A user line mining method and device, computer equipment and storage medium

Abstract

The embodiment of the application belongs to the field of artificial intelligence and big data, and relates to a user clue mining method, which comprises the following steps: obtaining user call text, performing data cleaning on the user call text, inputting the user call text into a preset label processing large model for text keyword extraction processing to obtain first user data labels, performing data recall processing on potential user call text through the first user data labels to obtain weak label data, taking the weak label data as clue training data set, training a general model to obtain a clue model, inputting the potential user call text into the clue model for clue search processing to generate second user data labels, uploading the first user data labels and the second user data labels to a label information library, performing classification processing on transaction keywords and clue keywords, and obtaining user clue information. The application establishes a closed-loop system for clue mining, and can improve the clue mining efficiency and the amount of clues mined from inventory customers.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and big data technology, and in particular to user clue mining methods, devices, computer equipment and storage media. Background Technology

[0002] In the field of insurance customer service, how to encourage customers to repurchase new insurance products, that is, to discover potential new customers from existing customers, is an important research direction. Compared with externally discovering potential new users, the cost of discovering leads from existing customers is lower and the transaction success rate is higher.

[0003] The traditional lead mining process mainly involves collecting keywords from business stakeholders, retrieving data based on those keywords, manually labeling the data, training a lead model, and then using that model to mine leads.

[0004] However, this approach is limited by the fact that the labeled data relies entirely on manual labor, making it impossible to directly extract clues from the call text, and it cannot proactively discover new clues or keywords. Clue mining is inefficient and cannot handle large amounts of historical call data. Summary of the Invention

[0005] The purpose of this application is to provide a user clue mining method, apparatus, computer equipment, and storage medium to solve the technical problems of traditional clue mining processes being unable to actively discover new clues, having low efficiency in clue mining, and being unable to handle large amounts of historical data.

[0006] To address the aforementioned technical problems, this application provides a user clue mining method, which employs the following technical solution: Obtain user call texts, including call texts of potential users and call texts of completed users; The text of the completed user's call is cleaned to obtain the target text data. The target text data is input into a preset tag processing model for text keyword extraction to obtain a first user data tag, which includes several transaction keywords. Weakly labeled data is obtained by retrieving the call text of the potential user using the first user data label. The weakly labeled data is used as the cue training dataset to train the general model, thus obtaining the cue model. The potential user's call text is input into the clue model for clue search processing to generate a second user data tag, which includes several clue keywords. The first user data tag and the second user data tag are uploaded to the tag information database, and the transaction keywords and the lead keywords are classified and processed to obtain user lead information.

[0007] Furthermore, the large-scale tag processing model includes an LLM model, a BCE embedding model, and a text clustering model. The step of inputting the target text data into the preset large-scale tag processing model for text keyword extraction to obtain the first user data tag specifically includes: The target text keywords are obtained by executing the prompt command on the target text data using the LLM model. The target text keywords are vectorized using the BCE embedding model to obtain the target text vector. The target text vector is clustered using the text clustering model to obtain cluster categories, each containing several transaction keywords. The transaction keywords are categorized and interpreted using the LLM model to generate the first user data tag.

[0008] Furthermore, the step of performing data retrieval processing on the potential user's call text using the first user data tag to obtain weakly labeled data specifically includes: Randomly filter all transaction keywords within the first user data tag to obtain the target transaction keywords; Based on the target transaction keywords, the context of the potential user's call text is located to obtain the text position information of the target transaction keywords in the potential user's call text; Based on the text location information, the context field of the potential user's call text is extracted to obtain text recall data; The text recall data is input into the LLM model for data annotation processing to obtain the weak label data.

[0009] Furthermore, the step of processing the text recall data through the LLM model to obtain the weakly labeled data specifically includes: The text recall data is processed by the text clustering model to segment intent and obtain category samples. The LLM model is used to perform label classification on the category samples to generate target label data. Calculate the text confidence score of the target label data; If the text confidence level is greater than or equal to a preset confidence threshold, the target label data is labeled to obtain the weak label data, and the target label data is saved to the label information database. If the confidence level of the text is less than the preset confidence level threshold, the target tag data is directly saved to the tag information database.

[0010] Furthermore, the step of using the weakly labeled data as a cue training dataset to train the general model and obtain the cue model specifically includes: Text feature extraction is performed on the clue training dataset to obtain the text label features of the clue training dataset; The model parameters of the clue model are estimated using the text label features of the clue training dataset; The model parameters of the clue model are imported into the general model, and the general model is trained using the clue training dataset to obtain the clue model.

[0011] Furthermore, the step of uploading the first user data tag and the second user data tag to the tag information database, and classifying the transaction keywords and the lead keywords to obtain user lead information, specifically includes: The target lead dataset is obtained by iterating through all transaction keywords in the first user data tag and all lead keywords in the second user data tag. Obtain the text confidence scores of all keywords within the target clue dataset, classify them based on the text confidence scores, and obtain the user clue information.

[0012] Furthermore, after the step of uploading the first user data tag and the second user data tag to the tag information database to obtain user clue information, the method further includes: The tag information database is subscribed to and connected to the monitoring and scheduling platform through a message queue. The monitoring and scheduling platform includes a monitoring sub-platform and a scheduling sub-platform. The monitoring and scheduling platform interacts with the clue model, the preset tag processing model and the tag information database through the message queue. The monitoring sub-platform performs data monitoring on the user clue information to obtain clue monitoring data, which includes the number of valid clues, the number of valid clues tagged, and the confidence level of the tags. Based on the aforementioned clue monitoring data, the tag information database is scheduled using the routing rules of the scheduling sub-platform to obtain clue scheduling data.

