A multi-label fraud identification method and system for financial reporting

Abstract

The application discloses a kind of multi-label fraud identification method and system for financial report, with the management discussion and analysis MD&A text and structured financial indicators in the annual financial report of enterprise as input. Utilize large language model to generate mechanism explanation from three perspectives of emotional tone, text readability and statement clarity, split the explanation into sub-units and interact with MD&A paragraph with bidirectional attention, determine the usefulness weight of each perspective according to the original text response;Through cross network and deep network, financial features are extracted, and relevant evidence in MD&A paragraph is queried by financial state representation. The text risk representation, financial state representation and financial state driven text enhancement representation are fused, and the probability of each type of fraud is output, and an asymmetric loss optimization model is used. The application can reduce the interference of irrelevant explanation and irrelevant text, realize the identification of multiple fraud types of the same enterprise, and provide a basis for enterprise risk screening and fraud type judgment.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of financial technology and natural language processing, specifically involving a method and system for identifying multiple types of fraud risks in the financial reports of listed companies. Background Technology

[0002] Identifying financial fraud in listed companies is a core requirement of capital market regulation. In recent years, researchers have explored AI-based methods for identifying financial fraud from various perspectives, but all have significant limitations.

[0003] 1. Traditional binary classification methods cannot support fine-grained diagnosis.

[0004] Existing research on financial fraud detection mainly revolves around enterprise-level binary classification (fraud / non-fraud). For example, existing technology 1 (Wang G, Ma J, Chen G. Attentive statement fraud detection: Distinguishing multimodal financial data with fine-grained attention[J]. Decision Support Systems, 2023, 167: 113913.) proposes a multimodal deep learning method RCMA based on an attention mechanism. It integrates financial indicators and annual report text through a triple attention mechanism of ratio awareness, chapter awareness, and modality awareness. However, its output is only a single fraud probability, which cannot reveal what type of fraudulent behavior the enterprise is involved in. Such methods are difficult to meet the regulatory requirement of "not only knowing who has problems, but also knowing what problems they have".

[0005] 2. The multi-label method only uses structured financial indicators and does not utilize textual information.

[0006] Existing technology 2 (Tragouda M, Doumpos M, Zopounidis C. Identification of fraudulent financial statements through a multi-label classification approach[J]. Intelligent Systems in Accounting, Finance and Management, 2024, 31(2):e1564.) uses a multi-label classification method for financial fraud detection; existing technology 3 (Chen Z, Chen H, Liu Y, et al. Multi-label financial statement fraud detection based on long short-term memory and multilayer perceptron hybrid model[J]. Engineering Applications of Artificial Intelligence, 2026, 170: 114188.) further classifies fraudulent behavior into five categories: profit overstatement, asset overstatement, false statements, delayed disclosure, and material omissions, and uses an LSTM-MLP hybrid model for multi-label classification. However, the above methods only use structured financial indicators, macroeconomic characteristics, and industry characteristics, and do not utilize the textual information in the annual report at all. Annual report texts (especially the MD&A section) contain a wealth of management intentions and risk signals; discarding this textual information would result in the loss of a large number of valuable clues about fraud.

[0007] 3. Existing large language model-based financial fraud detection methods mainly adopt the method of directly fusing the generated results.

[0008] The FSFDLLM framework proposed in Existing Technology 4 (Dai Y, Qian Y, Wang X, et al. FSFDLLM: Financial statement fraud detection aided by large language models[J]. Intelligent Computing,2026, 5: 0274.) utilizes a large language model to perform deep feature extraction and interpretable reasoning on financial text through a structured prompting mechanism. It uses the pseudo-labels and inference chains generated by LLM as enhanced features to jointly train the classification model with structured financial data. Existing Technology 5 (Zhang Lu, Song Changhao, Dong Chuanghao, et al. A framework for financial fraud detection integrating large model prompting learning and efficient parameter fine-tuning[J / OL]. Computer Engineering and Applications, 1-19[2026-09-10.]) also adopts a similar approach, using the SAGE framework and thought chain prompts to generate a principle explanation from MD&A text and then concatenating it with operating indicators. However, this "direct fusion after generation" paradigm has inherent flaws: First, there is a lack of interactive verification mechanism between the generated description and the original text, making it impossible to confirm whether the analytical conclusions are based on the original text; second, all samples use the analytical conclusions from each perspective equally, making it impossible to dynamically determine which perspectives are truly useful for the current sample; third, there is a lack of directed correlation between financial information and textual information, failing to establish a directed information correlation path of "financial anomalies → finding explanations / concealment evidence in the text".

[0009] 4. The integration of financial and textual information lacks a directional connection.

[0010] Existing methods for integrating financial indicators and textual information often employ feature concatenation, equal attention fusion, or JSON text parallelism. While the invention application CN117994055A, titled "Method, Apparatus, Device, Medium, and Program Product for Identifying Corporate Financial Fraud," uses a self-attention mechanism to process financial indicators and MD&A section text, its fusion method remains equal fusion. These solutions all treat the two types of information as equals, failing to consider the asymmetric relationship between them—abnormal states of financial indicators should be used as query conditions, actively guiding the model to search for corresponding explanations or evidence of concealment within the text.

