A voice outbound call tag retrieval method and system based on reinforcement learning

Abstract

This application provides a method and system for voice outbound call tag retrieval based on reinforcement learning. The method includes: identifying business targets requiring initial exploration, which have low frequency of occurrence or have not appeared in the system's historical decision-making experience; allocating a predetermined proportion of specific resources to the business target, these specific resources being unaffected by the system's predicted success rate for the business target; utilizing these specific resources to conduct initial outreach to the target group of the business target and acquiring customer interaction information generated from the initial outreach; performing intensity enhancement processing on the learning signal generated from the customer interaction information; adjusting the recommendation strategy for the business target based on the intensity-enhanced learning signal; evaluating the initial outreach effect of the business target; and adjusting the allocation proportion of the specific resources based on the initial outreach effect. This application improves the accuracy of tag retrieval and the rationality of resource allocation.

Description

Technical Field

[0001] This application relates to the field of voice outbound calling technology, and more specifically, to a voice outbound calling tag retrieval method and system based on reinforcement learning. Background Technology

[0002] In the daily operation of intelligent voice outbound calling systems, traditional tag retrieval technology often falls short when faced with a large customer base and the diverse behavioral information they generate. This is mainly reflected in slow system response speed, the inability to provide outbound callers with accurate customer information in a timely manner, and the low efficiency of customer behavior data utilization, with much valuable information not being fully explored. Existing data storage methods are often fixed and unchanging, making it difficult to adapt to changes in business needs. This directly limits the efficiency of tag matching, thereby affecting the timeliness and accuracy of outbound calling services.

[0003] As the market environment rapidly changes and customer behavior patterns become increasingly complex, banks continuously launch new financial products and services. Simultaneously, external economic conditions, social hot topics, and regulatory policy adjustments are subtly altering customer demand for financial products and their behavioral preferences. This shift in customer behavior patterns causes the predictive effectiveness of many previously learned tag combinations to decline, and even become biased. Tag combinations previously deemed valuable by the system no longer achieve the expected outbound call success and conversion rates, while some previously overlooked tag combinations may become potentially valuable due to new market trends. Faced with this continuous and rapid conceptual drift, the system's internal reinforcement learning mechanism, trained and optimized based on past data distributions, begins to lag in adaptation. It requires more time to identify new effective patterns and adjust its tag retrieval strategy.

[0004] This lag in adaptation further impacted the system's efficiency in executing new outbound calling tasks. When the marketing department urgently launched a cryptocurrency investment education campaign targeting a specific young customer group, the target customer profile for this campaign differed significantly from traditional financial product marketing. It primarily focused on tags such as being aged 20-35, frequently using mobile payments, following technology news, and having high social media activity. In the system's past learning mechanisms, these tag combinations had relatively low weight and relevance, belonging to niche or emerging tag combinations. Because the system was in an adaptation period to the shift in overall customer behavior patterns, its ability to assess the value of these emerging tag combinations was weakened. It struggled to quickly and accurately determine the potential returns of these new tag combinations, leading the system to conservatively allocate more outbound calling resources to traditional marketing tasks that it considered more secure and had performed better in the past, even if the actual returns of these traditional tasks had begun to decline. This risk-averse decision-making mechanism resulted in the newly launched cryptocurrency education campaign failing to gain sufficient exposure, a large number of potential target customers being overlooked, and the campaign's effectiveness falling far short of expectations. Summary of the Invention

[0005] This application discloses a voice outbound call tag retrieval method based on reinforcement learning, which aims to solve the problems of slow response speed, low efficiency of customer behavior data utilization, difficulty in adapting data storage methods to changes in business needs, lag in system adaptation to concept drift, and insufficient allocation of resources for the promotion of emerging businesses when existing intelligent voice outbound call systems face complex and ever-changing customer behavior patterns and market environments.

[0006] To achieve the above objectives, this application adopts the following technical solution: In the first aspect, this application discloses a voice outbound call tag retrieval method based on reinforcement learning, including the following steps: Identify business objects that require initial exploration, which have a low frequency of occurrence or have not appeared in the system's historical decision-making experience; A specific proportion of resources is allocated to this business object, and these specific resources are not affected by the system's prediction of the success rate of this business object. Utilize this specific resource to conduct initial outreach to the target group of this business entity and obtain customer interaction information generated from this initial outreach; The learning signal generated from the customer interaction information is enhanced in intensity. Based on the learning signal after the intensity enhancement processing, adjust the recommendation strategy for this business object; Assess the initial outreach effectiveness to the target business; Based on the initial reach effect, adjust the allocation ratio of this specific resource.

[0007] Secondly, this application also discloses a voice outbound call tag retrieval system based on reinforcement learning, the system comprising: The identification module is used to identify business objects that need to be explored initially, which have a low frequency of occurrence or have not appeared in the system's historical decision-making experience. The resource allocation module is used to allocate a specific proportion of resources to the business object, and these specific resources are not affected by the system's prediction of the success rate of the business object. The outreach execution module is used to utilize the specific resources to conduct initial outreach to the target group of the business object and obtain customer interaction information generated by the initial outreach. The signal enhancement module is used to enhance the intensity of the learning signal generated by the customer interaction information. The strategy adjustment module is used to adjust the recommendation strategy for the business object based on the learning signal after the intensity enhancement processing. The effectiveness evaluation module is used to evaluate the initial reach of this business target. The resource adjustment module is used to adjust the allocation ratio of a specific resource based on the initial reach effect. Beneficial effects

[0008] This application effectively addresses the problem in existing technologies where emerging businesses struggle to obtain sufficient resources due to a lack of historical data by identifying business targets requiring initial exploration and allocating specific resources to them, unaffected by predicted success rates. By utilizing these specific resources to conduct initial outreach to the target group and acquire customer interaction information, this application obtains authentic and up-to-date customer feedback data, overcoming the lag in adaptation to concept drift inherent in traditional systems. Furthermore, the learning signals generated from customer interaction information undergo intensity amplification processing, enabling the system to more accurately identify valuable feedback and avoiding learning biases caused by data sparsity or noise. Adjusting the recommendation strategy for business targets based on the amplified learning signals allows the system to quickly adapt to market changes and shifts in customer behavior patterns, improving the accuracy and timeliness of tag retrieval. Finally, by evaluating the effectiveness of the initial outreach and adjusting the allocation ratio of specific resources accordingly, a closed-loop adaptive learning mechanism is formed, ensuring dynamic optimization of resource allocation and improved overall efficiency.

[0009] In summary, this application effectively solves the problems of existing intelligent voice outbound calling systems when facing complex and ever-changing customer behavior patterns and market environments, such as slow response speed of traditional tag retrieval technology, low efficiency of customer behavior data utilization, difficulty in adapting data storage methods to changes in business needs, lag in system adaptation to concept drift, and insufficient allocation of resources for promoting emerging businesses. It significantly improves the accuracy and conversion rate of voice outbound calling. Attached Figure Description

[0010] Figure 1 A flowchart illustrating a voice outbound call tag retrieval method based on reinforcement learning provided in this application; Figure 2 This is a schematic diagram of the structure of a voice outbound call tag retrieval system based on reinforcement learning provided in this application. Detailed Implementation

[0011] The technical solutions 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, and not all embodiments.

