User voice calling system and method based on big data processing

Abstract

The invention discloses a user voice calling system and method based on big data processing, and relates to the technical field of voice interaction, and the method comprises the steps: collecting multi-modal user data, and carrying out the preprocessing of the multi-modal user data; feature extraction is carried out on the preprocessed multi-modal user data, and user feature vectors are fused; constructing a user knowledge graph and a service according to the user feature vector through a feature clustering algorithm, and dynamically associating the user knowledge graph with a service graph to generate a comprehensive graph; reasoning the comprehensive graph through a graph neural network, calculating a matching degree between a user node and a service node, and generating a personalized calling scheme; and based on the personalized call scheme, generating a voice call script and synthesizing a voice file, and initiating a call to the user through the voice outbound device. According to the method, the user feature vectors are generated through feature extraction and fusion, and the comprehensive atlas is generated through dynamic association, so that the association capability of user requirements and service features is improved.

Description

Technical Field

[0001] The present invention relates to the field of voice interaction technology, and in particular to a user voice calling system and method based on big data processing. Background Art

[0002] With the development of information technology and artificial intelligence, voice interaction, as an important way of human-computer interaction, has been widely used in the fields of intelligent customer service, voice assistants and marketing. Voice call technology realizes efficient interaction between people and machines through technical means such as speech recognition, natural speech processing and speech synthesis. In this process, the introduction of big data technology has significantly improved the personalization ability of voice call systems. By analyzing user voice data, behavior data and business data, user needs and interest preferences are mined, thereby generating targeted call content and strategies. However, voice call technology still faces many challenges, such as the singleness of voice content, the rigidity of call strategies, and the inability to respond to dynamic changes in user needs in a timely manner, resulting in insufficient accuracy of call services and user satisfaction. Especially in terms of multimodal data fusion and user portrait construction, the existing technology lacks effective processing and analysis methods, resulting in a low match between user demand expression and business recommendations.

[0003] Existing voice call technology has deficiencies in the following two aspects: First, for the processing of multimodal user data (such as voice data, behavior data and business data), existing technologies often use a single data analysis method, which is unable to fully explore the correlation between multimodal data, resulting in incomplete user portraits and lack of personalization in call content; second, in terms of calculating the matching degree between users and services, existing methods are mostly based on traditional similarity calculations, and fail to effectively combine knowledge graphs and deep learning technologies, making it difficult to capture the complex semantic relationship between user needs and business characteristics, resulting in insufficient targeting of call content. Summary of the invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a user voice call method based on big data processing to solve the problems of insufficient multimodal data correlation mining and inaccurate calculation of user and business matching.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a user voice calling method based on big data processing, which includes collecting multimodal user data and preprocessing the multimodal user data; extracting features from the preprocessed multimodal user data and fusing them into a user feature vector; constructing a user knowledge graph and a service according to the user feature vector through a feature clustering algorithm, dynamically associating the user knowledge graph and the service graph to generate a comprehensive graph; reasoning the comprehensive graph through a graph neural network, calculating the matching degree between user nodes and service nodes, and generating a personalized call plan; based on the personalized call plan, generating a voice call script and synthesizing a voice file, and initiating a call to the user through a voice outbound calling device.

[0008] As a preferred solution of the user voice call method based on big data processing described in the present invention, wherein: the multimodal user data includes voice data, historical behavior data and business data.

[0009] As a preferred solution of the user voice call method based on big data processing of the present invention, wherein: the multimodal user data is preprocessed, and the specific steps are as follows:

[0010] Remove noise from speech data through noise reduction algorithm;

[0011] Use data cleaning to fill missing values ​​and remove duplicate data from historical behavior data;

[0012] Eliminate irrelevant information and standardize the format of business data through data filtering and conversion rules;

[0013] The normalization method is used to standardize the multimodal user data.

[0014] As a preferred solution of the user voice call method based on big data processing described in the present invention, wherein: the feature extraction of the pre-processed multimodal user data is performed and fused into a user feature vector, the specific steps are as follows:

[0015] Extract semantic feature vectors from user speech data through BER;

[0016] Use CNN to extract acoustic feature vectors from speech data;

[0017] Use LSTM to extract behavior feature vectors from historical behavior data;

[0018] Extract business feature vectors from business data through Word2Vec;

[0019] Use the Attention mechanism to align time series of different feature vectors;

[0020] The geometric interaction of different modal feature vectors is carried out dimension by dimension to capture the high-order nonlinear relationship between the modes. At the same time, the linear superposition of all modal feature vectors is supplemented and fused into the user feature vector, which is expressed as:

[0021]

[0022] Among them, F is the user feature vector, F i represents the i-th modal eigenvector, ||F i ||2 represents the L2 norm of the ith modal feature vector, γ is the weight coefficient controlling the linear superposition part in the final fusion result, α is the hyperparameter for adjusting the nonlinear effect of the modal feature, n represents the total number of modal features, and F j Represents the eigenvector of the jth mode.

