Telecommunication marketing strategy generation method and device and storage medium

Abstract

The invention provides a telecommunication marketing strategy generation method and device and a storage medium, relates to the field of telecommunication marketing, and can solve the problem of low marketing efficiency. The method comprises the following steps: performing intention recognition based on voice information of a user in a telecommunication customer service hotline to obtain intention probability distribution; feature extraction is carried out on the user data to obtain a user portrait, and the user portrait is used for representing objective attributes of a telecommunication service currently used by a user and subjective feelings of the user; based on the intention probability distribution and the user portrait, a user score is determined, and the user score is used for representing the intention degree of bandwidth service installation of the user; and under the condition that the user score is greater than the user score threshold, determining a marketing strategy of the broadband service corresponding to the user score, and pushing the marketing strategy to the terminal of the user. According to the method, different requirements of different users can be accurately matched, different marketing strategies are pushed to different users, the marketing efficiency is improved, and the user experience is improved.

Description

Technical Field

[0001] This application relates to the field of telecommunications marketing, and more particularly to a method, apparatus and storage medium for generating telecommunications marketing strategies. Background Technology

[0002] With the rapid development of communication technology and the widespread adoption of broadband services, telecom operators face the serious challenge of how to efficiently tap into potential user needs and improve marketing conversion rates.

[0003] Currently, marketing resources are typically delivered to users through keyword matching and other methods. However, the marketing strategies generated by these methods often fail to match user needs, resulting in low marketing efficiency. Summary of the Invention

[0004] This application provides a method, apparatus, and storage medium for generating telecommunications marketing strategies, which can improve marketing efficiency.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, this application provides a method for generating telecommunications marketing strategies. The method includes: performing intent recognition based on user voice information from a telecommunications customer service hotline to obtain an intent probability distribution, which indicates the probability that a user will subscribe to each type of telecommunications service based on the voice information; extracting features from user data to obtain a user profile, which characterizes the objective attributes of the telecommunications services currently being used by the user and the user's subjective feelings; determining a user score based on the intent probability distribution and the user profile, which characterizes the user's degree of intention to install bandwidth services; and, if the user score is greater than a user score threshold, determining a marketing strategy for the broadband service corresponding to the user score and pushing the marketing strategy to the user's terminal.

[0006] In conjunction with the first aspect mentioned above, one possible implementation involves performing intent recognition based on the user's voice information from the telecommunications customer service hotline to obtain the intent probability distribution. This includes: performing speech recognition on the voice information using the Wav2Vec model to convert the voice information into text information; and performing intent recognition based on the text information using a pre-trained intent recognition model to obtain the intent probability distribution.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, user data includes user location information, user service information, and user voice information. Feature extraction is performed on the user data to obtain a user profile, including: extracting features from the user's location information and service information to obtain explicit features, which characterize the user's objective attributes when using telecommunications services; extracting features from the user's voice information to obtain implicit features and emotional features, where implicit features characterize the user's potential needs when using telecommunications services, and emotional features characterize the user's level of satisfaction with telecommunications services; and finally, a weighted sum of the explicit features, implicit features, and emotional features is performed to obtain the user profile.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, user ratings are determined based on intent probability distribution and user profiles, including: determining user ratings based on intent probability distribution, user profiles, and market environment information, where the market environment information is used to characterize the marketing situation of broadband services in the user's region.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, user ratings are determined based on intent probability distribution, user profiles, and market environment information, including: processing intent probability distribution, user profiles, and market environment information through an ensemble learning model based on gradient boosting decision trees to obtain user ratings.

[0010] In conjunction with the first aspect above, in one possible implementation, determining the marketing strategy for broadband services corresponding to user ratings includes at least one of the following: determining the marketing strategy for broadband services corresponding to user ratings based on a first mapping relationship, wherein the first mapping relationship is used to characterize the correspondence between user ratings and marketing strategies for broadband services; determining the marketing strategy for broadband services based on user ratings and the probability of installing broadband services in the intent probability distribution; determining the marketing strategy for broadband services based on user ratings and the economic development level of the user's region; and determining the marketing strategy for broadband services based on whether there is competitor information in user ratings and voice information.

[0011] Secondly, this application provides a telecommunications marketing strategy generation device, comprising: an identification unit for identifying intent based on user voice information from a telecommunications customer service hotline, obtaining an intent probability distribution, the intent probability distribution indicating the probability of a user subscribing to each type of telecommunications service based on the voice information; an extraction unit for extracting features from user data to obtain a user profile, the user profile representing the objective attributes of the telecommunications services currently being used by the user and the user's subjective feelings; a determination unit for determining a user score based on the intent probability distribution and the user profile, the user score representing the user's degree of intention to install bandwidth services; and a generation unit for determining a marketing strategy for the broadband service corresponding to the user score if the user score is greater than a user score threshold, and pushing the marketing strategy to the user's terminal.

[0012] In conjunction with the second aspect above, in one possible implementation, the recognition unit is used to: perform speech recognition on the speech information using the Wav2Vec model, converting the speech information into text information; and based on the text information, perform intent recognition using a pre-trained intent recognition model to obtain the intent probability distribution.

[0013] In conjunction with the second aspect mentioned above, in one possible implementation, user data includes user location information, user service information, and user voice information; the extraction unit is used to: extract features from the user's location information and user service information to obtain the user's explicit features, which are used to characterize the user's objective attributes when using telecommunications services; extract features from the user's voice information to obtain the user's implicit features and user emotional features, which are used to characterize the user's potential needs when using telecommunications services, and the emotional features are used to characterize the user's satisfaction with telecommunications services; and perform a weighted summation of the explicit features, implicit features, and emotional features to obtain a user profile.

[0014] In conjunction with the second aspect above, in one possible implementation, the determining unit is used to: determine a user score based on the intent probability distribution, user profile, and market environment information, wherein the market environment information is used to characterize the marketing situation of broadband services in the user's region.

[0015] In conjunction with the second aspect above, in one possible implementation, the determining unit is used to: process the intent probability distribution, user profile, and market environment information through an ensemble learning model based on gradient boosting decision trees to obtain user ratings.

[0016] In conjunction with the second aspect above, in one possible implementation, the generation unit is used to: determine the marketing strategy for broadband services corresponding to user ratings based on a first mapping relationship, wherein the first mapping relationship is used to characterize the correspondence between user ratings and marketing strategies for broadband services; determine the marketing strategy for broadband services based on user ratings and the probability of installing broadband services in the intent probability distribution; determine the marketing strategy for broadband services based on user ratings and the economic development level of the user's region; and determine the marketing strategy for broadband services based on whether there is competitor information in the user ratings and voice information.

[0017] Thirdly, this application provides an electronic device, including: a processor and a communication interface; the communication interface and the processor are coupled, and the processor is used to run computer programs or instructions to implement the telecommunications marketing strategy generation method as described in the first aspect and any possible implementation of the first aspect.

[0018] Fourthly, this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the telecommunications marketing strategy generation method as described in the first aspect and any possible implementation thereof.

[0019] Fifthly, this application provides a computer program product containing instructions that, when run on a computer, cause the computer to execute the telecommunications marketing strategy generation method as described in the first aspect and any possible implementation thereof.

[0020] In a sixth aspect, this application provides a chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run computer programs or instructions to implement the telecommunications marketing strategy generation method as described in the first aspect and any possible implementation thereof.

[0021] Specifically, the chip provided in this application also includes a memory for storing computer programs or instructions.

