Product recommendation risk identification method and device, equipment and medium

Abstract

The application relates to the technical field of artificial intelligence, is suitable for the risk control scene of financial technology and medical health, and discloses a product recommendation risk identification method, device, equipment and medium, which comprises the following steps: obtaining a social chat record of a target group chat; performing sentence processing on the social chat record to obtain chat sentences; performing text similarity calculation on the chat sentences and fraud texts in a preset fraud text library to obtain fraud similarity; performing interaction sentiment analysis on the chat sentences to obtain a stress score of a product recommender to a user and an anxiety emotion score of the user to the product recommender; performing risk determination on the fraud similarity, the stress score and the anxiety emotion score to obtain a risk determination result; and in response to the social chat record being a high-risk conversation, pushing a compliance prompt information in the target group chat through a customer service robot. The application can accurately identify risks in social conversation, so that the compliance of the social conversation is improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and is applicable to the fields of fintech and healthcare, and in particular to a method, apparatus, device and medium for identifying product recommendation risks. Background Technology

[0002] Community-based product recommendation refers to recommending products through online community dialogues. For example, in the financial insurance sector of fintech, financial products (such as credit cards and consumer loans) are often recommended through external financial insurance agents or third-party customer service representatives in online communities, using dialogue to introduce these products to users. Similarly, in the healthcare sector, health insurance products (such as health insurance) can be recommended through health insurance agents in online communities.

[0003] However, product recommendation methods via social media chat are prone to misleading users through agents during conversations, resulting in poor compliance. For example, in the financial insurance scenario within fintech, agents responsible for selling insurance may employ non-compliant tactics and fraudulent inducements (such as promising high returns or implying illegal operations) during conversations. Current risk control technologies, typically relying on keyword filtering to identify non-compliant conversations, are ineffective at identifying such content, leading to poor accuracy in risk control. Summary of the Invention

[0004] This application provides a product recommendation risk identification method, apparatus, equipment, and medium to address the technical problem of how to optimize the compliance of social media conversations. The aim is to accurately identify risks in social media conversations, improve the compliance of social media conversations, and thus enhance the accuracy of risk control.

[0005] Firstly, a method for identifying product recommendation risks is provided, including: Obtain the chat history of the target group chat; wherein, the target group chat contains a customer service robot, a product recommender, and users; The community chat history is segmented into sentences to obtain chat statements; The text similarity is calculated based on the chat statements and fraudulent texts in the preset fraudulent text library to obtain the fraud similarity. Based on the chat statements, interactive sentiment analysis was performed to obtain the stress score of the product recommender to the user and the anxiety score of the user to the product recommender. Risk assessment is performed based on the fraud similarity, stress score, and anxiety score to obtain a risk assessment result; wherein, the risk assessment result includes whether the community chat record is a high-risk conversation or a low-risk conversation; If the chat log is found to be a high-risk conversation, a compliance alert will be sent to the target chat group via the customer service robot.

[0006] Secondly, a product recommendation risk identification device is provided, including: The record acquisition module is used to acquire the social chat history of the target group chat; wherein, the target group chat contains a customer service robot, a product recommender, and a user; The sentence segmentation module is used to segment the community chat records into sentences to obtain chat statements; The similarity calculation module is used to calculate the text similarity between the chat statement and the fraudulent text in the preset fraudulent text library to obtain the fraud similarity. The sentiment analysis module is used to perform interactive sentiment analysis based on the chat statements to obtain the stress score of the product recommender to the user and the anxiety score of the user to the product recommender. The risk assessment module is used to assess risk based on the fraud similarity, the stress score, and the anxiety score, and obtain a risk assessment result; wherein, the risk assessment result includes whether the community chat record is a high-risk conversation or the community chat record is a low-risk conversation; The information push module is used to push compliance reminder information to the target group chat in response to the fact that the social chat history is a high-risk conversation.

[0007] Thirdly, a computer device is provided, 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 steps of the above-described method.

[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described method.

