A multi-dimensional sensitive content identification and complaint prediction method and system

Abstract

This application relates to a method, system, computer device, and computer-readable storage medium for multi-dimensional sensitive content identification and complaint prediction. It constructs a multi-dimensional feature set by collecting financial order-related business data, utilizes a financial compliance-specific knowledge base and a pre-trained language model for contextual analysis, and generates multi-dimensional fused feature vectors. Simultaneously, it constructs a multi-layer collaborative model and dynamically adjusts weights based on order status to output tiered labels. Finally, it combines business characteristics with a predictive model to output tiered complaint risks. This application solves the problems of sensitive content identification being detached from business context, insufficient detection capabilities of single models, and severe lag in complaint prediction in traditional technologies. It achieves proactive complaint risk warning based on context awareness and multi-dimensional association, significantly improving the accuracy of sensitive content identification and the collaborative efficiency of compliance review.

Description

Technical Field

[0001] This application relates to the field of content recognition, analysis and prediction, and in particular to a method, system, computer device and computer-readable storage medium for multi-dimensional sensitive content recognition and complaint prediction. Background Technology

[0002] With the online development of finance, securities, and internet businesses, enterprises generate massive amounts of unstructured communication data, including text and voice messages, in customer service, marketing, and complaint handling. In the financial sector, this type of communication data is a core basis for compliance audits and is directly related to user complaint risks; the appearance of sensitive content is often a significant precursor to complaints. Currently, technologies related to financial compliance and sensitive content identification mainly face the following limitations: First, existing context-aware and feature processing methods fail to adequately adapt to the unique characteristics of the financial sector. Traditional keyword matching methods rely solely on comparisons using a fixed sensitive word library, failing to understand context, metaphors, or variant expressions, leading to numerous false positives and false negatives. For example, when a user expresses "something's a bit strange," existing systems cannot identify the potential complaint risk; simultaneously, current technologies struggle to accurately identify expressions in financial scenarios that are "seemingly compliant but actually non-compliant." Furthermore, at the vectorization level, existing solutions primarily extract features from a single modality, failing to integrate text with financial business attributes across multiple dimensions, resulting in vector representations lacking business semantic relevance.

[0003] Second, there is a disconnect between sensitive content identification and complaint prediction, lacking proactive early warning capabilities. Existing complaint prediction methods are mostly based on general user business characteristics (such as activity level and registration duration), failing to establish a direct correlation between sensitive content types, emotional change trends, and complaint risk. This makes it impossible to proactively alert and prevent users from explicitly expressing sensitive content.

[0004] Third, the multi-model collaboration mechanism lacks dynamic adaptation to financial scenarios. Existing multi-model solutions mostly adopt a fixed process integration method, and there is no effective collaborative judgment mechanism between the various models or rule bases. They fail to dynamically adjust the collaboration logic according to the entire life cycle status of financial orders, resulting in an imbalance between recognition efficiency and accuracy at different business stages.

[0005] Currently, no effective solutions have been proposed to address the problems of weak context awareness in sensitive content recognition, single feature dimensions, and lack of scenario adaptation in multi-model collaboration, which lead to lagging complaint prediction and low accuracy. Summary of the Invention

[0006] This application provides a method, system, computer device, and computer-readable storage medium for multi-dimensional sensitive content identification and complaint prediction, in order to at least solve the problems of poor sensitive content identification and inability to predict complaint risks in advance in related technologies.

[0007] In a first aspect, embodiments of this application provide a method for multi-dimensional sensitive content identification and complaint prediction, the method comprising: Collect communication traces and related business data of financial orders, extract text features and build an index to obtain a multi-dimensional feature set containing text content, business attributes and risk tags; Based on the multi-dimensional feature set, the text content context is parsed using a pre-trained language model and a financial compliance knowledge base to obtain text semantic features that represent financial compliance semantics. The text semantic features, business attribute features, and sensitive features are then fused and dimensionality reduced to generate a multi-dimensional fused feature vector. After constructing a multi-layer collaborative model and adjusting the weights of each layer of the multi-layer collaborative model according to the order status of financial orders, sensitive content hierarchical tags are output based on the multi-dimensional fused feature vector. By integrating the sensitive content classification tags with the business features in the multi-dimensional feature set, a classification complaint risk prediction result is output through a complaint prediction model.

