A document compliance decision method, system and apparatus

Abstract

The application discloses a kind of document compliance decision method, system and equipment, comprising: obtaining heterogeneous policy data in multi-source knowledge base, text segmentation is carried out to the heterogeneous policy data, obtain a plurality of semantic block fragments, and index construction is carried out based on the semantic block fragment, corresponding vector index is generated;Based on the vector index, the received marketing document under review is carried out in the semantic block fragment using ColBERT retrieval model and metadata filtering mechanism fusion retrieval, obtain the highest correlation target compliance basis fragment;The target compliance basis fragment and the marketing document under review are input into the preset expert decision model to carry out compliance analysis, generate the compliance analysis result including risk classification result and modification suggestion.It can improve the efficiency and quality of examination, ensure the compliance of marketing document.

Description

Technical Field

[0001] The present invention relates to the field of text processing technology, and in particular to a document compliance decision-making method, system and device. Background Art

[0002] In the financial industry, compliance reviews of marketing documents face increasingly complex challenges. However, existing compliance review technologies have significant shortcomings in several areas, such as:

[0003] When it comes to data processing, existing technologies struggle to effectively integrate and process heterogeneous data from various channels. Regulatory policy documents are often published in a variety of formats, such as official PDF documents, HTML web announcements, and RSS feeds. Traditional methods for processing this heterogeneous data often require cumbersome format conversions and manual intervention, resulting in inefficient data processing and an inability to provide timely and accurate information support for compliance reviews.

[0004] When it comes to risk identification, existing technologies are unable to accurately identify and judge ambiguous terms in marketing documents. Terms commonly used in financial marketing, such as "high yield," "zero risk," and "expected returns," are vague and ambiguous. Traditional methods based on keyword matching or simple rule engines struggle to accurately understand and determine whether these terms comply with regulatory requirements, leading to potential misjudgments or omissions.

[0005] In terms of adaptability to policy updates, existing technologies struggle to keep pace with and adapt to the rapid pace of regulatory policy changes. Regulatory policies are often adjusted and updated based on market conditions and regulatory requirements, with updates occurring three or more times per month. However, existing compliance review systems lag in updating and integrating new policies, often requiring regular manual intervention and failing to automatically incorporate new policies into review rules in real time. This results in the system being unable to accurately assess the compliance of marketing documents during policy updates, increasing compliance risks for financial institutions.

[0006] In summary, the existing financial marketing document compliance review technology has many shortcomings in processing heterogeneous data, identifying ambiguous expressions and adapting to policy updates, and it is difficult to meet the financial industry's urgent needs for efficient, accurate and real-time compliance review. Summary of the Invention

[0007] In view of this, the purpose of the present invention is to propose a document compliance decision-making method, system and equipment, aiming to solve the problems in the existing technology such as inaccurate compliance review of marketing documents and delayed response due to the inability to efficiently process heterogeneous data and accurately identify fuzzy expressions.

[0008] To achieve the above objectives, the present invention provides a document compliance decision-making method, the method comprising:

[0009] Obtaining heterogeneous policy data from a multi-source knowledge base, performing text segmentation on the heterogeneous policy data to obtain multiple semantic block segments, and constructing an index based on the semantic block segments to generate corresponding vector indexes;

[0010] Based on the vector index, the ColBERT retrieval model and metadata filtering mechanism are used to perform a fusion search on the received marketing document to be reviewed in the semantic block segments to obtain the target compliance evidence segment with the highest relevance;

[0011] The target compliance basis fragment and the marketing document to be reviewed are input into a preset expert decision model for compliance analysis to generate compliance analysis results including risk grading results and modification suggestions.

[0012] Preferably, the method further comprises:

[0013] The ColBERT retrieval model and the expert decision model are optimized and updated based on a preset loss function and the compliance analysis results.

[0014] Preferably, the optimizing and updating of the ColBERT retrieval model and the expert decision model based on a preset loss function in combination with the compliance analysis result includes:

[0015] Optimizing and updating the ColBERT retrieval model using a contrastive learning loss function;

[0016] The expert decision model is optimized and updated using the cross entropy loss function combined with the compliance analysis results.

[0017] Preferably, the text segmentation of the heterogeneous policy data to obtain a plurality of semantic block segments, and index construction based on the semantic block segments to generate corresponding vector indexes include:

[0018] Calculating entity density coefficients of corresponding financial entities in the heterogeneous policy data, and determining text block lengths according to the entity density coefficients;

[0019] Performing text segmentation on the heterogeneous policy data according to the length of the text block to obtain a plurality of semantic block segments;

[0020] The semantic block segment is input into the FinBERT model to generate a semantic vector of a preset dimension, and the semantic vector is inserted into the HNSW index to generate the vector index.