[0023] As a preferred solution of the user voice call method based on big data processing described in the present invention, wherein: the user knowledge graph and business are constructed according to the user feature vector through the feature clustering algorithm, the knowledge graph and the business graph are dynamically associated to generate a comprehensive graph, and the specific steps are as follows:

[0024] Based on the user feature vector F, the Euclidean distance between the user feature vector and the cluster center is calculated by the feature clustering algorithm, and the expression is:

[0025]

[0026] Among them, d(F i ,C k ) represents the user feature vector F i and the i-th cluster center C k The Euclidean distance between k represents the user feature vector of the kth cluster center, F i,r represents the value of the i-th user feature vector on the r-th feature dimension, C k,r represents the value of the k-th cluster center on the r-th feature dimension, and m represents the number of dimensions of the user feature vector;

[0027] Based on the Euclidean distance d(F i ,C k ), dynamically assign users to k cluster groups until the cluster centers converge, and finally generate k cluster groups;

[0028] Each user F iAs an independent node, each cluster group is converted into a group node, and the user node and the group node are connected through the belonging relationship. The interest tags and behavior patterns in the user feature vector are extracted to generate feature nodes, which are associated with the user node and the group node through the owning feature relationship to form a structured user knowledge graph;

[0029] Extract entities from business data and generate business nodes;

[0030] Generate a business feature node according to the business feature vector;

[0031] Connect the business node with its feature node through attributes to establish a business relationship, and store the business node, business feature node and business relationship in a structured manner to form a business knowledge graph;

[0032] Based on user behavior data, the user knowledge graph and the business knowledge graph are semantically associated using user features and business features to generate a comprehensive graph.

[0033] As a preferred solution of the user voice call method based on big data processing described in the present invention, wherein: the comprehensive graph is inferred by the graph neural network, the matching degree between the user node and the service node is calculated, and a personalized call plan is generated. The specific steps are as follows:

[0034] HetGNN is used to initialize the feature vectors of user nodes and business nodes in the comprehensive graph, and the attention mechanism is used to dynamically aggregate the feature information of neighboring nodes to generate a high-dimensional embedding vector for each node;

[0035] According to the high-dimensional embedding vector of each node, combined with the semantic features of the path in the comprehensive graph, the matching degree between the user node and the business node is calculated, and the expression is:

[0036]

[0037] Among them, S(u,e) represents the matching degree between user node u and service node e, H u [i] represents the value of the embedding vector of user node u in the xth dimension, H e [x] represents the value of the embedding vector of business node e in the xth dimension, H u represents the normalized norm of user node u, H e represents the normalized norm of the business node e, d represents the dimension of the embedding vector, and x is the index variable of the dimension;

[0038] Define a relevance threshold T based on historical data;

[0039] When S(u,e)≥T, it is considered that the current user u is highly related to the current service e, and the call content focuses on the content related to the current service;

[0040] When S(u,e)<T, it is considered that the current user u has a low correlation with the current service e, and a general voice call solution is adopted.

[0041] As a preferred solution of the user voice call method based on big data processing of the present invention, wherein: based on the personalized call solution, a voice call script is generated and a voice file is synthesized, and a call is initiated to the user through a voice outbound call device. The specific steps are as follows:

[0042] Generate a voice call script including opening remarks, business key points, interactive guidance and closing remarks based on the personalized call plan;

[0043] The voice script is converted into a voice file through the TTS engine, and the speaking speed, intonation and timbre are adjusted based on the interest tags in the user feature vector, and a call is initiated to the user using an outbound calling device.

[0044] In the second aspect, the present invention provides a user voice call system based on big data processing, including a data acquisition module, a feature fusion module, a graph construction module, a call plan generation module and a voice call module; the data acquisition module is used to collect multimodal user data and preprocess the multimodal user data; the feature fusion module is used to extract features from the preprocessed multimodal user data and fuse them into user feature vectors; the graph construction module is used to construct a user knowledge graph and business according to the user feature vector through a feature clustering algorithm, dynamically associate the user knowledge graph and the business graph, and generate a comprehensive graph; the plan generation module is used to infer the comprehensive graph through a graph neural network, calculate the matching degree between user nodes and business nodes, and generate a personalized call plan; the voice call module is used to generate a voice call script and synthesize a voice file based on the personalized call plan, and initiate a call to the user through a voice outbound call device.

