A large model-based securities intelligent customer service man-machine interaction question and answer risk control system

Abstract

The application belongs to the field of artificial intelligence, and particularly relates to a security intelligent customer service man-machine interaction question and answer risk control system based on a large model, which comprises a semantic understanding and compliance embedding module, an implicit investment suggestion semantic detection module, a compliance state tracking management module and a generated content source tracing module, the semantic understanding and compliance embedding module pre-embeds a constraint prompt vector into a large model inference link, guides the model to follow the compliance boundary in the generation stage, the implicit investment suggestion semantic detection module constructs an investment suggestion semantic space vector representation model to perform semantic level identification on the implied investment tendency, the compliance state tracking management module adopts a hierarchical index structure to continuously maintain an investor compliance state archive, can realize embedding of regulatory constraints into the content generation process from the architecture level, realizes a technical change from passive defense to active compliance control, and significantly improves the coverage capability of the risk control system on the regulatory red line.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically a securities intelligent customer service human-computer interaction question-and-answer risk control system based on a large model. Background Technology

[0002] The securities intelligent customer service system is an important digital application of artificial intelligence in the financial services field. Relying on large language models, it enables intelligent interaction and business consultation for investors, which can effectively improve the efficiency of financial services and reduce the cost of manual services. The content generation quality and compliance control capabilities of the large language model directly determine the service quality and operational security of the securities intelligent customer service system.

[0003] Existing intelligent customer service systems for securities generally adopt a serial architecture that separates content generation from compliance verification. They prioritize generating dialogue response content through a large language model, and then rely on an independent rule engine or classifier to complete post-event compliance screening and interception, thereby achieving basic intelligent consultation services and passive risk prevention and control, which can meet the service and risk control needs of conventional standardized consultation scenarios.

[0004] The existing post-verification separation architecture has a fundamental system flaw. It cannot embed regulatory compliance constraints into the model content generation stage, can only perform passive binary screening on the generated results, has difficulty understanding the semantic boundaries of regulatory rules, cannot resist the implicit compliance risks brought about by users’ leading questions, and is prone to problems such as illegal content bypassing the verification mechanism.

[0005] Therefore, this invention provides a securities intelligent customer service human-computer interaction question-and-answer risk control system based on a large model. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0007] The technical solution adopted by the present invention to solve its technical problem is as follows: The securities intelligent customer service human-computer interaction question and answer risk control system based on a large model, as described in the present invention, includes a semantic understanding and compliance embedding module, an implicit investment advice semantic detection module, a compliance status tracking and management module, and a generated content source tracing module.

[0008] In the semantic understanding and compliance embedding module, the system receives securities consultation questions submitted by users, performs semantic parsing on the text to extract user intent, relevant investment targets, and the current dialogue context. Based on the semantic parsing results, the system generates a constraint prompt vector and embeds the constraint prompt vector as a precondition into the inference chain of the large language model. The constraint prompt vector includes prohibited expression types in securities regulatory rules, semantic boundaries of deterministic judgments, and thresholds for determining the tendency of investment advice, guiding the large language model to follow the preset compliance boundaries during the content generation stage rather than performing post-generation filtering.

[0009] The implicit investment advice semantic detection module is communicatively connected to the semantic understanding and compliance embedding module. The system performs multi-level semantic analysis on the candidate response text generated by the large language model. This module constructs an investment advice semantic space vector representation model and maps the candidate responses to the investment advice semantic space for implicit bias determination. Specifically, this includes extracting implicit buy or sell tendencies from technical analysis statements, identifying implicit positive or negative performance implications from information disclosure statements, and judging implicit deterministic prediction components from conditional assumption statements. The detection module determines whether the candidate response contains implicit investment advice based on the relationship between the implicit bias score and a preset threshold.

[0010] The compliance status tracking and management module is communicatively connected to the semantic understanding and compliance embedding module and the implicit investment advice semantic detection module. This module continuously maintains and updates the investor's compliance status profile during multi-round dialogue interactions. The compliance status profile includes the investor's risk level, records of risk warnings that have been disclosed, and the mapping relationship between the product type and risk level involved in the current dialogue. The tracking and management module performs matching calculations based on the product information involved in the current response and the compliance status profile to determine whether there are any omissions in risk warnings or duplicate disclosures, and generates corresponding compliance status update instructions.

[0011] The content generation source tracing module is communicatively connected to the implicit investment advice semantic detection module and the compliance status tracking management module. After each round of dialogue, this module extracts the information sources on which the large language model is based when generating the current response and performs structured annotation. Specifically, this includes the securities regulatory provisions number and specific content cited in the response, the fundamental information or market data sources of the investment targets involved in the response, and the confidence score and supporting basis of the definitive statements in the response. The source tracing module stores the annotation results together with the dialogue log to meet the requirements of full-process traceability and compliance auditing in securities regulation.

