 0.80, showing a statistically significant 4-percentage-point underperformance (p = 0.032, paired t-test) due to cultural-linguistic factors that make detection of bias features less clear. Business discourse in Mandarin tends to use euphemistic expressions where features of bias are less clear—the phrase ‘‘cau-tiously optimistic’’ paradoxically may be an indicator of overconfidence bias, as the text contains an inherent duality that acknowledges uncertainty and future positive outcomes simultaneously.

The initial performance of the Chinese test set was an F1 = 0.78, improving to an F1 = 0.80 after inclusion of domain-specific fine-tuning data from 50,000 sentences of Chinese financial news corpora. Because the overall English-vs-Chinese gap is non-significant after FDR correction (p = 0.032), the following per-category breakdown is descriptive and should be read as suggestive. The gap is uneven across categories: largest for overconfidence (English 0.84 vs Chinese 0.78), where Mandarin disclosures favor hedged wording (e.g., ‘‘cautiously optimistic’’) that mutes absolutist markers, followed by bandwagon (0.83 vs 0.78) and anchoring (0.80 vs 0.76) with more implicit cues; sunk cost shows the smallest gap (0.80 vs 0.79) after the 28 Chinese rules, and confirmation stays strong (0.86 vs 0.83). Degradation thus concentrates in categories whose markers are culturally softened in Chinese.

Further, cross-lingual transfer experiments were con-ducted, which quantified the degree of generalization across languages, as presented in Table 6. Specifically, zero-shot transfer, where a model is only exposed to English documents and tested on Chinese, resulted in a low F1 measure of 0.68, a 14-point drop from the full bilingual training scenario. This significant drop in performance underscores that, while a shared semantic space between languages is possible, as achieved through XLM-RoBERTa, bias-related linguistic effects have language-specific manifestations that are not entirely representational-level equivalent. The key observa-tion, however, is that adding just 100 Chinese documents, which comprised just 8.3% of the entire dataset, resulted in a Chinese F1 measure of 0.78, which recovered 71.4% of the lost performance (gap narrowed from 0.14 to 0.04).

This rapid learning curve suggests that even a small amount of target-language supervision may have a disproportionately large effect on learning bias detection patterns, likely due to its effect on calibrating decision boundaries within a shared multilingual space. As this few-shot adaptation effect does not survive Benjamini–Hochberg correction (p = 0.041), it is reported as suggestive rather than confirmatory.

Language-specific bias manifestation patterns called for the development of specific detection strategies. English bandwagon effects, for instance, predominantly manifest through industry comparison terminology, where ‘‘industry leading innovation,’’ ‘‘adoption of best practices,’’ and ‘‘market consensus’’ indicate the presence of uncritical trend-follower behavior. The ML branch’s TF-IDF features were particularly effective for English bandwagon detec-tion, achieving an F1 of 0.83, owing to the consistent keyword-based manifestation patterns. In the case of the Chinese sunk cost fallacy, the expression patterns were less direct. They typically involve phrases meaning ’hav-ing invested substantial resources’ and ‘cannot give up halfway,’ which embed cultural values around perseverance.

Distinguishing rational persistence from fallacious reasoning in such phrasing is therefore challenging. Following the language-specific rule augmentation in the Chinese ML branch (Section III-D), sunk cost fallacy detection improved from F1 = 0.76 to F1 = 0.79. On an independent external set from sources not used in training—40 London-Stock-Exchange English reports and 40 Shenzhen Stock Exchange Chinese reports (about 2,000 sentences), annotated by the stage-one experts without model suggestions—CogDeBias attained weighted F1 = 0.78 (English 0.79, Chinese 0.76), a 4-point drop from the internal 0.82 reflecting cross-market shift, while the strongest baseline reached 0.74. Given the modest size of this external set (80 documents, ∼2,000 sentences), this is preliminary evidence of cross-market trans-fer rather than a generalizability guarantee.