[0013] To address the aforementioned technical problems, this application also provides a user clue mining device, which employs the following technical solution: The acquisition module is used to acquire user call texts, which include call texts of potential users and call texts of completed transactions. The data cleaning module is used to clean the call text of the transacting users to obtain target text data. The processing module is used to input the target text data into a preset tag processing model for text keyword extraction processing to obtain a first user data tag, which includes several transaction keywords. The data retrieval module is used to perform data retrieval processing on the potential user's call text using the first user data tag to obtain weak tag data; The training module is used to train the general model using the weakly labeled data as the cue training dataset to obtain a pre-trained cue model. The input module is used to input the potential user's call text into the clue model for clue search processing and generate a second user data tag, which includes several clue keywords. The classification module is used to upload the first user data tag and the second user data tag to the tag information database, and to classify the transaction keywords and the lead keywords to obtain user lead information.

[0014] To address the aforementioned technical problems, this application also provides a computer device that employs the following technical solution: A computer device includes a memory and a processor, the memory storing computer-readable instructions, the processor executing the computer-readable instructions to implement the steps of the user clue mining method.

[0015] To address the aforementioned technical problems, this application also provides a computer-readable storage medium, employing the technical solution described below: A computer-readable storage medium, characterized in that the computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the user clue mining method.

[0016] Compared with the prior art, the embodiments of this application have the following main advantages: This application embodiment obtains user call text, preprocesses the call text of converted users, and inputs it into a preset tag processing model for text keyword extraction to obtain a first user data tag. The first user data tag is then used to retrieve potential user call texts to obtain weakly labeled data. This weakly labeled data is used as a lead training dataset to train a general model, resulting in a lead model. Potential user call texts are then input into the lead model for lead search processing to generate a second user data tag. The first and second user data tags are uploaded to a tag information database. Converted keywords and lead keywords are then classified to obtain user lead information. This completes a closed loop of the lead mining process, fully utilizing the capabilities of the large model to maximize its advantages in each sub-task of lead mining, effectively reducing labeling costs, improving lead mining efficiency, and proactively discovering new leads or keywords, reducing lead omissions, and increasing the number of leads mined from existing customers. Attached Figure Description

[0017] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an exemplary system architecture diagram to which this application can be applied; Figure 2 A flowchart of an embodiment of the user clue mining method according to this application; Figure 3 This is a schematic diagram of a structure of one embodiment of the user clue mining device according to this application; Figure 4 This is a schematic diagram of the structure of one embodiment of the computer device according to this application; Figure 5 A flowchart of an embodiment of the user clue mining method according to this application. Detailed Implementation

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0020] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0022] like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, and 103, a network 104, and a server 105. Network 104 serves as the medium for providing communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables.

[0023] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0024] Terminal devices 101, 102, and 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 players (Moving Picture Experts Group Audio Layer IV), laptops, and desktop computers.

[0025] Server 105 can be a server that provides various services, such as a backend server that supports the pages displayed on terminal devices 101, 102, and 103.

[0026] It should be noted that the user clue mining method provided in this application embodiment is generally executed by a server / terminal device, and correspondingly, the user clue mining device is generally set in the server / terminal device.

[0027] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0028] Continue to refer to Figure 2 A flowchart illustrating an embodiment of a user clue mining method according to this application is shown. The user clue mining method includes the following steps: Step S201: Obtain the user's call text.

[0029] In this embodiment, the user clue mining method described above can be deployed in a user clue mining platform. This platform can be constructed using the aforementioned server or server cluster. The server or server cluster can be any electronic device with functions such as text recognition, text processing, data transmission, and data storage. The user clue mining method runs on this electronic device (e.g., Figure 1 The server / terminal device shown can receive the user's call text via wired or wireless connection. It should be noted that the wireless connection methods may include, but are not limited to, 3G / 4G / 5G connections, WiFi connections, Bluetooth connections, WiMAX connections, Zigbee connections, UWB (ultrawideband) connections, and other currently known or future wireless connection methods.

[0030] In this embodiment, the aforementioned user call text includes call text from potential users and call text from successful transactions.

[0031] Specifically, the aforementioned transaction texts of customers who have already purchased the company's products or services may include call records of customers who have already purchased the company's products or services. These transaction texts may contain text information such as customer satisfaction, product feedback, and repurchase intentions. The aforementioned potential user texts may include call records of customers who have not yet purchased but have shown interest. These potential user texts may contain clues about customer needs, questions, or purchase intentions.

[0032] In one possible embodiment, taking the promotion of non-auto insurance products for existing auto insurance customers as an example, the aforementioned transaction user call text can be the call text of some existing auto insurance customers who have transacted non-auto insurance, and the aforementioned potential user call text can be the historical call text of all auto insurance customers (who have not transacted non-auto insurance). The call text can be automatically recorded by the customer relationship management (CRM) system and call center technology integrated into the insurance company platform, and the voice can be converted into text information (i.e., the aforementioned user call text).

[0033] Step S202: Clean the text of the completed user's call to obtain the target text data. In this embodiment, the above-mentioned data cleaning process can be text processing such as standardization of format, removal of stop words, and stemming.

[0034] By cleaning the text of transactions from users, errors and inconsistencies, such as duplicate records, spelling mistakes, or formatting errors, can be removed or corrected. This helps improve the accuracy of the transaction texts, better supports the decision-making process, and reduces problems caused by erroneous data.

[0035] Step S203: Input the target text data into the preset tag processing model for text keyword extraction processing, and obtain the first user data tag.

[0036] In this embodiment, the aforementioned preset tag processing model may include an LLM model, a BCE embedding model, and a text clustering model.