[0012] Preferably, in the semantic understanding and compliance embedding module, the constraint prompt vector is trained and optimized using a reinforcement learning alignment framework. By constructing a fine-tuned dataset containing demonstration samples from compliance experts, the large language model actively avoids responses containing deterministic predictions, implicit suggestion tendencies, or absolute statements during the generation stage using human feedback reinforcement learning methods. The reward function of the reinforcement learning alignment framework is designed with a dual constraint structure. The first constraint term negatively penalizes the type of non-compliant expression, while the second constraint term positively incentivizes the professionalism and accuracy of the response. Through multiple rounds of policy optimization, the probability distribution of the model output is shifted towards compliance.

[0013] Preferably, in the implicit investment advice semantic detection module, the investment advice semantic space vector representation model adopts a bidirectional Transformer encoding architecture to perform deep semantic encoding on candidate responses, extract the implicit intent feature vectors in the response text and project them into a preset investment advice judgment subspace. The construction of the judgment subspace is based on compliance guidance documents issued by securities regulatory authorities and historical violation case annotation data. By comparative analysis, the distribution boundaries of implicit advice and compliance statements in the semantic space are determined. The detection module also integrates a temporal context analysis unit, which identifies the implicit investment advice tendency accumulated across rounds by analyzing the semantic correlation strength between the current round and the previous round.

[0014] Preferably, in the compliance status tracking and management module, the compliance status file is stored and managed using a hierarchical index structure. The first-level index is used for targeted retrieval based on investor identity, the second-level index is used for categorization and aggregation based on product type, and the third-level index is used for fine-grained marking based on risk warning items. Before each response is generated, the tracking and management module performs a compliance status pre-check operation, retrieves relevant information from the compliance status file based on the current dialogue context, and generates a compliance gap report. The compliance gap report clearly indicates the list of risk warning items that need to be covered in this response, avoiding omissions or duplicate notifications.

[0015] Preferably, in the content source tracing module, the information source labeling process is implemented using a retrieval-enhanced generation technology framework. When the large language model generates a response, the internal knowledge retrieval module is activated simultaneously to match relevant content entries from a preset securities knowledge base. Based on the matching results, source tags are assigned to key expressions in the response. The securities knowledge base covers multiple independent data sources, such as a securities laws and regulations database, a listed company information disclosure database, and a securities company compliance Q&A knowledge base. The source tags include the data source type, original entry identifier, timestamp, and confidence score, and support source tracing and evidence restoration for each response in compliance audit scenarios.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. The securities intelligent customer service human-computer interaction question-and-answer risk control system based on a large model described in this invention embeds regulatory constraints into the generative inference chain of the large language model in an embedded manner through semantic understanding and compliance embedding modules. This fundamentally changes the traditional architectural paradigm in which compliance verification and content generation are separated. The pre-embedding of constraint prompt vectors enables the large model to actively avoid illegal expressions during the generation stage rather than passively filtering them afterward. This effectively solves the fundamental technical problem between the generation tendency of the large model and the inability to say certain things in securities regulation.

[0017] 2. The securities intelligent customer service human-computer interaction question-and-answer risk control system based on a large model described in this invention constructs an investment advice semantic space vector representation model through an implicit investment advice semantic detection module. This enables semantic-level recognition of implicit buy / sell tendencies, performance expectation implications, and certain prediction components in technical analysis statements, information disclosure statements, and conditional assumption statements. Compared with traditional technologies such as keyword matching or explicit advice classifiers, this solution can accurately identify implicit investment advice disguised as neutral statements, significantly improving the risk control system's ability to cover regulatory red lines. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 This is a structural block diagram of a securities intelligent customer service human-computer interaction question-and-answer risk control system based on a large model, as described in this invention. Figure 2 This is a schematic diagram of the annotation and compliance audit process of the content source tracing module in this invention. Detailed Implementation

[0020] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0021] Example 1: This embodiment is applied to a risk control scenario in a securities company's intelligent customer service human-computer interaction Q&A system. When investors submit securities inquiries through a securities company's mobile application, web page, or telephone channels, the risk control system of this invention performs full-process compliance risk control processing on this interaction.

[0022] like Figure 1 and Figure 2 As shown, the securities intelligent customer service human-computer interaction question-and-answer risk control system based on a large model according to the present invention includes: a semantic understanding and compliance embedding module, an implicit investment advice semantic detection module, a compliance status tracking and management module, and a generated content source tracing module.