[0021] Preferably, the method of performing a fusion search on the received marketing document to be reviewed in the semantic block segments using the ColBERT retrieval model and metadata filtering mechanism based on the vector index to obtain the target compliance evidence segment with the highest relevance includes:

[0022] Calculating the semantic similarity between the text block in the marketing document to be reviewed and the semantic block fragment using the ColBERT retrieval model to obtain a maximum similarity value;

[0023] Filtering the metadata of the semantic block fragments within a preset time period using a metadata filtering mechanism to obtain a matching value;

[0024] Calculating the maximum similarity value and the matching value to obtain a fusion score;

[0025] The semantic block fragments are screened according to the fusion score to obtain multiple target semantic block fragments. When it is determined that the multiple target semantic block fragments are continuous clauses in the corresponding heterogeneous policy data, the continuous target semantic block fragments are merged to obtain the target compliance basis fragments.

[0026] Preferably, the calculating the maximum similarity value and the matching value to obtain a fusion score includes:

[0027] The tanh function is used to calculate the average monthly update frequency of policy terms to obtain the policy heat factor;

[0028] The maximum similarity value and the matching value are fused and calculated based on the policy heat factor to obtain the fusion score.

[0029] Preferably, the step of inputting the target compliance basis fragment and the marketing document to be reviewed into a preset expert decision model for compliance analysis to generate compliance analysis results including risk grading results and modification suggestions includes:

[0030] Segmenting the marketing document to be reviewed, counting the number of fuzzy words according to a preset risk word library, and calculating the fuzzy word index based on the number of fuzzy words combined with a preset risk weight;

[0031] Calculating the corresponding expert routing probability by combining the fuzzy word index, the vector mean of the marketing document to be reviewed, the vector mean of the target compliance basis fragment, and the trainable parameter matrix corresponding to the expert decision model, wherein the expert decision model includes a policy analysis expert model, a semantic risk expert model, and an empirical reasoning expert model;

[0032] Predicting the marketing document to be reviewed by each expert decision model to obtain a corresponding prediction probability, and calculating based on the prediction probability and the expert routing probability to obtain a risk probability;

[0033] The risk probability is matched with the preset risk level to obtain the risk grading result and the modification suggestion.

[0034] To achieve the above objectives, the present invention further provides a document compliance decision system, comprising:

[0035] An index construction unit is used to obtain heterogeneous policy data from a multi-source knowledge base, perform text segmentation on the heterogeneous policy data to obtain multiple semantic block segments, and perform index construction based on the semantic block segments to generate corresponding vector indexes;

[0036] A fusion retrieval unit, configured to perform a fusion retrieval of the received marketing document to be reviewed in the semantic block segments based on the vector index using a ColBERT retrieval model and a metadata filtering mechanism to obtain a target compliance evidence segment with the highest relevance;

[0037] The compliance analysis unit is used to input the target compliance basis fragment and the marketing document to be reviewed into a preset expert decision model for compliance analysis, and generate compliance analysis results including risk grading results and modification suggestions.

[0038] Preferably, the system further comprises:

[0039] An optimization and updating unit is used to optimize and update the ColBERT retrieval model and the expert decision model based on a preset loss function and the compliance analysis result.

[0040] In order to achieve the above objectives, the present invention also proposes a document compliance decision-making device, including a processor, a memory, and a computer program stored in the memory, wherein the computer program is executed by the processor to implement the steps of a document compliance decision-making method as described in the above embodiment.

[0041] In order to achieve the above objectives, the present invention also proposes a computer-readable storage medium, on which a computer program is stored. The computer program is executed by a processor to implement the steps of a document compliance decision-making method as described in the above embodiment.

[0042] In order to achieve the above objectives, the present invention also proposes a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of a document compliance decision-making method as described in the above embodiment.

[0043] Beneficial effects:

[0044] The above solution, by acquiring heterogeneous policy data from multi-source knowledge bases and performing text segmentation and vector index construction, can efficiently process policy documents of different formats and sources, solving the compatibility and efficiency problems of traditional methods when processing heterogeneous data, and realizing unified management and rapid retrieval of data; using the ColBERT retrieval model and metadata filtering mechanism for fusion retrieval, not only considers the semantic similarity of the text, but also combines metadata information such as the timeliness, importance and scope of application of the policy, improving the accuracy and relevance of the retrieval, and can quickly locate the compliance basis fragment that best matches the marketing document to be reviewed; by inputting the target compliance basis fragment and the marketing document to be reviewed into the expert decision model for compliance analysis, it can automatically generate risk grading results and specific modification suggestions, providing clear guidance for compliance reviewers, reducing the workload and subjectivity of manual review, improving the efficiency and quality of review, and ensuring the compliance of marketing documents.