[0008] In some embodiments, the associated business data includes order attribute data, personnel attribute data, and user attribute data; The collected communication traces and related business data of financial orders are analyzed, and textual features are extracted and indexed to obtain a multi-dimensional feature set, including: After segmenting the communication trace data, text semantic features, sensitive word features, and sentiment polarity features are extracted. Based on the order ID, personnel ID, and user ID corresponding to the associated business data, a multi-dimensional association index is constructed to obtain a three-dimensional feature set containing text content, business attributes, and risk tags.

[0009] In some embodiments, generating a multi-dimensional fused feature vector includes: Obtain the text semantic features of the pre-trained language model based on the parsed output of the financial compliance knowledge base; The associated business data is transformed into a business attribute vector, and the sensitive word feature encoding is transformed into a sensitive feature vector. The sensitive word features include: sensitive word matching degree, sentiment polarity score, and violation type. The text semantic vector, the business attribute vector, and the sensitive feature vector are concatenated, and dimensionality reduction is performed through principal component analysis to generate a multi-dimensional fused feature vector that combines semantic information and business relevance.

[0010] In some embodiments, the multi-layer collaborative model includes: A rule matching model is used to construct a regular expression rule base based on the financial compliance knowledge base. Through the regular expression rule base, the text content in the multi-dimensional feature set is initially screened and a high-confidence violation label is output. A semantic recognition model is used to input the multi-dimensional fused feature vector into the fine-tuned LightGBM model to identify hidden violations and output a medium-confidence violation label. The behavioral association model combines order status with users' historical sensitive records and uses a logistic regression model to analyze the correlation between sensitive content and user behavior, outputting low-confidence violation labels.

[0011] In some embodiments, the weights of each layer of the multi-layer collaborative model are dynamically adjusted according to the current order status of the financial order, and sensitive content classification tags are output based on the multi-dimensional fused feature vector, including: When the order status is in the creation stage, the rule matching model is assigned the highest weight to focus on efficient initial screening; When the order status is pending review, the semantic recognition model is assigned the highest weight to emphasize accurate recognition. When the order status is in the completed stage, the behavior association model is assigned the highest weight to focus on risk review; The output results of the three-layer model are integrated using a dynamic weighted voting method, and the final sensitive content classification labels are determined based on the preset confidence threshold range.

[0012] In some embodiments, the sensitive content classification tags are integrated with extracted business features, and a classification complaint risk prediction result is output through a complaint prediction model, including: The sensitive content classification tags are transformed into sensitive content features including violation type, confidence level and frequency of occurrence, and business features including order information, user information and personnel information are extracted based on the multi-dimensional feature set; The extracted features are input into a complaint prediction model that integrates LightGBM and logistic regression. LightGBM captures the non-linear relationships between features, and logistic regression improves interpretability. The weighted output values ​​of the LightGBM and logistic regression models are calculated to obtain the original risk prediction score; the original risk prediction score is compared with the dynamically adjusted complaint risk threshold to determine the corresponding risk range. The graded complaint risk prediction results are output based on the risk range, wherein the prediction results include multi-level complaint risk labels and corresponding probabilities.

[0013] In some embodiments, after outputting the tiered complaint risk prediction results through the complaint prediction model, the method further includes: The actual complaint data obtained is compared with the tiered complaint risk prediction results. If an unforeseen actual complaint is detected, the corresponding sensitive content characteristics are traced, and the training data of the financial compliance knowledge base and model are updated to perform model self-learning iteration. The communication trace data, multi-dimensional feature vectors, model recognition and prediction results are associated and stored using the order ID.

[0014] Secondly, embodiments of this application provide a multi-dimensional sensitive content identification and complaint prediction system, the system comprising: a collection module, a parsing module, and a prediction module, wherein: The acquisition module is used to collect communication trace data and related business data of financial orders, extract text features and build an index to obtain a multi-dimensional feature set containing text content, business attributes and risk tags. The parsing module is used to perform text content context parsing based on the multi-dimensional feature set using a pre-trained language model and a financial compliance knowledge base to obtain text semantic features that represent financial compliance semantics, and to fuse and reduce the dimensionality of the text semantic features, business attribute features and sensitive features to generate a multi-dimensional fused feature vector. Furthermore, a multi-layer collaborative model is constructed, and after adjusting the weights of each layer of the multi-layer collaborative model according to the order status of the financial order, sensitive content hierarchical tags are output based on the multi-dimensional fused feature vector. The prediction module is used to integrate the sensitive content classification tags with the business features in the multi-dimensional feature set, and output the classification complaint risk prediction results through the complaint prediction model.