[0022] It should be noted that the aforementioned computer instructions may be stored, in whole or in part, on a computer-readable storage medium. This computer-readable storage medium may be packaged together with the processor of the device, or it may be packaged separately from the processor of the device; this application does not impose any limitation on this.

[0023] In a seventh aspect, this application provides a telecommunications marketing strategy generation system, comprising: a telecommunications marketing strategy generation apparatus and a terminal, wherein the telecommunications marketing strategy generation apparatus is used to perform the telecommunications marketing strategy generation method as described in the first aspect and any possible implementation thereof.

[0024] The descriptions of aspects two through seven in this application can be referenced to the detailed description of aspect one; and the beneficial effects of the descriptions of aspects two through seven can be referenced to the analysis of the beneficial effects of aspect one, which will not be repeated here.

[0025] In this application, the name of the aforementioned telecommunications marketing strategy generation device does not limit the device or functional module itself. In actual implementation, these devices or functional modules may appear under other names. As long as the functions of each device or functional module are similar to those in this application, they fall within the scope of the claims of this application and their equivalents.

[0026] These or other aspects of this application will become more readily apparent in the following description.

[0027] The telecommunications marketing strategy generation method provided in this application, by recognizing user intent from voice information, can meet the need for real-time acquisition of user intent and improve user satisfaction. By determining user ratings through intent probability distribution and user profiles, and then generating broadband service marketing strategies based on these ratings, it can accurately match the different needs of different users, push different marketing strategies to different users, prevent user churn, dynamically optimize resource allocation, improve marketing efficiency, and enhance user experience. Attached Figure Description

[0028] Figure 1 A flowchart illustrating a method for generating a telecommunications marketing strategy, as provided in this application embodiment; Figure 2 An architecture diagram of a pre-trained intent recognition model provided for embodiments of this application; Figure 3 A schematic diagram illustrating a telecommunications marketing strategy generation method provided in an embodiment of this application; Figure 4 A schematic diagram illustrating the module division of a telecommunications marketing strategy generation system provided in this application embodiment; Figure 5 A schematic diagram illustrating the hierarchical division of a telecommunications marketing strategy generation system provided in this application embodiment; Figure 6 A schematic diagram of Prometheus monitoring provided in an embodiment of this application; Figure 7 A schematic diagram illustrating the working stages of a telecommunications marketing strategy generation system provided in this application embodiment; Figure 8 This is a schematic diagram of a telecommunications marketing strategy generation device provided in an embodiment of this application. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0031] The terms "first" and "second," etc., used in the specification and drawings of this application are used to distinguish different objects or to distinguish different treatments of the same object, rather than to describe a specific order of objects.

[0032] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0033] It should be noted that in the embodiments of this application, the words "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0034] In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0035] With the rapid development of communication technology and the widespread availability of broadband services, telecom operators face the serious challenge of efficiently identifying potential user needs and improving marketing conversion rates. Traditional methods for identifying user needs mainly rely on historical data analysis, keyword matching, and small-scale machine learning models. However, these methods have significant shortcomings in terms of real-time performance, accuracy, and intelligence. Especially in customer service hotline scenarios (e.g., the 10010 customer service hotline), the ability to achieve real-time identification and accurate response to user needs has become crucial for improving marketing efficiency.

[0036] 1. Analysis model based on historical data Traditional methods of user demand analysis often rely excessively on historical data from T-1 (the previous day). This model suffers from significant latency and cannot meet the needs of real-time marketing. For example, in the scenario of new broadband installations, after a user calls customer service to express their needs, the relevant information takes at least 24 hours to be processed. This long wait leads to lost business opportunities. Furthermore, historical data has poor timeliness and cannot accurately reflect the user's current real-time intent, resulting in marketing decisions that lack timeliness and targeting.

[0037] Specifically, existing solutions rely on historical data for user intent analysis, using historical data to construct a spatiotemporal behavior tensor, and mapping tensor elements to semantics. Sentiment feature vector; semantics The method of performing temporal logic analysis on emotional feature vectors to construct an intent state transition graph results in the inability to identify new user needs in a timely manner.

[0038] 2. Small model technology based on keyword matching In existing solutions, some systems employ small machine learning models combined with predefined keywords (such as "broadband," "new installation," and "router") for intent recognition. However, this approach has several limitations: Low recall and precision: Relying solely on a limited set of keywords is simply insufficient to fully cover the diverse ways users express themselves. For example, relying solely on keywords fails to identify colloquial expressions (such as "My home internet speed is too slow, I want to change my plan"), dialectal expressions (such as dialectal variations of "install broadband"), and implicit intentions (such as "Is there a faster internet service?"). This results in a large number of potential users being missed or misjudged, making it impossible to accurately target marketing resources.

[0039] Lack of semantic understanding: Small models struggle to accurately distinguish subtle differences in user intent. For example, while "new broadband installation," "broadband relocation," and "broadband repair" all involve broadband services, user needs differ. Small models easily confuse the two, leading to misjudgments and severely impacting marketing effectiveness (e.g., pushing a new installation package to a user experiencing repair issues), thus reducing user satisfaction.

[0040] Specifically, some existing solutions generate marketing leads by analyzing keywords in historical call recordings and combining them with basic user information. However, this approach suffers from data latency of up to 24 hours, failing to meet the demands of real-time marketing. Furthermore, the model accuracy is low, with both recall and precision below 25%, leading to wasted marketing resources and user complaints.

[0041] Currently, some companies in the industry have recognized the shortcomings of traditional methods and are attempting to optimize user intent mining by leveraging artificial intelligence (AI) technology. However, these solutions still have some problems: Data delay: Although some solutions use real-time data, they do not achieve end-to-end process continuity, resulting in delays in data delivery.

[0042] Model limitations: Some solutions use general language models without specific optimization for customer service scenarios. The accuracy and adaptability of general models are insufficient to meet actual marketing needs.

[0043] Insufficient multimodal data fusion: Existing solutions rarely integrate multi-dimensional information such as user profiles (e.g., historical packages, spending power), real-time behavioral data (e.g., recent search records), and business characteristics (e.g., regional broadband coverage). This results in models making judgments based solely on isolated text fragments, limiting their accuracy. For example, when a user asks "Are there any cost-effective broadband options suitable for large apartments?", the lack of correlation analysis with the user's housing area and spending level prevents the model from accurately recommending a suitable package.

[0044] In summary, existing technologies have significant shortcomings in terms of real-time performance, accuracy, and intelligence. They fail to effectively address key issues such as real-time data collection, accurate semantic understanding, and seamless end-to-end process integration. Therefore, the industry urgently needs a highly efficient intent mining solution that can combine real-time data, large language models, and multi-dimensional user profiles to improve the marketing conversion rate and user experience of customer service hotlines.

[0045] The telecommunications marketing strategy generation method provided in this application, by recognizing user intent from voice information, can meet the need for real-time acquisition of user intent and improve user satisfaction. By determining user ratings through intent probability distribution and user profiles, and then generating broadband service marketing strategies based on these ratings, it can accurately match the different needs of different users, push different marketing strategies to different users, prevent user churn, dynamically optimize resource allocation, improve marketing efficiency, and enhance user experience.

[0046] It should be noted that the various embodiments of this application can be referenced or learned from each other. For example, the same or similar steps, method embodiments, system embodiments and device embodiments can be referenced from each other without limitation.

[0047] Figure 1 A flowchart illustrating a method for generating a telecommunications marketing strategy, as provided in an embodiment of this application. Figure 1 As shown, the method for generating this telecommunications marketing strategy includes the following steps: S101. Based on the user's voice information in the telecommunications customer service hotline, perform intent recognition to obtain the intent probability distribution.