[0023] The semantic understanding and compliance embedding module serves as the system's entry point, responsible for receiving securities consultation questions submitted by users. Its input end receives user consultation texts from various channels through a standardized user interface, while its output end communicates with the input ends of the implicit investment advice semantic detection module and the compliance status tracking and management module.

[0024] The input of the implicit investment advice semantic detection module receives the processing results from the semantic understanding and compliance embedding module as well as the candidate response text generated by the large language model. Its output is connected to the input of the content source tracing module.

[0025] The input end of the compliance status tracking and management module receives data from both the semantic understanding and compliance embedding module and the implicit investment advice semantic detection module. Its output end is connected to the input end of the content source tracing module and the control end of the large language model generation module.

[0026] The input end of the content source tracing module receives the processing results from the implicit investment advice semantic detection module and the compliance status tracking management module, and its output end is connected to the dialogue log storage database and the compliance audit archive system.

[0027] The modules interact with each other through internal communication interfaces, and use a standard message queue mechanism to achieve asynchronous data transmission, ensuring the system's response speed and stability.

[0028] In the semantic understanding and compliance embedding module, the system receives securities consultation questions submitted by users, performs semantic parsing on the text to extract user intent, relevant investment targets, and the current dialogue context. The semantic parsing rules are based on a pre-trained language model to extract user intent and investment targets, and dynamically adjust the weight allocation of each constraint term in the constraint prompt vector according to the extraction results.

[0029] When a user's inquiry involves stock investment value assessment, the system increases the weight coefficient of the investment advice tendency constraint to 0.95, and at the same time increases the weight coefficient of the deterministic judgment constraint to 0.90, in order to guide the large language model to avoid giving deterministic judgments and investment advice during the response generation stage.

[0030] When users inquire about fund product subscription and redemption rules, the system increases the weighting factor of the product rule interpretation constraint item to 0.85, and at the same time increases the weighting factor of the risk warning notification constraint item to 0.92.

[0031] Based on the semantic parsing results, the system generates a constraint prompt vector, which is then embedded as a precondition into the inference chain of the large language model. The constraint prompt vector includes prohibited expression types in securities regulatory rules, semantic boundaries of deterministic judgments, and thresholds for determining the bias of investment advice. This guides the large language model to follow the preset compliance boundaries during the content generation stage, rather than performing post-generation filtering.

[0032] The training and optimization of the constraint prompt vector employs a reinforcement learning alignment framework. This involves constructing a fine-tuned dataset containing demonstration samples from compliance experts, and utilizing human feedback reinforcement learning to enable the large language model to proactively avoid responses containing deterministic predictions, implicit suggestion biases, or absolute statements during the generation phase. The reward function of the reinforcement learning alignment framework is designed with a dual-constraint structure: the first constraint negatively penalizes the type of non-compliant expression, while the second constraint positively incentivizes the professionalism and accuracy of the response. Through multiple rounds of policy optimization, the probability distribution of the model's output is shifted towards compliance.

[0033] The implicit investment advice semantic detection module communicates with the semantic understanding and compliance embedding module. The system performs multi-level semantic analysis on the candidate response text generated by the large language model. This module constructs an investment advice semantic space vector representation model, mapping candidate responses to the investment advice semantic space for implicit bias determination.

[0034] The investment advice semantic space vector representation model employs a bidirectional Transformer encoding architecture to perform deep semantic encoding on candidate responses, extracting implicit intent feature vectors from the response text and projecting them into a predefined investment advice judgment subspace. The judgment subspace is constructed based on compliance guidelines issued by securities regulatory authorities and historical violation case annotation data, determining the distribution boundaries of implicit advice and compliant statements in the semantic space through comparative analysis.

[0035] The specific analysis process of this detection module includes three dimensions: The first dimension involves extracting implicit buy or sell tendencies from technical analysis statements. The system first identifies statements in candidate responses that involve technical analysis elements, including candlestick pattern descriptions, technical indicator values, and moving average alignments. Then, it calculates an implicit tendency score based on the semantic similarity between these technical analysis elements and preset buy or sell signal patterns. When the semantic similarity between a technical analysis element and a buy signal pattern exceeds a first preset threshold, the implicit tendency score accumulates positively; when the semantic similarity exceeds a second preset threshold, the implicit tendency score accumulates negatively.