[0045] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the user voice call method based on big data processing as described in the first aspect of the present invention is implemented.

[0046] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the user voice call method based on big data processing as described in the first aspect of the present invention.

[0047] The beneficial effects of the present invention are as follows: by collecting and preprocessing multimodal user data, the high quality of data input is ensured; by generating user feature vectors through feature extraction and fusion, the correlation of multimodal data is fully explored; by utilizing feature clustering to construct user knowledge graphs and business graphs, and dynamically associating to generate comprehensive graphs, the ability to associate user needs with business features is improved; by inferring the comprehensive graph through graph neural networks, the matching degree between users and businesses is accurately calculated, and a personalized call plan is generated; based on the plan, a voice script is generated and a voice file is synthesized, which is efficiently reached through outbound calling equipment, solving the problems of insufficient personalization of call content and inaccurate matching, and improving the accuracy of voice call services and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0049] Figure 1 This is a flow chart of the user voice calling method based on big data processing in Example 1.

[0050] Figure 2 This is a schematic diagram of a user voice call system based on big data processing in Example 1. DETAILED DESCRIPTION

[0051] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0052] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0053] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0054] Example 1, reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, and provides a user voice calling method based on big data processing, comprising the following steps:

[0055] S1: Collect multimodal user data and preprocess the multimodal user data.

[0056] S1.1: Multimodal user data includes voice data, historical behavior data, and business data.

[0057] The purpose of collecting voice data, historical behavior data and business data is to comprehensively obtain multi-dimensional information about users, laying a data foundation for subsequent feature extraction, user portrait construction and personalized solution generation.

[0058] S1.2: Remove noise from speech data using a noise reduction algorithm;

[0059] For example, noisy background sounds (such as wind or traffic noise) in user call recordings are filtered out, retaining only clear voice content to facilitate subsequent semantic feature extraction.

[0060] S1.3: Use data cleaning to fill missing values ​​and remove duplicate data in historical behavior data;

[0061] For example, fill in missing transaction date information in user purchase records and remove duplicate records of the same product to ensure data integrity and accuracy.

[0062] S1.4: Eliminate irrelevant information and standardize the format of business data through data filtering and conversion rules;

[0063] For example, remove operation records that are not related to users from business logs, and convert amount fields into a standard two-decimal format to facilitate subsequent calculations.

[0064] S1.5: Use normalization methods to standardize multimodal user data.

[0065] For example, the user age value (such as 18-60) and the consumption amount (such as 100-10,000 yuan) are normalized to the interval [0,1] to eliminate the dimensional differences between the features and facilitate model processing.

[0066] S2: Extract features from the preprocessed multimodal user data and fuse them into a user feature vector.

[0067] S2.1: Extract semantic feature vectors from user speech data through BER;

[0068] Specifically, the BERT model is used to perform semantic analysis on user voice data, converting the voice content into a high-dimensional semantic feature vector to represent user intentions and semantic information.

[0069] S2.2: Extract acoustic feature vectors from speech data using CNN;

[0070] Specifically, the speech data is converted into a representation in the time-frequency domain, such as a spectrogram or Mel-frequency cepstral coefficients (MFCC), as the input of the CNN; then, the convolution layer is used to extract features of the local area of ​​the input data to identify key acoustic patterns in the speech signal (such as pitch, timbre, frequency changes, etc.); then, the pooling layer is used to reduce the dimension of the convolution features to retain important information while reducing redundancy and computational complexity; finally, high-level acoustic features are further extracted through multi-layer convolution and pooling operations, and finally a high-dimensional acoustic feature vector is output to characterize the sound characteristics of the speech data and provide support for subsequent user profiling and voice interaction.

[0071] S2.3: Use LSTM to extract behavior feature vectors from historical behavior data;

[0072] Specifically, the process of extracting behavioral feature vectors from historical behavior data using LSTM is: input the user's historical behavior data into the LSTM model in chronological order, use its memory and forgetting mechanism to capture the time dependency and dynamic changes of the behavior, and finally output a high-dimensional feature vector to comprehensively characterize the user's behavioral characteristics.

[0073] S2.4: Extract business feature vectors from business data through Word2Vec;

[0074] Specifically, the process of extracting business feature vectors from business data through Word2Vec is: input business-related text data (such as product descriptions or service tags) into the Word2Vec model, learn the semantic associations of vocabulary through training, and convert business keywords into low-dimensional dense feature vectors to represent the semantic relationships and characteristics between businesses.