[0048] The intent probability distribution is used to indicate the probability that a user will subscribe to each type of telecommunications service based on voice information.

[0049] In some embodiments, a recording system connected to a telecommunications customer service hotline in real time is used to collect users' voice information in real time via a real-time transport protocol (RTP). For example, the sampling rate is 16kHz and the encoding format is pulse code modulation (PCM).

[0050] For example, the intent probability distribution includes six types of intent probabilities. These include the probability of intent to install new broadband, the probability of intent to upgrade broadband, the probability of filing a complaint, the probability of broadband renewal, the probability of modifying broadband packages, and the probability of broadband suspension or resumption.

[0051] In some embodiments, S101 can be specifically implemented as follows: recognizing speech information through an intent recognition model to obtain the intent distribution probability.

[0052] S102. Extract features from user data to obtain user profiles.

[0053] Among them, user profiles are used to characterize the objective attributes of the telecommunications services currently being used by users and the users' subjective feelings.

[0054] For example, user data includes both objective and subjective data. Objective data includes user location information and consumption records. Subjective data includes user complaints.

[0055] In some embodiments, S102 can be specifically implemented as: extracting features from the user's objective data and the user's subjective data respectively to obtain a user profile.

[0056] S103. Determine user ratings based on intent probability distribution and user profiles.

[0057] Among them, user ratings are used to characterize the degree of user intention to install bandwidth services.

[0058] S104. If the user rating is greater than the user rating threshold, determine the marketing strategy for the broadband service corresponding to the user rating and push the marketing strategy to the user's terminal.

[0059] For example, a user rating threshold is used to determine whether a user urgently needs to install broadband service. For instance, user rating thresholds could be 0.6, 0.85, etc.

[0060] Marketing strategies can be pushed to users' terminals through various means, including business operation support systems, outbound calls, SMS push notifications, app push notifications, and WeChat for Enterprise. For example, marketing strategies can be pushed to the provincial telecommunications marketing system in real time via API, and then further distributed to users' terminals in real time by the provincial telecommunications marketing system. To ensure the real-time nature of the marketing strategies, data latency can be controlled within 5 minutes to support immediate response from frontline personnel.

[0061] In some embodiments, the marketing strategy for determining broadband services corresponding to user ratings as described above includes at least one of the following: (1) Based on the first mapping relationship, determine the marketing strategy of broadband services corresponding to user ratings. The first mapping relationship is used to characterize the correspondence between user ratings and marketing strategies of broadband services.

[0062] For example, different user ratings correspond to different marketing strategies for broadband services.

[0063] For example, when a user rating is greater than a first user rating threshold, the marketing strategy for the corresponding broadband service is determined to be "half price for the first month".

[0064] For example, when a user rating is less than or equal to a first user rating threshold and greater than a second user rating threshold, the marketing strategy for the corresponding broadband service is determined to be "providing gigabit broadband + free router".

[0065] Determining broadband service marketing strategies through the first mapping relationship can quickly identify marketing strategies and improve marketing efficiency.

[0066] (2) Determine the marketing strategy for broadband services based on the probability of installing broadband services in the user rating and intent probability distribution.

[0067] For example, when a user's rating is greater than a third user rating threshold, and the probability of installing broadband service in the intent probability distribution is greater than a first probability, the marketing strategy for the broadband service is determined to be "immediate outbound call, senior customer service manager connection, offering a 50% discount for the first month." For instance, the third user rating threshold is 0.85, and the first probability is 0.9.

[0068] (3) Determine the marketing strategy for broadband services based on user ratings and the level of economic development in the user's region.

[0069] For example, the level of economic development of a user's location can be determined based on income and / or consumption information in the user's region. For instance, the level of economic development of the user's region can be determined based on factors such as GDP per capita or per capita consumption expenditure in the user's region.

[0070] Understandably, when the economic development level of a user's region is high, the user is more likely to apply for broadband services. In other words, the success rate of the business will be higher when the marketing strategy for broadband services is pushed to the user.

[0071] (4) Determine the marketing strategy for broadband services based on whether there is competitor information in user ratings and voice information.

[0072] For example, keyword matching can be used to determine whether competitor information exists in the voice information.

[0073] Understandably, when a user's voice message contains information about competing products, it indicates that the user is interested in that type of telecommunications service, meaning there is a high probability that the user will subscribe to the service. Generating marketing strategies in this way and pushing them to users can prevent user churn.

[0074] For example, the following is an example of a marketing strategy for a broadband service provided in this application.

[0075] If user rating > 0.85 AND the probability of installing broadband service in the intent probability distribution > 0.9, then... Broadband service marketing strategy: "Immediate outbound calls, senior customer service manager connection, offering a 50% discount for the first month." ELSE IF User rating > 0.75 AND the probability of installing broadband service in the intent probability distribution > 0.8 AND the user's location in a high-value area THEN Broadband service marketing strategy: "Contact us via SMS and phone within 24 hours, and offer free on-site speed testing." ELSE IF User rating > 0.65 AND Competitive keywords exist in the voice message THEN Broadband service marketing strategy: Send SMS messages highlighting the competitive advantages of our products compared to competitors within 48 hours. ... The telecommunications marketing strategy generation method provided in this application, by recognizing user intent from voice information, can meet the need for real-time acquisition of user intent and improve user satisfaction. By determining user ratings through intent probability distribution and user profiles, and then generating broadband service marketing strategies based on these ratings, it can accurately match the different needs of different users, push different marketing strategies to different users, prevent user churn, dynamically optimize resource allocation, improve marketing efficiency, and enhance user experience.

[0076] In some embodiments, the process of obtaining the intent probability distribution described above includes the following steps: Step 1: Use the Wav2Vec model to perform speech recognition on the speech information and convert it into text information.

[0077] For example, the Wav2Vec model can be an improved Wav2Vec 2.0 model.

[0078] For example, the Wav2Vec model consists of a feature extractor and a context network. The feature extractor comprises a 7-layer 1D convolutional neural network, each with a stride of 2 and 512 channels respectively. The context network comprises a 24-layer Transformer encoder, each layer containing 1024-dimensional hidden states and 16 attention heads.

[0079] For example, the training process of the Wav2Vec model described above includes two stages: self-supervised pre-training and supervised fine-tuning. The self-supervised pre-training involves training on 100,000 hours of unlabeled speech information. Based on this self-supervised pre-training, the model is then fine-tuned using 100,000 hours of labeled speech information.

[0080] Furthermore, considering the large number of technical terms in telecommunications scenarios (e.g., fiber to the home (FTTH), gigabit-capable passive optical network (GPON), etc.), to avoid recognition errors in the Wav2Vec model due to unfamiliar technical terms, commonly used technical dictionaries in the telecommunications scenario are added to the Wav2Vec model's dictionary or decoder. Adding these technical dictionaries improves speech recognition accuracy by 15.8%.

[0081] Furthermore, considering that communication between users and customer service representatives in telecommunications customer service hotlines takes the form of dialogue, an endpoint detection algorithm based on voiceprint features is used to accurately identify the moment of speaker switching. For example, the voice data is segmented into multiple short frames of 10ms each, and 39-dimensional Mel-frequency cepstral coefficients (MFCC) features are extracted from each frame. Then, based on the MFCC features, a pre-trained BiLSTM network is used to predict the speaker switching probability. When the switching probability exceeds a threshold of 0.85, it is determined as a dialogue segmentation point. The pre-trained BiLSTM network consists of two layers, each containing 256 neurons.