[0011] In summary, existing intelligent financial report analysis methods lack fine-grained semantic consistency verification between externally generated interpretations and the original long text. This results in noise information in the generated interpretations that is inconsistent with or cannot be supported by the original text directly entering the classification model, polluting subsequent feature representations. At the same time, the lack of a state-conditional directional feature extraction mechanism between structured financial states and unstructured text paragraphs causes irrelevant text information to interfere with classification decisions during the fusion of heterogeneous features, reducing the accuracy of effective feature extraction. Summary of the Invention

[0012] The purpose of this invention is to provide a multi-label fraud identification method and system for financial reports. By decomposing the multi-perspective interpretation generated by a large language model into semantic sub-units and constructing a bidirectional interactive attention with the original MD&A text, fine-grained consistency verification between the generated interpretation and the original text and sample-level interpretation reliability assessment are achieved. At the same time, using structured financial status as query conditions, textual evidence related to the current financial status is extracted in a directed manner from the MD&A paragraph, realizing directed association and conditional fusion of heterogeneous data, thereby improving the accuracy of multi-label fraud type identification.

[0013] Technical solution:

[0014] This invention first discloses a multi-label fraud detection method for financial reporting, comprising the following steps:

[0015] Step T1: Obtain the management discussion and analysis (MD&A) text and structured financial indicators of the company to be identified, and input them into the multi-label fraud detection model;

[0016] Step T2: The multi-label fraud detection model uses a large language model to perform multi-perspective semantic analysis on the MD&A text through structured prompt templates, generating... A mechanistic explanation of the text from a different perspective;

[0017] Step T3: The MD&A text is segmented into paragraph sequences, and the MD&A paragraph vector sequences are encoded using a pre-trained encoder and projected into a common semantic space to obtain the MD&A paragraph semantic vector sequence of the sample to be identified. The superscript T indicates the text modality. The interpretation text from each perspective is encoded into a whole interpretation vector and a sub-unit vector sequence, wherein the sub-unit vector sequence is obtained by splitting the interpretation text into no less than 2 sub-units by sentence, encoding them separately, stacking them, and then projecting them onto a common semantic space.

[0018] Step T4: For each explanatory perspective, using its entire explanatory vector as the query and the MD&A paragraph vector sequence as the key and value, calculate the scaled dot product attention and pool it to obtain the evidence aggregation representation; using the MD&A paragraph vector sequence as the query and the sub-unit vector sequence of that perspective as the key and value, calculate the scaled dot product attention and mask pool it to obtain the restatement representation; generate usefulness weights from the restatement representation through a learnable network, and weight the evidence aggregation representation with the usefulness weights to obtain a weighted representation; The weighted representations from each perspective are aggregated after multi-head self-attention interaction and added to the mask mean of the MD&A paragraph vector to obtain the final text risk representation.

[0019] Step T5: Input the structured financial indicators into a dual-encoding network, one of which is an explicit high-order interactive cross network, and the other is an implicit nonlinear encoding deep network. The outputs of the two networks are concatenated and projected to obtain the financial status representation.

[0020] Step T6: Using the financial status representation as the query and the MD&A paragraph vector sequence as the key and value, perform cross-modal multi-head attention, and output the financial status-driven text augmentation representation after summing the projected residuals with learnable scaling coefficients;

[0021] Step T7: The text risk representation, financial status representation, and financial status-driven text augmentation representation are concatenated, fused, projected, and input into a multi-label classification head to obtain... Output the probability of each fraud type.

[0022] Preferably, the multi-label fraud detection model is trained through the following steps:

[0023] Step S1: Obtain the Management Discussion and Analysis (MD&A) text and structured financial indicators from the company's annual financial report, and construct a structure containing... A multi-label training dataset for non-mutually exclusive fraud behaviors, in which For the total number of fraudulent behavior categories, the first The label vector of each sample is , Indicates sample There exists the first Fraudulent behavior, This indicates that it does not exist;

[0024] Step S2: Utilize a large language model to perform multi-perspective semantic analysis on the MD&A text using structured prompt templates, and generate... A mechanistic explanation of the text from multiple perspectives, in which To analyze the total number of perspectives, let the first one be denoted as the first. The first sample The text is interpreted from one perspective. , ;

[0025] Step S3: Segment the MD&A text into paragraph sequences, and encode them into paragraph vector sequences using a pre-trained encoder. ,in Represents the real number field. For the first The total number of MD&A paragraphs in each sample. Let be the output dimension of the pre-trained encoder; project the paragraph vector sequence onto the common semantic space to obtain the _th ... Sequence of MD&A paragraph semantic vectors from training samples ,in For the public semantic space dimension; the first The textual interpretation from each perspective is encoded into a vector. And projected onto the public semantic space to obtain the entire interpretation vector. ; will the first The explanatory text from each perspective is broken down into sentences. Each subunit is encoded and stacked to obtain the original subunit encoding sequence. ,Will Projecting onto the common semantic space yields a sequence of sub-unit vectors. ,in For the first The number of subunits of the explanation and ;

[0026] Step S4: For the first Each interpretive perspective is used to construct a bidirectional interactive attention mechanism between itself and the MD&A paragraph vector sequence. Evidence aggregation representations of the original text interpreting the query and restatement representations of the original text query interpretations are calculated separately. These restatement representations are then used to generate usefulness weights for that perspective on the current sample through a learnable network. ,in It is the Sigmoid activation function. For the first Each perspective corresponds to a two-layer fully connected network. This is a restatement of the original query explanation; with The aggregated evidence representation is weighted; The weighted representations from each perspective are aggregated after multi-head self-attention interaction and added to the mask mean of the MD&A paragraph vectors to obtain the final text risk representation. ;