[0012] like Figure 1 As shown, this application proposes a voice outbound call tag retrieval method based on reinforcement learning, including: Identify business objects that require initial exploration, which have a low frequency of occurrence or have not appeared in the system's historical decision-making experience; Allocate a specific proportion of resources to a business object, and these specific resources are not affected by the system's prediction of the success rate of the business object. Utilize specific resources to conduct initial outreach to the target group of the business entity and obtain customer interaction information generated from the initial outreach; Intensity enhancement processing is applied to the learning signals generated from customer interaction information; Adjust the recommendation strategy for business objects based on the learning signal after intensity enhancement processing; Evaluate the effectiveness of initial outreach to business targets; Adjust the allocation ratio of specific resources based on the initial reach effect.

[0013] The method in this application aims to solve the problems of slow response speed, low data utilization efficiency and poor adaptability of traditional tag retrieval technology when facing complex and ever-changing market environments and customer behavior patterns, thereby improving the accuracy and efficiency of voice outbound calls.

[0014] To better understand the technical solution proposed in this application, some key terms involved will be explained first.

[0015] The business targets in this application refer to banks or financial institutions launching emerging businesses, including newly launched financial products, services, or marketing activities, such as new wealth management products, credit card promotions, and loan services. These business targets typically have specific target customer groups and marketing objectives.

[0016] Initial exploration refers to the process by which a system proactively allocates resources to conduct small-scale exploratory marketing or outreach when it lacks sufficient historical data or experience with a particular business target, in order to quickly collect feedback information.

[0017] Specific resources can be understood as marketing budgets, outbound call agent time, SMS or email sending quotas, etc. for initial outreach. Their characteristic is that their allocation ratio is not affected by the system's prediction of the success rate of the target business in the initial exploration phase, which is intended to encourage the exploration of new businesses.

[0018] Customer interaction information refers to various behavioral data generated by customers after receiving initial contact, such as call duration, SMS reply content, link click behavior, and product page browsing history.

[0019] Learning signals are quantitative indicators extracted from customer interaction information to guide adjustments to system strategies, such as customer interest and conversion intention.

[0020] Intensity enhancement processing refers to amplifying or weighting the learning signal to highlight its importance to system decision-making, especially in the initial exploration phase, in order to identify potential value more quickly.

[0021] Recommendation strategy refers to the methods and content that the system adjusts based on learned signals to recommend to customers based on business targets. Examples include adjusting outbound call scripts, recommending product combinations, and optimizing outreach channels.

[0022] The core of the voice outbound call tag retrieval method based on reinforcement learning proposed in this application lies in achieving effective exploration and strategy optimization for emerging or low-frequency business targets through a series of steps.

[0023] Specifically, this method first identifies business entities that require initial exploration. These entities typically appear infrequently or never in the system's historical decision-making experience, such as newly launched financial products or new services targeting specific niche markets. Identifying these entities can be achieved in several ways. For example, the system can periodically scan the list of newly launched businesses or identify them based on innovative business or pilot project identifiers submitted by business departments. Alternatively, historical data can be analyzed to identify businesses that have been reached by the system less than a preset threshold in the past period, or whose related tag combinations have never been effectively utilized by the system. For instance, when a bank launches a brand-new green energy-themed fund, because its investment direction and target customer group differ significantly from traditional funds, the system's historical data may lack effective predictions of its success rate; in this case, the fund is identified as a business entity requiring initial exploration.

[0024] Subsequently, specific resources are allocated to each business entity according to a predetermined proportion. These specific resources are unaffected by the system's predicted success rate for the business entity. This resource allocation mechanism aims to break away from the inertia of traditional resource allocation based on historical success rates, providing opportunities for trial and error for emerging businesses. Resource allocation can be initially set based on factors such as the type of business entity, potential market size, and risk level. For example, for an innovative business identified as having high potential, a relatively high initial exploration resource quota can be allocated, such as 1,000 outbound call agent hours or 5,000 SMS sending quotas. These resources will not be reduced during the initial exploration phase, even if the system predicts a low success rate for the business, thus ensuring the smooth progress of the initial exploration.

[0025] Next, specific resources are used to conduct initial outreach to the target group of the business entity and to obtain customer interaction information generated during this initial outreach. Initial outreach is a crucial step in collecting feedback and can be conducted through various channels and methods. For example, allocated outbound call center resources can be used to make phone calls to the target customer group to introduce the business entity and record customer feedback; alternatively, business introduction information can be sent via digital channels such as SMS, email, or app push notifications, and customer clicks, browsing, and responses can be monitored. For instance, regarding the aforementioned green energy-themed fund, the system can send SMS messages containing a link to the fund introduction to a group of initially screened potential customers and record interaction data such as whether customers clicked the link, the duration of their stay on the page, and whether they downloaded related materials.

[0026] Furthermore, the learning signals generated from customer interaction information are enhanced. In the initial exploration phase, due to limited data, the original learning signals may be weak and insufficient to drive effective strategy adjustments. Therefore, these signals need to be enhanced to amplify their impact on decision-making. For example, if customers exhibit positive interactive behaviors after initial contact (such as proactively inquiring or browsing product pages for extended periods), the system can assign higher weights to the learning signals corresponding to these positive feedbacks, allowing them to play a greater role in subsequent strategy adjustments. This enhancement helps the system more quickly identify potential business targets and effective outreach strategies.

[0027] Based on the enhanced learning signals, the system adjusts its recommendation strategy for specific business targets. After enhancement, the system uses these signals to optimize the recommendation strategy for that business target. This might include adjusting outbound call scripts to better align with customer interests; optimizing the combination of recommended products to improve conversion rates; or adjusting outreach channels to select communication methods preferred by customers. For example, if the enhanced learning signals indicate a strong customer interest in the environmental philosophy of a green energy-themed fund, the system can adjust subsequent outbound call scripts to emphasize the fund's social responsibility and environmental benefits, rather than solely focusing on returns.

[0028] Next, assess the initial outreach effectiveness. This assessment is crucial for determining the business's potential and adjusting resource allocation. The assessment can be conducted from multiple dimensions, such as customer interest, initial conversion rate, and market adaptability. For example, it can be calculated what percentage of initially reached customers showed clear purchase intent or completed initial registration. Simultaneously, external market data can be used to evaluate the business's alignment with current market trends.

[0029] Finally, the allocation ratio of specific resources is adjusted based on the initial outreach results. According to the evaluation results of the initial outreach, the system will dynamically adjust the proportion of specific resources allocated to that business target. If the initial outreach results are good, indicating that the business target has high potential, the system can increase its resource quota for larger-scale promotion; conversely, if the results are poor, resources can be appropriately reduced to avoid unnecessary investment. For example, if the initial outreach evaluation of the green energy-themed fund shows high customer interest and an initial conversion rate exceeding expectations, the system can increase its resource quota from 1,000 outbound call hours to 2,000 and expand its promotion scope to a wider range of potential customers.