[0075] S2.5: Use the Attention mechanism to align time series of different feature vectors;

[0076] Specifically, the process of using the Attention mechanism to align different feature vectors in time series is: by calculating the importance weight of each feature vector in the time series, dynamically focusing on key features, adjusting the temporal order relationship between modalities, and achieving alignment and fusion of different feature vectors, providing consistency and accuracy support for subsequent analysis.

[0077] S2.6: Geometric interaction is performed dimension by dimension on the eigenvectors of different modes to capture the high-order nonlinear relationship between modes. At the same time, the linear superposition of all modal eigenvectors is supplemented and fused into the user eigenvector. The expression is:

[0078]

[0079] Among them, F is the user feature vector, F irepresents the i-th modal eigenvector, ||F i ||2 represents the L2 norm of the ith modal feature vector, γ is the weight coefficient controlling the linear superposition part in the final fusion result, α is the hyperparameter for adjusting the nonlinear effect of the modal feature, n represents the total number of modal features, and F j Represents the eigenvector of the jth mode.

[0080] F i and F j represent the eigenvectors of the i-th and j-th modes respectively. The difference between them is: F i is the feature vector under a specific mode, representing the features generated after feature extraction of the i-th mode (such as voice, behavior or business data); and F j is the eigenvector of another mode, representing the characteristics of the jth mode. The geometric interaction in the formula (such as the product operation) realizes the information fusion between the eigenvectors of different modes. i and F j The relationship between the modalities captures the high-order nonlinear relationship between the modalities, thereby integrating the characteristics of different modalities to form the final user feature vector F.

[0081] S3: Through the feature clustering algorithm, the user knowledge graph and business are constructed according to the user feature vector, and the user knowledge graph and business graph are dynamically associated to generate a comprehensive graph.

[0082] S3.1: Based on the user feature vector F, the Euclidean distance between the user feature vector and the cluster center is calculated by the feature clustering algorithm, and the expression is:

[0083]

[0084] Among them, d(F i ,C k ) represents the user feature vector F i and the i-th cluster center C k The Euclidean distance between k represents the user feature vector of the kth cluster center, F i,r represents the value of the i-th user feature vector on the r-th feature dimension, C k,r represents the value of the k-th cluster center on the r-th feature dimension, and m represents the number of dimensions of the user feature vector;

[0085] It should be noted that the user feature vector F is calculated by the feature clustering algorithm i With cluster center C kThe Euclidean distance between them can quantify the similarity between user characteristics and different cluster centers, so as to dynamically divide users into the most matching cluster groups. This process improves the accuracy and reliability of user grouping by accurately capturing the differences and similarities between user characteristics and group characteristics, and provides a clear structured foundation for the subsequent construction of user knowledge graphs, ultimately achieving more accurate user demand analysis and personalized matching effects.

[0086] S3.2: Based on the Euclidean distance d(F i ,C k ), dynamically assign users to k cluster groups until the cluster centers converge, and finally generate k cluster groups;

[0087] It should be noted that by dynamically assigning users to k cluster groups based on the Euclidean distance between the user feature vector and the cluster center until the cluster center converges, users can be efficiently and adaptively clustered. This process effectively integrates the feature information of similar users, reveals the similarities within the group and the differences between groups, and provides a structural basis for the refined construction of user portraits. Ultimately, this dynamic clustering method improves the accuracy of user demand classification and the hierarchy of knowledge graph construction, providing reliable support for personalized services and recommendations.

[0088] S3.3: For each user F i As an independent node, each cluster group is converted into a group node, and the user node and the group node are connected through the belonging relationship. The interest tags and behavior patterns in the user feature vector are extracted to generate feature nodes, which are associated with the user node and the group node through the owning feature relationship to form a structured user knowledge graph;

[0089] Furthermore, on the basis of constructing a structured user knowledge graph, the user nodes, group nodes and feature nodes are connected through association relationships using the interest tags and behavior patterns in the user feature vector to form a multi-level graph structure. This graph can not only intuitively represent the subordinate relationship between individual users and groups, but also show the deep correlation between user characteristics and behavior patterns, providing rich semantic information for user demand analysis. Ultimately, this structured graph lays the foundation for accurate recommendations, user behavior prediction and personalized services.

[0090] S3.4: Extract entities from business data and generate business nodes;

[0091] Specifically, business data (such as product descriptions, service information, etc.) are parsed through natural speech processing technology, key entities (such as product names, service types, attributes, etc.) are identified, and these entities are converted into business nodes as the basic units of the business knowledge graph, providing structured data support for subsequent business feature modeling and user matching.