[0037] Specifically, the aforementioned LLM model can be a large-scale pre-trained language model, including large models such as OpenAI's GPT series, Google's BERT, or T5. This LLM model can learn the wide range of uses of language by processing large amounts of text data, enabling tasks such as text generation, translation, summarization, and keyword extraction. The aforementioned BCE embedding model can be a text processing model trained on a transformer structure to generate numerical representations of text data, allowing machine learning models to process text data. The aforementioned text clustering model can be a data processing model trained on a neural network clustering algorithm. The aforementioned clustering methods can include K-means and DBSCAN clustering algorithms. This text clustering model can cluster text vectors into different categories based on text similarity. The aforementioned preprocessing can be a series of techniques for cleaning and preparing data, including data cleaning processes such as format standardization, stop word removal, and stemming.

[0038] In this embodiment, within the aforementioned preset tag processing model, the LLM model, BCE embedding model, and text clustering model can be connected via message queues to achieve data exchange processing. The message queues can be middleware such as RocketMQ, Redis, and IBM MQ.

[0039] In one possible embodiment, such as Figure 5 As shown, the preprocessed transaction user call text is stored on the server. First, keywords are extracted using an LLM model. The extracted keywords are then used as input data for a BCE embedding model. The BCE embedding model vectorizes the keywords to obtain keyword vector data. This keyword vector data is then used as input data for a text clustering model. The text clustering model performs clustering to obtain text categories and clustered keywords (the aforementioned transaction keywords). Finally, the text categories are input back into the LLM model for category interpretation to obtain the first user data label. The first user data label is then stored in a label information database.

[0040] Specifically, the aforementioned first user data tag can be a tag obtained based on the call text of a transacting user through a series of processing and model analysis. The aforementioned transaction keywords can be keywords or phrases extracted from the call text of a transacting user. The aforementioned first user data tag contains multiple of the aforementioned transaction keywords. For example, taking non-auto insurance products that promote existing auto insurance customers as an example, the aforementioned first user data tag is the tag of the target customer of the non-auto product, and the aforementioned transaction keywords are the keywords under the first user data tag. Taking student accident insurance as an example, the target customer is a customer with school-age children, the first user data tag is student, child, and the transaction keywords are respectively child / kid / child, after school / primary school / middle school / school, etc.

[0041] Specifically, the implementation process of preprocessing the transaction text of users and then inputting it into a preset tag processing model for text keyword extraction to obtain the first user data tag will be described in more detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0042] Step S204: Data retrieval processing is performed on the potential user call text using the first user data tag to obtain weak tag data.

[0043] In this embodiment, the aforementioned weakly labeled data can be a reliable labeled data used to quickly generate and expand the training dataset. The aforementioned training dataset can be used to train the machine model, and can also be used for actual predictions in the iterative optimization of the machine model. Through feedback loops, the model is gradually corrected and optimized to improve its accuracy and robustness.

[0044] Specifically, the detailed implementation process of retrieving weakly labeled data from potential user call texts using the first user data tag will be described in further detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0045] Step S205: Use the weakly labeled data as the cue training dataset to train the general model and obtain the cue model.

[0046] In this embodiment, the general model described above can be a basic deep learning model with broad application capabilities. It can be pre-trained on large-scale datasets to learn a wide range of data representations and patterns. It is not optimized for specific domains or tasks, but maintains sufficient generality to adapt to a variety of different data and tasks. The general model can be trained in a specific way to make it a deep learning model that can efficiently achieve specific tasks in specific domains.

[0047] The aforementioned clue model can be used to mine user tags that match the product's target customer group, enabling the mining of clues from large volumes of potential user call texts.

[0048] Specifically, the implementation process of using weakly labeled data as the training dataset for clues to train the general model and obtain the clue model will be described in more detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0049] Step S206: Input the potential user's call text into the clue model for clue search processing to generate a second user data label.

[0050] In this embodiment, the second user data tag can be a user tag with a target customer group for the product. The second user data tag includes multiple clue keywords, which can be elements constituting the second user tag. The clue keywords can reflect the specific needs, problems, or interests of potential customers.

[0051] In one possible embodiment, taking non-auto insurance products that promote existing auto insurance customers as an example, the aforementioned clue keywords may include demand clue keywords such as "insurance coverage", "long-term investment returns", and "emergency service needs", emotional clue keywords such as "worry", "dissatisfaction", and "excitement", and behavioral characteristic clue keywords such as "asking for details" and "comparing prices".

[0052] Specifically, the above-mentioned clue search processing can be based on the identification of tag categories from the aforementioned potential user call text.

[0053] In one possible embodiment, taking the promotion of non-auto insurance products for existing auto insurance customers as an example, the clue model can also incorporate a self-attention mechanism. The input layer of the clue model extracts initial feature vectors from the sequence information of potential user call texts and standardizes these initial feature vectors to obtain standard feature vectors. These standard feature vectors are then input into the hidden layer of the clue model. The hidden layer enhances the sensitivity to user needs, user emotions, and user behavior characteristics, thereby mapping them into the corresponding latent space. Within this latent space, the clue model focuses on extracting clue keywords related to the potential needs of existing auto insurance customers for non-auto insurance products from the standard feature vectors, obtaining a clue keyword matrix. This clue keyword matrix can serve as a second user data label. Finally, the clue model outputs the second user data label of the potential user call text through the output layer.

[0054] Step S207: Upload the first user data tag and the second user data tag to the tag information database, classify and process the transaction keywords and lead keywords, and obtain user lead information.

[0055] In this embodiment, the aforementioned user lead information may be a new message dataset obtained by performing a union processing on the aforementioned transaction keywords and the aforementioned lead keywords. The aforementioned new message dataset includes all data information within the aforementioned first user data tag and the aforementioned second user data tag.