[0030] The voice outbound call tag retrieval method proposed in this application, through a series of closely linked steps, forms a closed-loop exploration, learning, and optimization mechanism. This method first identifies emerging business targets that have low frequency of occurrence or have not appeared in the system's historical experience, breaking the traditional system's conservative attitude towards niche businesses. By allocating specific resources to these business targets that are unaffected by predicted success rates, this application encourages exploration of unknown areas, providing valuable trial-and-error opportunities for new products and services. In the initial outreach phase, the system proactively collects customer interaction information and enhances the intensity of the generated learning signals. This allows the system to quickly capture valuable feedback even with limited data, accelerating the identification of new business potential. Subsequently, the recommendation strategy is adjusted based on the enhanced learning signals, ensuring the accuracy and effectiveness of marketing activities. Through a comprehensive evaluation of the initial outreach results, the system can objectively judge the market performance of business targets and dynamically adjust the allocation ratio of specific resources accordingly, achieving optimal resource allocation.

[0031] Compared to traditional tag retrieval technologies, the advantage of this application lies in its ability to rapidly adapt to and explore emerging businesses. Traditional systems often rely on large amounts of historical data for prediction and decision-making. For new businesses lacking historical data, their performance is often poor, and they may even miss market opportunities due to conservative resource allocation strategies. This application proactively creates opportunities for new businesses to obtain feedback by introducing an initial exploration and specific resource allocation mechanism. Furthermore, by enhancing the learning signal, the system can extract stronger decision signals from limited interaction data in the initial stage, accelerating the learning process. This mechanism enables the system to more quickly identify changes in market trends and shifts in customer behavior patterns, adjusting recommendation strategies in a timely manner, thus effectively solving the problem of adaptation lag when facing concept drift. For example, when facing emerging businesses such as digital currency investment education, the method of this application can allocate the necessary initial exploration resources, quickly collect customer feedback through small-scale outreach, and amplify positive feedback to quickly identify the potential value of emerging tag combinations such as age 20-35, frequent use of mobile payment, and attention to technology information. The recommendation strategy can then be adjusted in a timely manner to ensure that the campaign gains sufficient exposure and effective reach to the target customers, ultimately achieving the expected goals of the campaign.

[0032] In some of the above embodiments, the step of enhancing the intensity of the learning signal generated by customer interaction information includes: When the system obtains initial positive interaction information from customers through initial outreach, it generates an initial positive feedback record containing customer identification, task identification, original reward value, and feedback timestamp. Generate a digital tracking code for this initial outreach and embed the digital tracking code into the digital profile link that will be sent to the customer; Based on customer preferences, send customers links to digital profiles containing embedded digital tracking codes via digital channels; Mark the customer's association status with this initial contact as pending confirmation and start a confirmation waiting timer; Continuously monitor the bank's digital channels to track whether customers click on links with digital tracking codes, visit specific product pages, download related documents, or actively search for related keywords within a preset time period; If a customer is detected to have engaged in the aforementioned proactive behavior within a preset time period, the association status between the customer and the initial contact will be updated to "intent confirmed". Only when a customer's intention is confirmed as confirmed, the original reward value in the initial positive feedback record is enhanced to obtain an enhanced learning signal. If the waiting timer times out and no proactive confirmation is detected, the original reward value in the initial positive feedback record will not be enhanced.

[0033] Specifically, initial positive interaction information can be understood as the initial interest shown by the customer during the outbound voice call, such as the customer expressing a desire to learn more or requesting materials. Upon receiving such information, the system immediately generates an initial positive feedback record, which includes the customer's unique identifier, the identifier of this outbound call task, an initial reward value (e.g., set to 1 or 0.5 to represent initial positive feedback), and a timestamp of the feedback occurrence. The digital tracking code is a unique, traceable identifier designed to precisely link the customer's subsequent actions on digital channels to this initial contact. This digital tracking code is embedded in the digital profile link sent to the customer, for example, as a URL parameter or part of a short link. The digital profile link can point to the bank's product introduction page, brochure, online application form, or related videos. In practice, digital channels can include, but are not limited to, SMS, email, bank app push notifications, and WeChat official account messages. The system will select the most appropriate digital channel to send the digital profile link containing the digital tracking code to the customer based on their historical preferences, registration information, or the current interaction context.

[0034] Furthermore, marking the customer's association status with this initial contact as "intent pending confirmation" means the system needs to wait for further customer actions on digital channels to confirm their true intention. A confirmation wait timer is used to set a reasonable waiting period, such as 24 hours, 48 ​​hours, or 72 hours, during which the system will continuously monitor the customer's subsequent behavior. Continuously monitoring the bank's digital channels to see if the customer clicks on links with digital tracking codes, visits specific product pages, downloads related documents, or actively searches for related keywords within a preset timeframe means the system obtains real-time data on customer behavior on these channels by integrating with the bank's digital platform data interface. For example, it can monitor whether the customer clicked on a link with a specific tracking code, visited a specific page related to the product initially contacted, downloaded related product documents, or actively searched for keywords related to the business within the bank's app or website. If the system detects that the customer has engaged in the above-mentioned proactive behaviors within the preset timeframe, such as clicking links, visiting pages, downloading documents, or actively searching for keywords, these behaviors are considered strong evidence of the customer's true intention. At this point, the customer's association status with this initial contact will be updated to "intent confirmed." The original reward value in the initial positive feedback record is only enhanced when the customer's intention is confirmed. For example, the original reward value can be multiplied by an enhancement factor (such as 1.5 or 2), or directly increased to a higher fixed value (such as 5 or 10), resulting in an enhanced learning signal. This enhancement ensures that only positive feedback actively confirmed by the customer is given higher weight for subsequent recommendation strategy adjustments. If the confirmation wait timer times out and no active confirmation is detected, it is assumed that the customer's initial positive interaction may not reflect genuine intention, or that their interest has waned. In this case, the original reward value in the initial positive feedback record is not enhanced; it remains unchanged or is set to a lower value to avoid introducing inaccurate learning signals.

[0035] This application's solution effectively addresses the inaccuracy that may result from relying solely on initial interaction information to enhance the learning signal by introducing an intent confirmation mechanism. Specifically, upon receiving initial positive interaction information from a customer, the system does not immediately enhance the learning signal. Instead, it first generates an initial positive feedback record and a digital tracking code for this outreach, embedding it into the digital material link sent to the customer. This mechanism allows the system to send materials to the customer via digital channels and continuously monitor the customer's subsequent proactive behavior within a preset timeframe. By monitoring whether the customer clicks links, visits pages, downloads documents, or actively searches for keywords, the system can determine the customer's true intent. Only when the customer confirms their intent through these proactive behaviors is the original reward value enhanced, thus ensuring the quality and accuracy of the learning signal. This delayed and conditional enhancement process avoids noise caused by superficial or uncertain customer interactions, enabling the reinforcement learning model to adjust its strategy based on more reliable feedback.

[0036] Through the above technical solution, this application can significantly improve the accuracy and reliability of learning signals. Compared with directly enhancing initial interaction information, the intention confirmation mechanism introduced in this solution effectively filters out superficial or uncertain customer interactions, ensuring that only feedback that truly reflects customer interests and intentions is given higher learning weight. This allows the reinforcement learning model to receive higher-quality training data, thereby more accurately adjusting the recommendation strategy for business targets and avoiding resource waste and decision bias caused by misleading signals. Therefore, it not only improves the efficiency and accuracy of voice outbound call tag retrieval but also optimizes the customer experience and reduces the risk of invalid outreach.