[0045] By using the contrastive learning loss function to optimize and update the ColBERT retrieval model, the model's ability to judge the semantic similarity between policy texts and marketing documents can be continuously improved, making the retrieval results more accurate; using the cross-entropy loss function combined with compliance analysis results to optimize and update the expert decision-making model, the model's decision parameters can be dynamically adjusted according to the actual review situation, improving the accuracy of risk grading and modification suggestions.

[0046] By calculating the entity density coefficient of the corresponding financial entity in the heterogeneous policy data to determine the length of the text block, the size of the text block can be adaptively adjusted according to the importance and density of the text content, so that each semantic block fragment contains relatively complete and meaningful financial entity information, avoiding excessive segmentation or omission of information, and helping to improve the accuracy of subsequent retrieval and analysis; the semantic block fragment is input into the FinBERT model to generate a semantic vector of preset dimensions, and inserted into the HNSW index to generate a vector index. The FinBERT model can capture the deep semantic features of text in the financial field, and the HNSW index provides efficient approximate nearest neighbor search capabilities, so that compliance basis fragments related to the marketing documents to be reviewed can be found quickly and accurately in large-scale policy data, greatly improving the retrieval efficiency and effectiveness of the system.

[0047] The ColBERT retrieval model calculates the semantic similarity between text blocks and semantic block segments in marketing documents under review, enabling a deeper understanding of the text's semantic meaning and uncovering potential connections that are difficult to detect using traditional keyword matching methods. Furthermore, a metadata filtering mechanism filters policy clauses within a pre-set timeframe within the metadata of semantic block segments, further selecting more relevant clauses based on time and policy type, thereby enhancing retrieval targeting. By fusing the maximum similarity and match values ​​to generate a fusion score, the fusion score is used to select and merge target semantic block segments. This comprehensive consideration of semantic similarity and metadata matching allows for more accurate identification of complete compliance evidence related to marketing documents. Specifically, a policy popularity factor is calculated using the tanh function based on the average monthly update frequency of policy clauses. This factor is then used to fuse the maximum similarity and match values. This allows the retrieval process to dynamically adjust weights based on policy update popularity, prioritizing policy clauses with recent, frequent updates. This ensures the timeliness and accuracy of retrieval results, providing a more reliable foundation for subsequent compliance analysis.

[0048] By segmenting marketing documents for review and counting fuzzy words in a pre-set risk vocabulary, a fuzzy word index is calculated. This quantifies potential fuzzy expressions and compliance risk points in marketing documents, providing an important quantitative basis for subsequent risk assessment. The corresponding expert routing probability is calculated by combining the fuzzy word index, the vector mean of the marketing documents for review, the vector mean of the target compliance evidence fragment, and the trainable parameter matrix of the expert decision model. This process fully considers the semantic and risk characteristics of the text, as well as the parameter characteristics of the expert model, and dynamically assigns weights to different expert models, allowing each expert model to fully leverage its strengths in a specific area and improving the accuracy and comprehensiveness of risk assessment. Predictions are made for marketing documents using the policy parsing expert model, the semantic risk expert model, and the empirical reasoning expert model. Risk probabilities are calculated based on the predicted probabilities and the expert routing probabilities, further refining the risk assessment process and making it more nuanced. Ultimately, by matching the risk probability with the preset risk level, the risk grading results and specific modification suggestions are obtained, providing compliance reviewers with clear and actionable review results and modification guidance, which helps to promptly discover and correct compliance issues in marketing documents, reduce the compliance risks of financial institutions, and improve the efficiency and effectiveness of compliance reviews. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 A flowchart of a document compliance decision-making method provided by one embodiment of the present invention.

[0051] Figure 2 A schematic diagram of the structure of a document compliance decision-making system provided by one embodiment of the present invention.

[0052] The realization of the objectives of the invention, the functional features and advantages will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention for which protection is sought, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0054] The present invention is described in detail below with reference to the embodiments.

[0055] Reference Figure 1 The figure is a flow chart of a document compliance decision-making method provided by an embodiment of the present invention.

[0056] In this embodiment, the method includes:

[0057] S11, obtaining heterogeneous policy data from a multi-source knowledge base, performing text segmentation on the heterogeneous policy data to obtain multiple semantic block segments, and performing index construction based on the semantic block segments to generate corresponding vector indexes.