[0036] The second dimension identifies implicit positive or negative implications for company performance from informational statements. The system first identifies informational statements in candidate responses that involve company operations, industry trends, and market conditions. Then, it calculates an implicit bias score based on the semantic relevance between these informational statements and preset patterns of positive or negative performance implications. When the semantic relevance between an informational statement and a pattern of positive performance implications exceeds a third preset threshold, the implicit bias score accumulates positively; when the semantic relevance exceeds a fourth preset threshold, the implicit bias score accumulates negatively.

[0037] The third dimension identifies implicit deterministic prediction components from conditional statement types. The system first identifies conditional statement prompts and their implied conclusions in candidate responses, then calculates the implicit propensity score based on the degree of match between the conclusion and the deterministic prediction pattern. When the degree of match between the conclusion and the deterministic prediction pattern exceeds a fifth preset threshold, the implicit propensity score accumulates positively.

[0038] The detection module performs a weighted summation of the implicit propensity scores across three dimensions to obtain a comprehensive score indicating the implicit investment advice tendency of the candidate responses. This comprehensive score is then compared with a preset judgment threshold. When the comprehensive score exceeds the judgment threshold, a qualitative label indicating implicit investment advice is output; when the comprehensive score does not exceed the judgment threshold, a qualitative label indicating no implicit investment advice is output.

[0039] This detection module also integrates a temporal context analysis unit, which identifies implicit investment recommendation tendencies accumulated across rounds by analyzing the semantic correlation strength between the current round and previous rounds. The temporal context analysis unit first extracts the semantic vector representation of the candidate responses in the current round, then calculates the cosine similarity between this semantic vector representation and the semantic vector representations of responses from previous rounds, and calculates the cumulative tendency score based on the cosine similarity. When the cumulative tendency score exceeds a preset cross-round accumulation threshold, it is determined that the candidate responses in the current round exhibit an implicit investment recommendation tendency accumulated across rounds.

[0040] The compliance status tracking and management module communicates with the semantic understanding and compliance embedding module, as well as the implicit investment advice semantic detection module. This module continuously maintains and updates the investor's compliance status profile during multi-round dialogue interactions.

[0041] The compliance status file is stored and managed using a hierarchical index structure: the first-level index is used for targeted retrieval based on investor identity, the second-level index is used for categorization and aggregation based on product type, and the third-level index is used for fine-grained labeling based on risk warning items.

[0042] Before each response is generated, the tracking and management module performs a compliance status pre-check, retrieving relevant information from the compliance status file based on the current dialogue context and generating a compliance gap report. The compliance gap report clearly identifies the list of risk warnings that need to be covered in this response and the list of risk warnings that have already been communicated. By comparing the two lists, it determines whether there are any omissions or duplicate notifications of risk warnings.

[0043] When the determination result indicates that there is an omission, the tracking and management module generates a compliance status update instruction, adds the risk warnings involved in this response to the completed notification record in the compliance status file, and sends the compliance status update instruction to the big language model generation module, which instructs the big language model to include the omitted risk warnings in the response content when generating the response.

[0044] When the determination result indicates that there is duplicate notification, the tracking and management module generates a compliance status update instruction, which removes the risk warning items involved in this response from the list of risk warning items that should be covered in this response, thereby driving the big language model to avoid repeatedly notifying the same risk warning item when generating a response.

[0045] The tracking and management module matches the product information involved in the current response with the compliance status file to determine whether there are any omissions in risk warnings or duplicate notifications, and generates corresponding compliance status update instructions.

[0046] The content source tracing module communicates with the implicit investment advice semantic detection module and the compliance status tracking and management module. After each round of dialogue, this module extracts the information sources on which the current response is based using the large language model and performs structured annotation.

[0047] The information source labeling process employs a retrieval-enhanced generation technology framework. When the large language model generates a response, the internal knowledge retrieval module is simultaneously activated to match relevant content entries from a pre-defined securities knowledge base. Based on the matching results, source tags are assigned to key expressions in the response. The securities knowledge base encompasses multiple independent data sources, including a securities laws and regulations database, a listed company information disclosure database, and a securities company compliance Q&A knowledge base. Source tags include the data source type, original entry identifier, timestamp, and confidence score, and support source tracing and evidence restoration for each response in compliance audit scenarios.

[0048] The specific annotation content of the content source tracing module includes three aspects: The first aspect involves labeling the securities regulatory provisions cited in the responses with their numbers and specific content. The system identifies the legal references in candidate responses, determines the corresponding provisions' numbers and specific content based on the matching results of the internal knowledge retrieval module, and records them in the source tag data structure.

[0049] The second aspect involves identifying the source of fundamental information or market data related to the investment targets mentioned in the responses. The system identifies statements in candidate responses that involve fundamental data or market data related to the investment targets, determines the corresponding data entry identifier and source database type based on the matching results of the internal knowledge retrieval module, and records them in the source tag data structure.