[0027] Step S5: The structured financial indicators Input a dual-channel coding network, where The total number of financial indicators is represented by two paths: one is an explicit high-order interactive cross-network, and the other is an implicit non-linear coding deep network. The outputs of the two paths are concatenated and projected to obtain the financial status representation. ;

[0028] Step S6: Represent the financial status For the query, the MD&A paragraph vector sequence is used as the key and value. Cross-modal multi-head attention is performed, and the output is summed after being projected by learnable scaling factors to obtain a financial state-driven text augmentation representation. ;

[0029] Step S7: [The sentence is incomplete and requires more context to be translated accurately.] , and After stitching and fusion projection, the data is input into a multi-label classification head to obtain... Each independent logical value is activated by a Sigmoid function to represent the probability of each fraud type. ;

[0030] Step S8: Calculate the sum of the losses for each label using asymmetric loss, and update the model parameters based on backpropagation of the total loss until the model converges, thus obtaining the trained multi-label fraud detection model.

[0031] Preferably, the The prompt template contains definition and reasoning instructions for three analysis dimensions:

[0032] Emotional tone analysis: Identify whether management's language contains excessively optimistic, pessimistic, or irrationally emotional expressions;

[0033] Text readability analysis: assesses the syntactic complexity, terminology redundancy, and logical coherence of a text;

[0034] Statement clarity analysis: Determine whether the statement is clear and specific, and whether there are any vague, evasive, or self-contradictory expressions;

[0035] The prompt template further includes output format constraints to ensure that the explanatory text from each perspective can be accurately parsed.

[0036] Preferably, in step S4 or T4:

[0037] In the bidirectional interactive attention mechanism, when interpreting the original query text, the entire interpretation vector is used. For querying, MD&A paragraph vector sequences Given keys and values, compute scaled dot product attention and then pool it to obtain the evidence aggregation representation. :

[0038]

[0039] in, This represents the normalization exponential function, used to normalize the scaled dot product result along the MD&A paragraph dimension, obtaining the attention weight of each paragraph relative to the current interpretive perspective. This indicates a pooling operation, specifically mean pooling, indicated by the superscript. Represents the transpose of a matrix;

[0040] When searching for explanations in the original text, use the MD&A paragraph vector sequence. For query, the first Explanation of sub-unit vector sequences from various perspectives For the key and value, compute the scaled dot product attention and obtain the restated representation through masked pooling. :

[0041]

[0042] in This indicates that pooling is performed only on valid paragraph positions in MD&A;

[0043] The learnable network For the first The two fully connected networks corresponding to each perspective will Mapped to interval;

[0044] Usefulness weight Weighted Each perspective is represented as ,Will The weighted representation of each perspective is stacked as follows: After multi-head self-attention interaction, the average value is taken along the viewpoint dimension. :

[0045]

[0046] in For the sake of the bulls' self-attention, Mean pooling is performed along the viewpoint dimension;

[0047] The MD&A basis is represented as the mask mean of the MD&A paragraph vector sequence. :

[0048]

[0049] in For the first The effective mask for each paragraph, for The OK;

[0050] The final text risk is represented as .

[0051] Preferably, in step S5 or T5:

[0052] The cross-network uses financial indicators For input , its first layer for:

[0053]

[0054] in and For the first Learnable parameters of the layer The total number of financial indicators, with the number of cross-layers being [number missing]. Let the output of the cross-network be... ;

[0055] The deep network is a multi-layer fully connected network, with each layer activated by GELU. The output of the deep network is denoted as... ;

[0056] The financial status is obtained by splicing the two outputs and then linearly projecting them. :

[0057]

[0058] in This is a vector concatenation operation. For linear projection layers, This represents a dimension indicating financial status.

[0059] Preferably, in step S6 or T6:

[0060] MD&A paragraph vector sequence As a key and value, the financial status is represented Projected to dimension Then, as a query, cross-modal multi-head attention is performed. :

[0061]

[0062] in To project the financial status representation to Linear layers of dimension This is the effective paragraph mask vector for MD&A. This indicates a multi-head attention mechanism, specifically cross-modal multi-head attention.

[0063] Will Linear projection to After dimension, learnable scaling factor Adding the residuals of the financial status representation to the sum, and then performing layer normalization, yields the financial status-driven text augmentation representation:

[0064]

[0065] in For cross-dimensional projection layer, For layer normalization, the learnable scaling factor The initial value is 0.1.

[0066] Preferably, in step S8, the asymmetric loss is... The first sample Loss per label for:

[0067]

[0068] in The model predicts the first Fraud probability, For real labels, The truncated negative class probability. To truncate the threshold, and They are positive sample focus indices and negative sample focus indices, respectively, and satisfy the following conditions: , For the first The weights of positive samples for each label; total loss The sum of the losses for all samples and all labels:

[0069]

[0070] in This represents the total number of training samples.

[0071] Preferably, the output layer bias of the multi-label classification head is based on the prior occurrence rate of each label in the training set. initialization for:

[0072]

[0073] in, Represents the natural logarithm. For the first The prior incidence of fraudulent behavior in the training set. This represents the total number of training samples.

[0074] The present invention also discloses a multi-label fraud detection system for financial reporting, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the multi-label fraud detection method for financial reporting disclosed in the present invention.

[0075] The beneficial effects of this invention are:

[0076] 1. Achieve fine-grained fraud type diagnosis, breaking through the limitations of binary classification.

[0077] By simultaneously outputting the probabilities of multiple fraud types through a multi-label classification framework, it can not only answer "whether fraud exists" but also "what kind of fraud exists," providing a clear direction for subsequent investigations.