[0082] For example, to improve the accuracy of speech recognition, the speech information can be preprocessed to obtain preprocessed speech information. Then, speech recognition can be performed on the preprocessed speech information to obtain text information. For example, preprocessing of the speech information can include noise reduction, channel separation, and format standardization. For instance, the RNNoise algorithm can be used to reduce noise in the speech information.

[0083] Understandably, the collected speech information contains diverse expressions such as dialects. To improve the accuracy of speech recognition, the collected speech information can first be translated into dialects. Then, the translated speech information is used for speech recognition through the Wav2Vec model to convert it into text information.

[0084] For example, AI translation technology can be used to translate speech information into dialects. For instance, open-source speech recognition frameworks such as Mozilla DeepSpeech and Kaldi can be used to translate speech information into dialects. This improves the dialect recognition accuracy from 62.3% to 86.2%, an increase of 38.5%.

[0085] Step 2: Based on the text information, perform intent recognition using a pre-trained intent recognition model to obtain the intent probability distribution.

[0086] For example, the pre-trained intent recognition model can be a large language model (LLM), SVM, random forest, etc.

[0087] For example, intent recognition can be performed using a fine-tuned ChatGLM model, GPT model, etc., combined with a rule engine (e.g., Drools).

[0088] For example, the pre-trained intent recognition model can be an intent recognition model based on the improved RoBERTa-large model framework. Figure 2 This is an architecture diagram of a pre-trained intent recognition model provided in an embodiment of this application. Figure 2 As shown, the pre-trained intent recognition model includes an input layer, an improved RoBERTa-large model framework, a domain adaptation layer, and an intent classification head.

[0089] The input layer is used to input text information, such as text segmentation, positional encoding, and [CLS] tags. The improved RoBERTa-large model framework includes a 24-layer Transformer encoder, each with 1024-dimensional hidden states and 16 attention heads. The domain adaptation layer consists of a 3-layer feedforward neural network with a 1024-dimensional input, 512-dimensional intermediate layers, and a 256-dimensional output. The intent classification head includes a 256-dimensional fully connected layer, ReLU activation, a 64-dimensional fully connected layer, and a Softmax layer.

[0090] For example, the training process of a pre-trained intent recognition model includes a data preparation phase, a pre-training phase, a fine-tuning phase, a continuous learning phase, and an inference optimization phase.

[0091] In the data preparation phase, the training data for the model came from 20,000 voice messages from telecom customer service representatives. These voice messages were independently annotated by three professional annotators, achieving a Krippendorff's α coefficient of 0.82 to ensure annotation consistency and avoid different interpretations of the same dialogue.

[0092] During the pre-training phase, the system was pre-trained on a corpus of 1 billion tokens from the telecommunications industry. Masked language modeling (MLM) and next sentence prediction (NSP) tasks were employed to learn language habits specific to the telecommunications customer service domain. The training parameters were batch size = 256 and learning rate = 2e-5.

[0093] During the fine-tuning phase, since each dialogue may contain multiple intentions, the model needs to output multi-label probabilities for multi-label intent recognition. To further address the imbalanced sample problem, the weight γ of easily classified samples in the loss function is set to 2.0, and the ratio of positive to negative samples is set to 0.75 to prevent the model from neglecting the minority class due to the majority class dominance. The model parameters are: batch size = 32, learning rate = 5e-6, warmup ratio = 0.1, total training steps = 15000; Dropout = 0.2, weight decay = 0.01, gradient clipping threshold = 1.0.

[0094] During the continuous learning phase, after deploying the pre-trained intent recognition model, incremental training is performed every two weeks using 500 newly collected labeled data points to prevent model degradation.

[0095] To ensure stable model operation in low-latency, high-concurrency scenarios and meet the real-time requirements of the customer service system, dynamic batching technology was used during the inference optimization phase to group dialogue texts of different lengths into groups of similar length. Furthermore, TensorRT was used for model quantization, converting FP32 precision to INT8, resulting in a 3.5x speedup inference while keeping accuracy loss within 1%. Four inference instances were deployed for load balancing, achieving a single instance query per second (QPS) of 42 (on an NVIDIA T4 GPU).

[0096] For example, to improve the accuracy of intent recognition, the text information can be preprocessed to obtain preprocessed text information. Then, a pre-trained intent recognition model can be used to perform intent recognition on the preprocessed text information to obtain the intent recognition probability.

[0097] For example, text information can be filtered using preset keywords to reduce the load on subsequent processing. The preset keywords can be one of 128 keywords, such as "broadband," "slow internet speed," and "poor signal." The Aho-Corasick automaton algorithm is used for keyword filtering, with a time complexity of O(n), where n is the length of the text information.

[0098] Thus, in the field of telecommunications customer service, combining the Wav2Vec model with a pre-trained intent recognition model to obtain the intent probability distribution can improve the accuracy and robustness of intent recognition.

[0099] In some embodiments, user data includes the user's location information, the user's business information, and the user's voice information. In this case, the process of obtaining a user profile includes the following steps: Step 1: Extract features from the user's location information and business information to obtain the user's explicit features.

[0100] Among them, explicit features are used to characterize the objective attributes of users when using telecommunications services.

[0101] For example, the user's location information includes the district / county where the user is located.

[0102] For example, a user's business information includes the user's historical consumption information, the user's current package type, the user's historical consultation records, the user's historical billing information, and so on.

[0103] For example, explicit features can include 12 dimensions such as user location, historical spending amount, current package type, and number of complaints.

[0104] For example, explicit features are obtained by using natural language processing techniques such as TF-IDF and word embedding.

[0105] For example, explicit features are obtained by processing with deep packet inspection (DPI) technology.

[0106] Step 2: Extract features from the user's voice information to obtain the user's latent features and emotional features.

[0107] Among them, latent features are used to characterize users' potential needs when using telecommunications services, while emotional features are used to characterize users' satisfaction with telecommunications services.

[0108] For example, 28-dimensional latent features are extracted from voice information. Examples include "urgency of internet access needs." This "urgency of internet access needs" can be calculated based on the frequency and emotional intensity of words such as "online classes" and "remote work" mentioned in the voice information.

[0109] For example, sentiment analysis models are used to process speech information to obtain sentiment features. For instance, a sentiment analysis model could be a BERT-based model fine-tuned for use in the telecommunications customer service field. Another example is that sentiment analysis features could be values ​​in the range [-1, 1].

[0110] For example, latent features are obtained by processing the user's speech information through graph neural networks (e.g., GNN) and temporal prediction models (e.g., LSTM).

[0111] Step 3: Perform a weighted summation of explicit features, implicit features, and emotional features to obtain the user profile.

[0112] For example, after normalizing the dominant features, latent features, and emotional features, a weighted sum is then performed.

[0113] For example, the weights of dominant features, latent features, and emotional features can be preset values, or they can be values ​​calculated based on the user's historical information or the user's voice information. This application embodiment does not limit this. For example, the weight of dominant features is 0.4, the weight of latent features is 0.35, and the weight of emotional features is 0.25.

[0114] Understandably, when the dimensions of explicit features, implicit features, and emotional features are different, the dimensions of explicit features, implicit features, and emotional features can be reduced or aligned so that the dimensions of explicit features, implicit features, and emotional features are the same. Then, a weighted sum is performed to obtain the user profile.

[0115] For example, the following is an algorithm implementation process for obtaining a user profile provided in an embodiment of this application.