[0056] Specifically, the detailed implementation process of uploading the first user data tag and the second user data tag to the tag information database, classifying and processing transaction keywords and lead keywords to obtain user lead information will be further described in subsequent specific embodiments of this application, and will not be elaborated on here.

[0057] This application achieves a closed loop in the lead mining process by acquiring user call texts, preprocessing the call texts of successful customers, inputting them into a pre-defined tagging model for text keyword extraction, obtaining first user data tags, using the first user data tags to retrieve potential user call texts, obtaining weakly labeled data, using the weakly labeled data as a lead training dataset to train a general model, obtaining a lead model, inputting potential user call texts into the lead model for lead search processing, generating second user data tags, uploading the first and second user data tags to a tag information database, classifying and processing transaction keywords and lead keywords to obtain user lead information, realizing a complete closed loop in the lead mining process, fully utilizing the capabilities of the large model to maximize its advantages in each sub-task of lead mining, effectively reducing annotation costs, improving lead mining efficiency, and proactively discovering new leads or keywords, reducing lead omissions, and increasing the number of leads mined from existing customers.

[0058] In some alternative implementations, step S203 includes the following steps: The target text keywords are obtained by executing the prompt command on the target text data using an LLM model. The target text keywords are vectorized by embedding using the BCE embedding model to obtain the target text vector; The target text vector is clustered using a text clustering model to obtain cluster categories, each containing multiple transaction keywords. The LLM model is used to perform category interpretation processing on transaction keywords to generate the first user data tags.

[0059] In this embodiment, the prompt instruction mentioned above can refer to an input text built into the LLM model. The purpose of the input text is to guide or instruct the model to make a specific answer or output. The target text keyword can be the transaction keyword mentioned above.

[0060] Specifically, by proposing a prompt text that "lists all product features and user-expressed needs mentioned in the target text data," the LLM model can be guided to focus on the most important information points in the text and extract the keywords from the transaction text of the user who made the purchase that are most helpful in understanding the user's intent and the transaction behavior.

[0061] In this embodiment, the above-mentioned embedding vectorization process can be to use pre-trained word embedding models such as BERT, Word2Vec, or GloVe to convert target text keywords into numerical vectors. These numerical vectors can capture the semantic characteristics of words and enable the model to process text data based on mathematical methods.

[0062] In this embodiment, the above clustering process can be performed using clustering algorithms such as Faisal mean clustering, K-means, hierarchical clustering, and DBSCAN to process the target text vector, obtain keywords for each cluster category, and finally randomly select 20 keywords for each category. The above cluster categories may include categories such as "insurance needs", "product questions", and "price sensitivity". Each cluster category can reflect different user concerns based on the transaction keywords it contains.

[0063] In this embodiment, the aforementioned category interpretation processing can be achieved by defining a prompt task. This task aims to guide the LLM model to deeply interpret the meaning of each category. Specifically, taking the promotion of non-auto insurance products for existing auto insurance customers as an example, if the cluster category contains transaction keywords such as "emergency service needs" and "24-hour response," the prompt might be defined as "explaining the customer needs represented by this set of keywords." The prompt task can be automatically generated, and the prompt content can be adjusted according to the dynamic changes of the cluster category. Then, the cluster category and the corresponding prompt are combined to form the input data of the LLM model, and the LLM model outputs one or more user labels (the aforementioned first user data labels).

[0064] In one possible implementation, taking auto insurance as an example, the above-mentioned data cleaning processes, such as standardizing the format of the transaction user's call text, removing stop words, and stemming, are performed. An LLM model is then used to execute a prompt instruction to "list all product features mentioned in the text data and the user's expressed needs." The extracted transaction keywords are used as input data for the BCE embedding model. Pre-trained word embedding models such as BERT, Word2Vec, or GloVe built into the BCE embedding model are used to convert the transaction keywords into vector data. This vector data is then used as input data for a text clustering model. Faisal mean clustering is performed to obtain text categories such as "insurance needs," "product questions," and "price sensitivity." Finally, an automated prompt generation task is used, adjusting the prompt content based on the dynamic changes in text categories. The text categories and corresponding prompts are then combined to form the input data for the LLM model, which outputs one or more user tags.

[0065] This application combines automated data processing with various neural network models for analysis, enabling precise extraction and clustering of customer call keywords and generating detailed user tags, thereby improving lead generation efficiency.

[0066] In some alternative implementations, step S204 includes the following steps: Randomly filter all transaction keywords within the first user data tag to obtain the target transaction keywords; Contextual localization of potential user call text based on target sales keywords yields the text location information of target sales keywords in potential user call text. Text location information is used to extract context fields from the call text of potential users to obtain text recall data. Text recall data is input into an LLM model for data labeling to obtain weakly labeled data.

[0067] In this embodiment, the above-mentioned random selection can be a simple random sampling method for all transaction keywords within the first user data tag, or a weighted random selection method based on keyword frequency, relevance score, etc.

[0068] In this embodiment, the aforementioned contextual localization can involve identifying and locating the selected target conversion keywords within the potential user's call text. Specifically, this can be achieved using natural language processing techniques such as regular expression search and keyword matching algorithms, or more advanced text analysis tools, to ensure accurate location of the keywords.

[0069] In this embodiment, the above-mentioned context field extraction process can be based on the above-mentioned target transaction keyword, extracting it into 10 sentences as a fragment. The above-mentioned text location information can be the specific word position of the target transaction keyword in the potential user's call text or the sentence or paragraph number where the above-mentioned target transaction keyword is located.

[0070] In this embodiment, the aforementioned text recall data can be relevant text fragments extracted from potential user call texts based on contextual location information. Specifically, the detailed implementation process of inputting the text recall data into an LLM model for data annotation to obtain weakly labeled data will be further described in subsequent specific embodiments of this application, and will not be elaborated upon here.