[0078] 2. Introduce domain knowledge from large language models to guide in-depth text analysis.

[0079] By leveraging a large language model and using structured prompts to generate mechanistic explanations from multiple perspectives such as emotional tone, readability, and clarity, domain knowledge is injected into the text analysis process, effectively uncovering deep-seated clues to fraud.

[0080] 3. Establish a fine-grained consistency verification and dynamic weighting mechanism between the interpretation conclusions and the original evidence.

[0081] Unlike existing technologies that simply encode or generate text and then directly merge it, this invention breaks down each interpretation into a sequence of sub-units by sentence, and then... The paragraph-sub-unit matching matrix enables MD&A paragraphs to query each sub-unit of the explanation in reverse, realizing fine-grained consistency verification between the generated explanation and the original text; based on the reverse interaction results, the explanation perspective that is inconsistent with or irrelevant to the current sample is dynamically suppressed through learnable sample-level usefulness weights, reducing the pollution of the classification model by the noise of the generated explanation.

[0082] 4. Enable targeted textual evidence extraction driven by financial status.

[0083] Using financial status as the query condition, semantic evidence most relevant to the current financial status is extracted from the MD&A paragraph. This realizes the conditional constraint of financial status on text feature extraction, so that the extraction of text information during heterogeneous feature fusion is guided by the current financial status, and the interference of irrelevant text information is reduced.

[0084] 5. Effectively address the distribution of long-tail tags.

[0085] By combining the negative sample strong focusing index and the positive sample class weight in the asymmetric loss function, it can still maintain the ability to effectively identify various types of fraudulent behavior even under the condition that the positive samples are extremely sparse. Attached Figure Description

[0086] Figure 1 This is a general framework diagram of the multi-label fraud detection method for financial reporting according to the present invention.

[0087] Figure 2 This is a schematic diagram of the "interpretation-original text bidirectional interactive attention mechanism" in this invention. Detailed Implementation

[0088] The present invention will be further described below with reference to embodiments, but the scope of protection of the present invention is not limited thereto:

[0089] This invention proposes a multi-label fraud detection method for financial reporting, comprising an offline training phase and an online inference phase. See the appendix for the overall process. Figure 1 .

[0090] Phase 1: Offline training phase.

[0091] Step S1: Obtain the MD&A text and structured financial indicators from the company's annual financial report and construct a multi-label training dataset.

[0092] Extract the Management Discussion and Analysis (MD&A) section text from listed companies' annual reports and extract corresponding structured financial indicators from financial databases. Assume there are a total of Class-specific fraudulent behavior, for the first For each enterprise-year sample, the label vector is:

[0093]

[0094] in Indicates sample There exists the first Fraudulent behavior, This indicates that the sample does not exist. Different categories are not mutually exclusive; a single company can be involved in multiple types of fraud simultaneously. Each sample also contains a vector of structured financial indicators. ( (Total financial indicators) and unstructured MD&A text .

[0095] Step S2: Utilize a large language model to generate multi-perspective mechanistic explanations using structured prompt templates.

[0096] Targeting MD&A text The Large Language Model (LLM) is used to generate mechanistic explanations through structured prompt templates. The prompt templates contain definitions and inference instructions for three analytical dimensions:

[0097] Emotional tone analysis: Identify whether management's language contains excessively optimistic, pessimistic, or irrationally emotional expressions;

[0098] Text readability analysis: assesses the syntactic complexity, terminology redundancy, and logical coherence of a text;

[0099] Statement clarity analysis: Determine whether the statement is clear and specific, and whether there are any vague, evasive, or self-contradictory expressions.

[0100] The prompt template further includes output format constraints, forcing the large language model to output the three-part analysis conclusions in a specified format, ensuring that the explanatory text from each perspective can be accurately parsed.

[0101] For the For each sample, the mechanistic explanation is generated as follows:

[0102]

[0103] in This is a structured prompt template. The total number of analytical perspectives (preferred in this invention) ), For the first Explanatory text generated from multiple perspectives.

[0104] Step S3: Construct vector representations of MD&A text and multi-perspective interpretations.

[0105] The MD&A text is divided into paragraph sequences based on natural paragraphs. ,in This represents the number of paragraphs. Each paragraph is encoded into a vector using a pre-trained language encoder (such as Chinese RoBERTa-wwm-ext), resulting in the MD&A paragraph vector sequence:

[0106]

[0107] in The output dimension of the pre-trained encoder (preferred in this invention) ).

[0108] Will The representation of the MD&A paragraph sequence in the common semantic space is obtained after linear projection and layer normalization. :

[0109]

[0110] in For learnable projection parameters, For learnable paragraph position encoding, LN represents layer normalization, and GELU is the activation function. For the public semantic space dimension (preferred in this invention) ).

[0111] At the same time, explanatory texts for each perspective Two representations are constructed respectively:

[0112] (1) Whole paragraph interpretation: Encode the entire interpretation text as a vector. After linear projection and layer normalization, we get:

[0113]

[0114] (2) Explanation of sub-unit sequence representation: The explanation text is split into sentences. Subunit Each of these is encoded into a sequence of original sub-unit vectors:

[0115]

[0116] Will After linear projection and layer normalization, the sub-unit representations in the common semantic space are obtained:

[0117]

[0118] in These are learnable projection parameters.

[0119] The number of sub-units obtained by segmenting sentences If there are fewer than two, further divide them into clauses or a preset maximum length to ensure... This allows for multiple sub-units to be selected when performing reverse lookup interpretation of the original text, thus avoiding attention degradation.