[0058] Furthermore, in step S11, the heterogeneous policy data is subjected to text segmentation to obtain a plurality of semantic block segments, and index construction is performed based on the semantic block segments to generate corresponding vector indexes, including:

[0059] S11-1, calculating the entity density coefficient of the corresponding financial entity in the heterogeneous policy data, and determining the text block length according to the entity density coefficient;

[0060] S11-2, performing text segmentation on the heterogeneous policy data according to the length of the text block to obtain a plurality of semantic block segments;

[0061] S11-3, input the semantic block segment into the FinBERT model to generate a semantic vector of a preset dimension, insert the semantic vector into the HNSW index to generate the vector index.

[0062] In this example, a wealth management product from a bank is used as an example. A marketing document corresponding to the bank's CMS system is received as a marketing document for review (a sentence in this document includes the text "Invest in wealth management products and enjoy preferential exchange rates with an expected annualized return of 8%!"). This approach integrates heterogeneous policy data from multiple sources, including PDF / Word files from an internal policy repository and HTML / RSS source files from external policy repositories (e.g., "Wealth Management Business Notice" from an external policy repository on a bank's official website RSS feed as a reference for regulatory policy terms, and penalty cases related to "principal protection commitments" from an internal policy repository as a reference for historical violation data). This data is then stream-processed to clean the text. Furthermore, the Apache NiFi engine is used to capture real-time update events (e.g., triggers within 5 minutes of a policy term being published).

[0063] The dynamic block segmentation algorithm is used to segment the heterogeneous policy data into semantic blocks. The size of the semantic block is adaptively adjusted according to the entity density of the corresponding financial entity. That is, L = 512 × (1 + α) is used to segment the heterogeneous policy data. In the formula, L represents the length of the text block (unit: number of words), α represents the entity density coefficient, and α = N entity / word_count,N entity Indicates the number of financial entities in the text (for example, the financial entities identified in the corresponding sentence in the aforementioned financial marketing document include financial products, exchange rates, and an 8% yield). word_count is the total number of words in the text. Financial entities can be automatically extracted using a pre-trained entity recognition model (such as the FinBERT-NER model). The entity density coefficient ensures that text blocks with high entity density are processed first to reduce information fragmentation.

[0064] After segmentation, multiple semantic blocks are obtained. The 768-dimensional semantic vector V of each semantic block is generated by using the domain-adapted FinBERT model. i , and the semantic vector V iInsert data into the HNSW (Hierarchical Navigable SmallWorld) index in real time to generate a vector index. Furthermore, the vector index update complexity is O(logN) (N is the number of index entries). By using WebSocket to monitor the official regulatory website, we ensure that the index is rebuilt within 1 hour after the update or release of new policy terms.

[0065] S12, based on the vector index, using the ColBERT retrieval model and metadata filtering mechanism to perform a fusion search on the received marketing document to be reviewed in the semantic block segments to obtain the target compliance basis segment with the highest relevance.

[0066] Furthermore, in step S12, based on the vector index, the ColBERT retrieval model and the metadata filtering mechanism are used to perform a fusion search on the semantic block segments of the received marketing document to be reviewed, and the target compliance evidence segment with the highest relevance is obtained, including:

[0067] S12-1, using the ColBERT retrieval model to calculate the semantic similarity between the text block in the marketing document to be reviewed and the semantic block segment, to obtain a maximum similarity value;

[0068] S12-2, using a metadata filtering mechanism to filter policy clauses of the metadata of the semantic block segment within a preset time period to obtain a matching value;

[0069] S12-3, calculating the maximum similarity value and the matching value to obtain a fusion score;

[0070] S12-4, screening the semantic block fragments according to the fusion score to obtain multiple target semantic block fragments, and when it is determined that the multiple target semantic block fragments are continuous clauses in the corresponding heterogeneous policy data, merging the continuous target semantic block fragments to obtain the target compliance basis fragments.

[0071] Furthermore, in step S12-3, the maximum similarity value and the matching value are calculated to obtain a fusion score, including:

[0072] S12-3-1, use the tanh function to calculate the average monthly update frequency of policy terms and obtain the policy heat factor;

[0073] S12-3-2, performing a fusion calculation on the maximum similarity value and the matching value based on the policy heat factor to obtain the fusion score.