[0092] S3.5: Generate a business feature node according to the business feature vector;

[0093] Specifically, by quantizing the key entities extracted from business data (such as product attributes, service types, etc.), a low-dimensional vector space representation representing business features is generated using feature extraction methods (such as Word2Vec or other embedding models). Subsequently, these business feature vectors are converted into business feature nodes, and connected to the corresponding business nodes through attribute relationships to construct the feature layer structure in the business knowledge graph, providing accurate feature support for the refined representation of business content and user matching.

[0094] S3.6: Connect the business node with its feature node through attributes to establish a business relationship, and store the business node, business feature node and business relationship in a structured manner to form a business knowledge graph;

[0095] It should be noted that according to the characteristics of business nodes (such as product categories, service types, etc.), key attributes are extracted, and these attributes are associated with corresponding business feature nodes to establish attribute relationships between nodes; then, business nodes, business feature nodes and the relationships between them are stored in a structured form to form a business knowledge graph with clear relationships and clear levels. This graph can comprehensively describe business entities and their characteristic attributes, providing efficient support for business analysis, association mining and user matching.

[0096] S3.7: Based on user behavior data, the user knowledge graph and the business knowledge graph are semantically associated using user features and business features to generate a comprehensive graph.

[0097] It should be noted that by analyzing user behavior data (such as browsing, purchasing or evaluation records), the correlation between user feature vectors and business feature vectors is extracted, and the connection between user nodes and business nodes is established by combining semantic similarity calculation. Subsequently, user nodes, business nodes and their feature nodes are connected through interactive relationships to generate a comprehensive graph that integrates user interests and business features, fully reflecting the correlation between users and businesses, and providing support for accurate recommendations and personalized services.

[0098] S4: The comprehensive graph is inferred through the graph neural network, the matching degree between the user node and the business node is calculated, and a personalized call plan is generated.

[0099] S4.1: HetGNN is used to initialize the feature vectors of user nodes and business nodes in the comprehensive graph, and the attention mechanism is used to dynamically aggregate the feature information of neighbor nodes to generate a high-dimensional embedding vector for each node;

[0100] Furthermore, the initial feature vectors of user nodes and business nodes are encoded through the HetGNN model, and the basic representation is completed using the structural information and node features in the graph. Then, the neighbor nodes of each node are dynamically aggregated based on the attention mechanism, and different weights are assigned to different neighbor nodes to highlight the influence of key nodes. Finally, a high-dimensional embedding vector of each node is generated by combining the feature information of the neighbor nodes and its own features, which accurately represents the semantic relationship and structural information of the node in the comprehensive graph, providing support for further analysis and matching.

[0101] S4.2: Based on the high-dimensional embedding vector of each node and the semantic features of the path in the comprehensive graph, the matching degree between the user node and the business node is calculated. The expression is:

[0102]

[0103] Among them, S(u,e) represents the matching degree between user node u and service node e, H u [i] represents the value of the embedding vector of user node u in the xth dimension, H e [x] represents the value of the embedding vector of business node e in the xth dimension, H u represents the normalized norm of user node u, H e represents the normalized norm of the business node e, d represents the dimension of the embedding vector, and x is the index variable of the dimension;

[0104] It should be noted that by calculating the matching degree S(u,e) between user nodes and business nodes, the correlation between user needs and business features can be quantified based on the high-dimensional embedding vector of the node and the semantic features in the comprehensive graph. This method combines the dimension-by-dimensional nonlinear transformation and normalization operation of the embedding vector to fully capture the complex semantic and structural relationship between users and businesses, avoiding the limitations of single feature matching. Ultimately, this matching degree calculation method significantly improves the accuracy of user needs and business recommendations, providing higher efficiency and reliability for personalized services and intelligent recommendations.

[0105] S4.3: Define a relevance threshold T based on historical data;

[0106] When S(u,e)≥T, it is considered that the current user u is highly related to the current service e, and the call content focuses on the content related to the current service;

[0107] For example, if it is found after calculation that a user's matching degree for the "new smartphone" service is higher than the threshold T, the relevant discounts and functional highlights of the smartphone will be highlighted in the call.

[0108] When S(u,e)<T, it is considered that the current user u has a low correlation with the current service e, and a general voice call solution is adopted.

[0109] For example, if it is detected that the matching degree of a certain user for the "financial management service" business is lower than the threshold T, a general voice call solution is adopted, and only regular greetings and basic business introductions are performed.