As a simplified single-label diagnostic, Figure 4 reduces each sentence to its top-probability predicted and primary annotated category to show dominant inter-category confusion; the most prominent is the 28-instance overconfidence–confirmation overlap. Sentences such as ‘‘Market results validate our strategic judgment’’ simultaneously exhibit self-assured predictions and the interpretation of selected evidence, reflecting genuine taxonomic overlap rather than purely classifier weakness — multi-label re-annotation of these 28 cases by the three domain experts indicated that 17 of the 28 (61%) carry both bias signals concurrently, suggesting that joint occurrence is a property of the discourse itself. A second, smaller overlap occurs between recency bias and anchoring: both reference a temporal value point, so a sentence citing a recent figure as a decision basis resembles both.

This produced roughly 10 bidirectional confusions in Figure 4, about 38 percent carrying both signals—weaker than the confirmation–overconfidence pair (61 percent). Recency bias, the lowest-frequency category (6.7 percent), has limited training signal, consistent with its lowest F1 (0.76) and anchoring’s 0.78.

Two architectural strategies were considered to address this overlap. First, a two-stage hierarchical detector was prototyped, in which Stage 1 separates ‘‘self-assertion-related biases’’ (overconfidence + confirmation bias) from the remaining four categories, and Stage 2 performs a fine-grained binary discrimination within the self-assertion cluster. On the test set this design improved within-cluster discrimination by 2.1 percentage points (overconfidence F1 from 0.81 to 0.83; confirmation F1 from 0.85 to 0.87) but slightly increased Stage-1 errors that propagate to Stage 2, yielding only a 0.3-percentage-point net improvement in overall weighted F1 (0.82 → 0.823). Second, allowing soft multi-label outputs at inference (consistent with the multi-label formulation in Section III-C) eliminated 11 of the 28 confusions by accepting both labels jointly.

Given the marginal gain of the hierarchical design and the conceptual fit of multi-label outputs with the underlying discourse, we adopted the multi-label-output strategy in the final system. Future taxonomic refinement could merge overconfidence and confirmation bias into a single ‘‘self-assertion bias’’ super-category at coarse granularity, with finer subdivision applied only when downstream analysis requires it.

C. BIAS MITIGATION QUALITY ASSESSMENT

The quality of the generated mitigation suggestions was automatically evaluated through quantitative metrics, which measured the lexical and semantic quality of the generated text against the quality of the text produced by humans, i.e., the reference text. In this regard, the BLEU-4 score, which measures n-gram precision between the generated text and the reference text, was 0.58 for 600 test cases, while the ROUGE-L score, which measures the recall of the longest common subsequence between the generated text and the reference text, was 0.64. As these scores are computed against only two reference rewrites over 200 sentences, they are lexical-overlap indicators rather than quality measures, and the low absolute values are expected given the many valid debiasings of a sentence; the human evaluation below is therefore the primary basis for assessing mitigation quality.

Finally, the efficiency of the generation model was appropriate for production use, as Mixtral-8x7B was able to produce three alternative suggestions for each detected bias instance in an average of 1.2 seconds, per bias instance. With an average of 6.7 bias-positive sentences per document in the test set, the complete detection-and-mitigation pipeline processed documents in approximately 11 seconds each (3.0s detection + 8.0s mitigation generation). With batch processing (batch size 32), this throughput supports high-volume batch analysis rather than per-document real-time use, which is well matched to earnings-season workloads where hundreds of 10-K reports are released over short periods.

Significantly, the human evaluation also offered critical assessment criteria for the various dimensions of practical utility, which the automated metrics could not provide, as shown in Table 7. We recruited 30 business analysts from investment research firms and corporate finance departments, requiring a minimum of 5 years of experience in analyz-ing management discourse. All 30 analysts independently evaluated the same set of 20 randomly selected mitigation outputs (stratified to cover all six bias categories), yielding a fully overlapping 30 × 20 rating matrix per dimension and 600 evaluations per dimension overall, on which Fleiss’ kappa was computed.