[0074] In this embodiment, based on the generated vector index, the text block d in the marketing document to be reviewed is iPerform hybrid search, text blocks in pending marketing documents i The corresponding segmentation process is similar to the text segmentation process for heterogeneous policy data, which can be referred to the above process. Specifically: Use the ColBERT retrieval model to calculate d i and semantic block fragment K j The semantic similarity of , output the maximum similarity score Score ColBERT , where Embed(t) and Embed(k) represent the embedding vectors of word t in the marketing document to be reviewed and word k in the semantic block fragment, respectively. The metadata filtering engine filters the metadata of the semantic block fragment (such as policy type, policy issuance date, issuing agency, and scope of application) based on predefined rules (such as timeliness and policy type) and policy clauses within a preset time (such as "only search for commercial bank policies in the past three years") to generate a matching degree M meta (value range [0,1]), which represents the degree of compliance of metadata, where metadata is extracted from the corresponding heterogeneous policy data in the multi-source knowledge base. Through the dynamic weight formula The semantic similarity score and the matching score are integrated to obtain the final fusion score, where β represents the policy heat factor (the calculation formula is ), update_frequency represents the average monthly update frequency of policy terms (unit: times / month). The policy heat factor is based on the dynamic adjustment of the weight of financial policy terms. When policy terms are updated frequently (such as update_frequency=5), β is close to 0.7, giving a higher weight to semantic similarity. Semantic block fragments are screened according to the fusion score, that is, Score is selected. final Semantic block fragments exceeding the threshold (such as Top 5) are used to obtain multiple target semantic block fragments. If the multiple target semantic block fragments are continuous clauses (such as adjacent clauses) in the original policy document, they are merged and spliced ​​(such as "Articles 5-7 of the "XX Notice") to output highly relevant target compliance basis fragments.

[0075] S13, inputting the target compliance basis fragment and the marketing document to be reviewed into a preset expert decision model for compliance analysis, and generating compliance analysis results including risk grading results and modification suggestions.

[0076] Furthermore, in step S13, the target compliance basis fragment and the marketing document to be reviewed are input into a preset expert decision model for compliance analysis to generate compliance analysis results including risk grading results and modification suggestions, including:

[0077] S13-1, segmenting the marketing document to be reviewed, counting the number of fuzzy words according to a preset risk word library, and calculating the fuzzy word index based on the number of fuzzy words combined with a preset risk weight;

[0078] S13-2, calculating a corresponding expert routing probability by combining the fuzzy word index, the vector mean of the marketing document to be reviewed, the vector mean of the target compliance basis segment, and a trainable parameter matrix corresponding to the expert decision model, wherein the expert decision model includes a policy analysis expert model, a semantic risk expert model, and an empirical reasoning expert model;

[0079] S13-3, predicting the marketing document to be reviewed using each of the expert decision models to obtain a corresponding prediction probability, and calculating based on the prediction probability and the expert routing probability to obtain a risk probability;

[0080] S13-4, matching the preset risk level according to the risk probability to obtain the risk grading result and the modification suggestion.

[0081] In this embodiment, the target compliance basis fragment and the pending marketing document are input into a preset expert decision model for risk analysis. The expert decision model includes a policy analysis expert model, a semantic risk expert model, and an empirical reasoning expert model. The policy analysis expert model detects violations of policy terms based on legal BERT and outputs a predicted probability of violation (e.g., detecting "expected rate of return" → probability of violation 0.9). The semantic risk expert model identifies fuzzy expressions based on DeBERTa (fuzzy words such as "discount" and "high risk"; after identification, the output score is 0.85 for the ambiguity score of "discount"). The empirical reasoning expert model generates modification suggestions based on ChatGLM3 (e.g., outputs the confidence level of the suggestion (e.g., the historical adoption rate of the modification suggestion)). The call of the corresponding expert decision model e is controlled by a dynamic routing mechanism, namely:

[0082] Where Pe represents the selection probability of the expert decision model e (expert routing probability), that is, the expert model routing weight (the decision weight assigned to each expert decision model, reflecting which expert decision model is most needed for the current problem), W1 and W2 represent the trainable parameter matrices of the corresponding expert decision models, is the vector mean of the marketing documents to be reviewed, is the vector mean of the target compliance basis fragment, amb_index is the fuzzy word index, that is, , where count vague_termsIndicates the number of ambiguous words (such as guaranteed principal, discount, etc.), word_count is the total number of words, and risk_weight represents the risk weight obtained from the historical violation database (value range [0.1, 1]). This index quantifies the ambiguity of the expression. Furthermore, the calculation process of the vector mean of the marketing document to be reviewed includes: inputting the text block of the marketing document to be reviewed into the domain-adapted FinBERT model, outputting a 768-dimensional embedding vector for each word, and the output is as follows: {Embed(t1),Embed(t2),...,Embed(t n )} (n is the number of words in the block), and then average all word vectors in the block, that is, The calculation process of the vector mean of the target compliance segment is similar. Among them, the fuzzy words of the marketing documents to be reviewed are dynamically matched through the predefined financial risk vocabulary, that is, the marketing documents to be reviewed are segmented, the frequency of the words hitting the risk vocabulary is counted, and the number of fuzzy words is obtained. Then, Calculate the comprehensive risk probability (i.e., the probability of non-compliance). (pe is the predicted probability obtained by the corresponding expert decision model e for the marketing document under review, corresponding to the three expert decision models: the policy analysis expert model, the semantic risk expert model, and the empirical reasoning expert model.) Then, determine the corresponding risk level based on the p-value (e.g., p < 0.3 is low risk; 0.3 ≤ p < 0.7 is medium risk, requiring manual review; p ≥ 0.7 is high risk, requiring mandatory modification). Finally, generate natural language modification suggestions (e.g., changing "expected rate of return" to "historical performance is for reference only"), such as the output:

[0083] [High Risk]: Violation of Article 10 of the "Notice on Certain Wealth Management Services";

[0084] [Revision Suggestions]: 1. Delete "Expected annualized rate of return 8%", 2. "Enjoy preferential exchange rate" → "Must bear the risk of exchange rate fluctuations".

[0085] In another embodiment, the method further comprises:

[0086] S14: Optimize and update the ColBERT retrieval model and the expert decision model based on a preset loss function and the compliance analysis results.

[0087] Furthermore, in step S14, the ColBERT retrieval model and the expert decision model are optimized and updated based on the preset loss function and the compliance analysis result, including:

[0088] S14-1, optimizing and updating the ColBERT retrieval model using a contrastive learning loss function;

[0089] S14-2, optimizing and updating the expert decision model using a cross entropy loss function combined with the compliance analysis results;

[0090] S14-3, when the policy terms in the multi-source knowledge base are monitored to be updated, index reconstruction is triggered to generate an updated vector index.

[0091] In this embodiment, based on the compliance analysis results obtained above and the manual review mark (e.g., it is judged to be medium risk, but it is manually revised to high risk), the update is initiated, including fine-tuning the ColBERT retrieval model at a preset time, such as every 24 hours, using the contrastive learning loss function. , where sim(d i , K + ) represents the similarity of positive samples, sim(d i , K — ) represents the negative sample similarity, d i Represents a block of text in a marketing document to be reviewed, K + represents a positive sample (such as a relevant target compliance basis fragment), K — represents negative samples (such as irrelevant target compliance evidence fragments), Represents the negative sample sum, and this loss optimizes the recall rate; at a preset time, such as every week, the expert decision model is optimized, and the loss function is , where CE(y,p) is the cross entropy loss (y is the true risk label (0 = low risk, 1 = high risk), p is the comprehensive risk probability (the higher the value, the greater the risk)), represents the diversity regularization term (λ=0.1, P ei and P ej is the routing probability of different experts), The feedback loss (γ = 0.05, calculated based on manually labeled samples) prevents decision ossification. Furthermore, by monitoring the frequent updates of policy sources (such as policy clauses in a multi-source knowledge base), the system triggers index reconstruction and generates updated vector indexes when policy clauses in the multi-source knowledge base are updated. This ensures that the system can promptly reflect the latest policy changes, maintains the timeliness and effectiveness of retrieval and decision-making, and reduces compliance risks caused by untimely policy updates. The updated model and vector index are then output for the next round of processing, forming a self-evolving closed loop.

[0092] Reference Figure 2 FIG2 is a schematic diagram showing the structure of a document compliance decision-making system provided by an embodiment of the present invention.

[0093] In this embodiment, the system 20 includes:

[0094] An index construction unit 21 is configured to obtain heterogeneous policy data from a multi-source knowledge base, perform text segmentation on the heterogeneous policy data to obtain a plurality of semantic block segments, and perform index construction based on the semantic block segments to generate corresponding vector indexes;

[0095] A fusion retrieval unit 22 is configured to perform a fusion search on the received marketing document to be reviewed in the semantic block segments based on the vector index using the ColBERT retrieval model and the metadata filtering mechanism to obtain the target compliance evidence segment with the highest relevance;

[0096] The compliance analysis unit 23 is configured to input the target compliance basis fragment and the marketing document to be reviewed into a preset expert decision model for compliance analysis, and generate a compliance analysis result including a risk grading result and modification suggestions.

[0097] In another embodiment, the system further comprises:

[0098] An optimization and updating unit is used to optimize and update the ColBERT retrieval model and the expert decision model based on a preset loss function and the compliance analysis result.