[0116] Python edit def build_user_profile(explicit_features, implicit_features,sentiment_score): # Weight allocation (optimized via grid search) w_explicit = 0.4 w_implicit = 0.35 w_sentiment = 0.25 # Feature Normalization explicit_norm = min_max_normalize(explicit_features) implicit_norm = min_max_normalize(implicit_features) # Weighted summation profile_vector = ( w_explicit explicit_norm + w_implicit implicit_norm + w_sentiment np.array([sentiment_score)) len(explicit_features)) ) # Applying Sigmoid Transform profile_vector = 1 / (1+np.exp(-profile_vector)) return profile_vector In this way, user profiles are obtained based on explicit features, implicit features, and emotional features. This avoids the user profiles being limited by static data and allows the user profiles to more realistically reflect user attributes from multiple perspectives, providing richer supporting information for the generation of subsequent marketing strategies.

[0117] In some embodiments, determining user ratings based on intent probability distribution and user profiles includes: determining user ratings based on intent probability distribution, user profiles, and market environment information.

[0118] Among them, market environment information is used to characterize the marketing situation of broadband services in the user's region.

[0119] For example, market environment information includes broadband penetration rate in the user's area, competitor coverage, promotional activity time, etc.

[0120] For example, the user rating value can be between 0 and 1.

[0121] In some embodiments, determining user ratings based on intent probability distribution, user profiles, and market environment information includes: processing intent probability distribution, user profiles, and market environment information using an ensemble learning model based on gradient boosting decision trees to obtain user ratings.

[0122] For example, the ensemble learning model based on gradient boosting decision trees consists of an XGBoost model, a LightGBM model, and a logistic regression model.

[0123] In the XGBoost model, the parameters are max_depth=6, learning_rate=0.1, and n_estimators=200; in the LightGBM model, the parameters are num_leaves=64, learning_rate=0.05, and n_estimators=300. The outputs of both the XGBoost and LightGBM models are used as inputs to the logistic regression model. This gradient boosting decision tree-based ensemble learning model has an area under the curve (AUC) of 0.923, a precision (top 10%) of 85.7%, and a recall (top 20%) of 76.4%.

[0124] For example, in the training process of an ensemble learning model based on gradient boosting decision trees, the training data consists of historical data from the last 6 months, including 24,538 labeled samples. Among them, 1,207 are positive samples that have actually been converted, and 23,331 are negative samples that have not been converted.

[0125] By using an ensemble learning model based on gradient boosting decision trees to obtain user ratings, user classification performance can be improved.

[0126] In this way, determining user ratings based on intent probability distribution, user profiles, and market environment information can fully take into account the current market environment, making the resulting marketing strategies more in line with the market situation, timely and flexible, and increasing the likelihood of users subscribing to broadband services.

[0127] In addition, after pushing marketing strategies to users' terminals, the above method also includes: obtaining feedback data, which is used to indicate the user's installation status after the marketing strategies are pushed; and optimizing the telecommunications marketing strategy generation model based on the feedback data.

[0128] For example, feedback data includes statistical user conversion rates, user feedback, return on investment (ROI), and how long it takes for users to install broadband services after receiving the marketing strategy. For instance, feedback data can be obtained through a marketing campaign management system to optimize the telecommunications marketing strategy generation model.

[0129] For example, based on feedback data, a reinforcement learning algorithm is used to optimize the telecommunications marketing strategy generation model, maximizing the conversion rate and resource investment ratio. For instance, the reinforcement learning algorithm could be a PPO algorithm.

[0130] For example, it is also possible to calculate the conversion rate of users installing broadband services based on feedback data through different methods of pushing telecommunications marketing strategies, so as to optimize subsequent push methods.

[0131] To enhance the effectiveness of telecom marketing strategies, the optimal push time can be determined based on users' call duration and historical active periods. The marketing strategies can then be pushed out at this optimal time. Furthermore, anti-harassment protection measures are implemented, limiting outbound calls to the same user to no more than twice and SMS messages to no more than five times within seven days. This resulted in a 31.2% conversion rate for users installing broadband services within seven days, a 232% increase compared to the existing 9.4%, a 65% reduction in outbound calls, a 216% increase in marketing resource utilization, and a 78% decrease in user complaints. After three months, marketing costs decreased by 65%, and new broadband users increased by 178%.

[0132] For example, Figure 3 This is a schematic diagram illustrating a telecommunications marketing strategy generation method provided in an embodiment of this application. Figure 3As shown, user voice information is acquired through real-time recording. The acquired voice information is then processed by speech recognition to obtain text information. A pre-trained intent recognition model is used to perform intent recognition on the text information to obtain an intent probability distribution. User location and service information are used to extract explicit user features, while the user's voice information is used to extract implicit and emotional features. These explicit, implicit, and emotional features are then weighted and summed to obtain a user profile. An ensemble learning model based on gradient boosting decision trees is used to process the intent probability distribution and user profile to obtain a user rating. Based on the user rating, a corresponding broadband service marketing strategy is generated.

[0133] In addition, the above-mentioned method for generating telecommunications marketing strategies also includes an exception handling mechanism, which includes at least one of the following: (1) If the delay is greater than or equal to the delay threshold, the network shall be expanded or downgraded.

[0134] For example, P95 / P99 latency thresholds can be set at various stages to monitor latency, and automatic scaling up or downgrading can be implemented if the latency threshold is exceeded. This could also be achieved by increasing bandwidth or the number of servers. Other examples include disabling certain functions, rate limiting, or circuit breaking.

[0135] (2) When the speech recognition accuracy is less than the speech recognition accuracy threshold, the speech information is re-recognized.

[0136] For example, the quality and recognition accuracy of voice information can be evaluated in real time, and unqualified data can be automatically reprocessed.

[0137] (3) If the KL divergence of the model is greater than the divergence threshold, retrain the model, which includes the Wav2Vec model, the pre-trained intent recognition model, and the ensemble learning model based on gradient boosting decision tree.

[0138] For example, monitoring changes in the predicted distribution and triggering model retraining when the KL divergence > 0.15.

[0139] For example, ensuring 99.96% service availability, exceeding the target value of 99.9%, and an average recovery time of less than 3.2 minutes.

[0140] In this way, not only can real-time monitoring of the entire chain from voice collection to marketing execution be achieved, but also self-healing from anomalies can be realized.

[0141] As a specific example, taking Mr. Zhang's complaint about slow home internet speed affecting his child's online classes as an example, the telecommunications marketing strategy generation method provided in this application is used. The voice information is converted into text information within 30 seconds of the user's voice recording, with a voice recognition accuracy of 96.3%. Analyzing the text information through a pre-trained intent recognition model, the probability distribution shows "intent to install new broadband" at 0.92 and "intent to upgrade network" at 0.78. In the user profile, explicit features indicate the user's location is in a key school district of a certain city, their historical monthly consumption is an average of 128 yuan, and their sentiment feature is -0.75, indicating extreme dissatisfaction. The ensemble learning model based on gradient boosting decision trees outputs a user score of 0.89 (exceeding the user score threshold of 0.85). The generated broadband service marketing strategy is "Immediately assign a senior customer service manager to make an outbound call, providing a gigabit broadband + free router + first month half-price discount combination." The feedback data shows that the customer service manager successfully made the outbound call 20 minutes later, and the user completed a new broadband service installation order that day.

[0142] The performance metrics in this process are as follows: the average end-to-end latency from the end of a 10-minute call to the generation of a marketing strategy is 4.7 minutes, meeting real-time marketing needs. The daily processing capacity of the customer service hotline is 12,000 calls. In terms of intent recognition accuracy, the F1 score for broadband new installation intent recognition is 0.887, a 69.3% improvement compared to the traditional keyword matching method (0.524). Regarding marketing conversion rate, among users with a user rating greater than 0.85, the 7-day conversion rate reaches 31.2%, a 226% improvement compared to traditional methods. In terms of resource savings, precise screening reduces outbound call volume by 65% ​​and increases customer service human resource utilization by 3.1 times.