[0120] Simultaneously record the valid paragraph mask of MD&A. ,in Indicates the first One paragraph is valid. This indicates an empty paragraph.

[0121] Step S4: Combining Figure 2 We construct a two-way interactive attention mechanism between the explanation and the original text to generate text risk representations.

[0122] For the first We construct a bidirectional interactive attention mechanism between the explanatory perspective and the MD&A paragraph sequence, thereby enabling mutual verification between the explanatory conclusions and the original text evidence.

[0123] (1) Explain the query original text ( ): Explaining the vector as a whole For query, MD&A paragraph sequence For the key and value, compute the scaled dot product attention to obtain the response weight of each paragraph to the current interpretive perspective. This weighted summation and pooling yields the aggregated evidence representation.

[0124]

[0125] in For pooling operations (mean pooling is preferred in this invention), this representation aggregates the evidence information in the original text that is most relevant to the current interpretation.

[0126] (2) Explanation of the original text query ( ): MD&A paragraph sequence For querying, the first Explanatory sub-unit sequence from a single perspective Using keys and values, calculate the response of each paragraph to each interpretation subunit:

[0127]

[0128] in This indicates that pooling is performed only on valid paragraph positions in MD&A. This expression is achieved through... The paragraph-subunit matching matrix characterizes the overall response pattern of the original paragraph sequence on each subunit of the explanatory perspective, reflecting the "explanatory power" of the explanatory perspective on the current sample.

[0129] (3) Weighting of explanatory perspective usefulness: Not all explanatory perspectives are equally important for every sample. The restatement represents... By inputting a two-layer fully connected network and activating it with a sigmoid function, we obtain the usefulness weights of this perspective for the current sample:

[0130]

[0131] in It is the Sigmoid activation function. For the first Each perspective corresponds to a two-layer fully connected network (structure is as follows) This weight is a sample-level dynamic weight, generated based on the reverse response of the original text to each sub-unit of explanation, which can effectively suppress explanatory perspectives that are irrelevant to the current enterprise text features.

[0132] Weighting is applied to the aggregated evidence representation of the original query text:

[0133]

[0134] (4) Multi-view aggregation: The weighted representation of each perspective is stacked as follows: Add learnable viewpoint position encoding Then, multi-head self-attention is executed to facilitate inter-view interaction:

[0135]

[0136] in This is multi-head self-attention. The mean is taken along the viewpoint dimension:

[0137]

[0138] in Mean pooling is performed along the view dimension.

[0139] The basic representation of MD&A is the mask mean of a sequence of paragraph vectors:

[0140]

[0141] The final text risk is represented by the sum of the two:

[0142]

[0143] This design allows the text risk representation to retain the basic semantics of the original text while incorporating enhanced information from multiple perspectives, after reverse verification of the original text and dynamic weighting.

[0144] Step S5: Construct a dual-path coding network for financial indicators to generate a financial status representation.

[0145] Structured financial indicators This invention employs dual-path parallel encoding:

[0146] (1) Explicit high-order cross network: For input , its first The layers are:

[0147]

[0148] in and For the first Learnable parameters of the layer, number of cross layers: (Preferred version of the present invention) Let the output of the cross-network be... This network explicitly models high-order interactions between financial indicators using bounded order.

[0149] (2) Implicit Nonlinear Coding Network (DNN): This is a multi-layer fully connected network with the following structure: The output is (Preferred version of the present invention) ).

[0150] The financial status is obtained by concatenating the two outputs and then linearly projecting them.

[0151]

[0152] in This is a vector concatenation operation. For linear projection layers, For the dimension representing financial status (preferred in this invention) ).

[0153] Step S6: Financial Status-Driven Directed Text Evidence Extraction (Cross-Modal Directed Attention).

[0154] To implement conditional constraints on the extraction of textual evidence based on financial status, a directed information association path is established: "financial status → targeted extraction of relevant evidence in MD&A," represented by financial status. For querying, MD&A paragraph vector sequence Perform cross-modal multi-head attention for keys and values.

[0155] MD&A paragraph vector sequence As a key and value, the financial status is represented Projected to dimension Then used as a query:

[0156]

[0157] in To project the financial status representation to Linear layers of dimension This is the valid paragraph mask for MD&A.

[0158] Will Linear projection to After dimension, learnable scaling factor (Initial value 0.1) Added to the residual representing the financial status, then normalized:

[0159]

[0160] in For cross-dimensional projection layer, For layer normalization, this design enables the model to selectively extract relevant explanatory evidence from MD&A based on a company's financial status, such as liquidity, profitability, and leverage levels, thereby achieving conditional constraints and directional guidance of financial status on text feature extraction.

[0161] Step S7: Multi-source representation fusion and multi-label classification output.

[0162] Representing textual risks Financial Status Indication Enhanced text representation with financial guidance After splicing, the images are blended and projected in two layers:

[0163]

[0164]

[0165] in Will 1-dimensional mapping to 256-dimensional, Map 256 dimensions to Dimensions (preferred in this invention) ).

[0166] Fusion representation after classification head (structure is) ) mapped to A single logical value:

[0167]

[0168] Each label is activated independently using a Sigmoid function to obtain a probability:

[0169]

[0170] The same enterprise can output high probabilities of multiple fraud types at the same time, achieving fine-grained fraud diagnosis.

[0171] Step S8: Asymmetric loss optimization and model parameter update for long-tailed label distribution.