[0015] Thirdly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect above.

[0016] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect above.

[0017] Compared to related technologies, this application provides a multi-dimensional sensitive content identification and complaint prediction method, system, computer device, and computer-readable storage medium. It constructs a multi-dimensional feature set by collecting financial order-related business data, utilizes a financial compliance-specific knowledge base and a pre-trained language model for contextual analysis, and generates a multi-dimensional fused feature vector. Simultaneously, it constructs a multi-layer collaborative model and dynamically adjusts weights based on order status to output tiered labels. Finally, it combines business characteristics with a prediction model to output tiered complaint risks. This application solves the problems of sensitive content identification being detached from business context, insufficient detection capabilities of single models, and severe lag in complaint prediction in traditional technologies. It achieves proactive complaint risk warning based on context awareness and multi-dimensional association, significantly improving the accuracy of sensitive content identification and the collaborative efficiency of compliance review. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a multi-dimensional sensitive content identification and complaint prediction method according to an embodiment of this application; Figure 2 This is a structural block diagram of a multi-dimensional sensitive content identification and complaint prediction system according to an embodiment of this application; Figure 3 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0020] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0021] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0022] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, product, or apparatus that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or apparatus. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0023] This embodiment provides a method for multi-dimensional sensitive content identification and complaint prediction. Figure 1 This is a flowchart of a multi-dimensional sensitive content identification and complaint prediction method according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps: Step S101: Collect communication trace data and related business data of financial orders, extract text features and build an index to obtain a multi-dimensional feature set containing text content, business attributes and risk tags.

[0024] In this embodiment, the execution entity first collects and integrates the communication trace data and associated business data throughout the entire life cycle of financial orders. Among them, the communication trace data includes text chat records, operation logs, and text content obtained through the speech-to-text module and image optical character recognition (OCR) technology; for example, the WeChat Work chat records, the logs of the built-in communication module in the APP, and the speech-to-text data of phone calls in securities brokerage business. The associated business data specifically covers order attribute data (such as order ID, order amount, order type, current order status), personnel attribute data (such as personnel ID, compliance qualification level score), and user attribute data (such as user ID, user risk tolerance level, number of historical complaint records).

[0025] After obtaining the above basic data, perform data preprocessing and cleaning operations on the communication trace data. Specifically, the jieba word segmentation tool can be used to remove invalid punctuation marks and non-domain general stop words, such as removing common words like "of", and excluding financial neutral words such as "buy, sell"; subsequently, a word segmentation processing tool is used to extract multi-dimensional text features, specifically including text semantic features, sensitive word features based on matching with a violation word library, and sentiment polarity features extracted based on text sentiment calculation; for example, the sentiment polarity score from -1 to 1 is extracted through the SnowNLP tool to distinguish explicit risk statements such as "Isn't it non-compliant for you to do this?" and implicit emotional fluctuation statements such as "This thing seems a bit strange". To construct the context basis in the time series dimension, a context window is constructed according to the communication timestamp to generate a cross-turn conversation sequence.

[0026] Finally, based on the extracted features, using the order ID, personnel ID, and user ID corresponding to the associated business data as a joint primary key to construct a multi-dimensional association index, a three-dimensional feature set containing text content, business attributes, and risk labels is obtained, ensuring the structured alignment of cross-modal data.

[0027] Through the above steps of feature extraction and preprocessing operation in S101, a bottom-layer multi-dimensional association mapping is established, providing basic feature support for subsequent context awareness, and overcoming the technical defect that traditional simple keyword matching cannot understand metaphorical or variant expressions.

[0028] Step S102, through a pre-trained language model and a financial compliance knowledge base, based on the multi-dimensional feature set, perform text content context parsing to obtain text semantic features representing financial compliance semantics, and fuse and reduce the dimensions of the text semantic features, business attribute features, and sensitive features to generate a multi-dimensional fusion feature vector.

[0029] To address the unique semantic complexity of financial scenarios, this step introduces a context-aware mechanism tailored to the financial context. First, a financial compliance knowledge base is constructed. Specifically, this knowledge base not only includes standard violation terms but also defines multiple compliance red lines and their variant expressions (including misspellings, homophones, and word decompositions). For example, we sorted out 18 regulatory documents, including the "Interim Provisions on Securities Investment Advisory Business", extracted 12 types of violation scenarios, more than 500 core violation terms, and more than 2,000 variant expressions (such as "guaranteed profit", "guaranteed return", and the variant expression "wei gui" for "violation"), and stored them in XML format to support dynamic updates.