[0009] The aforementioned product recommendation risk identification method, device, equipment, and medium can acquire community chat records from target group chats involving customer service robots, product recommenders, and users. These records are then segmented into sentences to obtain chat statements. Text similarity is calculated between these chat statements and pre-set fraudulent texts in a fraudulent text library to obtain fraud similarity scores. This allows for precise matching of the product recommender's statements to identify content matching known fraudulent inducement tactics and non-compliant words, thus identifying illegal and fraudulent texts. Furthermore, interactive sentiment analysis is performed on the chat statements to obtain the product recommender's pressure score towards the user and the user's anxiety score towards the product recommender. This allows for the capture of hidden violation characteristics from the emotional dimension of dialogue interaction, identifying the risk of inducing marketing through emotional pressure. Risk assessment results are obtained by analyzing fraud similarity, stress score, and anxiety score. In cases where there are high-risk conversations in the community chat history, compliance reminders are pushed to the target group chat via customer service robots. This proactively triggers compliance intervention mechanisms, rather than relying solely on post-event user complaints to discover non-compliant marketing tactics. This allows for accurate identification of risks in community conversations, improving the compliance of community conversations and thus enhancing the accuracy of risk control. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of an application environment for a product recommendation risk identification method in one embodiment of this application; Figure 2 This is a flowchart illustrating a product recommendation risk identification method in one embodiment of this application; Figure 3 yes Figure 2 A schematic diagram of a specific implementation of step S104; Figure 4 yes Figure 3 A flowchart illustrating a specific implementation of step S202; Figure 5 yes Figure 3 A schematic diagram of a specific implementation method for step S203; Figure 6 This is another flowchart illustrating the product recommendation risk identification method in one embodiment of this application; Figure 7 yes Figure 6 A schematic diagram of a specific implementation of step S502; Figure 8 yes Figure 2 A schematic diagram of a specific implementation of step S105; Figure 9 This is another flowchart illustrating the product recommendation risk identification method in one embodiment of this application; Figure 10 This is a schematic diagram of the structure of a product recommendation risk identification device in one embodiment of this application; Figure 11 This is a schematic diagram of the structure of a computer device according to one embodiment of this application; Figure 12 This is another structural schematic diagram of a computer device in one embodiment of this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] The product recommendation risk identification method provided in this application embodiment can be applied to, for example, Figure 1 In this application environment, the client communicates with the server via a network. The server can receive the chat logs of the customer service robot, product recommender, and user in the target group chat through the client, and use a product recommendation risk identification method to identify the risk of the chat logs, determining whether they are low-risk or high-risk conversations. The server can then feed the risk assessment result back to the client. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The following detailed description of specific embodiments further illustrates this application.

[0014] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the product recommendation risk identification method provided in this application embodiment includes the following steps: S101: Obtain the community chat history of the target group chat; the target group chat contains customer service robots, product recommenders, and users; S102: Segment the community chat history to obtain the chat statements; S103: Calculate the text similarity between the chat statements and the fraudulent texts in the preset fraudulent text library to obtain the fraud similarity; S104: Perform interactive sentiment analysis based on chat statements to obtain the stress score of the product recommender to the user and the anxiety score of the user to the product recommender. S105: Risk assessment is conducted based on fraud similarity, stress score, and anxiety score to obtain risk assessment results; among which, the risk assessment results include whether the social chat record is a high-risk conversation or a low-risk conversation. S106: If a social chat log is found to be a high-risk conversation, a compliance reminder will be sent to the target group chat via a customer service robot.

[0015] The beneficial effects of this application's embodiments include, but are not limited to: obtaining community chat records from target group chats containing customer service robots, product recommenders, and users; segmenting the community chat records into sentences to obtain chat statements; and calculating text similarity between the chat statements and fraudulent texts in a pre-set fraudulent text library to obtain fraud similarity. This allows for precise matching of whether the product recommender's statements contain content matching known fraudulent inducement tactics and non-compliant words, thus identifying illegal fraudulent texts. Furthermore, interactive sentiment analysis is performed on the chat statements to obtain the product recommender's pressure score towards the user and the user's anxiety score towards the product recommender. This allows for capturing hidden violation characteristics from the emotional dimension of dialogue interaction, identifying the risk of inducing marketing through emotional pressure. Risk assessment results are obtained based on fraud similarity, pressure score, and anxiety score. In cases where the community chat records indicate high-risk dialogue, compliance reminders are pushed to the target group chat via customer service robots. This proactively triggers compliance intervention mechanisms, rather than relying solely on post-event user complaints to discover non-compliant marketing methods, thereby accurately identifying risks in community dialogues, improving the compliance of community dialogues, and enhancing the accuracy of risk control.

[0016] In step S101 of some embodiments, the target group chat is a group chat whose chat history needs to be risk-analyzed. The target group chat contains a customer service robot, a product recommender, and users. It should be noted that the product recommender is the speaker promoting a product to the user, while the user is the speaker expressing a potential purchase intention. The customer service robot is an intelligent agent used to monitor the compliance of the dialogue between the product recommender and the user.

[0017] In some embodiments, such as customer service robots, the agent can be an intelligent agent based on an artificial intelligence model (e.g., a large language model). It should be noted that the large language model includes embedding layers and Transformer layers, and is trained in a supervised manner on large-scale natural language inference datasets and fraud similarity datasets. The large language model learns sentence-level semantic representations through contrastive learning. Its training objective is to maximize the similarity of positive sample pairs (sentences that are semantically similar or inferentially related) and minimize the similarity of negative sample pairs (sentences that are semantically unrelated), that is, to ensure that semantically similar sentences are close in distance in the vector space, while semantically dissimilar sentences are far apart.

[0018] In some embodiments, such as in the financial product promotion scenario in the fintech field, the product recommender in the target group chat can be an external financial insurance agent or a third-party customer service representative, who can recommend financial products (such as credit cards and consumer loan products) to users through the conversation in the target group chat.

[0019] In another embodiment, for example, in a medical insurance sales scenario in the healthcare field, the product recommender in the target group chat can be a medical insurance agent, who can recommend medical insurance products (such as health insurance) to users through the conversation in the target group chat.

[0020] It should be noted that community chat logs refer to the recorded speech data in the target group chat. Specifically, community chat logs may include speech data from product recommenders and user speech data. Community chat logs may also include speech data from customer service chatbots. In some embodiments, community chat logs may include text data or voice data; this embodiment does not limit this. For voice-based community chat logs, they need to be uniformly converted into text data for subsequent processing. For example, speech conversion can be performed using a Large Language Model (LLM).