[0071] In this embodiment, text location information is obtained through context positioning and interception, and data retrieval is performed using the text location information. Then, the retrieved data is labeled, which improves the automation and efficiency of data labeling and enables the rapid discovery of potential customers' interests and needs in large-scale data.

[0072] In some alternative implementations, the step of inputting text recall data into an LLM model for data annotation to obtain weakly labeled data also includes the following steps: The text recall data is processed by using a text clustering model to segment intents and obtain category samples. The LLM model is used to classify the category samples and generate target label data. Calculate the text confidence score of the target label data; If the text confidence level is greater than or equal to the preset confidence level threshold, the target label data is labeled to obtain weak label data, and the target label data is saved to the label information database. If the text confidence level is less than the preset confidence level threshold, the target label data will be directly saved to the label information database.

[0073] In this embodiment, the above intent segmentation process can be performed by first vectorizing the text recall data through the BCE embedding model, and then performing Faiss mean clustering through the above text clustering model to divide the text recall data of a single sentence into several category samples. The above category samples can be labeled. If a certain category sample is labeled as a positive sample, the sample of that category will be preferentially entered into the large model for label classification. The above positive sample can be a category sample with a clear intent.

[0074] In one possible embodiment, the clustering quality assessment indexes Davies-Bouldin Index (DBI) and Calinski-Harabasz Index (CHI) can be used to evaluate whether a cluster sample is a positive sample. The DBI can be used to assess the intra-cluster density and inter-cluster separation of cluster samples to reflect the quality of the cluster samples. It can determine the similarity of elements within each cluster sample and the distinguishability from other cluster samples to assess whether a cluster sample is a positive sample. The CHI can be used to measure the quality of the cluster samples by calculating the ratio of inter-cluster dispersion to intra-cluster dispersion.

[0075] Specifically, the average distance from all points within each category sample to the center of the category sample can be obtained, the distance between the centers of different categories can be calculated, the DBI formula can be applied to evaluate the relative quality of each category sample with the nearest category sample, and the average value can be taken to determine the overall quality of the category sample (the DBI value of the category sample). At the same time, the inter-class dispersion (the sum of the distances from the centers of different categories to the global center) and the intra-class dispersion (the sum of the distances from the points within each category sample to the center of the category sample) can be calculated. The CHI formula can be applied, and the CHI value of the category sample can be obtained by comparing the ratio of inter-class and intra-class dispersion. Finally, the DBI value and CHI value of each category sample can be collected and compared. When the DBI value of the category sample is less than the preset DBI threshold and the CHI value of the category sample is greater than the preset CHI threshold, the above category sample is marked as a positive sample.

[0076] In this embodiment, the aforementioned label classification process can be achieved by designing specific prompts to guide the LLM model in generating relevant user labels for each category sample. Specifically, taking the vehicle insurance field as an example, for the category of "car insurance product inquiry," the prompt might be "to indicate which type of car insurance product the user is interested in," while for the category of "service complaint," the prompt might be "to identify the specific issue in the user's complaint."

[0077] In this embodiment, the aforementioned text confidence score can be used to describe the degree of confidence of the LLM model in its prediction results. Specifically, the text confidence score can be a numerical measure of the LLM model's prediction, usually expressed as a probability value ranging from 0 to 1, where a higher value indicates a higher degree of confidence. The text confidence score of the target label data can be calculated by using the softmax layer built into the LLM model. Alternatively, the text confidence score information can be obtained by calculating the relative magnitude of the difference between the highest and second-highest values ​​of the logits output by the LLM model (the original output before the softmax layer) and the logits output by the softmax layer. The aforementioned logits can be the output of the last linear layer in the LLM model.

[0078] In one possible embodiment, the logits output by the LLM model can be used to process data based on the confidence level. If the confidence level of the target label data is greater than or equal to a preset confidence threshold, the data is considered reliable, the target label data is marked as weak label data and saved in the label information library. If the text confidence level is less than the preset confidence threshold, the target label data is directly saved in the label information library.

[0079] This application combines text clustering models and LLM models to achieve automated intent segmentation and label classification of text recall data. It can filter out target label data with high confidence as weak label data, reduce reliance on manual annotation, and lower costs.

[0080] In some alternative implementations, step S205 includes the following steps: Text feature extraction is performed on the clue training dataset to obtain the text label features of the clue training dataset; Estimate the model parameters of the cue model using the text label features of the cue training dataset; The model parameters of the cue model are imported into the general model, and the general model is trained using the cue training dataset to obtain the cue model.

[0081] In this application, the aforementioned text label features may include term frequency (TF), inverse document frequency (IDF), word embeddings (such as Word2Vec, GloVe, FastText), syntactic features (such as dependency syntax tree information), and sentiment analysis scores. Specifically, high-dimensional features in the cue training dataset can be extracted using pre-trained models such as NLTK, SpaCy, or more advanced BERT through natural language processing toolkits. The model parameters of the cue model can be estimated using the text label features of the cue training dataset through parameter estimation methods such as gradient descent, regularization, and cross-validation.

[0082] In one possible embodiment, the clue training dataset can be input into the clue model for processing to obtain the processing result corresponding to the clue training data. The error loss function between the processing result corresponding to the clue training data and the label data corresponding to the clue training data can be calculated. With minimizing the error loss function as the optimization objective, the parameters of the clue model can be adjusted through the backpropagation algorithm. The parameter adjustment process is iterated until the error loss function converges at the minimum or the number of iterations reaches a preset value, at which point the training stops, and the trained clue model is obtained.