[0050] The third aspect involves annotating the confidence scores and supporting evidence for definitive statements in the responses. The system identifies content in candidate responses that contains definitive statements, calculates the confidence score for each definitive statement based on the matching results of the internal knowledge retrieval module, and records it along with the supporting evidence in the source tag data structure.

[0051] The content source tracing module stores the annotation results together with the dialogue logs to meet the full-process traceability and compliance audit requirements of securities regulation.

[0052] The system's workflow and the collaborative operation of its modules are as follows: When an investor submits a securities inquiry for the first time through a securities company's online channels, the system first obtains the investor's identity identifier through the user authentication interface, and then retrieves the investor's existing compliance status file from the compliance status file database based on this identifier. If this is the investor's first inquiry, the system creates a new compliance status file for them. This file includes fields for investor identity identifier, risk level, a list of risk warning records that have been disclosed, and a mapping relationship between the product type and risk level involved in the current conversation.

[0053] After receiving the securities consultation question text submitted by the user, the system sends it to the semantic understanding and compliance embedding module for semantic parsing. The module extracts the user's intent: to inquire about the investment value of a certain stock; the investment target: a listed company's stock; and the current dialogue context: first-time consultation. Based on the extracted results, the semantic parsing engine generates a constraint vector. The weight of the deterministic judgment constraint is set to 0.9 to emphasize avoiding deterministic statements; the weight of the investment advice bias constraint is set to 0.95 to emphasize avoiding investment advice bias; and the weight of the absolute statement constraint is set to 0.85 to emphasize avoiding absolute statements.

[0054] The constraint prompt vectors are trained and optimized using a reinforcement learning alignment framework. The first constraint term of the reward function negatively penalizes the type of violation, while the second constraint term positively incentivizes the professionalism and accuracy of the responses. After embedding the constraint prompt vectors into the inference chain of the large language model, the system submits user queries to the large language model, which then generates candidate responses based on the constraint prompt vectors.

[0055] After candidate responses are generated, the system sends them to the implicit investment advice semantic detection module for multi-level semantic analysis. The detection module uses a bidirectional Transformer encoding architecture to perform deep semantic encoding on the candidate responses, extracting the implicit intent feature vectors from the response text and projecting them into a preset investment advice judgment subspace. The detection module extracts implicit buy or sell tendencies from technical analysis statements, calculating the semantic similarity between technical analysis elements and preset buy and sell signal patterns; it identifies implicit positive or negative earnings implications from information disclosure statements, calculating the semantic correlation between information disclosure statements and preset positive and negative earnings implications patterns; and it determines implicit deterministic prediction components from conditional assumption statements, calculating the degree of matching between the assumed conclusions and deterministic prediction patterns.

[0056] The detection module also integrates a temporal context analysis unit, which identifies implicit investment advice tendencies accumulated across rounds by analyzing the semantic correlation strength between the current round and previous rounds. The detection module performs a weighted summation of the implicit tendency scores across three dimensions, obtaining a comprehensive score of 0.35 for the implicit investment advice tendency of the candidate response. After comparing this score with the preset judgment threshold of 0.5, the candidate response is determined to not contain implicit investment advice.

[0057] The system simultaneously sends candidate responses to the compliance status tracking and management module to perform a compliance status pre-check, retrieving relevant information from the compliance status file based on the current dialogue context. The tracking and management module retrieves the investor's existing risk warning notification records from the compliance status file. Based on the product information involved in this response (i.e., a listed company's stock), it determines that the risk warnings to be covered in this response are related to stock investment risk disclosure and listed company information disclosure. The tracking and management module compares the risk warnings to be covered in this response with the list of previously disclosed risk warnings to determine if there are any omissions or duplicates. The tracking and management module generates a compliance gap report, which shows that all risk warnings to be covered in this response are already included in the compliance status file, requiring no additional risk warning content. The tracking and management module sends the compliance gap report to the large language model generation module, instructing the large language model to maintain the current risk warning coverage status in subsequent response generation.

[0058] The system sends candidate responses to the content source tracing module for information source labeling. The tracing module employs a retrieval-enhanced generation technology framework, simultaneously activating the internal knowledge retrieval module when generating responses using the large language model. The internal knowledge retrieval module retrieves information entries related to a listed company's stock from the securities knowledge base, including relevant regulatory provisions, company fundamentals, and market data. Based on the matching results, the tracing module assigns source tags to key statements in the candidate responses, including data source type (securities laws and regulations database or listed company information disclosure database), original entry identifier (a unique identifier for the corresponding data entry), timestamp (a time stamp for data collection or update), and confidence score (a score indicating the reliability of the information source). The tracing module records the labeling results in a source tag data table, which includes fields for source tag identifier, labeled content, data source type, original entry identifier, timestamp, and confidence score.