[0172] To address the significant differences in positive sample rates across different labels in financial fraud data, this invention employs asymmetric loss (ASL) for model optimization. For labels... Let the predicted probability be... The real label is Define the truncated negative class probability:

[0173]

[0174] in For the truncation threshold (preferred in this invention) The loss for a single label is:

[0175]

[0176] in, Represents the natural logarithm; and These are the positive sample focus index and the negative sample focus index, respectively, satisfying... (Preferred version of the present invention) , ); For the first The positive sample weights of each label (preferred in this invention) (Set according to the reciprocal of the positive sample rate for each label). The total loss is the sum of the losses for all samples and all labels:

[0177]

[0178] Based on total loss Perform backpropagation to update the learnable parameters of the model until the model converges, thus obtaining a trained fraud detection model.

[0179] Phase Two: Online Reasoning Phase.

[0180] Step S9: Use the trained model to identify multi-label fraud types in the samples to be analyzed.

[0181] For the sample to be analyzed, extract its MD&A text. and structured financial indicators Load the model parameters obtained from steps S1 to S8, and perform the following operations in sequence:

[0182] (1) Using a large language model through prompt templates Generate multi-perspective mechanistic explanations ;

[0183] (2) Encode the MD&A paragraph as Each perspective interpretation is encoded into a complete representation. and subunit sequence ;

[0184] (3) Generate text risk representation through the bidirectional interactive attention mechanism described in step S4. ;

[0185] (4) Generate a financial status representation through the dual-path coding network described in step S5. ;

[0186] (5) Generate enhanced representations through financial status-driven targeted text attention as described in step S6. ;

[0187] (6) Output the multi-label classification head as described in step S7. The probability of each type of fraud:

[0188]

[0189] Step S10: Output the multi-label fraud detection results.

[0190] The system outputs the probability of each type of fraud. On the validation set, the classification threshold for each label is searched based on maximizing the F1 score of a single label. Thresholding is applied to each type:

[0191]

[0192] The output includes a structured report containing the probability of each type of fraud and thresholding criteria.

[0193] Example 1: Detailed Implementation of the Model Training Method

[0194] 1. Hardware environment: Servers equipped with NVIDIA GPUs (24GB or more of video memory recommended).

[0195] 2. Data preparation and label definition:

[0196] Annual reports of listed companies and corresponding penalty announcements were obtained from financial databases such as CSMAR. Based on publicly available penalty decisions from regulatory agencies, annual reports of companies suspected of financial fraud were selected as positive samples. Fraudulent activities were categorized according to violation type codes. Categories: Inflated profits, inflated assets, false statements.

[0197] Label vector for each sample Labels are assigned according to the actual categories involved, allowing multiple labels to coexist. Stratified sampling is used to match non-fraudulent samples by industry and asset size.

[0198] Extract the MD&A section text and segment it into paragraphs, with a maximum of 200 paragraphs. Extract structured financial indicators (including ratios across dimensions such as solvency, profitability, operational efficiency, growth potential, and risk level, totaling 30 raw indicators), and construct explicit cross features and missing indicator features to ultimately form... Dimensional financial input vector.

[0199] 3. Generation of multi-perspective mechanistic explanations from large language models:

[0200] Use a large language model (such as Deepseek) to generate three mechanistic explanations for the MD&A text of each sample.

[0201] The design includes a prompt template containing the following content:

[0202] Scenario setting: "You are a senior financial analyst. Please analyze the following MD&A text from three dimensions."

[0203] Analysis Dimensions Definition: Define the specific analytical points and reasoning steps for sentiment tone, readability, and clarity.

[0204] Output format constraints: The analysis conclusions are required to be output in a fixed format, with each paragraph containing approximately 150 words.

[0205] The temperature parameter is set to 0.1 during interpretation generation to ensure the determinism and consistency of the output. The generated interpretation is parsed and saved together with the MD&A text.

[0206] 4. Offline semantic encoding:

[0207] Use the Chinese RoBERTa-wwm-ext encoder (dimension) All MD&A paragraphs and three-part explanations are encoded offline.

[0208] For MD&A text: Encode as a sequence of paragraph vectors Obtained by linear projection And generate an MD&A valid paragraph mask. .

[0209] For each interpretation: perform two encodings:

[0210] 1) Whole-segment encoding: Encoding the entire explanatory text into a single vector. Obtained by linear projection ;

[0211] 2) Sub-unit encoding: The explanation is split into sub-unit sequences by sentence, each sub-unit is encoded separately, and the sequences are stacked to form the original sub-unit encoding sequence. Sub-unit representations in the common semantic space are obtained through linear projection. .

[0212] When the number of sub-units obtained from sentence segmentation is less than two, further division is performed according to clauses or preset lengths to ensure... .

[0213] The encoder parameters are frozen and will not participate in subsequent training.

[0214] 5. Model Structure and Training Configuration:

[0215] Public semantic space projection dimension Financial status dimension , fusion representation dimension .

[0216] Explanation of perspective number , 4 heads of attention.

[0217] Cross-network layers The DNN has a hidden layer dimension of 256 and a structure of... .

[0218] Blended Projection: First Layer The second layer .

[0219] Category Header: .

[0220] Asymmetric loss parameters: , truncated threshold Positive sample weights .

[0221] Optimizer: AdamW ), base learning rate The learning rate of the attention interaction module (bidirectional attention, usefulness weight network, and viewpoint self-attention in step S4) Weight decay .

[0222] Batch size 64 (gradient accumulation 2 steps, effective batches 128), maximum training epochs 50, warm-up 3 epochs, cosine annealing to .

[0223] Gradient clipping norm 1.0, model Dropout 0.25.