[0037] In some of the above embodiments, the step of evaluating the initial reach effect of the business target in the above reinforcement learning-based voice outbound call tag retrieval method includes: Receive product descriptions and target customer tags from business objects, and identify the themes and investment directions of business objects through semantic analysis to obtain the semantic features of business objects; Continuously monitor the external financial market environment, obtain market hot topics, and market sentiment and frequency of mention related to the topics of business targets; Based on market sentiment and frequency of mention, the weight of customer interest and initial conversion rate in the overall evaluation is adjusted in real time. The market adaptability score is calculated based on the degree to which the semantic features of the business object match the hot topics in the market, as well as the degree of match between the initial reach and market sentiment. The initial reach effect on business targets is judged based on the adjusted evaluation indicator weights and market adaptability scores.

[0038] Specifically, receiving product descriptions and target customer tags for business objects refers to the system acquiring detailed information about the business object to be evaluated, such as textual descriptions of its functions, features, and target audience, as well as pre-defined target customer profile tags. Through semantic analysis, these descriptive texts can be deeply understood, extracting the core themes of the business object (e.g., green investment, technology innovation funds) and its investment directions (e.g., new energy, artificial intelligence), thus forming a set of structured semantic features of the business object. These semantic features form the basis for subsequent assessments of its market fit.

[0039] Continuous monitoring of the external financial market environment can be understood as the system tracking and analyzing financial market dynamics in real time through various data sources, such as news media, social media, industry reports, and financial data platforms. In this process, the system can identify current market hot topics, such as industries or concepts that have recently received widespread attention. Simultaneously, the system also monitors keywords related to the business's target topics to obtain public market sentiment (e.g., positive, negative, neutral) regarding these topics and the frequency of their mention in the market. This real-time market data provides a basis for dynamically adjusting evaluation strategies.

[0040] Based on market sentiment and mention frequency, the system adjusts the weighting of customer interest and initial conversion rate metrics in the overall evaluation in real time. Specifically, when the market shows highly positive sentiment or high mention frequency for a particular topic, it indicates that the topic has high market attention and potential appeal. In this case, the system will correspondingly increase the weighting of customer interest, initial conversion rate, and other evaluation metrics in the overall evaluation to more sensitively capture positive market feedback to the business target. Conversely, if market sentiment is low or mention frequency decreases, the weighting of these metrics will be reduced to avoid overinterpreting inactive market signals.

[0041] A market adaptability score is calculated based on the degree to which the semantic features of the business object align with current market trends, and the degree of match between initial outreach and market sentiment. The alignment between the semantic features of the business object and current market trends refers to quantifying the degree of alignment between the business object and market trends by comparing the similarity or correlation between the semantic features of the business object and current market trends. The match between initial outreach and market sentiment refers to analyzing the consistency between customer interaction information triggered by the initial outreach and current market sentiment. For example, if the market is generally optimistic about a certain topic, and the initial outreach also received positive feedback, the match is considered high. Combining these two factors, a market adaptability score can be calculated, which aims to reflect the potential success rate of the business object in the current market environment.

[0042] Therefore, the initial outreach effectiveness of the business target is assessed based on the adjusted evaluation indicator weights and market adaptability score. This means that the final evaluation result is no longer based on fixed rules, but rather integrates dynamically adjusted evaluation indicator weights (reflecting the market's sensitivity to customer interest and conversion rates) and market adaptability score (reflecting the business target's fit with the market environment). In this way, a more comprehensive, accurate, and forward-looking assessment of the initial outreach effectiveness of the business target can be made.

[0043] This application effectively addresses the static and lagging issues that may exist in the initial outreach effectiveness assessment by introducing a multi-dimensional and dynamic evaluation mechanism. Specifically, firstly, by performing semantic analysis on the product description and target customer tags of the business object, the core themes and investment directions of the business object can be accurately captured, forming its unique semantic features and laying the foundation for subsequent matching with market hotspots. Secondly, continuous monitoring of the external financial market environment and real-time acquisition of market hotspots, market sentiment, and mention frequency ensure that the evaluation process closely follows the market pulse and avoids evaluation bias caused by market changes. It is precisely because of the ability to perceive market dynamics in real time that the system can dynamically adjust the weight of evaluation indicators such as customer interest and initial conversion rate based on market sentiment and mention frequency, ensuring that more attention is paid to positive signals when the market is active and avoiding misjudgments when the market is sluggish. On this basis, by calculating the degree of fit between the semantic features of the business object and market hotspots and the matching degree between the initial outreach and market sentiment, a market adaptability score is generated. This score intuitively reflects the potential success potential of the business object in the current market environment. Ultimately, by combining these dynamically adjusted weights and market adaptability scores, the effectiveness of initial outreach is assessed, ensuring that the evaluation results not only reflect direct customer interaction but also incorporate considerations of the macro-market environment, making the evaluation results more instructive.

[0044] Through the aforementioned technical solution, this application significantly improves the accuracy and real-time performance of initial outreach evaluations of business targets. Compared to evaluation methods relying solely on fixed indicators, this solution introduces semantic analysis, real-time market monitoring, and dynamic weight adjustments, enabling the evaluation results to more accurately reflect the fit between the business target and the current market environment, as well as the true interests of customers. This dynamic and adaptive evaluation mechanism effectively avoids evaluation distortion caused by changes in the market environment, thus providing a more reliable basis for adjusting subsequent recommendation strategies and allocating specific resources. Especially for emerging or low-frequency business targets, whose market performance often fluctuates significantly, this solution can more sensitively capture market signals, adjust evaluation focus in a timely manner, and thereby optimize resource allocation, improving the success rate and efficiency of initial exploration.

[0045] In some of the above embodiments, the steps of continuously monitoring the bank's digital channels to track whether customers click on information links with digital tracking codes, visit specific product pages, download related documents, or actively search for related keywords within a preset time period include: When the system detects that a customer actively searches for keywords related to business objects through the bank's digital channels, it performs text processing on the search keywords and extracts core words; Based on a pre-defined financial intent dictionary and rule set, core words are matched and categorized to distinguish intents such as proactive understanding, risk assessment, or competitor comparison. Search behavior that actively seeks to understand intent will be included in the criteria for intent confirmation; Searches with other intentions will not be included in the criteria for confirming intent.

[0046] Specifically, when the system detects that a customer is actively searching for keywords related to the bank's business through its digital channels, the first step is to process these search keywords. Text processing can include word segmentation, part-of-speech tagging, and stop word removal. The goal is to extract key information representing the customer's core intent—the core vocabulary—from the original search query. For example, for the search phrase "how much does a certain wealth management product yield?", after text processing, the core vocabulary might be extracted as "wealth management product" and "yield".

[0047] Furthermore, to accurately determine a customer's true intent, it is necessary to perform intent matching and classification on core terms based on a pre-defined financial intent dictionary and rule set. The financial intent dictionary can contain terms related to financial products, services, risks, and competitors, along with their corresponding intent tags, such as purchase, investment, risk, and comparison. The rule set can define the mapping relationship between word combinations, syntactic structures, and specific intents. In this way, customer search intent can be categorized into various types, such as proactive understanding, risk assessment, or competitor comparison. Proactive understanding intent typically indicates that the customer shows a clear interest in a product or service, aiming to obtain more detailed information for consideration; risk assessment intent focuses on understanding potential risks or adverse factors; and competitor comparison intent indicates that the customer is weighing the pros and cons of different products or services.