[0030] Furthermore, a language model pre-trained on a financial domain corpus (such as FinBERT; for example, this model was fine-tuned using 50,000 annotated texts on financial violations during training, with a batch size of 32, 10 rounds, and a cross-entropy loss function) is used in conjunction with cross-round dialogue sequences constructed using context windows for contextual analysis. For example, inputting a continuous question sequence without explicit sensitive words, such as "It was clearly agreed... why is it different from before?" During the analysis, the model can apply a contextual attention mechanism to analyze the sentiment and semantic change trends of the user's statements, thereby accurately identifying and outputting textual semantic feature vectors that are "literally compliant but substantially non-compliant."

[0031] After obtaining independent feature vectors, a multi-dimensional feature fusion operation is performed. The associated business data is transformed into a numerical business attribute vector (constructed as a 32-dimensional vector containing normalized order amount values, personnel compliance qualification levels, and user risk levels), and the sensitive word matching degree, sentiment polarity score, and violation type code are combined into a 64-dimensional sensitive feature vector.

[0032] Subsequently, the text semantic vector (the 768-dimensional vector output by the FinBERT [CLS] token), business attribute vector, and sensitive feature vector are directly concatenated along the feature dimension. In this embodiment, considering the impact of high-dimensional sparse features on the computational power of subsequent prediction models, principal component analysis (PCA) is further introduced to perform dimensionality reduction on the concatenated ultra-high-dimensional vector. This process retains the principal feature components with the largest variance in the data, reducing the dimensionality to 128 dimensions and storing it in a Redis vector database. This generates a multi-dimensional fused feature vector that combines deep semantic information with clear business relationships, thereby solving the problem of a single text vector lacking global business semantic constraints.

[0033] Through this step, the context-aware model, combined with the compliance knowledge base, can improve the accuracy of identifying financial-specific violations by more than 30% compared to the general BERT model. The multi-dimensional fusion vector takes into account both semantic information and financial business attributes, solving the problem of single-modal vectors lacking contextual relevance. Experiments have shown that this step can improve vector retrieval efficiency by 40%.

[0034] Step S103: Construct a multi-layer collaborative model, and after adjusting the weights of each layer of the multi-layer collaborative model according to the order status of the financial order, output sensitive content classification tags based on multi-dimensional fusion feature vectors.

[0035] In this step, to improve recognition accuracy and processing efficiency, a three-layer collaborative model architecture consisting of a rule matching model, a semantic recognition model, and a behavior association model is constructed. Specifically: The first layer is a rule-based matching model, which uses regular expressions and direct matching and semantic extension matching rules built from the financial compliance knowledge base to perform a rapid initial screening of communication trace data. For example, it uses Python's re module to build regular expressions. When it identifies "recommend [0-9A-Za-z]{6}", it matches the stock code recommendation violation. When it identifies "yield ≥|guaranteed profit|guaranteed minimum", it matches the return promise violation and outputs a high-confidence violation label.

[0036] The second layer is a semantic recognition model, which inputs the dimensionality-reduced multi-dimensional fusion feature vector generated in step S102 into the LightGBM model after hyperparameter fine-tuning. For example, the model parameters are set to a learning rate of 0.03, a number of trees of 1000, and a maximum depth of 8, and are optimized by 5-fold cross-validation to capture implicit violations and metaphorical expressions, and output medium-confidence violation labels and specific confidence scores.

[0037] The third layer is the behavioral association model, which uses a logistic regression model (e.g., setting the regularization parameter C=1.0) to combine order status with user's historical sensitive records, emotional time series and other behavioral characteristics to analyze the intrinsic correlation between the current sensitive content and the user's long-term behavioral sequence, and outputs low-confidence violation labels.

[0038] After obtaining the independent discrimination results of the three-layer model, a scenario-adaptive dynamic weighted fusion is performed. The current status field of the financial order is read in real time: when the order status is in the creation stage, the rule matching model is assigned the highest weight (e.g., 60%) to focus on efficient initial screening under high concurrency; when the order status is in the pending review stage, the semantic recognition model is assigned the highest weight (e.g., 70%) to focus on accurate identification and risk interception; when the order status is in the completion stage, the behavioral association model is assigned the highest weight (e.g., 50%) to focus on post-event risk review and potential complaint mining. A dynamic weighted voting method is used to calculate the fusion score, and the final sensitive content classification label of the current communication text is determined according to the preset high, medium, and low confidence threshold ranges.