[0224] The exponential moving average (EMA) decay coefficient is 0.995, the early stop indicator is the validation set Macro-AUPRC, and the patience rounds are 10.

[0225] 6. Threshold search and model saving:

[0226] After training, perform testing on the validation set on each label within the probability interval. Twenty candidate thresholds are selected evenly within the range, and the one with the highest F1 score for a single label is used as the criterion for determination. At the same time, constraints are imposed on the number of predicted positive examples based on the number of true positive examples in the validation set (minimum / maximum prediction / true ratio limit) to prevent the threshold from being too low or too high.

[0227] Save model weights, EMA weights, and threshold values ​​for each label. Training set statistics and configuration metadata.

[0228] Example 2: Detailed Implementation of the Multi-Label Fraud Detection Method for Financial Reports (Online Inference)

[0229] 1. Hardware environment: Inference server equipped with NVIDIA GPU.

[0230] 2. Reasoning process:

[0231] Data Input: Receive the annual report PDF or extracted MD&A text of the company to be analyzed. and corresponding structured financial indicators .

[0232] Mechanistic Explanation Generation: Using the same large language model and prompt templates as in the training phase, generate three-view mechanistic explanations. .

[0233] Text encoding: MD&A is segmented into paragraphs and encoded to obtain... The three explanations are encoded as a whole segment to obtain the results. After being split into sub-units according to the sentence, it is encoded to obtain .

[0234] Forward reasoning: , , , , In the input model, a single forward propagation is performed to calculate the probabilities of three fraud types:

[0235]

[0236] Thresholding determination: Apply threshold values ​​for each label. ,get .

[0237] Output results: Output a JSON format report, including the probability of each type of fraud and hard detection.

[0238] Given that the explanatory text has been generated and encoded, the subsequent classification model can perform fast forward inference.

[0239] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A multi-label fraud detection method for financial reporting, characterized in that, Includes the following steps: Step T1: Obtain the management discussion and analysis (MD&A) text and structured financial indicators of the company to be identified, and input them into the multi-label fraud detection model; Step T2: The multi-label fraud detection model uses a large language model to perform multi-perspective semantic analysis on the MD&A text through structured prompt templates to generate... A mechanistic explanation of the text from a different perspective; Step T3: The MD&A text is segmented into paragraph sequences, and the MD&A paragraph vector sequences are encoded using a pre-trained encoder and projected into a common semantic space to obtain the MD&A paragraph semantic vector sequence of the sample to be identified. The superscript T indicates the text modality. The interpretation text from each perspective is encoded into a whole interpretation vector and a sub-unit vector sequence, wherein the sub-unit vector sequence is obtained by splitting the interpretation text into no less than 2 sub-units by sentence, encoding them separately, stacking them, and then projecting them onto a common semantic space. Step T4: For each explanatory perspective, using its entire explanatory vector as the query and the MD&A paragraph vector sequence as the key and value, calculate the scaled dot product attention and pool it to obtain the evidence aggregation representation; using the MD&A paragraph vector sequence as the query and the sub-unit vector sequence of that perspective as the key and value, calculate the scaled dot product attention and mask pool it to obtain the restatement representation; generate usefulness weights from the restatement representation through a learnable network, and weight the evidence aggregation representation with the usefulness weights to obtain a weighted representation; The weighted representations from each perspective are aggregated after multi-head self-attention interaction and added to the mask mean of the MD&A paragraph vector to obtain the final text risk representation. Step T5: Input the structured financial indicators into a dual-encoding network, one of which is an explicit high-order interactive cross network, and the other is an implicit nonlinear encoding deep network. The outputs of the two networks are concatenated and projected to obtain the financial status representation. Step T6: Using the financial status representation as the query and the MD&A paragraph vector sequence as the key and value, perform cross-modal multi-head attention, and output the financial status-driven text augmentation representation after summing the projected residuals with learnable scaling coefficients; Step T7: The text risk representation, financial status representation, and financial status-driven text augmentation representation are concatenated, fused, projected, and input into a multi-label classification head to obtain... Output the probability of each fraud type.