[0099] Each unit module of the system 20 can respectively execute the corresponding steps in the above method embodiment, so each unit module will not be described in detail here. Please refer to the description of the corresponding steps above for details.

[0100] The embodiment of the present invention further provides a document compliance decision device, which includes the document compliance decision system as described above, wherein the document compliance decision system can adopt Figure 2 The structure of the embodiment can be executed accordingly. Figure 1 The technical solution of the method embodiment shown has similar implementation principles and technical effects. For details, please refer to the relevant records in the above embodiments and will not be repeated here.

[0101] The device includes: a mobile phone, digital camera, tablet computer, or other device with a camera function, or a device with an image processing function, or a device with an image display function. The device may include components such as a memory, a processor, an input unit, a display unit, and a power supply.

[0102] Among them, the memory can be used to store software programs and modules, and the processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory may mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as an image playback function, etc.), etc.; the data storage area can store data created according to the use of the device, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage device. Accordingly, the memory may also include a memory controller to provide the processor and the input unit with access to the memory.

[0103] The input unit can be used to receive input digital, character, or image information, and generate keyboard, mouse, joystick, optical, or trackball signal input related to user settings and function control. Specifically, the input unit of this embodiment can include not only a camera, but also a touch-sensitive surface (such as a touch display) and other input devices.

[0104] The display unit can be used to display information input by the user or information provided to the user and various graphical user interfaces of the device, which can be composed of graphics, text, icons, videos and any combination thereof. The display unit may include a display panel. Optionally, the display panel can be configured in the form of an LCD (Liquid Crystal Display), an OLED (Organic Light-Emitting Diode), etc. Furthermore, the touch-sensitive surface can cover the display panel. When the touch-sensitive surface detects a touch operation on or near it, it is transmitted to the processor to determine the type of touch event. The processor then provides a corresponding visual output on the display panel based on the type of touch event.

[0105] The embodiment of the present invention further provides a computer-readable storage medium, which may be a computer-readable storage medium included in the memory in the above embodiment; or a computer-readable storage medium that exists independently and is not assembled into a device. The computer-readable storage medium stores at least one instruction, which is loaded and executed by a processor to implement Figure 1 The document compliance decision method shown. The computer-readable storage medium can be a read-only memory, a magnetic disk or an optical disk.

[0106] The embodiment of the present invention further provides a computer program product, including a computer program / instruction, which is loaded and executed by a processor to implement Figure 1 A document compliance decision-making method is shown.

[0107] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. For similar or identical parts between the various embodiments, reference can be made to each other. For system embodiments, device embodiments, and storage medium embodiments, since they are generally similar to the method embodiments, their descriptions are relatively simple. For relevant parts, reference can be made to the descriptions of the method embodiments.

[0108] Furthermore, in this document, the terms "comprises," "comprising," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not preclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0109] While the foregoing description shows and describes preferred embodiments of the present invention, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. Rather, the present invention can be used in various other combinations, modifications, and environments, and can be modified within the scope of the present invention by the teachings herein or by techniques or knowledge in the relevant art. Modifications and variations made by those skilled in the art without departing from the spirit and scope of the present invention are intended to be within the scope of the appended claims.

Claims

1. A document compliance decision-making method, characterized in that: The method comprises: Obtaining heterogeneous policy data from a multi-source knowledge base, performing text segmentation on the heterogeneous policy data to obtain a plurality of semantic block segments, and constructing an index based on the semantic block segments to generate a corresponding vector index; Based on the vector index, the ColBERT retrieval model and metadata filtering mechanism are used to perform a fusion search on the received marketing documents for review in the semantic block segments to obtain the target compliance evidence segment with the highest relevance, including: Calculating the semantic similarity between the text block in the marketing document to be reviewed and the semantic block fragment using the ColBERT retrieval model to obtain a maximum similarity value; Filtering the metadata of the semantic block fragments within a preset time period using a metadata filtering mechanism to obtain a matching value; Calculating the maximum similarity value and the matching value to obtain a fusion score; Filtering the semantic block segments according to the fusion score to obtain a plurality of target semantic block segments, and when determining that the plurality of target semantic block segments are continuous clauses in the corresponding heterogeneous policy data, merging the continuous target semantic block segments to obtain the target compliance basis segment; Input the target compliance basis fragment and the marketing document to be reviewed into a preset expert decision model for compliance analysis, generating compliance analysis results including risk grading results and modification suggestions; including: Segmenting the marketing document to be reviewed, counting the number of fuzzy words according to a preset risk word library, and calculating the fuzzy word index based on the number of fuzzy words combined with a preset risk weight; Calculating the corresponding expert routing probability by combining the fuzzy word index, the vector mean of the marketing document to be reviewed, the vector mean of the target compliance basis fragment, and the trainable parameter matrix corresponding to the expert decision model, wherein the expert decision model includes a policy analysis expert model, a semantic risk expert model, and an empirical reasoning expert model; Predicting the marketing document to be reviewed by each expert decision model to obtain a corresponding prediction probability, and calculating based on the prediction probability and the expert routing probability to obtain a risk probability; The risk probability is matched with the preset risk level to obtain the risk grading result and the modification suggestion.