[0110] S5: Based on the personalized call plan, generate a voice call script and synthesize a voice file, and initiate a call to the user through a voice outbound calling device.

[0111] S5.1: Generate a voice call script including opening remarks, business key points, interactive guidance and closing remarks based on the personalized call plan;

[0112] It should be noted that a friendly opening statement should be used to attract the user's attention; then, based on the matching degree between the user and the business, the business content related to the user's needs should be introduced in detail; then, interactive guidance should be designed to encourage the user to respond or ask questions; finally, a polite closing statement should be used to end the call, ensuring that the call process is smooth and attractive, and improving the user experience and communication effect.

[0113] S5.2: The voice script is converted into a voice file through the TTS engine, the speech speed, intonation and timbre are adjusted based on the interest tags in the user feature vector, and a call is initiated to the user using an outbound calling device.

[0114] It should be noted that the generated voice script is converted into a natural and fluent voice file through the TTS (text-to-speech) engine, and the speech speed, intonation and timbre are adjusted in combination with the interest tags in the user feature vector to make the voice style more in line with user preferences and enhance the personalized experience of the call. Finally, a voice call is initiated to the user through the outbound calling device to ensure accurate information transmission and improve user acceptance of the service.

[0115] This embodiment also provides a user voice call system based on big data processing, including: a data acquisition module, a feature fusion module, a graph construction module, a call plan generation module and a voice call module; the data acquisition module is used to collect multimodal user data and pre-process the multimodal user data; the feature fusion module is used to extract features from the pre-processed multimodal user data and fuse them into user feature vectors; the graph construction module is used to construct a user knowledge graph and business according to the user feature vector through a feature clustering algorithm, dynamically associate the user knowledge graph and the business graph, and generate a comprehensive graph; the plan generation module is used to infer the comprehensive graph through a graph neural network, calculate the matching degree between user nodes and business nodes, and generate a personalized call plan; the voice call module is used to generate a voice call script and synthesize a voice file based on the personalized call plan, and initiate a call to the user through a voice outbound call device.

[0116] This embodiment also provides a computer device, which is suitable for the case of a user voice call method based on big data processing, including: a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the user voice call method based on big data processing proposed in the above embodiment.

[0117] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0118] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the user voice call method based on big data processing proposed in the above embodiment is implemented; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0119] In summary, the present invention collects and preprocesses multimodal user data to ensure high quality of data input; generates user feature vectors through feature extraction and fusion to fully explore the correlation of multimodal data; uses feature clustering to construct user knowledge graphs and business graphs, and dynamically associates to generate comprehensive graphs, thereby improving the ability to associate user needs with business features; uses graph neural network to infer the comprehensive graph, accurately calculates the matching degree between users and businesses, and generates personalized call plans; generates voice scripts and synthesizes voice files based on the plan, and efficiently reaches them through outbound calling equipment, thereby solving the problems of insufficient personalization of call content and inaccurate matching, and improving the accuracy of voice call services and user experience.

[0120] Example 2, referring to Table 1, is the second example of the present invention. To further verify the technical solution of the present invention, experimental simulation data of a user voice call method based on big data processing is provided.

[0121] In order to verify the effectiveness of the user voice call method based on big data processing and compare the shortcomings of the existing voice call system, an experimental scenario was designed to simulate the actual business application process. This experiment collected multimodal data from 5 users, including user voice data, historical behavior data, and business interaction data. The experimental process is divided into the following steps:

[0122] First, the experiment collected multimodal user data, including voice data, historical behavior data, and business data, through voice collection equipment, user behavior logs, and business interaction records, and preprocessed these data. The preprocessing steps include noise reduction of voice data, missing value filling and duplicate data cleaning of historical behavior data, as well as format standardization of business data and elimination of irrelevant information. Finally, the multimodal data is unified into the same feature space through normalization methods.

[0123] Secondly, the preprocessed multimodal data is subjected to feature extraction and fusion. BERT is used to extract speech semantic feature vectors, CNN to extract acoustic feature vectors, LSTM to extract historical behavior feature vectors, Word2Vec to extract business feature vectors, and the multimodal features are aligned in time series through the Attention mechanism. Subsequently, geometric interaction and linear superposition methods are used to fuse these modal features into user feature vectors to fully characterize the user's interests and needs.