[0021] In some embodiments, the product recommendation risk identification method provided in this application can be applied to intelligent question-and-answer engines such as intelligent customer service or intelligent assistants in various application scenarios. Intelligent question-and-answer engines are typically implemented through a server-side component, which can receive user questions in real time. For example, in the insurance application field, users often communicate with intelligent customer service regarding insurance products through group chats, frequently requiring risk identification of the communication process between intelligent customer service and users. For instance, the product recommendation risk identification method can be used to identify risks in the communication records between intelligent customer service and users in group chats to prevent intelligent customer service from using non-compliant insurance recommendation scripts, thereby ensuring a better user experience. Specifically, the server-side component can be an insurance company's server.

[0022] In step S102 of some embodiments, the chat statement is a statement sent by a single speaker, obtained after segmenting the community chat record into sentences. It should be noted that the community chat record may contain statements sent by different speakers (such as product recommenders and users), therefore it is necessary to split the community chat record corresponding to multiple speakers into chat statements corresponding to a single speaker. In some embodiments, the community chat record can be segmented using a large language model. Other methods can also be used for segmenting the community chat record, and are not limited to these.

[0023] In step S103 of some embodiments, the fraudulent text library is a database used to store fraudulent text. Fraudulent text refers to text that does not comply with product recommendation guidelines. Specifically, fraudulent text in the fraudulent text library may include implicit terms that suggest illegal operations. For example, fraudulent text may include text promising high returns, evading risk warnings, or implying illegal operations (such as assisting with cash-out). In some embodiments, for example, product recommenders (such as agents of insurance products) may use fraudulent dialogue (such as exaggerating returns, applying pressure to induce a sale), risk-avoidance expressions (such as "guaranteed profit," "no threshold for loan approval"), and other fraudulent text to promote products.

[0024] It should be noted that fraud similarity refers to the semantic similarity between chat statements and fraudulent text. In some embodiments, text features can be extracted from both the chat statements and the fraudulent text separately, and the cosine similarity between the features of the chat statements and the fraudulent text can be calculated using a cosine similarity algorithm. This cosine similarity is then used as the fraud similarity. In another embodiment, a semantic similarity calculation model (such as a deep learning model) can also be used to calculate the text similarity between the chat statements and the fraudulent text to obtain the fraud similarity. Fraud similarity can also be calculated in other ways, and this application does not limit this method.

[0025] In step S104 of some embodiments, the stress score is a score used to characterize the degree of pressure exerted on the user by the product recommender. The anxiety score is a score used to characterize the degree of negative emotions such as anxiety or doubt exhibited by the user during the conversation with the product recommender. For example, in the context of financial product promotion in the fintech field, product recommenders of financial products such as credit cards and consumer loans may send urging, threatening, or inflammatory words in target group chats, such as "last chance" or "if you don't grab it now, it'll be gone."

[0026] In step S105 of some embodiments, the risk assessment result may be that the social chat record is a high-risk conversation, or the risk assessment result may be that the social chat record is a low-risk conversation.

[0027] In step S106 of some embodiments, if the community chat log contains high-risk conversations, a compliance alert is pushed to the target group chat via a customer service robot. This allows the customer service robot to intervene in the marketing process and address the risks promptly. Specifically, the compliance alert may include information to indicate the presence of non-compliant conversations in the target group chat. The compliance alert may also include information to clarify product marketing rules. For example, the compliance alert may be an official compliance statement. In another embodiment, punitive measures may also be taken against the product recommender, such as freezing the immediate marketing rewards of the product recommender's (e.g., an agent's) account.

[0028] In some embodiments, such as Figure 3 As shown, step S104, which involves performing interactive sentiment analysis based on chat statements to obtain the stress score of the product recommender towards the user and the anxiety score of the user towards the product recommender, includes the following steps: S201: Segment the chat statements by the category of the speaker to obtain the first chat statement of the product recommender and the second chat statement of the user; S202: Perform dialogue stress analysis on the first chat statement and obtain a stress score; S203: Perform emotional fluctuation analysis on the second chat statement to obtain an anxiety score.

[0029] The advantage of this embodiment lies in its ability to segment chat statements by the category of the speaker, resulting in the first chat statement from the product recommender and the second chat statement from the user. This allows for precise differentiation of the dialogue content between different subjects, avoiding sentiment analysis bias caused by confusion between statements from different roles. A stress score is obtained by performing dialogue stress analysis on the first chat statement, and an anxiety score is obtained by performing emotional fluctuation analysis on the second chat statement. This allows for the targeted extraction of marketing pressure characteristics exerted by the product recommender and emotional feedback characteristics of users when being marketed to, thereby accurately identifying risks in community dialogues and improving the compliance of community dialogues.

[0030] In step S201 of some embodiments, the speech object category segmentation refers to the processing operation of classifying chat statements from different speakers in the community chat history. The first chat statement is the content sent by the product recommender in the target group chat, and the second chat statement is the content sent by the user in the target group chat. For example, a large language model can be used to identify the sender's identity tag for each chat statement, classifying statements sent by the product recommender as first chat statements and statements sent by the user as second chat statements. In another embodiment, other identity recognition methods can also be used to achieve speech object category segmentation, and this is not limited to this.