[0039] Specifically, in this embodiment, the set confidence threshold ranges are as follows: when the fusion score is ≥ 0.85, it is determined to be a high-confidence violation; when 0.6 ≤ fusion score < 0.85, it is determined to be a medium-confidence violation; when 0.3 ≤ fusion score < 0.6, it is determined to be a low-confidence violation; if the fusion score < 0.3, it is determined to be compliant content. Through the aforementioned dynamic multi-model collaborative decision-making mechanism, the processing logic is adaptively adjusted based on the order status, effectively overcoming the performance deficiencies of single models in complex tasks and reducing high false positive and high false negative rates. According to the inventors' experiments, this mechanism ensures an accuracy rate of over 92% while reducing the manual cost of compliance review by 35%.

[0040] Step S104: Integrate sensitive content classification tags with business features from multi-dimensional feature sets, and output classification complaint risk prediction results through the complaint prediction model.

[0041] Optionally, this step is based on a model trained using historical complaint data in reverse to make predictive judgments before a customer explicitly files a complaint.

[0042] First, feature engineering is performed to extract features. For example, recursive feature elimination (RFE) is used to select 20 core features. Among them, sensitive content classification tags are transformed into sensitive content features including violation type, confidence level and frequency of occurrence. Business features including order information (order amount, type, status), user information (risk tolerance level, number of historical complaints, frequency of interaction with sensitive content) and personnel information (compliance qualification level, historical violation records) are extracted based on multi-dimensional feature sets. The extracted comprehensive feature set is input into a complaint prediction model that integrates LightGBM and logistic regression architectures. During the model building phase, model parameters are optimized using grid search; for example, the learning rate is set to 0.05, the tree depth to 6, and the regularization coefficient to 0.1. This fusion model leverages the advantages of the LightGBM tree model to capture non-linear relationships between complex features (e.g., capturing the high complaint probability caused by "high-value orders + violation of promised benefits"). Finally, the weighted output values ​​of the LightGBM and logistic regression models are calculated to obtain the original risk prediction score; the original risk prediction score is compared with the dynamically adjusted complaint risk threshold to determine the corresponding risk range; and the graded complaint risk prediction results are output based on the risk range, wherein the prediction results include multi-level complaint risk labels and corresponding probabilities.

[0043] For example, when the calculated risk probability exceeds the high-risk threshold, not only is a graphical pop-up warning triggered, but the high-risk prediction result is also pushed to the compliance audit via a message queue to automatically generate an audit work order. Furthermore, it will output interpretable reasons based on the importance of features within the model, such as: "Due to customer emotional fluctuations + high probability of complaint prediction + suspected profit promises, the risk level is determined to be high", thereby achieving a closed loop of "business communication - risk capture - pre-intervention".

[0044] The graded complaint risk prediction output in this step establishes a deep correlation between sensitive content and business characteristics, changing the traditional post-event handling model and realizing early warning of complaint risks. Experiments have shown that the average warning time is 2-3 business cycles earlier than the traditional method, improving the efficiency of complaint prevention and response by 50%, and achieving proactive prompts and avoidance before users make explicit complaints.

[0045] Step S105: Compare the actual complaint data with the graded complaint risk prediction results. If an actual complaint that was not predicted is detected, perform model self-learning iteration and associate it with the data for storage.

[0046] In this embodiment, to ensure long-term viability and accuracy, the method includes a closed-loop model self-learning and compliance audit mechanism after outputting the prediction results. Real complaint data occurring in actual business processes is periodically acquired and automatically compared with the graded complaint risk prediction results output in step S104.

[0047] If the comparison engine detects an actual complaint that was not successfully predicted by the model (i.e., a missed event, which is marked and reported to the interface by compliance auditors), an adaptive tracing process will be triggered to extract the sensitive content features and user contextual expression features corresponding to the complaint. These newly discovered variant features and contextual patterns will be automatically updated to the financial compliance knowledge base and transformed into positive samples to supplement the model's training dataset. This will drive LightGBM and the contextual model to perform regular self-learning and incremental iterative training. On the 1st of each month, the model will be automatically triggered to retrain using incremental training. After A / B testing to verify compliance, the online model will be replaced.