[0124] Take user A as an example:

[0125] The parameters and their values ​​involved in the expression are shown in Table 1:

[0126] Table 1

[0127] Parameter name symbol Numeric Phonetic and semantic matching <![CDATA[F1]]> 0.78 Behavioral characteristics matching <![CDATA[F2]]> 0.74 Business feature matching <![CDATA[F3]]> 0.83 Non-linear weights α 1 Linear stacking weights γ 0.5 Number of modal features n 3

[0128]

[0129] The calculation results of user AE are shown in Table 2:

[0130] Table 2

[0131] user User feature vector F User A 1.435 User B 1.012 User C 1.574 User D 0.921 User E 1.429

[0132] Next, based on the user feature vector and business feature vector, the user knowledge graph and business knowledge graph are constructed through the feature clustering algorithm, and the comprehensive graph is generated using semantic association. The embedding vectors of the user nodes and business nodes in the comprehensive graph are initialized by HetGNN, and the feature information of neighboring nodes is dynamically aggregated in combination with the attention mechanism to generate a high-dimensional embedding vector. The matching degree between the user node and the business node is calculated using the matching degree formula, and the high and low correlation users are distinguished according to the set threshold T.

[0133] Take user A as an example:

[0134] The parameters and their values ​​involved in the expression are shown in Table 3:

[0135] Table 3

[0136] symbol describe value Hu[1] User A's voice matching degree 0.78 Hu[2] User A's behavioral characteristics matching degree 0.74 Hu[3] User A's business feature matching degree 0.83 He[1] Voice feature matching degree of service nodes 0.75 He[2] Behavioral feature matching degree of business nodes 0.65 He[3] Business feature matching degree of business nodes 0.80

[0137]

[0138] The calculation results of the comprehensive matching degree S(u,e) are shown in Table 4:

[0139] Table 4

[0140]

[0141]

[0142] Finally, a personalized or universal call plan is generated based on the matching results. For users whose matching degree is higher than the threshold, a personalized call script containing the opening remarks, business key content, interactive guidance and closing remarks is generated; for users whose matching degree is lower than the threshold, a universal call script is generated. Finally, the script is converted into a voice file using the TTS engine, and the speed, tone and timbre of the voice are dynamically adjusted to complete the voice call through the outbound calling device, verifying the significant advantages of this method in user matching and call accuracy.

[0143] The details are shown in Table 5 below:

[0144] Table 5

[0145]

[0146] Through the analysis of the data in the above table, it can be seen that the present invention accurately calculates the comprehensive matching degree S(u,e) between users and services by integrating voice semantics, behavioral characteristics and business characteristics, and dynamically generates personalized or general call plans based on the matching degree. For example, the comprehensive matching degree of user C is 0.84, and a personalized plan is generated to accurately meet the needs, while the matching degree of user D is 0.55, and a general plan is used to optimize resource allocation. This method significantly improves call efficiency, user satisfaction and resource utilization, reflecting the innovation and practical value of the technology.

[0147] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A user voice calling method based on big data processing, characterized in that: include, Collect multimodal user data and pre-process the multimodal user data; Extract features from the preprocessed multimodal user data and fuse them into user feature vectors; Through the feature clustering algorithm, the user knowledge graph and business are constructed according to the user feature vector, and the user knowledge graph and business graph are dynamically associated to generate a comprehensive graph; The comprehensive graph is inferred through the graph neural network, the matching degree between the user node and the business node is calculated, and a personalized call plan is generated; Based on the personalized call plan, a voice call script is generated and a voice file is synthesized, and a call is initiated to the user through a voice outbound calling device.

2. The user voice calling method based on big data processing according to claim 1, characterized in that: The multimodal user data includes voice data, historical behavior data and service data.

3. The user voice calling method based on big data processing as claimed in claim 2, characterized in that: The specific steps of preprocessing the multimodal user data are as follows: Remove noise from speech data through noise reduction algorithm; Use data cleaning to fill missing values ​​and remove duplicate data from historical behavior data; Eliminate irrelevant information and standardize the format of business data through data filtering and conversion rules; The normalization method is used to standardize the multimodal user data.

4. The user voice calling method based on big data processing as claimed in claim 3, characterized in that: The pre-processed multi-modal user data is subjected to feature extraction and fused into a user feature vector. The specific steps are as follows: Extract semantic feature vectors from user speech data through BER; Use CNN to extract acoustic feature vectors from speech data; Use LSTM to extract behavior feature vectors from historical behavior data; Extract business feature vectors from business data through Word2Vec; Use the Attention mechanism to align time series of different feature vectors; The geometric interaction of different modal feature vectors is carried out dimension by dimension to capture the high-order nonlinear relationship between the modes. At the same time, the linear superposition of all modal feature vectors is supplemented and fused into the user feature vector, which is expressed as: Among them, F is the user feature vector, F i represents the i-th modal eigenvector, ||F i ||2 represents the L2 norm of the ith modal feature vector, γ is the weight coefficient controlling the linear superposition part in the final fusion result, α is the hyperparameter for adjusting the nonlinear influence of the modal feature, n represents the total number of modal features, and F j Represents the eigenvector of the jth mode.