[0143] The telecommunications marketing strategy generation method provided in this application combines real-time voice information with a large-scale intent recognition model, solving the problem in existing solutions that cannot capture implicit broadband needs in voice dialogues. This results in an F1 score of 0.85 or higher for broadband service installation intent recognition, an improvement over the existing 0.55. A dynamic data replacement mechanism is proposed, injecting the latest market environment information in real-time during the large-scale model inference process, improving user rating accuracy by 18.7%. A spatiotemporal feature fusion algorithm is designed to multi-dimensionally correlate user location information, user service information, and user voice information, improving the targeting of marketing strategies. The accuracy of high-value user identification is greater than or equal to 80%, higher than the 62% in existing solutions. Furthermore, the end-to-end processing latency is less than or equal to 5 minutes per call, far less than the 24 hours+ in existing solutions, providing possibilities for instant marketing. The model precision is improved to 56.83%, and the recall is improved to 51.69%. It also meets the daily processing capacity of ≥10,000 calls for telecommunications hotlines, satisfying the needs of provincial operators. Compared to traditional marketing methods, the user conversion rate is improved by greater than or equal to 150%. Marketing resource utilization is improved by ≥200%, reducing invalid outbound calls. Service availability is ≥99.9%, with response time fluctuation ≤15% under peak load. It achieves accurate and real-time identification of potential new broadband subscribers, solving problems such as low accuracy, poor timeliness, and significant resource waste in existing technologies. The entire process is automated, reducing manual intervention and supporting multi-channel collaboration and model iteration optimization. It overcomes the shortcomings of traditional new broadband subscriber identification in terms of timeliness, accuracy, and scalability, demonstrating significant technological advancement and commercial value.

[0144] In practical applications, the above-mentioned method for generating telecommunications marketing strategies can increase the number of broadband service orders, increase broadband service revenue, reduce the number of manual operations and the time spent on manual outbound calls, reduce labor costs, realize the automated flow of orders, reduce delivery time, and improve production efficiency.

[0145] For example, Figure 4 This is a schematic diagram illustrating the module division of a telecommunications marketing strategy generation system provided in an embodiment of this application. When the above-described telecommunications marketing strategy generation method is applied to a telecommunications marketing strategy generation system, the system includes a data acquisition module, a real-time data processing module, an intent recognition large-scale model module, a user profile construction module, a potential user scoring module, and a marketing strategy generation module.

[0146] The data acquisition module, real-time data processing module, intent recognition big data model module, user profile building module, potential user scoring module, and marketing strategy generation module are connected through application programming interface (API) and data bus to form a closed-loop data processing and decision-making system.

[0147] The data acquisition module collects user voice information, location information, service information, and market environment information. The real-time data processing module performs speech recognition on the voice information to generate text information. The intent recognition model module performs intent recognition on the text information to obtain the intent probability distribution. The user profile building module extracts features from user data to obtain user profiles. The potential user scoring module determines user scores based on the intent probability distribution and user profiles. The marketing strategy generation module determines the marketing strategy for broadband services.

[0148] For example, Figure 5 This is a schematic diagram illustrating the hierarchical division of a telecommunications marketing strategy generation system provided in an embodiment of this application. Figure 5 As shown, the telecommunications marketing strategy generation system includes a data acquisition layer, a data processing layer, an AI model layer, an application layer, and a monitoring and maintenance layer.

[0149] The data acquisition layer is used to acquire user voice information, user location information, user business information, market environment information, etc.

[0150] For example, the data processing layer subscribes to the real-time recording stream of the data platform through message queues such as Kafka, RabbitMQ, and NATS to obtain the user's voice information.

[0151] The data processing layer is used for preprocessing user voice information, performing voice recognition, and extracting user data features.

[0152] For example, the data processing layer calls an AI translation interface to generate structured text containing textual information. Another example is that the data processing layer deploys an AI translation service that supports multi-dialect translation and outputs textual information for subsequent processing by the AI ​​model layer. Yet another example is that the data processing layer fuses multiple features to obtain a user profile.

[0153] The AI ​​model layer includes a pre-trained intent recognition model and an ensemble learning model based on gradient boosting decision trees.

[0154] For example, the AI ​​model layer uses a pre-trained ChatGLM model to extract users' latent features and sentiment characteristics. Another example is that the AI ​​model layer outputs user ratings through a LightGBM classifier.

[0155] The application layer is used to generate marketing strategies for broadband services, push marketing strategies to users' terminals, obtain feedback data, and iteratively optimize models, etc.

[0156] For example, the application layer uses the real-time scenario center of the full-scale customer value operation platform to penetrate data down to the provincial grid. Another example is that the application layer supports multi-channel collaborative marketing (e.g., telecom customer service hotlines, WeChat for Business, etc.).

[0157] For example, iterative optimization of the model can be achieved through continuous learning, A / B testing, and other methods. For instance, A / B testing can be used to compare the model's precision, with a target value greater than or equal to 60%, and recall, with a target value greater than or equal to 55%.

[0158] The monitoring and operations layer is used for end-to-end monitoring, performance alerts, monitoring data quality, and model drift.

[0159] For example, when deploying Prometheus and / or Grafana at the monitoring and operations layer to monitor data source latency, set the threshold to less than or equal to ≤30 minutes.

[0160] For example, Figure 6 This is a schematic diagram of Prometheus monitoring provided in an embodiment of this application. Figure 6 As shown, Prometheus monitoring includes a data layer, a model layer, and an application layer.

[0161] In the data layer, data from telecom customer service hotlines, mobile internet service hall tracking, full voice recordings, interactive voice response (IVR) tracking, and mobile internet service hall chatbots are pushed to the model layer's refined and generalized scenarios via real-time Kafka. User profile tags, order data, and complaint tickets are pushed to the model layer's refined and generalized scenarios via API. CBSS incremental data, log center data, failure ticket data, network-side data, and client / network industry data are pushed to the model layer's full-scale problem monitoring via real-time Kafka.

[0162] Kafka is responsible for the transmission and caching of high-throughput data streams.

[0163] The model layer includes refined scenarios, generalized scenarios, and full-scale issue monitoring. In refined scenarios, a supervised refined intent recognition model is built, extracting semantic features through a large language model encoder to accurately identify user intent. In generalized scenarios, an unsupervised generalized intent recognition model is built, employing the RAG approach and combining a knowledge base with LLM for intent recognition. The results of intent recognition in both refined and generalized scenarios are pushed to the application layer for user intent recognition. Full-scale issue monitoring includes data processing, data storage, and monitoring results. Data processing includes real-time processing via Flink and offline processing via Spark. Data storage includes Hive structured storage and IceBerg data lake storage. Monitoring results include CUDB result storage and Prometheus metric monitoring. The monitoring results are pushed to the application layer for system root cause analysis.

[0164] Flink supports complex computations (such as window aggregation and state management), ensuring low latency and high fault tolerance, and enabling data updates in seconds.

[0165] The application layer includes user intent identification, system root cause analysis, and user issue management. User intent identification includes user intentions, such as inability to access the internet, inability to pay bills, unpaid bills, billing disputes, slow internet speeds, etc. System root cause analysis includes issues such as inconsistencies between business accounts (BOs), exceeding limits, service outages, unauthorized transactions, and contract expiration. Based on user intent identification and system root cause analysis, common and personalized capabilities are provided to headquarters and provincial branches to achieve smarter, more accurate, and faster user issue management. Examples include result verification, monitoring and alerts, public opinion monitoring, compliance control, system optimization, and work order management.