We also provided standardized instructions for the analysts, defining the 3-dimension evaluation criteria on 5-point Likert scales (1 = lowest, 5 = highest) with the following anchor descriptors: Relevance — 1 = does not address the identified bias or introduces unrelated issues; 3 = partially addresses the bias with some extraneous content; 5 = directly and fully addresses the identified bias with no extraneous content. Actionability — 1 = abstract or generic, no implementable guidance; 3 = moderately specific but missing concrete steps; 5 = immediately implementable, with concrete instructions for corporate communications teams.

Professionalism — 1 = inappropriate tone or terminology for business contexts; 3 = generally acceptable but with stylistic inconsistencies; 5 = fully suitable in tone, terminology, and rhetorical register for shareholder-facing disclosure. ‘‘Relevance’’ scored the highest, with a mean rating of 4.3/5 (σ = 0.6), indicating 86% of the suggestions were rated as ‘‘highly relevant’’ (4/5) or ‘‘completely relevant’’ (5/5), showing the strong match between the identified bias and the remediation strategies. ‘‘Actionability’’ scored 4.1/5 (σ = 0.7), suggesting the majority of the suggestions were actionable, providing implementation guidance rather than abstract principles, e.g., ‘‘commission independent third party technical audit’’ instead of ‘‘seek external validation.’’ ‘‘Professionalism’’ scored 4.0/5, but with slightly higher standard deviation

(σ =0.7), mainly due to the somewhat academic tone in some suggestions targeting shareholder communications. We also measured the inter-rater reliability using Fleiss’ kappa, which scored 0.71, showing substantial agreement among the analysts, thereby validating the consistency of the evaluation criteria applied by the analysts. CogDeBias far exceeded the no-mitigation control (relevance 1.6, actionability 1.4), approached the GPT-4 rewrite (4.5 / 4.3), and trailed the human upper bound (4.7 / 4.5), retaining all factual claims in 94 percent of cases (NLI entailment 0.91). For bias-reduction efficiency, re-running the detector on the rewritten sentences showed that 89 percent no longer triggered the original bias, confirming the revisions remove the targeted distortion while preserving facts.

Figure 5 shows the distribution of the ratings, which are right-skewed, with the majority of the ratings concentrated towards the higher ratings, but with 10-15% of the suggestions requiring refinement, particularly for the recency bias, where the complexity of temporal reasoning makes the quality of the automatically generated text less satisfactory.

In addition, qualitative case analysis showed the effec-tiveness of the system in providing contextually relevant and strategically sound mitigation suggestions.

A Chinese-language sunk cost fallacy case from the annual report of a tech firm stated: ‘‘Given the RMB 300 million already spent in R&D costs, management decided to continue advancing this project.’’ CogDeBias identified this as a sunk cost fallacy with 0.89 confidence and offered three relevant suggestions as follows: ‘‘Based on future market potential assessment, the project’s expected ROI is X%, with payback period of Y years...’’ to shift the focus from past costs to future value creation; ‘‘Independent third-party assessment indicates...’’ to utilize the power of independent verification to overcome the confirmation bias; and ‘‘Management has established staged exit criteria: if technical milestones are not met within 6 months...’’ to utilize the power of precommitment to overcome escalation of commitment.

The relevance and actionability of the case were rated as 4.6/5 and 4.4/5, respectively, as the suggestions offered ‘‘imme-diately applicable language for investor presentations’’ and ‘‘addresses the root cognitive distortion without appearing defensive.’’

To justify the chosen configuration, a 4 × 3 grid search was conducted over temperature ∈ {0.3, 0.5, 0.7, 0.9} and nucleus threshold p ∈ {0.7, 0.9, 0.95}, evaluated on a 60-sample stratified validation subset using BLEU-4, ROUGE-L, and a lightweight three-analyst relevance rating. As shown in Table 8, low-temperature settings (T = 0.3) produced lexically conservative outputs with the highest BLEU-4 (0.62) but reduced suggestion diversity (analyst relevance 3.7/5). High-temperature settings (T = 0.9) introduced semantic drift, lowering ROUGE-L from 0.64 to 0.58 and analyst relevance from 4.1/5 to 3.6/5. The configuration T = 0.7, p = 0.9 yielded the best joint trade-off (BLEU-4 = 0.58, ROUGE-L = 0.64, analyst relevance = 4.1/5) and was therefore adopted for the main experiments.