[0031] In step S202 of some embodiments, dialogue pressure analysis refers to the processing operation of evaluating the pressure attributes of the product recommender's statements. It should be noted that the pressure score is a numerical value used to quantify the degree of marketing pressure exerted on the user by the product recommender. Specifically, the pressure score reflects the intensity of the product recommender's use of pressure tactics such as urging, intimidation, and incitement in the dialogue. For example, a preset set of pressure-related vocabulary can be used to match pressure-related words in the first chat statement, and the pressure score can be determined based on the number of matched words. Alternatively, other sentiment analysis models can also be used for dialogue pressure analysis.

[0032] In step S203 of some embodiments, emotion fluctuation analysis refers to assessing the emotional state of the user's statements. It should be noted that the anxiety score is a numerical value used to quantify the degree of negative emotions such as anxiety and doubt exhibited by the user during the conversation. The anxiety score can reflect the user's emotional response to the product recommender's sales pitch. For example, the anxiety score can be obtained by analyzing the frequency of questions and the proportion of negative words in the second chat statement. Alternatively, other emotion recognition models can also be used to perform emotion fluctuation analysis on the second chat statement.

[0033] In some embodiments, such as Figure 4 As shown, step S202, which involves performing a dialogue stress analysis on the first chat statement to obtain a stress score, includes the following steps: S301: Match the first chat statement with the preset set of stress words to obtain the target stress words; S302: Determine the stress score based on the number of words used to apply the target stress; where the number of words is positively correlated with the stress score.

[0034] The advantage of this embodiment lies in that it uses a preset set of stress-related vocabulary to match the first chat statement to obtain target stress-inducing words. This allows for precise identification of specific stress-inducing words used by the product recommender in the conversation, quantifying the characteristics of their stress-inducing behavior. A stress score is determined based on the number of target stress-inducing words, and the number of words is positively correlated with the stress score. This objectively quantifies the degree of stress exerted by the product recommender, avoiding biases in subjective judgment, and thus accurately identifying risks in community conversations to improve their compliance.

[0035] In step S301 of some embodiments, the pressure vocabulary set refers to a preset data set containing various words with marketing pressure attributes. The target pressure vocabulary refers to the words in the first chat statement that match words in the pressure vocabulary set. In some embodiments, for example, the pressure vocabulary set may include urging, threatening, or persuasive words such as "last chance," "if you don't grab it now, it'll be gone," "only spots left," and "you must buy it." These words are typically used by product recommenders to guide users to quickly make a purchase. Assuming the first chat statement is "This is the last chance to buy this insurance product," then the target pressure vocabulary is "last chance." In some embodiments, a string matching algorithm or semantic matching model can be used to compare each word in the first chat statement with words in the pressure vocabulary set, and the successfully matched words can be used as the target pressure vocabulary.

[0036] In step S302 of some embodiments, the number of words refers to the total number of target pressure words matched in the first chat statement. The pressure score is a numerical value used to quantify the degree of marketing pressure exerted on the user by the product recommender. The number of words and the pressure score are positively correlated; that is, the more target pressure words there are, the higher the corresponding pressure score, indicating a stronger degree of pressure from the product recommender. In some embodiments, a mapping rule between the number of words and the pressure score can be preset. For example, the pressure score increases by 5 points for each target pressure word matched, and the final total score is used as the pressure score. In another embodiment, the pressure score can also be determined through other calculation methods, and is not limited to these.

[0037] In some embodiments, such as Figure 5 As shown, step S203, which involves analyzing the emotional fluctuations of the second chat statement to obtain an anxiety score, specifically includes the following steps: S401: Analyze the questions based on the second chat statement to obtain the question frequency; S402: Perform word matching on the second chat statement based on preset candidate negative words to obtain the proportion of negative words; S403: Anxiety score is obtained by fusing data based on question frequency and negative word ratio.

[0038] The advantage of this embodiment lies in its ability to analyze the question frequency based on the second chat statements, thereby capturing the uncertainty and anxiety users experience during marketing and quantifying the instability of their emotions. By matching the second chat statements with preset candidate negative words to obtain the proportion of negative words, the negative emotional content expressed by users can be accurately located, quantifying the level of anxiety. Then, by fusing the question frequency and negative word proportion, an anxiety score is obtained. This comprehensive approach, combining multiple emotional characteristics, fully reflects the user's anxiety state, avoiding the bias of single-dimensional assessments. This allows for accurate identification of risks in community dialogues, thereby improving the compliance of such dialogues.

[0039] In step S401 of some embodiments, the question frequency refers to the proportion of questions in the total number of second chat statements. In some embodiments, it should be noted that when faced with product recommendations, users often express their doubts by asking questions if they have questions about product information, distrust marketing rhetoric, or feel pressured. A higher question frequency usually indicates stronger anxiety or uncertainty from the user. For example, a sentence structure recognition model can be used to determine whether each second chat statement is a question, and the question frequency is obtained by dividing the total number of questions by the total number of second chat statements.