[0166] Prometheus monitoring enabled end-to-end monitoring, ensuring system stability.

[0167] For example, Figure 7 This is a schematic diagram illustrating the working stages of a telecommunications marketing strategy generation system provided in an embodiment of this application. Figure 7 As shown, the workflow of the telecommunications marketing strategy generation system includes a real-time data acquisition stage, a voice preprocessing stage, a voice-to-text stage, an intent recognition stage, a user profile building stage, a potential user rating stage, a strategy generation and execution stage, and an effect feedback closed-loop stage.

[0168] During the real-time data acquisition phase, the system monitors the customer service hotline recording server in real time and retrieves voice files at preset time intervals, processing voice information for a preset duration each time. For example, the preset time interval is 5 minutes; the preset duration is 3-5 hours.

[0169] In the speech preprocessing stage, the speech information is preprocessed by noise reduction, channel separation, and format standardization to obtain the preprocessed speech information.

[0170] In the speech-to-text stage, the preprocessed speech information is subjected to speech recognition to obtain text information.

[0171] In this stage, text information is output through methods such as the Wav2Vec model, adding a specialized dictionary, and dialogue segmentation. The output text information can be presented in JSON format as a list of dialogue segments.

[0172] In the intent recognition stage, the text information is processed through a pre-trained intent recognition model to obtain the intent probability distribution.

[0173] For example, by taking a list of dialogue segments as input, the intention probability distribution is obtained in the improved RoBERTa-large model.

[0174] In the user profile building phase, user profiles are obtained by extracting user data features.

[0175] For example, the process of generating a user profile during the user profile building phase takes less than or equal to 45 seconds.

[0176] The intent probability distribution and user profile can be in the form of structured JSON.

[0177] In the potential user rating stage, user ratings are determined using an ensemble learning model based on gradient boosting decision trees.

[0178] During the strategy generation and execution phase, marketing strategies for broadband services are generated for users whose user ratings exceed the user rating threshold, and these marketing strategies are pushed to the user's terminal.

[0179] For example, in addition to recommended content for users, the output marketing strategy may also include user IDs and other information. The marketing strategy is output in JSON format.

[0180] In the closed-loop effect feedback stage, feedback data is obtained, and the telecom marketing strategy generation model is optimized based on the feedback data.

[0181] For example, optimizing a telecommunications marketing strategy generation model includes weekly automatic model retraining, A / B testing verification, and performance metric evaluation.

[0182] In addition, the workflow of the aforementioned telecommunications marketing strategy generation system also includes quality checks on voice information and an automatic retry mechanism (e.g., retrying on failure).

[0183] For example, quality checks on voice information could involve improving the signal-to-noise ratio (SNR) to be greater than 15 dB. Liveness detection could be performed on the voice information, and if it does not meet requirements, manual review should be initiated.

[0184] During this process, data is stored and synchronized in real time. The real-time database stores users with ratings greater than a certain threshold using Redis, with a time-to-live (TTL) of 5 minutes. The analytics database uses ClickHouse to implement full-data columnar storage. Data synchronization is achieved through Debezium CDC to capture changed data.

[0185] For example, by monitoring database log changes through Debezium, incremental data can be synchronized in real time, ensuring data synchronization efficiency.

[0186] For example, data synchronization can be achieved through polling and long-lived connections. In low-frequency update scenarios, short polling is used for data synchronization. In high real-time scenarios (e.g., instant messaging), WebSocket and SSE are used for data synchronization.

[0187] For example, data stored in a database can be modified using the UPDATE and REPLACE commands. UPDATE is used for conditional updates; REPLACE handles primary key conflicts by overwriting existing data. Another example is using subqueries to retrieve data from other tables during data processing to dynamically adjust the target table's content, such as batch updating employee job information.

[0188] As a specific example, the actual deployment of the aforementioned telecommunications marketing strategy generation system includes hardware configuration, data flow configuration, and model deployment parameters.

[0189] The hardware configuration includes a voice processing server, an application server, and storage. Voice processing servers: 4 Dell PowerEdge R750 processors, each configured with 2 x Intel Xeon Silver 4314 CPUs, 256GB RAM, and 2 x NVIDIA T4 GPUs. Application servers: 8 virtual machines, each with 16 vCPUs and 64GB RAM. Storage: 30TB distributed storage (Ceph architecture), including 10TB of high-speed SSD cache.

[0190] In the data stream configuration, customer service hotline recordings are transmitted via a Kafka cluster, with 16 partitions and a throughput of 10MB / s per partition. The speech recognition service is deployed as a gRPC microservice, with an average response time of 280ms / 10 seconds of audio. The intent recognition API gateway is configured for automatic scaling, with a base of 4 instances, dynamically adjusting between 4 and 16 instances based on CPU utilization.

[0191] In the model deployment parameters, the Wav2Vec model batch size is 16, using FP16 precision, and the GPU memory usage is 4.2GB. The pre-trained intent recognition model is optimized using TensorRT, quantized with INT8, and has an inference latency of 127ms / batch(4). The gradient boosting decision tree-based ensemble learning model is deployed using ONNX Runtime, with CPU inference and a P99 latency of 48ms, much less than 5.2 minutes, ensuring stable response under high load.

[0192] The model service can be deployed to a Kubernetes cluster to support automatic scaling down, failover, and high-concurrency real-time prediction, ensuring high-concurrency processing capability of QPS=135≥1000.

[0193] The system throughput is 1200 calls / hour. The intent recognition F1 score is 0.887, exceeding the target value of 0.85. The accuracy rate for identifying high-value users is 83.6%, exceeding the target value of 80%. The false positive rate is 8.3%, a decrease of 17.4 percentage points compared to 25.7% in the existing solution.

[0194] Furthermore, the telecom marketing strategy generation system can also utilize an edge computing architecture to deploy lightweight models at provincial nodes for local inference and distribution. To enhance data credibility, blockchain technology can be used to achieve distributed storage and auditing of metrics such as data latency and model error rate.

[0195] In some embodiments, the above-mentioned telecommunications marketing strategy generation method can also be deployed in domestically produced systems. Vendors such as StarRing Technology provide smooth migration tools, improving the real-time processing capabilities of the system through parameter tuning and resource allocation, supporting domestic replacement requirements. In domestically produced systems, Apache NiFi is used to implement visual configuration data streams, supporting real-time synchronization and format conversion. Through Tencent Cloud's self-developed columnar storage engine THCoushou-C, UPSERT performance under unique key constraints is optimized, improving write efficiency by over 10 times. This not only allows for rapid location of existing data but also reduces update latency. THCoushou-C can be used to introduce a strong consistency protocol, avoiding data overwriting errors caused by replica lag. Automated data cleaning and replacement, such as missing value filling and regular expression matching, is achieved through Python's pandas library combined with SQLAlchemy. A balance between throughput and consistency is achieved through horizontal scaling using a distributed architecture. In domestically produced systems, Flink's Checkpoint mechanism is used to ensure the atomicity of stream processing tasks. Furthermore, it's necessary to convert from a NoSQL database to a relational database. This process requires restructuring the data structure (e.g., migrating Mapus from Firebase to Supabase) and solidifying the table creation process using SQL scripts (section 7). This addresses issues such as the conversion from NoSQL to relational databases and local deployment. During the localization process, compatibility testing is also required to verify application adaptability, adjust database connection configurations and query statements, optimize SQL performance, conduct stress tests for stability, and train business personnel to adapt to the new system. This localization approach balances performance optimization and business adaptability. The system also supports scenarios such as real-time large-screen monitoring and risk control early warning.