V. DISCUSSION.

A. KEY FINDINGS AND PRACTICAL IMPLICATIONS

Confirmation bias and overconfidence dominate the bias landscape revealed in Section IV-A, consistent with established behavioral-finance literature that these are the most prevalent cognitive biases in financial decision-making contexts. The dominance of overconfidence is particularly noteworthy given evidence that this bias materially degrades investment-decision quality and portfolio performance. The temporal analysis of the document corpus showed that cognitive biases are highly contextual. Documents filed during industry-specific stress periods (n = 420, e.g., 2024 reports from semiconductor firms amidst export-control uncertainty) contained on average 9.0 bias-positive sentences per document (95% CI: 8.5–9.5), compared with 6.4 sentences per document (95% CI: 6.2–6.6) in the baseline group (n = 780), a relative increase of about 40%.

A Mann–Whitney U test confirmed the difference is statistically significant (U = 209,114, p < 0.001), with rank-biserial correlation r = 0.34 indicating a medium effect size. After Benjamini–Hochberg FDR correction across all manuscript-level tests (Section III.E), this result remains significant. This is consistent with the established literature that financial factors exacerbate cognitive biases within analyst forecasts.

The CogDeBias framework facilitates a variety of high-value applications in investment, governance, regula-tion, and education. For example, in investment, automated bias annotation facilitates a move from qualitative disclosure analysis based on interpretation to a more systematic analysis based on risk flagging. This lets portfolio managers define thresholds for further investigation—for example, when a 10-K filing contains three or more overconfidence indicators. Internal audit functions in corporations use the system as a preemptive quality control tool, analyzing draft strategic reports prior to board approval in order to identify manifestations of groupthink.

Finally, in the realm of regulation technology, applications based on the system allow securities commissions to monitor the quality of disclosure reports from thousands of listed companies, an alternative to recent machine learning technology applications in ESG disclosure analysis that focus on cognitive risk factors rather than content analysis. The system’s ability to produce real-world cases that illustrate the use of language in reasoning fallacies represents a valuable tool in business education, providing MBA programs with a basis in behavioral finance theory that is illustrated through modern corporate practice. Across all these applications, the system’s outputs are intended to inform human reviewers rather than to replace their judgment.

B. COMPARISON WITH EXISTING APPROACHES

CogDeBias demonstrates substantial advantages over tra-ditional manual analysis across efficiency, consistency, and scalability dimensions. Processing the 1,200-document corpus required approximately 60 minutes on the four-GPU infrastructure, contrasting sharply with estimated human analyst throughput of 2-3 documents per day, which would necessitate 12-18 months for equivalent coverage. Beyond speed gains, automated detection eliminates inter-rater vari-ability that plagues subjective assessments—expert annotator agreement (Fleiss’ kappa = 0.74) inherently constrains the upper bound of human consistency, whereas CogDeBias applies labeling criteria deterministically and thus yields fully reproducible outputs.

However, the system remains vulner-able to complex metaphorical expressions and culturally-embedded idioms requiring contextual world knowledge, necessitating human review for approximately 8% of flagged instances exhibiting ambiguous rhetorical structures.

The comparative assessment with pure large language model (LLM) architectures serves to support the design rationale of the hybrid approach. In fact, research on irrationality patterns in large language models, as reported in, shows that frontier LLMs have residual cognitive biases despite the alignment, which may compromise the reliability of the detection outcome when LLMs are solely used to classify bias-laden text samples. In this respect, the integrated machine learning (ML) branch helps to overcome this limitation, as the explicit feature engineering of syntactical patterns and sentiment markers is able to reduce false positive detection rates by 5.1% with respect to pure LLM-based text classification. This is the concrete advantage over single-model multilingual classifiers.