[0059] The system integrates candidate responses, source tag data tables, and compliance gap reports to generate the final response and sends it to the user. The system stores the interaction record, candidate responses, source tag data tables, and compliance gap reports of this conversation in a dialogue log database to meet the requirements of full-process record keeping and compliance auditing for securities regulation. Simultaneously, the system updates the compliance status profile, adding the risk warnings related to this response to the investor's completed risk warning record list.

[0060] In subsequent dialogue rounds, the system continues to execute the above workflow. The semantic understanding and compliance embedding module performs semantic parsing on newly added user inquiries, extracting user intent, investment targets, and the current dialogue context state, generating new constraint prompt vectors, and embedding them into the inference chain of the large language model. The implicit investment advice semantic detection module performs multi-level semantic analysis on newly added candidate responses and integrates a temporal context analysis unit to identify implicit investment advice tendencies accumulated across rounds. The temporal context analysis unit first extracts the semantic vector representation of the candidate response in the current round, then calculates the cosine similarity between this semantic vector representation and the semantic vector representations of responses in previous rounds, and calculates the cumulative tendency score based on the cosine similarity.

[0061] When a user inquires about the investment value of the same stock in multiple consecutive rounds, the temporal context analysis unit accumulates and analyzes the implicit investment advice tendencies of the responses in each round. When the accumulated tendency score exceeds the preset cross-round accumulation threshold of 0.6, the detection module determines that the candidate responses in the current round have an implicit investment advice tendency formed by cross-round accumulation, outputs a qualitative label containing implicit investment advice, and triggers the compliance control process.

[0062] The compliance status tracking and management module updates the compliance status profile after each round of dialogue, recording the risk warnings and compliance gap reports involved in that round. The content source tracing module extracts the information source and performs structured annotation after each round of dialogue, storing the annotation results together with the dialogue log.

[0063] Through the coordinated operation of the above four modules, the system achieves full-process compliance and risk control for the human-computer interaction Q&A of securities intelligent customer service.

[0064] Example 2: Compared with Embodiment 1, this embodiment focuses on the construction method of the semantic space vector representation model of investment advice and the training and optimization process of the decision subspace.

[0065] The semantic space vector representation model for investment recommendations is implemented using a bidirectional Transformer encoding architecture. This model consists of an input layer, a bidirectional self-attention layer, a feedforward neural network layer, and an output layer. The input layer receives the word sequence of the candidate response text and maps each word to an initial vector representation. The core of the bidirectional context encoding lies in the bidirectional self-attention layer. This layer encodes the input sequence from left to right and from right to left using a forward encoder and a backward encoder, respectively, capturing the semantic information of each word in its complete context. The computation process of the self-attention mechanism is as follows: Multiply the input vector by the query weight matrix, key weight matrix, and value weight matrix respectively to obtain the query vector, key vector, and value vector; Calculate the dot product of the query vector and the key vector to obtain the attention weights; The attention weights are multiplied by the value vector to obtain a weighted contextual representation. The feedforward neural network layer performs a nonlinear transformation on the output of the self-attention layer to enhance the expressive power of the model. The output layer maps the output of the feedforward neural network layer to a vector representation in the investment advice semantic space.

[0066] The process of constructing the subspace is based on compliance guidelines issued by securities regulatory authorities and historical violation case annotation data; Specifically, the system first collects compliance guidelines issued by securities regulatory authorities, including clauses prohibiting the provision of definitive investment advice to investors, clauses prohibiting the prediction of securities price trends, and clauses prohibiting the implication of investment advice tendencies. The system then performs structured processing on the collected compliance guidelines and extracts the key constraints and judgment rules. The system collects annotated data on historical violation cases, including administrative penalty decisions from regulatory authorities and typical violation case analysis reports. The system annotates these historical violation cases, marking statements containing implicit investment advice and the type of violation. The system uses compliance guidelines and annotated historical violation case data as the basis for constructing the judgment subspace. Through comparative analysis, it determines the distribution boundaries of implicit advice and compliance statements in the semantic space. The training and optimization process of the judgment subspace includes: The system inputs the statements labeled as compliant and the statements labeled as non-compliant into the semantic space vector representation model of investment advice, respectively, to obtain the vector distribution of the two types of samples in the semantic space; The system uses a contrastive learning mechanism to adjust the model's parameters so that the vector distribution of compliant statements is as separate as possible from the vector distribution of non-compliant statements. The system determines the optimal location of the separation boundary and uses this separation boundary as the decision boundary for implicit investment recommendations.