[0083] In some alternative implementations, step S207 includes the following steps: The target lead dataset is obtained by iterating through all transaction keywords in the first user data tag and all lead keywords in the second user data tag. Obtain the text confidence scores of all keywords in the target clue dataset, classify them based on the text confidence scores, and obtain the user clue information.

[0084] In this embodiment, the target clue dataset can be a new message dataset obtained by performing a union processing on the transaction keywords and the clue keywords. The target clue dataset includes all data information within the first user data tag and the second user data tag.

[0085] In one possible embodiment, the aforementioned user lead information includes high-performing user lead information and backup user lead information. Specifically, the text confidence scores of all keywords in the target lead dataset can be calculated. These keywords can then be sorted in ascending order of text confidence score, and the median of the text confidence scores can be taken. If the text confidence score of a keyword is greater than or equal to the median text confidence score, it will be classified as high-performing user lead information. If the text confidence score of a keyword is less than the median text confidence score, it will be classified as backup user lead information.

[0086] This application classifies keywords based on the text confidence score of each keyword in the dataset, thereby accurately acquiring and organizing user clue information, extracting key information closely related to user behavior and intent, and improving the accuracy and value of user clue information.

[0087] In some alternative implementations, the following steps are included after step S207: The tag information database is subscribed to and connected to the monitoring and scheduling platform via a message queue; The user clue information is monitored through the monitoring sub-platform to obtain clue monitoring data; Based on the clue monitoring data, the data scheduling of the tag information database is performed through the routing rules of the scheduling sub-platform to obtain clue scheduling data.

[0088] In this embodiment, the monitoring and scheduling platform includes a monitoring sub-platform and a scheduling sub-platform. The monitoring sub-platform and scheduling sub-platform can be built on a backend platform using Django and a frontend webpage using Vue 3.0 to visualize the clue monitoring data. The visualization process can be to display the clue monitoring data on the frontend webpage as bar charts, line charts, and pie charts.

[0089] In this embodiment, the aforementioned lead monitoring data includes the number of valid leads, the number of valid leads tagged with labels, and the label confidence level. Specifically, the number of valid leads may be the number of transaction keywords contained in the current first user data label and the number of lead keywords contained in the current second user data label; the number of valid leads tagged with labels may be the number of weak-label data in the first user data label; and the label confidence level may be the sum of the text confidence levels of all transaction keywords in the first user data label and the sum of the text confidence levels of all lead keywords in the second user data label.

[0090] In this embodiment, the aforementioned subscription connection can bind the data input / output interface of the tag information database with the data input / output interface of the monitoring and scheduling platform and the message queue through middleware such as RocketMQ, Redis and IBM MQ to achieve data exchange. The aforementioned monitoring and scheduling platform can also interact with the clue model, the preset tag processing model and the tag information database through the message queue.

[0091] In this embodiment, the scheduling sub-platform can determine when to acquire data, invoke the tag processing large model, and the clue model based on the following routing rules, so as to efficiently utilize data and improve the output efficiency of clue samples. The main routing rules include: Rule 1: Prioritize the first user data tag with the most valid leads; Rule 2: Prioritize selecting the first user data label with the highest label confidence; Rule 3: If the number of valid leads for the first user data tag reaches N, then switch to sample mining for another tag; Rule 4: If the output ratio of the second user data tag of the current potential user call text is lower than the threshold P, then switch to another potential user call text for mining. Among them, rule1, rule2, and rule3 can be routing rules for calling the large model of tag processing, and rule4 can be a routing rule for calling the clue model.

[0092] This application implements a subscription connection between the tag information database and the monitoring and scheduling platform through a message queue, constructing an efficient data interaction mechanism. The monitoring sub-platform in the monitoring and scheduling platform can monitor user clue information in real time, providing accurate data support for subsequent clue mining methods. At the same time, based on this monitoring data, the scheduling sub-platform intelligently schedules the tag information database through preset routing rules to generate clue scheduling data, thereby realizing the optimized configuration and efficient utilization of clue resources.

[0093] It should be emphasized that, to further ensure the privacy and security of the aforementioned user clue information, this information can also be stored in a blockchain node.

[0094] The blockchain referred to in this application is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0095] 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 instructing related hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0096] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0097] Further reference Figure 3 As a response to the above Figure 2 To implement the method shown, this application provides an embodiment of a user clue mining device, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0098] like Figure 3 As shown, the user clue mining device 300 described in this embodiment includes: an acquisition module 301, a data cleaning module 302, a processing module 303, a data retrieval module 304, a training module 305, an input module 306, and a classification module 307. Wherein: The acquisition module 301 is used to acquire user call text, which includes call text of potential users and call text of completed transactions. Data cleaning module 302 is used to clean the call text of the transacting user to obtain target text data; Processing module 303 is used to input the target text data into a preset tag processing model for text keyword extraction processing to obtain a first user data tag, wherein the first user data tag includes several transaction keywords; Data retrieval module 304 is used to perform data retrieval processing on the potential user call text through the first user data tag to obtain weak tag data; Training module 305 is used to train the general model using the weakly labeled data as a cue training dataset to obtain a pre-trained cue model. Input module 306 is used to input the potential user call text into the clue model for clue search processing and generate a second user data tag, the second user data tag including several clue keywords; The classification module 307 is used to upload the first user data tag and the second user data tag to the tag information database, and to classify the transaction keywords and the lead keywords to obtain user lead information.

[0099] The processing module 303 includes: The execution submodule is used to execute the prompt instruction on the target text data through the LLM model to obtain the target text keywords; The vectorization submodule is used to perform embedding vectorization processing on the target text keywords through the BCE embedding model to obtain the target text vector; The clustering submodule is used to cluster the target text vector using the text clustering model to obtain cluster categories, wherein each cluster category contains several transaction keywords. The category interpretation submodule is used to perform category interpretation processing on the transaction keywords through the LLM model to generate the first user data tag.