2. A document compliance decision-making method according to claim 1, characterized in that: The method further comprises: The ColBERT retrieval model and the expert decision model are optimized and updated based on a preset loss function and the compliance analysis results.

3. A document compliance decision method according to claim 2, characterized in that: The optimizing and updating of the ColBERT retrieval model and the expert decision model based on a preset loss function and the compliance analysis result includes: Optimizing and updating the ColBERT retrieval model using a contrastive learning loss function; The expert decision model is optimized and updated using the cross entropy loss function combined with the compliance analysis results.

4. A document compliance decision method according to claim 1, characterized in that: The text segmentation of the heterogeneous policy data is performed to obtain a plurality of semantic block segments, and index construction is performed based on the semantic block segments to generate corresponding vector indexes, including: Calculating entity density coefficients of corresponding financial entities in the heterogeneous policy data, and determining text block lengths according to the entity density coefficients; Performing text segmentation on the heterogeneous policy data according to the length of the text block to obtain a plurality of semantic block segments; The semantic block segment is input into the FinBERT model to generate a semantic vector of a preset dimension, and the semantic vector is inserted into the HNSW index to generate the vector index.

5. A document compliance decision method according to claim 1, characterized in that: The calculating the maximum similarity value and the matching value to obtain a fusion score includes: The tanh function is used to calculate the average monthly update frequency of policy terms to obtain the policy heat factor; The maximum similarity value and the matching value are fused and calculated based on the policy heat factor to obtain the fusion score.

6. A document compliance decision system, characterized in that: The system comprises: An index construction unit is used to obtain heterogeneous policy data from a multi-source knowledge base, perform text segmentation on the heterogeneous policy data to obtain multiple semantic block segments, and perform index construction based on the semantic block segments to generate corresponding vector indexes; A fusion retrieval unit is configured to perform a fusion retrieval of the received marketing document for review in the semantic block segments based on the vector index using the ColBERT retrieval model and metadata filtering mechanism to obtain the target compliance evidence segment with the highest relevance; the fusion retrieval unit comprises: Calculating the semantic similarity between the text block in the marketing document to be reviewed and the semantic block fragment using the ColBERT retrieval model to obtain a maximum similarity value; Filtering the metadata of the semantic block fragments within a preset time period using a metadata filtering mechanism to obtain a matching value; Calculating the maximum similarity value and the matching value to obtain a fusion score; Filtering the semantic block segments according to the fusion score to obtain a plurality of target semantic block segments, and when determining that the plurality of target semantic block segments are continuous clauses in the corresponding heterogeneous policy data, merging the continuous target semantic block segments to obtain the target compliance basis segment; A compliance analysis unit is configured to input the target compliance basis fragment and the marketing document to be reviewed into a preset expert decision model for compliance analysis, and generate compliance analysis results including risk grading results and modification suggestions; including: Segmenting the marketing document to be reviewed, counting the number of fuzzy words according to a preset risk word library, and calculating the fuzzy word index based on the number of fuzzy words combined with a preset risk weight; Calculating the corresponding expert routing probability by combining the fuzzy word index, the vector mean of the marketing document to be reviewed, the vector mean of the target compliance basis fragment, and the trainable parameter matrix corresponding to the expert decision model, wherein the expert decision model includes a policy analysis expert model, a semantic risk expert model, and an empirical reasoning expert model; Predicting the marketing document to be reviewed by each expert decision model to obtain a corresponding prediction probability, and calculating based on the prediction probability and the expert routing probability to obtain a risk probability; The risk probability is matched with the preset risk level to obtain the risk grading result and the modification suggestion.

7. A document compliance decision system according to claim 6, characterized in that: The system further comprises: An optimization and updating unit is used to optimize and update the ColBERT retrieval model and the expert decision model based on a preset loss function and the compliance analysis result.

8. A document compliance decision-making device, characterized in that: The method comprises a processor, a memory and a computer program stored in the memory, wherein the computer program is executed by the processor to implement the steps of a document compliance decision method as described in any one of claims 1 to 5.

Images