The hybrid fusion (Table 5) lifts F1 from 0.79 (LLM-only) and 0.695 (ML-only) to 0.82, with each component supplying what the others lack. The LLM contributes contextual reasoning; the ML branch adds traceable features and a 5.1% false-positive reduction; and the XLM-RoBERTa embedding provides cross-lingual alignment. No stand-alone Llama-3 or XLM-RoBERTa classifier provides all three. Interpretability is another significant advantage of the hybrid approach, as although the neural attention-based mechanisms of LLMs are not easily interpretable, the ML branch provides traceable feature attributions, such as TF-IDF weights and parsing statistics, which may contribute to higher stakeholder trust, as is relevant in the context of increasing regulatory pressure on decision algorithms. Concretely, leave-one-out importance (Section IV-A) ranks the driving cues—TF-IDF

(−3.1 percent F1), syntactic parse (−2.4 percent), FinBERT sentiment (−1.8 percent)—while the LLM branch returns a per-category confidence and the matched bias phrase (e.g., overconfidence at 0.91, sunk cost at 0.89; Section IV-C), letting analysts audit why each sentence is flagged. In fact, research on Explainable AI methodologies for structured data shows that the use of transparency mechanisms may have a significant impact on the acceptance of AI-based decision systems in critical applications. Hybrid frameworks combining convolutional and recurrent neural networks for aspect-based sentiment analysis have shown promise in utilizing complementary semantic features. This supports the application of similar architectural paradigms for bias detection.

Feature extraction techniques also play an important role in making the system more robust, using features that are independent of learned feature spaces, while hybrid models using transformers and recurrent networks have shown promising results for better performance in financial text analysis. The hybrid paradigm also lowers per-document cost relative to a pure GPT-4 API solution, attributable to self-hosted Llama-3-70B and Mixtral-8x7B inference rather than per-token API billing (Section IV-C).

Deployment considerations. Five practical issues bound real use. False positives: guidance-heavy SEC Item 7 sen-tences yield an estimated 2.5 percent false-positive rate (Section V-C), so outputs are flagged for human review, not auto-applied. Domain adaptation: cross-market shift costs about 4 points (Section IV-B), recoverable with minimal target-domain fine-tuning (Section IV-B). Computational cost: about 0.065 per document self-hosted, 30 percent below a GPT-4 API solution at enterprise volume (Section III.E), though four A100 GPUs are required. Deployment mode: batch rather than per-document real-time use (Section IV-C). User acceptance: traceable feature attributions and con-fidence scores support analyst trust.

Both pathways are invoked in parallel for every prediction in line with the weighted-fusion mechanism in Section III-C; the framework is not a cascading architecture, and ‘selective invocation’ is not used.

Ethical considerations. Because flagging bias in disclo-sures carries regulatory and reputational consequences, the system is decision support requiring human sign-off, not automated judgment (Section IV-C). Detection is uneven across languages (Chinese 4 points below English) and may vary by industry, so flags need language- and sector-aware calibration before cross-firm comparison. Feature attributions and confidence scores (Section V-B) keep each flag inspectable and reviewable by a qualified analyst, who holds final determination. The framework should inform, not replace, professional and legal judgment.

C. LIMITATIONS AND FUTURE DIRECTIONS

There are several constraints affecting the present study’s scope and applicability. The linguistic support is restricted to

English and Chinese, failing to include other economically important languages like Japanese, German, and Spanish, which together encompass a considerable share of the world’s capital markets. Expanding support to other languages is hindered by data scarcity. XLM-RoBERTa is designed to support 100 languages in theory. However, detecting bias in additional languages requires annotated datasets that are currently unavailable. The mitigation strategies have a template-based nature; they do not include in-depth cus-tomization for various industries, which could have provided a distinction in the use of appropriate rhetorical styles in technology and pharmaceutical sectors.