[0100] The data retrieval module 304 includes: The random filtering submodule is used to randomly filter all transaction keywords within the first user data tag to obtain the target transaction keywords; The context positioning submodule is used to perform context positioning on the potential user's call text based on the target transaction keyword, and obtain the text position information of the target transaction keyword in the potential user's call text; The field extraction submodule is used to perform context field extraction processing on the potential user's call text based on the text location information to obtain text recall data. The data annotation submodule is used to input the text recall data into the LLM model for data annotation processing to obtain the weak label data.

[0101] The aforementioned data annotation submodule includes: The intent segmentation unit is used to perform intent segmentation processing on the text recall data through the text clustering model to obtain category samples; The label classification unit is used to perform label classification processing on the category samples using the LLM model to generate target label data; A calculation unit is used to calculate the text confidence score of the target label data; The first judgment unit is used to annotate the target label data to obtain the weak label data if the text confidence is greater than or equal to a preset confidence threshold, and save the target label data to the label information database. The second judgment unit is used to directly save the target tag data to the tag information database if the text confidence level is less than a preset confidence level threshold.

[0102] The training module 305 includes: The extraction submodule is used to extract text features from the clue training dataset to obtain the text label features of the clue training dataset; An estimation submodule is used to estimate the model parameters of the cue model using the text label features of the cue training dataset; The import submodule is used to import the model parameters of the clue model into the general model, and train the general model using the clue training dataset to obtain the clue model.

[0103] The classification module 307 includes: The traversal submodule is used to traverse all transaction keywords in the first user data tag and all clue keywords in the second user data tag to obtain the target clue dataset. The classification submodule obtains the text confidence scores of all keywords in the target clue dataset, performs classification based on the text confidence scores, and obtains the user clue information.

[0104] The following is included after the classification module 307: The connection module is used to subscribe to the tag information database and the monitoring and scheduling platform through a message queue. The monitoring and scheduling platform includes a monitoring sub-platform and a scheduling sub-platform. The monitoring and scheduling platform interacts with the clue model, the preset tag processing model and the tag information database through the message queue. The monitoring module is used to monitor the user clue information through the monitoring sub-platform to obtain clue monitoring data, which includes the number of valid clues, the number of valid clues with tags, and the tag confidence level. The scheduling module is used to schedule the tag information database based on the clue monitoring data and the routing rules of the scheduling sub-platform to obtain clue scheduling data.

[0105] In this embodiment, by acquiring user call text, preprocessing the call text of converted users, and inputting it into a preset tag processing model for text keyword extraction, a first user data tag is obtained. The first user data tag is then used to retrieve potential user call text, resulting in weakly labeled data. This weakly labeled data is used as a lead training dataset to train a general model, resulting in a lead model. The potential user call text is then input into the lead model for lead search processing, generating a second user data tag. The first and second user data tags are uploaded to a tag information database. Converted keywords and lead keywords are then classified to obtain user lead information. This completes a closed loop of the lead mining process, fully utilizing the capabilities of the large model to maximize its advantages in each sub-task of lead mining, effectively reducing labeling costs, improving lead mining efficiency, and proactively discovering new leads or keywords, reducing lead omissions, and increasing the number of leads mined from existing customers.

[0106] In this embodiment, the operations performed by the above-mentioned units or modules correspond one-to-one with the steps of the user clue mining method in the above-described embodiments, and will not be repeated here.

[0107] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed]. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0108] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected via a system bus. It should be noted that only the computer device 4 with components 41-43 is shown in the figure; however, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0109] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0110] The memory 41 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 41 may be an internal storage unit of the computer device 4, such as the hard disk or memory of the computer device 4. In other embodiments, the memory 41 may also be an external storage device of the computer device 4, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 4. Of course, the memory 41 may also include both the internal storage unit and its external storage device of the computer device 4. In this embodiment, the memory 41 is typically used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions for user clue mining methods. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or will be output.

[0111] In some embodiments, the processor 42 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. The processor 42 is typically used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to execute computer-readable instructions stored in the memory 41 or to process data, for example, to execute computer-readable instructions for the user clue mining method.

[0112] The network interface 43 may include a wireless network interface or a wired network interface, which is typically used to establish communication connections between the computer device 4 and other electronic devices.

[0113] This embodiment provides a computer device that acquires user call text, preprocesses the call text of converted users, and inputs it into a preset tag processing model for text keyword extraction to obtain a first user data tag. The first user data tag is then used to retrieve potential user call texts to obtain weakly labeled data. This weakly labeled data is used as a lead training dataset to train a general model, resulting in a lead model. Potential user call texts are then input into the lead model for lead search processing to generate a second user data tag. The first and second user data tags are uploaded to a tag information database. Converted keywords and lead keywords are then classified to obtain user lead information. This completes a closed loop of lead mining, fully utilizing the capabilities of the large model to maximize its advantages in each sub-task of lead mining, effectively reducing labeling costs, improving lead mining efficiency, and proactively discovering new leads or keywords, reducing lead omissions, and increasing the number of leads mined from existing customers.

[0114] This application also provides another embodiment, namely, providing a computer-readable storage medium storing computer-readable instructions that can be executed by at least one processor to cause the at least one processor to perform the steps of the user clue mining method described above.

[0115] This embodiment provides a computer-readable storage medium that acquires user call text, preprocesses the call text of converted users, and inputs it into a preset tag processing model for text keyword extraction to obtain a first user data tag. The first user data tag is then used to retrieve potential user call texts to obtain weakly labeled data. This weakly labeled data is used as a lead training dataset to train a general model, resulting in a lead model. Potential user call texts are then input into the lead model for lead search processing to generate a second user data tag. The first and second user data tags are uploaded to a tag information database. Converted keywords and lead keywords are then classified to obtain user lead information. This completes a closed loop of lead mining, fully utilizing the capabilities of the large model to maximize its advantages in each sub-task of lead mining, effectively reducing labeling costs, improving lead mining efficiency, and proactively discovering new leads or keywords, reducing lead omissions, and increasing the number of leads mined from existing customers.