[0040] In step S402 of some embodiments, candidate negative words refer to words collected and organized in advance that express various negative emotions, questioning attitudes, and anxieties. The negative word ratio refers to the proportion of negative words matched in the second chat statement to the total number of words in the second chat statement. A higher proportion of negative words in a user's speech usually indicates a more pronounced anxiety or questioning attitude. For example, a string matching algorithm can be used to compare each word in the second chat statement with candidate negative words, count the number of successfully matched negative words, and then divide this number by the total number of words in the second chat statement to obtain the negative word ratio.

[0041] In step S403 of some embodiments, the anxiety score is a numerical value used to quantify the degree of negative emotions such as anxiety and doubt exhibited by the user during the conversation. Specifically, the anxiety score can be obtained by weighted summation of question frequency and negative word proportion. Data fusion can also be achieved through other data fusion algorithms, and is not limited to this.

[0042] In some embodiments, such as Figure 6 As shown, after step S102, the product recommendation risk identification method may further include the following steps: S501: Obtain the compliance dialogue mode; where the compliance dialogue mode indicates that the risk warning node, product commitment node and transfer guidance node appear in sequence in the dialogue; S502: Perform statement pattern recognition on at least two chat statements according to the compliant dialogue mode, and obtain the statement pattern recognition result; S503: If the statement pattern recognition result indicates that at least two chat statements contain a risk warning node, a product commitment node, and a transfer guidance node in sequence, the community chat record is identified as a low-risk dialogue; otherwise, the community chat record is identified as a high-risk dialogue.

[0043] The advantage of this embodiment lies in acquiring a compliant dialogue pattern, where the pattern indicates the sequential appearance of risk warning nodes, product commitment nodes, and transfer guidance nodes in the dialogue. This clarifies the standard process for compliant dialogue in community marketing, providing a unified judgment benchmark for subsequent pattern matching. Based on the compliant dialogue pattern, at least two chat statements are subjected to statement pattern recognition. If the statement pattern recognition result indicates the sequential appearance of risk warning nodes, product commitment nodes, and transfer guidance nodes in at least two chat statements, the community chat record is identified as a low-risk dialogue; otherwise, it is identified as a high-risk dialogue. This allows for determining whether marketing behavior complies with compliance requirements based on the logical sequence of the dialogue process, thereby accurately identifying risks in community dialogues and improving their compliance.

[0044] In step S501 of some embodiments, the compliant dialogue pattern refers to a pre-defined standard dialogue process template that conforms to the requirements of community marketing regulations. This pattern clearly defines three key nodes that must appear sequentially in a compliant marketing dialogue: the risk warning node, the product commitment node, and the transfer guidance node. In some embodiments, the risk warning node is used to instruct the product recommender to inform the user of the product-related risks and disclaimers. The product commitment node is used to instruct the product recommender to introduce positive information such as product functions and benefits. The transfer guidance node is used to instruct the product recommender to guide the user to complete the purchase and transfer operation. It should be noted that a compliant marketing process should first warn of risks, then introduce the product, and finally guide the transaction, avoiding non-compliant situations where risk warnings are skipped and transfers are directly induced. For example, a compliant dialogue pattern can be formed by collecting samples of compliant marketing dialogues and extracting the process sequence from them.

[0045] In step S502 of some embodiments, the statement pattern recognition result refers to the conclusion that the identified chat statements match the three compliance nodes sequentially. Performing statement pattern recognition on at least two chat statements based on the compliance dialogue pattern can verify whether the current community dialogue process conforms to a pre-set compliance order. For example, preset risk warning words, product promise words, and transfer guidance words can be matched with chat statements to determine whether the content of the corresponding nodes exists, ultimately obtaining the statement pattern recognition result. Statement pattern recognition can also be performed using a semantic recognition model.

[0046] In step S503 of some embodiments, low-risk conversations are those that comply with community marketing compliance requirements, while high-risk conversations are those with non-compliant processes. In some embodiments, if the statement pattern recognition result indicates that risk warning nodes, product commitment nodes, and transfer guidance nodes appear sequentially in each chat statement, it indicates that the product recommender is strictly following the compliant process of first warning about risks, then introducing the product, and finally guiding the transaction. Therefore, the community chat record is identified as a low-risk conversation. In another embodiment, if the three nodes do not appear completely, or the order in which they appear does not meet the requirements, it indicates that the marketing process is in violation of regulations. Therefore, the community chat record is identified as a high-risk conversation. For example, if the product recommender sends the content of the transfer guidance node without showing the risk warning node, it can be directly determined as a high-risk conversation.

[0047] In some embodiments, such as Figure 7 As shown, step S502, which involves performing statement pattern recognition on at least two chat statements according to the compliant dialogue mode to obtain the statement pattern recognition result, specifically includes the following steps: S601: Extract the first current chat statement sequentially from at least two chat statements, perform a first detection on the first current chat statement according to the risk warning words corresponding to the risk warning node, and obtain the risk warning recognition result; wherein, the risk warning recognition result indicates whether the first current chat statement has or does not have a risk warning node; S602: When the first current chat statement contains a risk warning node, extract the second current chat statement for each chat statement after the first current chat statement, and perform a second detection on the second current chat statement according to the product commitment word corresponding to the product commitment node to obtain the product commitment identification result; wherein, the product commitment identification result indicates whether the second current chat statement contains or does not contain a product commitment node; S603: When the second current chat statement contains a product commitment node, the third current chat statement is extracted from each chat statement after the second current chat statement. The third current chat statement is then subjected to a third detection based on the transfer guidance word corresponding to the transfer guidance node to obtain the transfer guidance recognition result. The transfer guidance recognition result indicates whether the third current chat statement contains or does not contain a transfer guidance node. S604: If the third current chat statement contains a transfer guidance node, obtain the statement pattern recognition result, and the statement pattern recognition result indicates that at least two chat statements sequentially contain a risk warning node, a product commitment node, and a transfer guidance node.