2. The method according to claim 1, characterized in that, The multi-label fraud detection model is trained through the following steps: Step S1: Obtain the Management Discussion and Analysis (MD&A) text and structured financial indicators from the company's annual financial report, and construct a structure containing... A multi-label training dataset for non-mutually exclusive fraud behaviors, in which For the total number of fraudulent behavior categories, the first The label vector of each sample is , Indicates sample There exists the first Fraudulent behavior, This indicates that it does not exist; Step S2: Utilize a large language model to perform multi-perspective semantic analysis on the MD&A text using structured prompt templates, and generate... A mechanistic explanation of the text from multiple perspectives, in which To analyze the total number of perspectives, let the first one be denoted as the first. The first sample Interpreting the text from one perspective , ; Step S3: Segment the MD&A text into paragraph sequences, and encode them into paragraph vector sequences using a pre-trained encoder. ,in Represents the real number field. For the first The total number of MD&A paragraphs in each sample. Let be the output dimension of the pre-trained encoder; project the paragraph vector sequence onto the common semantic space to obtain the _th ... Sequence of MD&A paragraph semantic vectors from training samples ,in For the public semantic space dimension; the first The textual interpretation from each perspective is encoded into a vector. And projected onto the public semantic space to obtain the entire interpretation vector. ; will the first The explanatory text from each perspective is broken down into sentences. Each subunit is encoded and stacked to obtain the original subunit encoding sequence. ,Will Projecting onto the common semantic space yields a sequence of sub-unit vectors. ,in For the first The number of subunits of the explanation and ; Step S4: For the first Each interpretive perspective is used to construct a bidirectional interactive attention mechanism between itself and the MD&A paragraph vector sequence. Evidence aggregation representations of the original text interpreting the query and restatement representations of the original text query interpretations are calculated separately. These restatement representations are then used to generate usefulness weights for that perspective on the current sample through a learnable network. ,in It is the Sigmoid activation function. For the first Each perspective corresponds to a two-layer fully connected network. This is a restatement of the original query explanation; with The aggregated evidence representation is weighted; The weighted representations from each perspective are aggregated after multi-head self-attention interaction and added to the mask mean of the MD&A paragraph vectors to obtain the final text risk representation. ; Step S5: The structured financial indicators Input a dual-channel coding network, where The total number of financial indicators is represented by two paths: one is an explicit high-order interactive cross-network, and the other is an implicit non-linear coding deep network. The outputs of the two paths are concatenated and projected to obtain the financial status representation. ; Step S6: Represent the financial status For the query, the MD&A paragraph vector sequence is used as the key and value. Cross-modal multi-head attention is performed, and the output is a financial state-driven text augmentation representation obtained by summing the projected residuals after learning scaling factors. ; Step S7: [The sentence is incomplete and requires more context to be translated accurately.] , and After stitching and fusion projection, the data is input into a multi-label classification head to obtain... Each independent logical value is activated by a Sigmoid function to represent the probability of each fraud type. ; Step S8: Calculate the sum of the losses for each label using asymmetric loss, and update the model parameters based on backpropagation of the total loss until the model converges, thus obtaining the trained multi-label fraud detection model.

3. The method according to claim 2, characterized in that, The The prompt template contains definition and reasoning instructions for three analysis dimensions: Emotional tone analysis: Identify whether management's language contains excessively optimistic, pessimistic, or irrationally emotional expressions; Text readability analysis: assesses the syntactic complexity, terminology redundancy, and logical coherence of a text; Statement clarity analysis: Determine whether the statement is clear and specific, and whether there are any vague, evasive, or self-contradictory expressions; The prompt template further includes output format constraints to ensure that the explanatory text from each perspective can be accurately parsed.

4. The method according to claim 2, characterized in that, In step S4 or T4: In bidirectional interactive attention, when interpreting the query text, the entire interpretation vector is used. For querying, MD&A paragraph vector sequences Given keys and values, compute scaled dot product attention and then pool it to obtain the evidence aggregation representation. : in, This represents the normalization exponential function, used to normalize the scaled dot product result along the MD&A paragraph dimension, obtaining the attention weight of each paragraph relative to the current interpretive perspective. This indicates a pooling operation, specifically mean pooling, indicated by the superscript. Represents the transpose of a matrix; When searching for explanations in the original text, use the MD&A paragraph vector sequence. For query, the first Explanation of sub-unit vector sequences from various perspectives For the key and value, compute the scaled dot product attention and obtain the restated representation through masked pooling. : in This indicates that pooling is performed only on valid paragraph positions in MD&A; The learnable network For the first The two fully connected networks corresponding to each perspective will Mapped to interval; Usefulness weight Weighted Each perspective is represented as ,Will The weighted representation of each perspective is stacked as follows: After multi-head self-attention interaction, the average value is taken along the viewpoint dimension. : in For the sake of the bulls' self-attention, Mean pooling is performed along the viewpoint dimension; The MD&A basic representation is the mask mean of the MD&A paragraph vector sequence. : in For the first The effective mask for each paragraph, for The OK; The final text risk is represented as .

5. The method according to claim 2, characterized in that, In step S5 or T5: The cross-network uses financial indicators For input , its first layer for: in and For the first Learnable parameters of the layer The total number of financial indicators, with the number of cross-layers being [number missing]. Let the output of the cross-network be... ; The deep network is a multi-layer fully connected network, with each layer activated by GELU. The output of the deep network is denoted as... ; The financial status is obtained by splicing the two outputs and then linearly projecting them. : in This is a vector concatenation operation. For linear projection layers, This represents a dimension indicating financial status.

6. The method according to claim 2, characterized in that, In step S6 or T6: MD&A paragraph vector sequence As a key and value, the financial status is represented Projected to dimension Then, as a query, cross-modal multi-head attention is performed. : in To project the financial status representation to Linear layers of dimension This is the effective paragraph mask vector for MD&A. This indicates a multi-head attention mechanism, specifically cross-modal multi-head attention. Will Linear projection to After dimension, learnable scaling factor Adding the residuals from the financial state representation to the final normalized representation yields the financial state-driven text augmentation representation: in For cross-dimensional projection layer, For layer normalization, the learnable scaling factor The initial value is 0.

1.

7. The method according to claim 2, characterized in that, The asymmetric loss in step S8 is the first The first sample Loss per label for: in The model predicts the first Fraud probability, For real labels, The truncated negative class probability. To truncate the threshold, and They are positive sample focus indices and negative sample focus indices, respectively, and satisfy the following conditions: , For the first The weights of positive samples for each label; total loss The sum of the losses for all samples and all labels: in This represents the total number of training samples.

8. The method according to claim 2, characterized in that, The output layer bias of the multi-label classification head is based on the prior occurrence rate of each label in the training set. initialization for: in, Represents the natural logarithm. For the first The prior incidence of fraudulent behavior in the training set. This represents the total number of training samples.

9. A multi-label fraud detection system for financial reporting, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1 to 8.

Images