[0048] Based on this, this application explicitly stipulates that only search behavior with a positive intent to learn is included in the basis for intent confirmation. This means that only when a customer's search behavior is explicitly classified as positive learning is it considered a valid signal that the customer's intent has been confirmed. Conversely, search behavior classified as having other intents such as risk assessment or competitor comparison is not included in the basis for intent confirmation.

[0049] This application's solution addresses the potential for misjudgment based solely on keyword appearance by employing refined intent identification and classification of keywords actively searched by customers. Specifically, when a customer actively searches for keywords related to a business objective on digital channels, the system no longer simply interprets it as a positive signal. Instead, it first extracts core terms through text processing, ensuring the analysis focuses on key information. Subsequently, using a pre-defined financial intent dictionary and rule set, it performs intent matching and classification on these core terms, accurately distinguishing whether the customer's purpose is to actively seek information or for other purposes such as risk assessment or competitor comparison. This meticulous intent differentiation ensures that only search behaviors genuinely expressing a positive intent are adopted as the basis for intent confirmation, while other search behaviors that may not represent a strong positive intent are excluded. This ensures the accuracy and reliability of customer intent confirmation, avoiding resource waste and strategy deviations caused by misjudging customer intent.

[0050] The aforementioned technical solutions significantly improve the accuracy of customer intent confirmation. By finely categorizing customer search intent, it effectively avoids misclassifying non-positive search behaviors as positive intent, thereby reducing false positive feedback. This makes the learning signals used to enhance the initial reward values ​​in positive feedback records more authentic and reliable, thus improving the accuracy of adjusting business target recommendation strategies. Furthermore, more accurate intent confirmation helps optimize the allocation of specific resources, ensuring that resources are invested in truly potential customer groups, improving the overall efficiency and effectiveness of initial outreach.

[0051] In some of the above embodiments, the step of continuously monitoring the bank's digital channels to see if customers click on information links with digital tracking codes, visit specific product pages, or download related documents within a preset time period includes: When the system detects that a customer clicks on a link with a digital tracking code, visits a specific product page, or downloads a related document, it collects the customer's interaction data on digital channels. The interaction data includes page dwell time, scroll depth, number of clicks on interactive elements, and content reading progress. Based on preset engagement thresholds, it is determined whether the customer has engaged in in-depth browsing. Engagement thresholds include page dwell time threshold, scroll depth threshold, and number of clicks on interactive elements threshold. When a customer's interaction data meets the engagement threshold, the behavior will be included in the basis for confirming intent. If a customer's interaction data does not reach the engagement threshold, the behavior will not be included in the basis for confirming intent.

[0052] Interaction data refers to behavioral data generated when customers interact with links, product pages, or downloaded documents on digital channels. Specifically, page dwell time measures the time a customer spends paying attention to content; scroll depth reflects the extent to which a customer browses the page content; the number of clicks on interactive elements indicates the customer's interest in interactive elements such as buttons and links on the page; and content reading progress directly reflects the customer's completion rate of reading the document or article. This data is collected comprehensively to depict the actual level of customer engagement. Engagement thresholds are pre-set standards used to determine whether a customer has engaged in in-depth browsing. For example, a page dwell time threshold of 30 seconds, a scroll depth threshold of 80%, and an interactive element click threshold of 2 times can be set. A customer's in-depth browsing is only considered when their interaction data, such as page dwell time, scroll depth, and interactive element clicks, all reach or exceed these pre-set thresholds. The purpose is to filter out accidental or superficial interactions, ensuring that only customer behaviors that genuinely demonstrate high interest and engagement are considered valid evidence of intent.

[0053] This application addresses the problem of accurately determining a customer's true intention based solely on simple clicks or visits by further collecting and analyzing customer interaction data on digital channels while monitoring their proactive behavior, and combining this with preset engagement thresholds to determine whether the customer has engaged in in-depth browsing. Specifically, when a customer clicks on a link, visits a product page, or downloads a document, the system no longer simply considers this as confirmation of intention, but further collects interaction data such as page dwell time, scroll depth, number of clicks on interactive elements, and content reading progress. This interaction data provides a more detailed reflection of the customer's actual engagement with the content. Subsequently, by comparing this interaction data with preset engagement thresholds, the system can distinguish between customers who are merely superficially browsing and those who have truly engaged in in-depth understanding. Only when a customer's interaction data meets these engagement thresholds, indicating in-depth browsing, is the behavior counted as intent confirmation. Conversely, if the interaction data does not reach the thresholds, it is not counted, effectively avoiding misjudgments caused by accidental clicks or superficial browsing. This ensures the accuracy and reliability of customer intent confirmation, providing more precise input for subsequent learning signal strength enhancement processing.

[0054] Through the above technical solution, this application can significantly improve the accuracy of customer intent confirmation. By introducing interactive data collection and engagement threshold judgment, the system can effectively distinguish between shallow customer interactions and deep engagement, avoiding misjudging accidental or low-intent clicks as positive intent. This makes the customer intent confirmation based on the intensity enhancement processing of the original reward value in the initial positive feedback records more reliable, thereby generating more accurate and representative learning signals. More accurate learning signals help the system more effectively adjust the recommendation strategy for business objects, optimize resource allocation, and ultimately improve the overall efficiency and success rate of voice outbound call tag retrieval.

[0055] In some of the above embodiments, the step of allocating a specific proportion of resources to a business object, where the specific resources are not affected by the system's prediction of the business object's success rate, includes: Obtain information about the business type, risk level, and current market attention to the topic of the business object; Based on the business type, risk level, and market attention, and in conjunction with the preset resource allocation rules, an initial exploration resource quota is calculated. The initial exploration resource quota is not affected by the system's prediction of the success rate of business objects during the initial exploration phase; After the initial outreach, continuously monitor the market adaptability score of the target audience and the effectiveness of the initial outreach; Adjust the initial exploration resource quota based on market adaptability scores and initial reach effectiveness.

[0056] Specifically, business type refers to the industry category or product type of the business object, such as a bank's wealth management products, loan business, or insurance services. Risk level refers to the inherent risk attributes of the business object, such as the level of risk assessed based on its investment targets, return volatility, and other factors. Market attention to the business object's theme can be understood as indicators such as the current market's discussion volume, search volume, or media exposure related to the theme or concept, aiming to reflect the impact of the external environment on the potential attractiveness of the business object. The preset resource allocation rules are a series of predefined logics or algorithms used to determine a reasonable initial exploration resource quota based on input parameters such as business type, risk level, and market attention. For example, a high-risk but highly anticipated emerging business may be allocated relatively more exploration resources.

[0057] Initial exploration resource quota refers to the specific amount of resources allocated to a business target during the initial exploration phase. This could include, for example, the number of outbound call seats, a maximum number of outreach attempts, or a marketing budget. This quota is designed to be unaffected by the system's predicted success rate for the business target during the initial exploration phase, ensuring a fair exploration opportunity for new or low-frequency businesses. In practical applications, market adaptability score is an indicator that measures the degree to which a business target fits the current market environment. Its calculation can comprehensively consider the semantic characteristics of the business target and their relevance to trending market topics, as well as the match between initial outreach and market sentiment. Initial outreach effectiveness is an evaluation of the actual performance of the initial outreach activities, such as customer interaction rate and initial conversion rate. After the initial outreach begins, the system continuously monitors these indicators and dynamically adjusts the initial exploration resource quota based on their feedback to achieve optimal resource allocation.