[0196] This application embodiment can divide the telecommunications marketing strategy generation device into functional modules or functional units according to the above method example. For example, each function can be divided into a separate functional module or functional unit, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or in software functional modules or functional units. The module or unit division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0197] Figure 8 This application provides a schematic diagram of the structure of a telecommunications marketing strategy generation device 80, which includes: The identification unit 801 performs intent recognition based on the user's voice information from the telecommunications customer service hotline, obtaining an intent probability distribution. This intent probability distribution indicates the probability that the user will subscribe to each type of telecommunications service based on the voice information. The extraction unit 802 extracts features from the user data to obtain a user profile. This user profile represents the objective attributes of the telecommunications services the user is currently using and the user's subjective feelings. The determination unit 803 determines a user score based on the intent probability distribution and the user profile. This user score represents the user's degree of intention to install bandwidth services. The generation unit 804 determines a marketing strategy for the broadband service corresponding to the user score if the user score is greater than a user score threshold, and pushes the marketing strategy to the user's terminal.

[0198] In one possible implementation, the recognition unit 801 is used to: perform speech recognition on speech information using the Wav2Vec model, converting the speech information into text information; and based on the text information, perform intent recognition using a pre-trained intent recognition model to obtain the intent probability distribution.

[0199] In one possible implementation, user data includes user location information, user service information, and user voice information; the extraction unit 802 is used to: extract features from the user's location information and user service information to obtain the user's explicit features, which are used to characterize the user's objective attributes when using telecommunications services; extract features from the user's voice information to obtain the user's implicit features and user emotional features, which are used to characterize the user's potential needs when using telecommunications services and the emotional features are used to characterize the user's satisfaction with telecommunications services; and perform a weighted summation of the explicit features, implicit features, and emotional features to obtain a user profile.

[0200] In one possible implementation, the determining unit 803 is used to: determine a user score based on the intent probability distribution, user profile, and market environment information, wherein the market environment information is used to characterize the marketing situation of broadband services in the user's region.

[0201] In one possible implementation, the determining unit 803 is used to: process the intent probability distribution, user profile, and market environment information through an ensemble learning model based on gradient boosting decision trees to obtain user ratings.

[0202] In one possible implementation, the generation unit 804 is used to: determine a marketing strategy for broadband services corresponding to user ratings based on a first mapping relationship, wherein the first mapping relationship is used to characterize the correspondence between user ratings and marketing strategies for broadband services; determine a marketing strategy for broadband services based on user ratings and the probability of installing broadband services in the intent probability distribution; determine a marketing strategy for broadband services based on user ratings and the level of economic development of the user's region; and determine a marketing strategy for broadband services based on whether there is competitor information in the user ratings and voice information.

[0203] Through the above description of the embodiments, those skilled in the art will clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0204] This application provides a computer program product containing instructions that, when run on a computer, cause the computer to execute the telecommunications marketing strategy generation method described in the above method embodiments.

[0205] This application also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the telecommunications marketing strategy generation method in the method flow shown in the above method embodiments.

[0206] The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires; a portable computer disk drive; a hard disk drive; random access memory (RAM); read-only memory (ROM); erasable programmable read-only memory (EPROM); a register; a hard disk drive; an optical fiber; a compact disc read-only memory (CD-ROM); an optical storage device; a magnetic storage device; or any suitable combination thereof; or any other form of computer-readable storage medium known in the art. An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium may also be a component of the processor. The processor and the storage medium may reside in an application-specific integrated circuit (ASIC). In the embodiments of this application, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0207] Since the telecommunications marketing strategy generation device, computer-readable storage medium, and computer program product in the embodiments of this application can be applied to the above method, the technical effects that can be obtained can also be referred to the above method embodiments. The embodiments of this application will not be repeated here.

[0208] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0209] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0210] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0211] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating telecommunications marketing strategies, characterized in that, The method includes: Intent recognition is performed based on the user's voice information in the telecommunications customer service hotline to obtain an intent probability distribution, which is used to indicate the probability of the user engaging in each type of telecommunications service based on the voice information. Feature extraction is performed on user data to obtain user profiles, which are used to characterize the objective attributes of the telecommunications services currently being used by the user and the user's subjective feelings. Based on the intent probability distribution and the user profile, a user score is determined, which is used to characterize the user's degree of intention to install bandwidth services; If the user rating is greater than the user rating threshold, a marketing strategy for the broadband service corresponding to the user rating is determined, and the marketing strategy is pushed to the user's terminal.

2. The method according to claim 1, characterized in that, The process of performing intent recognition based on user voice information from telecommunications customer service hotlines to obtain an intent probability distribution includes: The speech information is recognized by the Wav2Vec model and converted into text information. Based on the text information, the intent is identified using a pre-trained intent recognition model to obtain the intent probability distribution.

3. The method according to claim 1, characterized in that, The user data includes the user's location information, the user's service information, and the user's voice information; The process of extracting features from user data to obtain a user profile includes: Feature extraction is performed on the user's location information and the user's service information to obtain the user's explicit features, which are used to characterize the objective attributes of the user when using the telecommunications service; Feature extraction is performed on the user's voice information to obtain the user's latent features and emotional features. The latent features are used to characterize the user's potential needs when using the telecommunications service, and the emotional features are used to characterize the user's satisfaction with the telecommunications service. The user profile is obtained by weighted summation of the explicit features, the implicit features, and the emotional features.

4. The method according to claim 1, characterized in that, The process of determining user ratings based on the intent probability distribution and the user profile includes: Based on the intent probability distribution, the user profile, and market environment information, the user score is determined, whereby the market environment information is used to characterize the marketing situation of broadband services in the user's region.

5. The method according to claim 4, characterized in that, The process of determining the user rating based on the intent probability distribution, the user profile, and market environment information includes: The user rating is obtained by processing the intent probability distribution, the user profile, and the market environment information through an ensemble learning model based on gradient boosting decision trees.

6. The method according to claim 1, characterized in that, The marketing strategy for determining the broadband service corresponding to the user rating includes at least one of the following: Based on the first mapping relationship, a marketing strategy for broadband services corresponding to the user rating is determined, wherein the first mapping relationship is used to characterize the correspondence between the user rating and the marketing strategy for broadband services. The marketing strategy for the broadband service is determined based on the user rating and the probability of installing the broadband service in the intent probability distribution. The marketing strategy for the broadband service is determined based on the user ratings and the level of economic development of the user's region. The marketing strategy for the broadband service is determined based on the user ratings and whether there is competitor information in the voice information.

7. A telecommunications marketing strategy generation device, characterized in that, The device includes: The recognition unit performs intent recognition based on the user's voice information in the telecommunications customer service hotline to obtain an intent probability distribution, which is used to indicate the probability that the user will handle each type of telecommunications service based on the voice information. The extraction unit is used to extract features from user data to obtain a user profile, which is used to characterize the objective attributes of the telecommunications service currently being used by the user and the user's subjective feelings. The determining unit is used to determine a user score based on the intent probability distribution and the user profile, wherein the user score is used to characterize the user's degree of intention to install bandwidth services; The generation unit is used to determine a marketing strategy for broadband services corresponding to the user rating when the user rating is greater than a user rating threshold, and to push the marketing strategy to the user's terminal.

8. An electronic device, characterized in that, include: A processor and a communication interface; the communication interface is coupled to the processor, the processor being configured to run computer programs or instructions to implement the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a computer, perform the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes computer instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1-6.

Images