[0048] The advantage of this embodiment lies in that, by detecting the risk warning in the first current chat statement based on the risk warning words corresponding to the risk warning nodes, the risk warning identification result can be obtained. This allows for prioritizing the identification of the starting node of the compliant dialogue process and confirming whether the product recommender has fulfilled its risk warning obligation. When the first current chat statement contains a risk warning node, the product commitment identification result is obtained by detecting the second current chat statement (i.e., the chat statement following the first current chat statement) based on the product commitment words corresponding to the product commitment node. This allows for analysis of whether the product recommender made product-related commitments after the risk warning. When the second current chat statement contains a product commitment node, the transfer guidance identification result is obtained by detecting the third current chat statement (i.e., the chat statement following the second current chat statement) based on the transfer guidance words corresponding to the transfer guidance node. If the third current chat statement contains a transfer guidance node, the statement pattern identification result indicating the sequential appearance of the three nodes is obtained. This rigorously verifies the order and logic of the compliant dialogue pattern, preventing non-compliant situations with missing or disordered nodes from being misjudged as low-risk, thereby accurately matching compliant full-process marketing dialogues. Therefore, it can accurately identify risks in community dialogues, thereby improving the compliance of community dialogues.

[0049] In step S601 of some embodiments, the first current chat statement refers to the chat statements to be detected extracted sequentially from multiple chat statements according to their speaking time. Risk warning words refer to specific words or phrases corresponding to risk warning nodes and used to identify risk notification content. For example, risk warning words may include words such as "investment risk," "return volatility," "disclaimer," and "please read the rules carefully." By detecting whether the first current chat statement contains such words, it can be determined whether the product recommender has fulfilled its risk warning obligation. In some embodiments, the first current chat statement can be compared with the risk warning words through string matching or semantic similarity calculation to obtain the risk warning identification result. It should be noted that the risk warning identification result refers to the judgment conclusion after detection regarding whether the first current chat statement contains a risk warning node.

[0050] In step S602 of some embodiments, the second current chat statement refers to the chat statement to be detected that follows the first current chat statement extracted in chronological order after the first current chat statement contains a risk warning node. Product commitment terms refer to specific words or phrases corresponding to product commitment nodes, used to identify product introductions and benefit descriptions. For example, product commitment terms may include phrases such as "coverage" or "no additional fees." It should be noted that the product commitment identification result refers to the determination of whether the second current chat statement contains a product commitment node after detection.

[0051] In step S603 of some embodiments, the third current chat statement refers to the chat statement to be detected that follows the second current chat statement extracted in chronological order after the second current chat statement contains a product commitment node. The transfer guidance word refers to a specific word or phrase corresponding to the transfer guidance node, used to identify and guide the user to make a transaction or transfer. For example, the transfer guidance word may include words such as "click the link to buy," "scan the code to pay," "transfer," and "place an order." It should be noted that the transfer guidance identification result refers to the determination of whether the third current chat statement, after detection, contains a transfer guidance node.

[0052] In step S604 of some embodiments, if the third current chat statement contains a transfer guidance node, it means that the risk warning node, product commitment node, and transfer guidance node appear sequentially in the chat statement in chronological order, which meets the requirements of a compliant dialogue mode, and the risk of skipping the risk warning and directly inducing a transfer is relatively small. Therefore, the statement pattern recognition result indicating the sequential appearance of the three nodes is obtained.

[0053] In some embodiments, such as Figure 8 As shown, step S105, which involves risk assessment based on fraud similarity, stress score, and anxiety score, specifically includes the following steps: S701: The total dialogue risk score is obtained by weighted summation based on fraud similarity, stress score, and anxiety score; S702: Compare the total dialogue risk score with the preset risk score threshold. If the total dialogue risk score is greater than or equal to the risk score threshold, obtain the risk judgment result and determine that the community chat record is a high-risk dialogue.

[0054] The advantage of this embodiment lies in its weighted summation of fraud similarity, stress score, and anxiety score to obtain a total dialogue risk score. This integrates multi-dimensional risk indicators into a unified, quantifiable risk assessment metric, avoiding the limitations of single-dimensional judgment. The total dialogue risk score is compared with a preset risk score threshold. If the total dialogue risk score is greater than or equal to the threshold, a risk determination result is obtained, and the community chat record is identified as a high-risk dialogue. This approach allows for risk assessment based on objective, quantifiable thresholds, reducing subjective bias and enabling accurate identification of risks in community dialogues, thereby improving the compliance of community conversations.