[0116] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0117] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A method for mining user clues, characterized in that, Includes the following steps: Obtain user call texts, including call texts of potential users and call texts of completed users; The text of the completed user's call is cleaned to obtain the target text data. The target text data is input into a preset tag processing model for text keyword extraction to obtain a first user data tag. The first user data tag includes several transaction keywords. The tag processing model includes an LLM model. The process of retrieving weakly labeled data from the potential user's call text using the first user data tags includes: randomly filtering all transaction keywords within the first user data tags to obtain target transaction keywords; performing contextual positioning on the potential user's call text based on the target transaction keywords to obtain the text position information of the target transaction keywords within the potential user's call text; performing context field truncation on the potential user's call text based on the text position information to obtain text retrieval data; and inputting the text retrieval data into the LLM model for data annotation to obtain the weakly labeled data. The weakly labeled data is used as the cue training dataset to train the general model, thus obtaining the cue model. The potential user's call text is input into the clue model for clue search processing to generate a second user data tag, which includes several clue keywords. The first user data tag and the second user data tag are uploaded to the tag information database, and the transaction keywords and the lead keywords are classified and processed to obtain user lead information.

2. The user clue mining method according to claim 1, characterized in that, The large-scale tag processing model includes an LLM model, a BCE embedding model, and a text clustering model. The step of inputting the target text data into the preset large-scale tag processing model for text keyword extraction to obtain the first user data tag specifically includes: The target text keywords are obtained by executing the prompt command on the target text data using the LLM model. The target text keywords are vectorized using the BCE embedding model to obtain the target text vector. The target text vector is clustered using the text clustering model to obtain cluster categories, each containing several transaction keywords. The transaction keywords are categorized and interpreted using the LLM model to generate the first user data tag.

3. The user clue mining method according to claim 1, characterized in that, The step of processing the text recall data through the LLM model to obtain the weakly labeled data specifically includes: The text recall data is processed by intent segmentation using a text clustering model to obtain category samples; The LLM model is used to perform label classification on the category samples to generate target label data. Calculate the text confidence score of the target label data; If the text confidence level is greater than or equal to a preset confidence threshold, the target label data is labeled to obtain the weak label data, and the target label data is saved to the label information database. If the confidence level of the text is less than the preset confidence level threshold, the target tag data is directly saved to the tag information database.

4. The user clue mining method according to claim 1, characterized in that, The step of using the weakly labeled data as a cue training dataset to train the general model and obtain the cue model specifically includes: Text feature extraction is performed on the clue training dataset to obtain the text label features of the clue training dataset; The model parameters of the clue model are estimated using the text label features of the clue training dataset; The model parameters of the clue model are imported into the general model, and the general model is trained using the clue training dataset to obtain the clue model.

5. The user clue mining method according to claim 1, characterized in that, The steps of uploading the first user data tag and the second user data tag to the tag information database, and classifying the transaction keywords and the lead keywords to obtain user lead information, specifically include: The target lead dataset is obtained by iterating through all transaction keywords in the first user data tag and all lead keywords in the second user data tag. Obtain the text confidence scores of all keywords within the target clue dataset, classify them based on the text confidence scores, and obtain the user clue information.

6. The user clue mining method according to claim 1, characterized in that, After the step of uploading the first user data tag and the second user data tag to the tag information database to obtain user clue information, the method further includes: The tag information database is subscribed to and connected to the monitoring and scheduling platform through a message queue. The monitoring and scheduling platform includes a monitoring sub-platform and a scheduling sub-platform. The monitoring and scheduling platform interacts with the clue model, the preset tag processing model and the tag information database through the message queue. The monitoring sub-platform performs data monitoring on the user clue information to obtain clue monitoring data, which includes the number of valid clues, the number of valid clues tagged, and the confidence level of the tags. Based on the aforementioned clue monitoring data, the tag information database is scheduled using the routing rules of the scheduling sub-platform to obtain clue scheduling data.

7. A user clue mining device, characterized in that, include: The acquisition module is used to acquire user call texts, which include call texts of potential users and call texts of completed transactions. The data cleaning module is used to clean the call text of the transacting users to obtain target text data. The processing module is used to input the target text data into a preset tag processing model for text keyword extraction to obtain a first user data tag. The first user data tag includes several transaction keywords. The tag processing model includes an LLM model. The data retrieval module is used to perform data retrieval processing on the potential user's call text using the first user data tag to obtain weakly labeled data. Specifically, it includes: randomly filtering all transaction keywords within the first user data tag to obtain target transaction keywords; performing context positioning on the potential user's call text based on the target transaction keywords to obtain the text position information of the target transaction keywords in the potential user's call text; performing context field truncation processing on the potential user's call text based on the text position information to obtain text retrieval data; and inputting the text retrieval data into the LLM model for data annotation processing to obtain the weakly labeled data. The training module is used to train the general model using the weakly labeled data as the cue training dataset to obtain a pre-trained cue model. The input module is used to input the potential user's call text into the clue model for clue search processing and generate a second user data tag, which includes several clue keywords. The classification module is used to upload the first user data tag and the second user data tag to the tag information database, and to classify the transaction keywords and the lead keywords to obtain user lead information.

8. A computer device, characterized in that, The method includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the user clue mining method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the user clue mining method as described in any one of claims 1 to 6.

Images