[0067] The implementation process of the temporal context analysis unit is as follows: The unit first receives the semantic vector representation of the candidate response in the current round as input, then extracts the semantic vector representation of the responses in previous rounds from the dialogue history, calculates the cosine similarity between the semantic vector of the current round and the semantic vectors of the previous rounds, and calculates the cumulative tendency score based on the cosine similarity. The formula for calculating the cumulative tendency score is to multiply the semantic similarity of each round by the implicit tendency score of the corresponding round, and then sum them to obtain the cumulative tendency score. When the cumulative tendency score exceeds the preset cross-round cumulative threshold, the unit determines that the candidate response in the current round has an implicit investment advice tendency formed by cross-round accumulation.

[0068] Compared with Embodiment 1, this embodiment further enhances the system's ability to identify implicit investment recommendations by elaborating in detail the construction method of the semantic space vector representation model of investment recommendations and the training and optimization process of the decision subspace.

[0069] Example 3: Compared with Embodiment 1 and Embodiment 2, this embodiment focuses on the storage management mechanism of the hierarchical index structure and the retrieval enhancement generation technology framework.

[0070] The compliance status file is stored and managed using a hierarchical index structure. This structure includes a three-level index design. The first-level index is used for targeted retrieval based on the investor's identity. This level of index is implemented using a hash table structure, with the investor's identity as the key and a pointer to the investor's compliance status file as the storage address corresponding to the key. When the system needs to obtain the compliance status file of a specific investor, it first calculates the hash value of the investor's identity identifier using a hash function; Then, based on the hash value, locate the corresponding storage address in the hash table, read the compliance status file pointer stored at that address, and finally read the complete compliance status file data based on that pointer. The second-level index categorizes and aggregates products by type. This level of index is implemented using a tree structure, with product type as the node key and a list of risk warnings related to that product type as the node data. When the system needs to retrieve risk warnings for a specific product type; First, locate the corresponding product type node in the second-level index, and then read the list of risk warning items stored in that node; The third-level index uses fine-grained labeling based on risk warning items. This level of index is implemented using a linked list structure, with the risk warning item as the key value of the linked list node and the notification status and timestamp of the item as the node data. When the system needs to track the notification status of a specific risk warning, it first locates the corresponding event node in the third-level index, and then reads the notification status and timestamp information stored in that node. The three-level indexes are connected through associated pointers to form a complete hierarchical index structure. The tracking and management module achieves efficient retrieval and accurate management of the investor's compliance status through the above hierarchical index structure.

[0071] The execution process of the compliance status pre-inspection operation is as follows: The tracking management module receives the semantic parsing results from the semantic understanding and compliance embedding module, including user intent, investment target and dialogue context status. The tracking management module determines the product type involved in this response based on the investment target. The second-level index retrieves the list of risk warnings corresponding to this product type. The tracking and management module compares the list of risk warnings with the investor's completed notification records to generate a compliance gap report. The compliance gap report includes a list of risk warnings that need to be covered in this response, a list of risk warnings that have been fully disclosed, a list of risk warnings that were omitted from the notification, and a list of risk warnings that were repeatedly disclosed. The tracking and management module sends the compliance gap report to the big language model generation module, which instructs the big language model to include any omitted risk warnings in the generated response and to avoid repeatedly disclosing risk warnings that have already been disclosed.

[0072] The information source labeling process is implemented using a retrieval-enhanced generation technology framework, which includes three sub-modules: a knowledge retrieval module, a knowledge matching module, and a source labeling module. After receiving the trigger signal from the large language model generating a response, the knowledge retrieval module simultaneously activates to retrieve relevant content items from the preset securities knowledge base. Based on key elements such as investment targets, regulatory laws and regulations, and market data involved in the candidate responses, the knowledge retrieval module constructs a retrieval query statement and matches relevant content items from multiple data sources such as the securities laws and regulations database, the listed company information disclosure database, and the securities company compliance Q&A knowledge base. The knowledge matching module scores and ranks the search results based on relevance, selecting the most relevant entries to the candidate responses. The source labeling module assigns source tags to key statements in the candidate responses based on the matching results. The data structure of the source tags includes a data source type field, an original entry identifier field, a timestamp field, and a confidence score field. The data source type field records the data source category to which the information source belongs, including securities laws and regulations databases, listed company information disclosure databases, or securities company compliance Q&A knowledge bases. The original entry identifier field records the unique identifier of the information source in the corresponding data source. The timestamp field records the publication time or latest update time of the information. The confidence score field records the reliability score of the information source, calculated by the knowledge matching module based on search relevance. The source labeling module associates and stores the source tag data with the candidate response content, supporting source tracing and evidence restoration for each response in compliance audit scenarios.