[0058] This application, when allocating specific resources, first comprehensively considers the business type, risk level, and current market attention to the business target's topic, and combines this with pre-defined resource allocation rules to calculate the initial exploration resource quota. This ensures that initial resource allocation is no longer a simple fixed ratio, but rather a preliminary intelligent decision based on multi-dimensional information. This solves the problem of resource allocation being too rigid and failing to fully consider the characteristics of the business target. Furthermore, after the initial outreach is implemented, by continuously monitoring the market adaptability score and initial outreach effectiveness of the business target, and dynamically adjusting the initial exploration resource quota based on this real-time feedback, this application's solution ensures the flexibility and adaptability of resource allocation. Thus, resource allocation can be optimized according to the actual performance of the business target in the market and changes in the external environment, avoiding resource waste and improving the efficiency and success rate of initial exploration.

[0059] Through the aforementioned technical solution, this application enables more refined and intelligent management of initial exploration resources. Firstly, by incorporating factors such as business type, risk level, and market attention during initial allocation, resource allocation becomes more targeted and rational, avoiding blind exploration. Secondly, by introducing market adaptability scores and initial reach results as the basis for subsequent adjustments, resource allocation gains dynamic feedback and self-optimization capabilities. It can adjust resource investment promptly based on actual market performance and customer feedback, thereby significantly improving resource utilization efficiency, reducing exploration costs, and effectively increasing the initial reach success rate and market conversion potential of new or low-frequency businesses.

[0060] In some of the above implementation methods, the step of calculating an initial exploration resource quota based on business type, risk level, and market attention, combined with preset resource allocation rules, includes: Receive information on the business type, risk level, and current market attention to the business object's topic; Perform consistency checks on business type information, cross-validate risk level information, and conduct real-time assessments of market attention information; Based on the business type after consistency check, the risk level after cross-validation, and the real-time market attention assessment results, combined with the preset resource allocation rules, the initial exploration resource quota is calculated.

[0061] Specifically, upon receiving information about the business type, risk level, and current market attention to the business object's topic, the system first preprocesses this raw data. This includes: a consistency check on business type information to ensure it conforms to the system's preset standard classification or format, avoiding information discrepancies caused by different data sources or input errors. For example, different business systems may have different names or codes for the same type of business; the consistency check aims to unify them into a standard format. Cross-validation of risk level information involves comparing multiple data sources or historical records to verify the accuracy and timeliness of the current risk level, ensuring it reflects the latest risk status of the business object. For example, the assessment time of the risk level can be checked, or it can be compared with historical risk trends. Real-time assessment of market attention information determines whether the acquired market attention data is recent enough to reflect current market dynamics and hot topics. For example, the data collection time can be assessed, or it can be compared with real-time market indicators. Through these rigorous data verifications and assessments, it is ensured that the input parameters used to calculate the initial exploration resource quota are high-quality, accurate, and real-time. Ultimately, the system will accurately calculate the initial exploration resource quota based on the business type after consistency checks, the risk level after cross-validation, and the market attention assessment results after real-time evaluation, combined with the preset resource allocation rules.

[0062] This application effectively addresses the limitations of the basic solution, which may suffer from inaccurate resource quota calculations due to data quality issues, by introducing consistency checks on business type information, cross-validation of risk level information, and real-time assessment of market attention information. Specifically, consistency checks ensure the standardization and uniformity of business type data, avoiding misjudgments caused by inconsistent data formats or definitions; cross-validation improves the accuracy and reliability of risk level assessment through multi-dimensional and multi-time-sensitive data comparison, enabling it to more realistically reflect the risk status of business entities; and real-time assessment ensures that market attention data closely tracks market trends and captures the latest market hotspots. It is precisely because of these data preprocessing and verification mechanisms that subsequent resource allocation rules can be calculated based on more reliable, accurate, and real-time input parameters, thereby significantly improving the accuracy of initial resource quota exploration.

[0063] Through the aforementioned technical solution, this application significantly improves the accuracy and reliability of initial exploration resource quota calculation. Compared to directly using raw data for calculation, this application effectively avoids resource misallocation problems caused by inconsistent, untimely, or inaccurate data by rigorously verifying and evaluating information on business type, risk level, and market attention. As a result, the system can allocate more reasonable and precise specific resources to emerging business targets, ensuring that initial outreach activities can more effectively reach the target group, improve the quality of customer interaction and the effectiveness of learning signals, ultimately optimizing the overall reinforcement learning process and accelerating the exploration and promotion of new businesses.

[0064] In some of the above embodiments, the steps for performing consistency checks on business type information include: When the system receives business type information, it obtains the source system identifier of the business type information; Based on the source system identifier, query the standard definition of the business type used by the source system from the preset business type mapping table; Compare the business type information with the standard business types defined in the mapping table; If the business type information is inconsistent with the standard business type, the business type information will be converted to the standard business type according to the preset conversion rules in the mapping table. If the business type information cannot be converted using the conversion rules, the business type information will be marked as pending manual review, and the manual review process will be initiated. After the manual review process is completed, the review results will be updated in the business type mapping table.

[0065] Business type information refers to structured data describing the category to which a business object belongs, such as financial product type or service type. The source system identifier uniquely identifies the specific system providing this business type information, such as different internal bank systems or external partner platforms. The business type mapping table is a pre-built database or configuration table that stores the correspondence between business types used by different source systems and the standard business types within this system, along with their conversion rules. Standard definitions refer to the business type classification standards uniformly recognized within this system or the industry. Conversion rules can be a series of predefined logical rules, lookup tables, or machine learning models used to automatically map non-standard business types to standard business types. The manual review process aims to handle business type information that cannot be automatically converted or is ambiguous, ensuring data accuracy and consistency. Updating the review results aims to continuously optimize and improve the business type mapping table, increasing the accuracy of subsequent automatic processing.

[0066] This application first obtains the source system identifier of the business type information and then queries a pre-defined business type mapping table based on this identifier to obtain the standard definition of the business type used by that source system. Subsequently, the received business type information is compared with the queried standard business type to identify any inconsistencies. If inconsistencies exist, an automatic conversion attempt is made using pre-defined conversion rules in the mapping table to standardize the business type information. If automatic conversion fails, the information is marked for manual review, and a manual review process is initiated to ensure that all business type information is ultimately accurately identified and standardized. After manual review, the review results are updated in the mapping table, thereby continuously optimizing and improving the accuracy of the mapping table and enhancing the system's ability to process business type information. It is precisely this multi-layered verification and conversion mechanism that enables the system to process business type data from different sources and with varying formats, unifying it into a standard format and providing a reliable data foundation for subsequent resource allocation and strategy adjustments.

[0067] The above technical solution effectively addresses the data chaos and processing difficulties caused by inconsistent definitions of business types across different systems or data sources. By introducing source system identifiers, business type mapping tables, and a mechanism combining automatic conversion and manual review, the accuracy, consistency, and standardization of business type information are ensured. This not only improves the automation and efficiency of data processing and reduces the error rate of manual intervention, but also provides high-quality input data for subsequent initial resource quota calculations, thereby enhancing the accuracy and reliability of the entire reinforcement learning-based voice outbound call tag retrieval method and avoiding resource mismatch or strategy deviations caused by inaccurate business type information.