[0048] Finally, to meet stringent financial compliance requirements, a unique order ID is used to structurally correlate the raw communication data throughout the entire lifecycle, the multi-dimensional feature vectors generated at each stage, the intermediate identification results of the three-layer model, and the final complaint prediction conclusion. Encrypted data is then stored, optionally in a hybrid storage architecture of MySQL and HDFS (MySQL stores the structured results, and HDFS stores the raw data and vectors). This supports fast retrieval by time and order, and the retention period is no less than 3 years.

[0049] Through steps S101 to S105 above, the method of this embodiment constructs a closed-loop self-learning ecosystem of "identification-prediction-feedback," rapidly adapting to changes in user expression and new business compliance regulations; meeting the requirements of financial regulatory audits and resolving issues of high misjudgment, high omission, and compliance lag in the business chain. It not only achieves a deep integration of text context awareness and dynamic multi-model collaboration, but also completely bridges the technical gap between compliance identification and proactive risk warning at the business level, significantly reducing the cost of manual compliance review and improving the agility of risk control response.

[0050] This application also provides a multi-dimensional sensitive content identification and complaint prediction system. Figure 2 This is a structural block diagram of a multi-dimensional sensitive content identification and complaint prediction system according to an embodiment of this application, such as... Figure 2 As shown, the system includes a data acquisition module 20, a data analysis module 21, and a prediction module 22, wherein: The data collection module 20 is used to collect communication trace data and related business data of financial orders, extract text features and build an index to obtain a multi-dimensional feature set containing text content, business attributes and risk tags. The parsing module 21 is used to perform text content context parsing based on a multi-dimensional feature set using a pre-trained language model and a financial compliance knowledge base, to obtain text semantic features that represent financial compliance semantics, and to fuse and reduce the dimensionality of text semantic features, business attribute features and sensitive features to generate a multi-dimensional fused feature vector. Furthermore, a multi-layer collaborative model is constructed, and after adjusting the weights of each layer of the multi-layer collaborative model according to the order status of the financial order, sensitive content hierarchical tags are output based on multi-dimensional fused feature vectors. The prediction module 22 is used to integrate sensitive content classification tags with business features from a multi-dimensional feature set, and output classification complaint risk prediction results through a complaint prediction model.

[0051] The system addresses the issues of sensitive content identification being detached from business context, insufficient detection capabilities of single models, and severe lag in complaint prediction in traditional technologies. It enables proactive complaint risk warning based on contextual awareness and multi-dimensional correlation, significantly improving the accuracy of sensitive content identification and the collaborative efficiency of compliance review.

[0052] In one embodiment, Figure 3 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application, such as... Figure 3 As shown, an electronic device is provided, which can be a server, and its internal structure diagram can be as follows. Figure 3 As shown, the electronic device includes a processor, a network interface, internal memory, and non-volatile memory connected via an internal bus. The non-volatile memory stores an operating system, computer programs, and a database. The processor provides computing and control capabilities, the network interface communicates with external terminals via a network, the internal memory provides an environment for the operating system, the computer programs are executed by the processor to implement a multi-dimensional sensitive content identification and complaint prediction method, and the database stores data.

[0053] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0054] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by hardware related to computer program instructions. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0055] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for multi-dimensional sensitive content identification and complaint prediction, characterized in that, The method includes: Collect communication traces and related business data of financial orders, extract text features and build an index to obtain a multi-dimensional feature set containing text content, business attributes and risk tags; Based on the multi-dimensional feature set, the text content context is parsed using a pre-trained language model and a financial compliance knowledge base to obtain text semantic features that represent financial compliance semantics. The text semantic features, business attribute features, and sensitive features are then fused and dimensionality reduced to generate a multi-dimensional fused feature vector. After constructing a multi-layer collaborative model and adjusting the weights of each layer of the multi-layer collaborative model according to the order status of financial orders, sensitive content hierarchical tags are output based on the multi-dimensional fused feature vector. By integrating the sensitive content classification tags with the business features in the multi-dimensional feature set, a classification complaint risk prediction result is output through a complaint prediction model.

2. The method according to claim 1, characterized in that, The associated business data includes order attribute data, personnel attribute data, and user attribute data; The collected communication traces and related business data of financial orders are analyzed, and textual features are extracted and indexed to obtain a multi-dimensional feature set, including: After segmenting the communication trace data, text semantic features, sensitive word features, and sentiment polarity features are extracted. Based on the order ID, personnel ID, and user ID corresponding to the associated business data, a multi-dimensional association index is constructed to obtain a three-dimensional feature set containing text content, business attributes, and risk tags.