[0073] Compared with Embodiment 1 and Embodiment 2, this embodiment further enhances the system's compliance status tracking and management capabilities and information source tracing capabilities.

[0074] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A large model-based securities intelligent customer service human-computer interaction question and answer risk control system, characterized in that, This includes a processing link, which is: The system obtains the text of a securities consultation question submitted by the user. After receiving the question text, it extracts the user's intent, the investment targets involved, and the current dialogue context state based on preset semantic parsing rules. It generates a constraint prompt vector based on the semantic parsing results and embeds the constraint prompt vector into the inference link of the large language model. The processing link performs multi-level semantic analysis on the candidate response text generated by the large language model, including extracting implicit buy or sell tendencies from technical analysis statements, identifying implicit performance expectation implications from information disclosure statements, and judging implicit deterministic prediction components from conditional assumption statements. The processing link maps the candidate responses to a preset investment advice semantic space, calculates an implicit bias score based on the mapping result, and outputs a qualitative label indicating whether the candidate response contains implicit investment advice based on the relationship between the implicit bias score and a preset judgment threshold. The processing link continuously maintains and updates the investor's compliance status file during multiple rounds of dialogue and interaction. It performs matching calculations with the compliance status file based on the product information involved in the current response to generate corresponding compliance status update instructions. After each round of dialogue, the processing link extracts the information sources on which the large language model generates the current response, including the securities regulatory clause number cited in the response, the source of fundamental information or market data of the investment targets involved in the response, and the confidence score and supporting basis of the deterministic statements in the response. The annotation results are stored together with the dialogue log.

2. The securities intelligent customer service human-computer interaction question-and-answer risk control system based on a large model according to claim 1, characterized in that, The training optimization of the constraint prompt vector adopts a reinforcement learning alignment framework, which includes constructing a fine-tuned dataset containing demonstration samples from compliance experts and using human feedback reinforcement learning methods to enable the large language model to actively avoid response content containing deterministic predictions, implicit suggestion tendencies, or absolute statements during the generation stage. The reward function of the reinforcement learning alignment framework is designed as a dual-constraint structure. The first constraint term negatively penalizes the type of violation, while the second constraint term positively incentivizes the professionalism and accuracy of the response.

3. The securities intelligent customer service human-computer interaction question-and-answer risk control system based on a large model according to claim 1, characterized in that, The investment advice semantic space vector representation model uses a bidirectional Transformer encoding architecture to perform deep semantic encoding on candidate responses, extract the implicit intent feature vectors in the response text and project them into a preset investment advice judgment subspace; The construction of the judgment subspace is based on compliance guidelines issued by securities regulatory authorities and historical violation case annotation data. The distribution boundaries of implicit advice and compliance statements in the semantic space are determined through comparative analysis.

4. The securities intelligent customer service human-computer interaction question-and-answer risk control system based on a large model according to claim 1, characterized in that, The processing link integrates a time-series context analysis unit, which identifies implicit investment recommendation tendencies accumulated across rounds by analyzing the semantic correlation strength between the current round and the previous round.

5. The securities intelligent customer service human-computer interaction question-and-answer risk control system based on a large model according to claim 1, characterized in that, The compliance status file is stored and managed using a hierarchical index structure; The first-level index performs targeted searches based on investor identity, the second-level index categorizes and aggregates by product type, and the third-level index provides fine-grained labeling based on risk warnings.

6. The securities intelligent customer service human-computer interaction question-and-answer risk control system based on a large model according to claim 1, characterized in that, Before each response is generated, the processing link performs a compliance status pre-check operation, retrieves relevant information from the compliance status file based on the current dialogue context, and generates a compliance gap report. The compliance gap report clearly indicates the list of risk warnings that need to be covered in this response.

7. The securities intelligent customer service human-computer interaction question-and-answer risk control system based on a large model according to claim 1, characterized in that, The information source labeling process is implemented using a retrieval-enhanced generation technology framework. When the large language model generates a response, the internal knowledge retrieval module is activated simultaneously to match relevant content entries from the preset securities knowledge base, and source tags are assigned to key expressions in the response based on the matching results.

8. The securities intelligent customer service human-computer interaction question-and-answer risk control system based on a large model according to claim 1, characterized in that, The semantic parsing rules extract user intent and investment targets based on a pre-trained language model, and dynamically adjust the weight allocation of each constraint term in the constraint prompt vector according to the extraction results.

Images