[0055] In step S701 of some embodiments, the total dialogue risk score is a quantitative risk value obtained by comprehensively evaluating fraud similarity, stress score, and anxiety score. Weighted summation integrates multi-dimensional risks into a unified evaluation value, avoiding the limitations of single-dimensional judgment. In some embodiments, for example, a weight of 0.4 can be assigned to fraud similarity, a weight of 0.3 to stress score, and a weight of 0.3 to anxiety score. The values ​​of the three indicators are multiplied by their respective weights and then summed to obtain the total dialogue risk score. The weights of each indicator can also be set or adjusted to other values, and are not limited to these.

[0056] In step S702 of some embodiments, the risk score threshold is a pre-set threshold value used to distinguish between high-risk and low-risk conversations. Specifically, if the total risk score of a conversation is greater than or equal to the risk score threshold, it indicates that the overall risk level has reached the standard requiring a warning, and therefore the risk assessment result is determined to be a high-risk conversation in the social media chat record. If the total risk score of a conversation is less than the risk score threshold, it is determined to be a low-risk conversation. For example, the risk score threshold can be set to 0.7 based on the risk score distribution of historical compliant and risky conversations.

[0057] In some embodiments, such as Figure 9 As shown, the product recommendation risk identification method may include the following steps: monitoring the community dialogue flow of the target group chat, obtaining the community chat records, and converting the voice in the community chat records into text. The text is then segmented and structured, and risk features are extracted from the segmented text. Risk feature extraction includes: performing semantic analysis on the segmented text to match a fraudulent text fingerprint database; performing sentiment analysis on the text to calculate agent pressure scores and customer anxiety scores; and performing pattern tracking on the text to track key dialogue nodes (including risk warning nodes, product promise nodes, and transfer guidance nodes). Real-time risk decision-making and handling are then performed. If the semantic similarity, agent pressure score, and customer anxiety score are all high, the dialogue risk is considered high, i.e., an inducement or fraudulent dialogue pattern has been identified. Alternatively, if the text does not conform to the pattern of key dialogue nodes, such as avoiding risk warning nodes, the dialogue risk is considered high, and automatic intervention is performed, such as inserting it into the target group chat. Otherwise, the dialogue risk is considered low.

[0058] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0059] In one embodiment, a product recommendation risk identification device is provided, which corresponds one-to-one with the product recommendation risk identification method described in the above embodiments. For example... Figure 10As shown, the product's recommended risk identification device includes a record acquisition module 801, a sentence segmentation processing module 802, a similarity calculation module 803, a sentiment analysis module 804, a risk determination module 805, and an information push module 806. Detailed descriptions of each functional module are as follows: The record acquisition module 801 is used to acquire the social chat history of the target group chat; the target group chat contains customer service robots, product recommenders, and users; The sentence segmentation module 802 is used to segment the social chat history into sentences to obtain chat statements; The similarity calculation module 803 is used to calculate the text similarity between the chat statement and the fraudulent text in the preset fraudulent text library to obtain the fraud similarity. The sentiment analysis module 804 is used to perform interactive sentiment analysis based on chat statements to obtain the stress score of the product recommender to the user and the anxiety score of the user to the product recommender. The risk assessment module 805 is used to assess risk based on fraud similarity, stress score, and anxiety score, and obtain risk assessment results. Among them, the risk assessment results include whether the social chat record is a high-risk conversation or a low-risk conversation. The information push module 806 is used to respond to high-risk conversations in social chat logs by pushing compliance reminders to the target group chat through a customer service robot.

[0060] This application provides a product recommendation risk identification device. It acquires community chat records from a target group chat containing customer service robots, product recommenders, and users. The chat records are segmented to obtain chat statements, and text similarity is calculated between these statements and fraudulent texts in a pre-set fraudulent text library to obtain a fraud similarity score. This allows for precise matching to determine if the product recommender's content matches known fraudulent inducement tactics and non-compliant words, thus identifying illegal and fraudulent texts. Furthermore, interactive sentiment analysis is performed on the chat statements to obtain the product recommender's stress score towards the user and the user's anxiety score towards the product recommender. This captures hidden violation characteristics from the emotional dimension of the dialogue interaction, identifying the risk of inducement marketing through emotional pressure. Risk assessment is performed based on fraud similarity, stress score, and anxiety score. In cases where the community chat records indicate high-risk dialogue, compliance reminders are pushed to the target group chat via customer service robots. This proactively triggers a compliance intervention mechanism, rather than relying solely on post-event user complaints to discover non-compliant marketing methods, thereby accurately identifying risks in community dialogues and improving the compliance of community conversations.

[0061] Specific limitations regarding the product recommendation risk identification device can be found in the limitations of the intelligent question-and-answer method described above, and will not be repeated here. Each module in the aforementioned product recommendation risk identification device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0062] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 11 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When executed by the processor, the computer program implements the functions or steps of a product recommendation risk identification method server-side.

[0063] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 12 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements the client-side functions or steps of a product recommendation risk identification method.

[0064] In one embodiment, a computer device is provided, 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 steps of a product recommendation risk identification method.

[0065] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0066] 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 a computer program instructing related hardware. The 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 a variety of 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.