Because system rewrites remain below the expert human upper bound and cor-porate disclosures carry legal and regulatory consequences, the mitigation component is intended as decision support requiring human sign-off, not as an automated correction tool. The focus on text-only content ignores the integration of other information types. Financial charts and tables, along with quantitative metrics, have a significant impact on the content and influence investors’ perception. However, this is beyond the capabilities of current architectures. This is similar to the sentiment analysis of complex social phenomena, as discussed in, where text-based cues are considered to be insufficient.

External validation on UK and Shenzhen disclosures (Section IV-B) showed a modest 4-point drop, but the external set is small (80 documents); the due-diligence, audit, and regulatory-monitoring applications in Section V-A should therefore be read as target use cases pending larger multi-market validation, not as validated deployments. Following reviewer recommendation, Benjamini–Hochberg FDR correction at q = 0.05 has been applied across the nine statistical tests reported in this work. Of these, five comparisons remain significant after FDR correction (the principal CogDeBias vs.

Few-shot Llama-3 advantage, the Zero-shot Transfer drop, the Remove LLM Branch ablation, and the mBERT replacement ablation, and the stress-period vs. baseline temporal comparison), while three results (Few-shot Adaptation, Remove ML Branch, English-vs-Chinese gap) become non-significant and are now interpreted as suggestive. Future confirmatory studies should additionally employ pre-registered testing protocols.

Future research directions that address these gaps include the following extensions: analysis of individual firm bias trajectories over time, or longitudinal analysis spanning fiscal years, to potentially uncover temporal effects, i.e., whether cognitive biases increase during periods of dis-tress or relate to subsequent performance decline. The development of multimodal fusion techniques that combine information from financial ratios, stock price volatility, and disclosure sentiment could potentially allow for a holistic approach to risk assessment, taking advantage of the machine learning capability to identify patterns in heterogeneous data sources.

The optimization of mitigation strategies through reinforcement learning offers opportunities to itera-tively optimize strategies based on outcome feedback in real-world scenarios, i.e., whether updated disclosure improves investor comprehension or reduces subsequent litigation risk. The extension of the range of included document types to earnings call transcripts, board meeting minutes, and internal emails could potentially allow for the detection of cognitive biases in organizational communication structures, similar to applications in sentiment analysis in the realm of social media political discourse, highlighting the flexibility of NLP techniques in dealing with different text genres.

A further taxonomic limitation concerns the inherent non-mutually-exclusive nature of cognitive biases. Overcon-fidence and confirmation bias by construction co-occur in self-assertion-laden discourse (61% joint-signal rate among the 28 misclassified cases observed in Section IV-B), and similar partial overlaps exist between other category pairs. This non-exclusivity places a structural ceiling on the maximum achievable single-label classification accuracy for any taxonomy that treats the six biases as disjoint classes — a limitation only partially offset by the multi-label output strategy adopted in this work (Section IV-B). Future work could explore a hierarchical ‘self-assertion super-category’ formulation as an alternative taxonomic structure.

Three failure classes warrant attention. First, sophisticated rhetorical obfuscation — layered hedging that dissolves overconfidence markers into qualified statements — accounts for roughly 4% of false negatives. Second, forward-looking SEC Item 7 disclosures sometimes share lexical patterns with overconfidence markers, generating an estimated 2.5% false-positive rate among guidance-heavy sentences. Third, adversarial robustness against deliberately crafted evasion attempts has not been formally tested. The 8% of flagged instances requiring human review (Section V-B) fall pri-marily into the first two categories. Human evaluation also remains inherently subjective.

Although a single-blind protocol with randomized order (Section IV-C) and Fleiss’ kappa = 0.71 mitigate part of this concern, analyst ratings remain susceptible to halo effects and rater fatigue; future evaluations should add a hypothesis-blinded analyst pool with counterbalanced human-authored controls. The evalua-tion also rests on 20 rated outputs and two reference rewrites per sentence; a larger rated sample and reference pool are needed before the mitigation quality estimates can be treated as tight.