3. The method according to claim 2, characterized in that, Generate multi-dimensional fused feature vectors, including: Obtain the text semantic features of the pre-trained language model based on the parsed output of the financial compliance knowledge base; The associated business data is transformed into a business attribute vector, and the sensitive word feature encoding is transformed into a sensitive feature vector. The sensitive word features include: sensitive word matching degree, sentiment polarity score, and violation type. The text semantic vector, the business attribute vector, and the sensitive feature vector are concatenated, and dimensionality reduction is performed through principal component analysis to generate a multi-dimensional fused feature vector that combines semantic information and business relevance.

4. The method according to claim 1, characterized in that, The multi-layer collaborative model includes: A rule matching model is used to construct a regular expression rule base based on the financial compliance knowledge base. Through the regular expression rule base, the text content in the multi-dimensional feature set is initially screened and a high-confidence violation label is output. A semantic recognition model is used to input the multi-dimensional fused feature vector into the fine-tuned LightGBM model to identify hidden violations and output a medium-confidence violation label. The behavioral association model combines order status with users' historical sensitive records and uses a logistic regression model to analyze the correlation between sensitive content and user behavior, outputting low-confidence violation labels.

5. The method according to claim 4, characterized in that, The weights of each layer of the multi-layer collaborative model are dynamically adjusted based on the current order status of the financial order, and sensitive content classification tags are output based on the multi-dimensional fused feature vector, including: When the order status is in the creation stage, the rule matching model is assigned the highest weight to focus on efficient initial screening; When the order status is pending review, the semantic recognition model is assigned the highest weight to emphasize accurate recognition. When the order status is in the completed stage, the behavior association model is assigned the highest weight to focus on risk review; The output results of the three-layer model are integrated using a dynamic weighted voting method, and the final sensitive content classification labels are determined based on the preset confidence threshold range.

6. The method according to claim 1, characterized in that, By integrating the aforementioned sensitive content classification tags with extracted business features, a complaint prediction model is used to output classified complaint risk prediction results, including: The sensitive content classification tags are transformed into sensitive content features including violation type, confidence level and frequency of occurrence, and business features including order information, user information and personnel information are extracted based on the multi-dimensional feature set; The extracted features are input into a complaint prediction model that integrates LightGBM and logistic regression. LightGBM captures the non-linear relationships between features, and logistic regression improves interpretability. The weighted output values ​​of the LightGBM and logistic regression models are calculated to obtain the original risk prediction score; the original risk prediction score is compared with the dynamically adjusted complaint risk threshold to determine the corresponding risk range. The graded complaint risk prediction results are output based on the risk range, wherein the prediction results include multi-level complaint risk labels and corresponding probabilities.

7. The method according to claim 1, characterized in that, After outputting the tiered complaint risk prediction results through the complaint prediction model, the method further includes: The actual complaint data obtained is compared with the tiered complaint risk prediction results. If an unforeseen actual complaint is detected, the corresponding sensitive content characteristics are traced, and the training data of the financial compliance knowledge base and model are updated to perform model self-learning iteration. The communication trace data, multi-dimensional feature vectors, model recognition and prediction results are associated and stored using the order ID.

8. A multi-dimensional sensitive content identification and complaint prediction system, characterized in that, The system includes: a data acquisition module, a data analysis module, and a prediction module, wherein: The acquisition module is used to collect communication trace data and related business data of financial orders, extract text features and build an index to obtain a multi-dimensional feature set containing text content, business attributes and risk tags. The parsing module is used to perform text content context parsing based on the multi-dimensional feature set using a pre-trained language model and a financial compliance knowledge base to obtain text semantic features that represent financial compliance semantics, and to fuse and reduce the dimensionality of the text semantic features, business attribute features and sensitive features to generate a multi-dimensional fused feature vector. Furthermore, a multi-layer collaborative model is constructed, and after adjusting the weights of each layer of the multi-layer collaborative model according to the order status of the financial order, sensitive content hierarchical tags are output based on the multi-dimensional fused feature vector. The prediction module is used to integrate the sensitive content classification tags with the business features in the multi-dimensional feature set, and output the classification complaint risk prediction results through the complaint prediction model.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images