[0067] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0068] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0069] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for identifying product recommendation risks, characterized in that, include: Obtain the chat history of the target group chat; wherein, the target group chat contains a customer service robot, a product recommender, and users; The community chat history is segmented into sentences to obtain chat statements; The text similarity is calculated based on the chat statements and fraudulent texts in the preset fraudulent text library to obtain the fraud similarity. Based on the chat statements, interactive sentiment analysis was performed to obtain the stress score of the product recommender to the user and the anxiety score of the user to the product recommender. Risk assessment is performed based on the fraud similarity, stress score, and anxiety score to obtain a risk assessment result; wherein, the risk assessment result includes whether the community chat record is a high-risk conversation or a low-risk conversation; If the chat log is found to be a high-risk conversation, a compliance alert will be sent to the target chat group via the customer service robot.

2. The method according to claim 1, characterized in that, The step of performing interactive sentiment analysis based on the chat statements to obtain the stress score of the product recommender to the user and the anxiety score of the user to the product recommender includes: The chat statements are categorized by the speaking audience to obtain the first chat statement of the product recommender and the second chat statement of the user. Perform a dialogue stress analysis on the first chat statement to obtain a stress score; Anxiety score was obtained by performing an emotional fluctuation analysis on the second chat statement.

3. The method according to claim 2, characterized in that, The step of performing dialogue stress analysis on the first chat statement to obtain a stress score includes: The first chat statement is matched with words based on a preset set of stress words to obtain the target stress words; The stress score is determined based on the number of words in the target stress vocabulary; wherein the number of words is positively correlated with the stress score.

4. The method according to claim 2, characterized in that, The step of performing emotional fluctuation analysis on the second chat statement to obtain an anxiety score includes: Based on the second chat statement, the question frequency is obtained through question analysis. The second chat statement is matched with preset candidate negative words to obtain the proportion of negative words; The anxiety score is obtained by fusing data based on the frequency of the questions and the proportion of negative words.

5. The method according to any one of claims 1 to 3, characterized in that, After segmenting the community chat log into sentences to obtain chat statements, the method further includes: Obtain a compliance dialogue mode; wherein, the compliance dialogue mode indicates that risk warning nodes, product commitment nodes, and transfer guidance nodes appear sequentially in the dialogue; Based on the compliant dialogue mode, perform statement pattern recognition on at least two of the chat statements to obtain statement pattern recognition results; If the statement pattern recognition result indicates that the risk warning node, the product commitment node, and the transfer guidance node appear sequentially in at least two of the chat statements, the community chat record is determined to be a low-risk conversation; otherwise, the community chat record is determined to be a high-risk conversation.

6. The method according to claim 5, characterized in that, The step of performing statement pattern recognition on at least two chat statements according to the compliant dialogue pattern to obtain statement pattern recognition results includes: A first current chat statement is obtained by sequentially extracting from at least two chat statements, and a first detection is performed on the first current chat statement according to the risk warning word corresponding to the risk warning node to obtain a risk warning identification result; wherein, the risk warning identification result indicates whether the risk warning node exists or does not exist in the first current chat statement; When the risk warning node exists in the first current chat statement, a second current chat statement is extracted from each chat statement after the first current chat statement. The second current chat statement is then subjected to a second detection based on the product commitment word corresponding to the product commitment node to obtain a product commitment identification result. The product commitment identification result indicates whether the product commitment node exists or does not exist in the second current chat statement. When the product commitment node exists in the second current chat statement, a third current chat statement is extracted from each chat statement after the second current chat statement. The third current chat statement is then subjected to a third detection based on the transfer guidance word corresponding to the transfer guidance node to obtain a transfer guidance identification result. The transfer guidance identification result indicates whether the transfer guidance node exists or not in the third current chat statement. If the third current chat statement contains the transfer guidance node, the statement pattern recognition result is obtained, and the statement pattern recognition result indicates that the risk warning node, the product commitment node, and the transfer guidance node appear sequentially in at least two of the chat statements.

7. The method according to any one of claims 1 to 3, characterized in that, The risk assessment based on the fraud similarity, the stress score, and the anxiety score, to obtain the risk assessment result, includes: The total dialogue risk score is obtained by weighting and summing the fraud similarity, stress score, and anxiety score. The total risk score of the conversation is compared with a preset risk score threshold. If the total risk score of the conversation is greater than or equal to the risk score threshold, the risk determination result is obtained, and the risk determination result is determined to be a high-risk conversation in the community chat record.

8. A product recommendation risk identification device, characterized in that, include: The record acquisition module is used to acquire the social chat history of the target group chat; wherein, the target group chat contains a customer service robot, a product recommender, and a user; The sentence segmentation module is used to segment the community chat records into sentences to obtain chat statements; The similarity calculation module is used to calculate the text similarity between the chat statement and the fraudulent text in the preset fraudulent text library to obtain the fraud similarity. The sentiment analysis module is used to perform interactive sentiment analysis based on the chat statements to obtain the stress score of the product recommender to the user and the anxiety score of the user to the product recommender. The risk assessment module is used to assess risk based on the fraud similarity, the stress score, and the anxiety score, and obtain a risk assessment result; wherein, the risk assessment result includes whether the community chat record is a high-risk conversation or the community chat record is a low-risk conversation; The information push module is used to push compliance reminder information to the target group chat in response to the fact that the social chat history is a high-risk conversation.

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 steps of the method as described in any one of claims 1 to 7.

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

Images