A further caveat concerns LLM-to-LLM label transfer bias, since 75% of training documents used GPT-4 candidate labels, and patterns favored by GPT-4 may be dispro-portionately easy for Llama-3 to detect. As quantified in Section IV-A, three checks bound this effect: a 2-point expert-only versus LLM-assisted gap (p = 0.087), a model-free expert re-annotation of half the test set that differs from the full-test F1 by only 1 point (0.81 vs 0.82; Cohen’s kappa 0.79), and a 76.4% raw acceptance rate of GPT-4 candidates. The residual effect is thus small but non-zero. GPT-4’s own stylistic preferences may have been partially imprinted despite two-tier verification, so the ground truth should be regarded as approximating, not fully equivalent to, an expert-authored gold standard.

Future work should validate the framework using annotators from structurally different models, and because only half the test set was re-annotated model-free, the 0.82 figure should not yet be treated as a definitive expert-grounded benchmark.

VI. CONCLUSION.

This study introduced CogDeBias, which to our knowledge is among the first frameworks to jointly perform detection and mitigation of cognitive biases in bilingual (English– Chinese) corporate disclosures, with the distinguishing combination shown in Table 1. By integrating transformer-based multilingual encoding, hybrid LLM-ML classification, and targeted suggestion generation, the system addresses critical gaps in behavioral finance tooling where manual analysis fails to scale and existing NLP applications overlook systematic reasoning distortions embedded in management discourse.

Three primary contributions advance the field. First, construction of a bilingual annotated corpus comprising 1,200 corporate documents (600 English, 600 Chinese) with 8,082 sentence-level bias annotations across six theoretically-grounded categories establishes foundational infrastructure for computational bias detection research. Sec-ond, empirical validation demonstrated F1 = 0.82 detection performance, surpassing the strongest baseline (Few-shot Llama-3, F1 = 0.785) by 3.5 percentage points (p = 0.008, McNemar’s test) while maintaining 0.42-second core model inference (3.0-second end-to-end) per document, supporting scalable batch deployment.

Third, mitigation quality assess-ment revealed that system-generated suggestions achieved mean human ratings of 4.1/5 for actionability, with 86% judged highly relevant to identified biases, indicating prac-tical utility beyond academic benchmarking. The generator is intended as decision support requiring human sign-off, not as an automated correction tool for regulated disclosures.

Methodologically, this work establishes a paradigm for applying natural language processing to decision science phenomena, demonstrating that linguistic patterns reliably signal cognitive distortions across languages and corporate contexts. Practically, the framework supports batch process-ing to assist investor due diligence, corporate governance audit, and regulatory monitoring workflows. Future work will pursue three concrete directions. First, the framework will be extended to Japanese (Tokyo Stock Exchange) and German (Frankfurt Stock Exchange) filings via few-shot adaptation, with pilot annotation within 12 months and full evaluation within 24 months of publication. Second, longitudinal validation across fiscal years 2020–2027 will examine whether bias intensity predicts subsequent earnings restatements and securities litigation.

Third, the de-identified annotation codebook and detection prompts will be made available from the corresponding author upon reasonable academic request, consistent with the Data Availability statement. As organizations increasingly rely on algorith-mic decision support, tools that surface human cognitive limitations embedded in strategic narratives become essential complements to traditional quantitative risk assessment.

ACKNOWLEDGMENT.

The authors would like to thank the financial analysts and domain experts who participated in the annotation and evaluation processes. The author Wang Yang acknowledges the support from the School of Business Administration, Huaqiao University.

DATA AVAILABILITY.

The annotated corpus and trained model checkpoints are not publicly available. The underlying corporate disclo-sure documents are subject to redistribution restrictions imposed by SEC EDGAR, the Shanghai Stock Exchange, the Hong Kong Stock Exchange, and OpenCorporates, and the sentence-level annotations contain verbatim excerpts of these documents. The human-evaluation data were collected under confidentiality agreements with the 30 participating analysts. Complete implementation details — taxonomy operationalization, codebook structure, detection prompts, training hyperparameters, and baseline configurations — are reported within the manuscript to support independent replication. The de-identified codebook is available from the corresponding author upon reasonable academic request.