[0068] In some of the above embodiments, the step of cross-validating the risk level information includes: When the system receives risk level information of a business object, the assessment time for obtaining the risk level information is as follows: Based on the assessment time, determine whether the risk level information is within the preset valid time limit; If the risk level information exceeds the validity period, the risk level update process will be initiated to obtain the latest risk level information of the business object from the real-time risk assessment module. Retrieve historical risk level change records for business objects; Analyze the trend of risk level changes for business targets based on historical risk level change records; Obtain market volatility indicators in the current financial market environment; Based on market volatility indicators, assess the potential impact of the current market environment on the risk level of business targets; The latest risk level information is dynamically verified by combining the trend of risk level changes and potential impacts; If the dynamic verification results show that the latest risk level information differs significantly from the risk level change trend or potential impact, the latest risk level information will be marked as requiring manual review.

[0069] Specifically, when the system receives risk level information for a business object, it first obtains the assessment time of that risk level information. This assessment time can be a precise timestamp, used to record the exact moment the risk level was calculated or last updated. Subsequently, the system compares the obtained assessment time with a preset valid time limit to determine whether the risk level information still has reference value. The preset valid time limit can be flexibly configured based on factors such as the characteristics of the business object, the frequency of market fluctuations, and the risk assessment update cycle. If the judgment result indicates that the risk level information has exceeded the preset valid time limit, the system will automatically initiate a risk level update process, interacting with the real-time risk assessment module to obtain the latest risk level information for the business object, ensuring that subsequent decisions are based on the latest data.

[0070] To conduct deeper cross-validation of risk level information, the system retrieves historical risk level change records for business entities. These historical records may include the business entity's risk level at different points in time, the assessment criteria, and any events that led to changes in the risk level. Based on these historical records, the system analyzes the risk level change trends of the business entity, for example, identifying whether the risk is continuously rising, falling, remaining stable, or exhibiting cyclical fluctuations. Simultaneously, the system also retrieves market volatility indicators from the current financial market environment, such as macroeconomic indices, industry sentiment indices, and volatility of specific asset classes. These indicators reflect the potential impact of the external environment on the risk level of the business entity.

[0071] Based on this, the system will dynamically verify the latest risk level information by combining the analyzed risk level change trends and the assessed potential market impact. This verification process aims to identify any inconsistencies or anomalies between the latest risk level information and historical trends or the current market environment. For example, if historical trends show a continuous decline in risk, but the latest risk level has increased significantly without any obvious negative market factors, there may be a data anomaly. If the dynamic verification results show a significant difference between the latest risk level information and the risk level change trends or potential impact, the system will not directly adopt the risk level but will mark it as requiring manual review. This manual review mechanism aims to introduce expert judgment to provide final confirmation for complex or abnormal situations, thereby avoiding potential misjudgments by automated systems.

[0072] This application's solution constructs a multi-dimensional cross-validation mechanism by introducing the timeliness assessment of risk level information, historical trend analysis, and dynamic market environment evaluation. First, by checking the assessment time of risk level information, the real-time nature of the data used is ensured, avoiding the risk of making decisions based on outdated information. Second, by analyzing the historical risk level changes of business objects, the system can understand the evolution of risk, providing an important basis for judging the rationality of the current risk level. Simultaneously, by combining market volatility indicators in the current financial market environment, the potential impact of external factors on the risk of business objects can be assessed, making risk assessment more comprehensive and dynamic. Finally, by combining these internal and external factors to dynamically verify the latest risk level information, abnormal or unreasonable risk assessment results can be promptly identified, and final confirmation is achieved through a manual review mechanism, thereby ensuring the accuracy and reliability of the risk level information and providing a solid data foundation for subsequent initial exploration of resource quota calculations.

[0073] Through the above technical solution, this application effectively addresses the issues of insufficient timeliness, lack of historical context, and lack of external environmental considerations in cross-validation of risk level information. Specifically, by introducing an assessment time judgment and real-time update mechanism, the up-to-dateness of risk level information is ensured; by analyzing historical trends, the understanding of risk evolution patterns is enhanced; and by combining market volatility indicators, risk assessment can dynamically adapt to changes in the external environment. As a result, the obtained risk level information is more accurate, comprehensive, and reliable, significantly improving the accuracy of initial resource quota calculations, thereby optimizing the efficiency of initial resource allocation and the overall success rate of business exploration, and reducing potential losses caused by inaccurate risk assessment.

[0074] Based on the same inventive concept, this application also discloses a voice outbound call tag retrieval system based on reinforcement learning, such as... Figure 2 As shown, the system includes: Identification module 1 is used to identify business objects that need to be explored initially, which have a low frequency of occurrence or have not appeared in the system's historical decision-making experience. Resource allocation module 2 is used to allocate a specific proportion of resources to a business object. The specific resources are not affected by the system's prediction of the success rate of the business object. The outreach execution module 3 is used to perform initial outreach to the target group of the business object using specific resources, and to obtain customer interaction information generated by the initial outreach. Signal enhancement module 4 is used to enhance the intensity of the learning signal generated by customer interaction information; Strategy adjustment module 5 is used to adjust the recommendation strategy for business objects based on the learning signal after intensity enhancement processing; Module 6, the effectiveness evaluation module, is used to evaluate the initial reach effect on business targets. Resource adjustment module 7 is used to adjust the allocation ratio of specific resources based on the initial reach effect.

[0075] The system proposed in this application aims to solve the problems of slow response speed, low data utilization efficiency and poor adaptability of traditional tag retrieval technology when facing complex and ever-changing market environments and customer behavior patterns, thereby improving the accuracy and efficiency of voice outbound calls.

[0076] This application's system, through its modular design, enables effective exploration and strategy optimization for emerging or low-frequency business targets. The identification module proactively discovers new businesses with low frequency or no presence in the system's historical decision-making experience, breaking away from the traditional system's conservative approach to niche businesses. The resource allocation module allocates specific resources to these business targets, unaffected by predicted success rates, providing valuable trial-and-error opportunities for new products and services. The outreach execution module utilizes these resources to conduct initial outreach to the target audience and collect customer interaction information. The signal enhancement module strengthens the generated learning signals, enabling the system to quickly capture valuable feedback even with limited data, accelerating the identification of new business potential. The strategy adjustment module adjusts the recommended strategy based on the enhanced learning signals, ensuring the accuracy and effectiveness of marketing campaigns. The performance evaluation module comprehensively assesses the initial outreach results, objectively judging the market performance of the business targets. Finally, the resource adjustment module dynamically adjusts the allocation ratio of specific resources based on the evaluation results, achieving optimal resource configuration.

[0077] The above are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A voice outbound call tag retrieval method based on reinforcement learning, characterized in that, include: Identify business objects that require initial exploration, which have a low frequency of occurrence or have not appeared in the system's historical decision-making experience; Allocate a specific proportion of resources to a business object, and these specific resources are not affected by the system's prediction of the success rate of the business object. Utilize specific resources to conduct initial outreach to the target group of the business entity and obtain customer interaction information generated from the initial outreach; Intensity enhancement processing is applied to the learning signals generated from customer interaction information; Adjust the recommendation strategy for business objects based on the learning signal after intensity enhancement processing; Evaluate the effectiveness of initial outreach to business targets; Adjust the allocation ratio of specific resources based on the initial reach effect.