5. The user voice calling method based on big data processing according to claim 4, characterized in that: The feature clustering algorithm is used to construct user knowledge graphs and services based on user feature vectors, and the knowledge graph and the business graph are dynamically associated to generate a comprehensive graph. The specific steps are as follows: Based on the user feature vector F, the Euclidean distance between the user feature vector and the cluster center is calculated by the feature clustering algorithm, and the expression is: Among them, d ( F i ,C k) Represents the user feature vector F and the i-th cluster center C k The Euclidean distance between k represents the user feature vector of the kth cluster center, F i,r represents the value of the i-th user feature vector on the r-th feature dimension, C k,r represents the value of the k-th cluster center on the r-th feature dimension, and m represents the number of dimensions of the user feature vector; Based on the Euclidean distance d between the user feature vector and the cluster center ( F i ,C k) , dynamically assign users to k cluster groups until the cluster centers converge, and finally generate k cluster groups; Each user F i As an independent node, each cluster group is converted into a group node, and the user node and the group node are connected through the belonging relationship. The interest tags and behavior patterns in the user feature vector are extracted to generate feature nodes, which are associated with the user node and the group node through the owning feature relationship to form a structured user knowledge graph; Extract entities from business data and generate business nodes; Generate a business feature node according to the business feature vector; Connect the business node with its feature node through attributes to establish a business relationship, and store the business node, business feature node and business relationship in a structured manner to form a business knowledge graph; Based on user behavior data, the user knowledge graph and the business knowledge graph are semantically associated using user features and business features to generate a comprehensive graph.

6. The user voice calling method based on big data processing according to claim 5, characterized in that: The graph neural network is used to infer the comprehensive graph, calculate the matching degree between the user node and the service node, and generate a personalized call plan. The specific steps are as follows: HetGNN is used to initialize the feature vectors of user nodes and business nodes in the comprehensive graph, and the attention mechanism is used to dynamically aggregate the feature information of neighboring nodes to generate a high-dimensional embedding vector for each node; According to the high-dimensional embedding vector of each node, combined with the semantic features of the path in the comprehensive graph, the matching degree between the user node and the business node is calculated, and the expression is: Among them, S ( u,e ) represents the matching degree between user node u and service node e, H u[ i ] represents the value of the embedding vector of user node u in the xth dimension, H e[ x ] represents the value of the embedding vector of business node e in the xth dimension, H u represents the normalized norm of user node u, H e represents the normalized norm of the business node e, d represents the dimension of the embedding vector, and x is the index variable of the dimension; Define a relevance threshold T based on historical data; When S ( u,e ) ≥T, the current user u is considered to be highly relevant to the current service e, and the call content focuses on the content related to the current service; When S ( u,e ) When <T, it is considered that the current user u has a low correlation with the current service e, and a general voice call solution is adopted.

7. The user voice calling method based on big data processing according to claim 6, characterized in that: Based on the personalized call plan, the voice call script is generated and the voice file is synthesized, and a call is initiated to the user through the voice outbound call device. The specific steps are as follows: Generate a voice call script including opening remarks, business key points, interactive guidance and closing remarks based on the personalized call plan; The voice script is converted into a voice file through the TTS engine, and the speaking speed, intonation and timbre are adjusted based on the interest tags in the user feature vector, and a call is initiated to the user using an outbound calling device.

8. A user voice call system based on big data processing, based on the user voice call method based on big data processing according to any one of claims 1 to 7, characterized in that: Including data collection module, feature fusion module, graph construction module, call plan generation module and voice call module; A data collection module, used to collect multimodal user data and pre-process the multimodal user data; The feature fusion module is used to extract features from the pre-processed multi-modal user data and fuse them into a user feature vector; The graph construction module is used to construct user knowledge graphs and businesses based on user feature vectors through feature clustering algorithms, dynamically associate user knowledge graphs with business graphs, and generate a comprehensive graph; The solution generation module is used to infer the comprehensive graph through the graph neural network, calculate the matching degree between the user node and the service node, and generate a personalized call solution; The voice call module is used to generate a voice call script and synthesize a voice file based on a personalized call plan, and initiate a call to the user through a voice outbound call device.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the user voice call method based on big data processing according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the user voice call method based on big data processing described in any one of claims 1 to 7 are implemented.

Images