2. The voice outbound call tag retrieval method based on reinforcement learning according to claim 1, characterized in that, The step of enhancing the intensity of the learning signal generated from customer interaction information includes: When the system obtains initial positive interaction information from customers through initial outreach, it generates an initial positive feedback record containing customer identification, task identification, original reward value, and feedback timestamp. Generate a digital tracking code for this initial outreach and embed the digital tracking code into the digital profile link that will be sent to the customer; Based on customer preferences, send customers links to digital profiles containing embedded digital tracking codes via digital channels; Mark the customer's association status with this initial contact as pending confirmation and start a confirmation waiting timer; Continuously monitor the bank's digital channels to track whether customers click on links with digital tracking codes, visit specific product pages, download related documents, or actively search for related keywords within a preset time period; If a customer is detected to have engaged in the aforementioned proactive behavior within a preset time period, the association status between the customer and the initial contact will be updated to "intent confirmed". Only when a customer's intention is confirmed as confirmed, the original reward value in the initial positive feedback record is enhanced to obtain an enhanced learning signal. If the waiting timer times out and no proactive confirmation is detected, the original reward value in the initial positive feedback record will not be enhanced.

3. The voice outbound call tag retrieval method based on reinforcement learning according to claim 1, characterized in that, The steps for evaluating the initial outreach effectiveness to the target audience include: Receive product descriptions and target customer tags from business objects, and identify the themes and investment directions of business objects through semantic analysis to obtain the semantic features of business objects; Continuously monitor the external financial market environment, obtain market hot topics, and market sentiment and frequency of mention related to the topics of business targets; Based on market sentiment and frequency of mention, the weight of customer interest and initial conversion rate in the overall evaluation is adjusted in real time. The market adaptability score is calculated based on the degree to which the semantic features of the business object match the hot topics in the market, as well as the degree of match between the initial reach and market sentiment. The initial reach effect on business targets is judged based on the adjusted evaluation indicator weights and market adaptability scores.

4. The voice outbound call tag retrieval method based on reinforcement learning according to claim 2, characterized in that, The steps of continuously monitoring the bank's digital channels to track whether customers click on links with digital tracking codes, visit specific product pages, download related documents, or actively search for related keywords within a preset time period include: When the system detects that a customer actively searches for keywords related to business objects through the bank's digital channels, it performs text processing on the search keywords and extracts core words; Based on a pre-defined financial intent dictionary and rule set, core words are matched and categorized to distinguish intents such as proactive understanding, risk assessment, or competitor comparison. Search behavior that actively seeks to understand intent will be included in the criteria for intent confirmation; Searches with other intentions will not be included in the criteria for confirming intent.

5. The voice outbound call tag retrieval method based on reinforcement learning according to claim 2, characterized in that, The steps of continuously monitoring the bank's digital channels to track whether customers click on links with digital tracking codes, access specific product pages, or download related documents within a preset time period include: When the system detects that a customer clicks on a link with a digital tracking code, visits a specific product page, or downloads a related document, it collects the customer's interaction data on digital channels. The interaction data includes page dwell time, scroll depth, number of clicks on interactive elements, and content reading progress. Based on preset engagement thresholds, it is determined whether the customer has engaged in in-depth browsing. Engagement thresholds include page dwell time threshold, scroll depth threshold, and number of clicks on interactive elements threshold. When a customer's interaction data meets the engagement threshold, the behavior will be included in the basis for confirming intent. If a customer's interaction data does not reach the engagement threshold, the behavior will not be included in the basis for confirming intent.

6. The voice outbound call tag retrieval method based on reinforcement learning according to claim 1, characterized in that, The step of allocating a specific proportion of resources to a business object, wherein the specific resources are not affected by the system's prediction of the success rate of the business object, includes: Obtain information about the business type, risk level, and current market attention to the topic of the business object; Based on the business type, risk level, and market attention, and in conjunction with the preset resource allocation rules, an initial exploration resource quota is calculated. The initial exploration resource quota is not affected by the system's prediction of the success rate of business objects during the initial exploration phase; After the initial outreach, continuously monitor the market adaptability score of the target audience and the effectiveness of the initial outreach; Adjust the initial exploration resource quota based on market adaptability scores and initial reach effectiveness.

7. The voice outbound call tag retrieval method based on reinforcement learning according to claim 6, characterized in that, The step of calculating an initial exploration resource quota based on business type, risk level, and market attention, combined with preset resource allocation rules, includes: Receive information on the business type, risk level, and current market attention to the business object's topic; Perform consistency checks on business type information, cross-validate risk level information, and conduct real-time assessments of market attention information; Based on the business type after consistency check, the risk level after cross-validation, and the real-time market attention assessment results, combined with the preset resource allocation rules, the initial exploration resource quota is calculated.

8. The voice outbound call tag retrieval method based on reinforcement learning according to claim 7, characterized in that, The steps for performing consistency checks on business type information include: When the system receives business type information, it obtains the source system identifier of the business type information; Based on the source system identifier, query the standard definition of the business type used by the source system from the preset business type mapping table; Compare the business type information with the standard business types defined in the mapping table; If the business type information is inconsistent with the standard business type, the business type information will be converted to the standard business type according to the preset conversion rules in the mapping table. If the business type information cannot be converted using the conversion rules, the business type information will be marked as pending manual review, and the manual review process will be initiated. After the manual review process is completed, the review results will be updated in the business type mapping table.

9. The voice outbound call tag retrieval method based on reinforcement learning according to claim 7, characterized in that, The steps for cross-validating the risk level information include: When the system receives risk level information of a business object, the assessment time for obtaining the risk level information is as follows: Based on the assessment time, determine whether the risk level information is within the preset valid time limit; If the risk level information exceeds the validity period, the risk level update process will be initiated to obtain the latest risk level information of the business object from the real-time risk assessment module. Retrieve historical risk level change records for business objects; Analyze the trend of risk level changes for business targets based on historical risk level change records; Obtain market volatility indicators in the current financial market environment; Based on market volatility indicators, assess the potential impact of the current market environment on the risk level of business targets; The latest risk level information is dynamically verified by combining the trend of risk level changes and potential impacts; If the dynamic verification results show that the latest risk level information differs significantly from the risk level change trend or potential impact, the latest risk level information will be marked as requiring manual review.

10. A voice outbound call tag retrieval system based on reinforcement learning, characterized in that, The system includes: The identification module is used to identify business objects that need to be explored initially, which are those that have a low frequency of occurrence or have not appeared in the system's historical decision-making experience. The resource allocation module is used to allocate a specific proportion of resources to a business object. These specific resources are not affected by the system's prediction of the success rate of the business object. The outreach execution module is used to perform initial outreach to the target group of the business object using specific resources, and to obtain customer interaction information generated by the initial outreach. The signal enhancement module is used to enhance the intensity of the learning signal generated by customer interaction information. The strategy adjustment module is used to adjust the recommendation strategy for business objects based on the learning signal after intensity enhancement processing. The effectiveness evaluation module is used to evaluate the initial reach of business targets; The resource adjustment module is used to adjust the allocation ratio of specific resources based on the initial reach effect.

Images