Intelligent generation method, device and storage medium of multimedia content under 5G new call

Abstract

The present invention discloses a method, device and storage medium for intelligently generating multimedia content under a 5G new call, and relates to the field of communication technology. The method includes: obtaining Q rounds of conversation information of a target call; performing semantic understanding analysis to determine Q semantic understanding feature sets; performing concentrated feature screening to determine a concentrated semantic understanding feature set; performing business association identification to obtain a first customer service scenario; performing iterative association identification on the Q rounds of conversation information according to the sequence of the first business keyword set and the Q rounds of conversation information to determine an iterative association feature set; identifying the iterative association feature set to determine the target multimedia content, and displaying the target multimedia content on the call page of the target call. The technical problems of low multimedia content generation efficiency and poor correlation between generated content and call semantics in the prior art are solved, and the technical effects of improving the efficiency of multimedia content generation and enhancing the correlation between generated content and call semantics are achieved.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a method, device and storage medium for intelligently generating multimedia content under 5G new calls. Background Art

[0002] With the rapid development of 5G communication technology, the new call function has gradually become an important way of interaction. Compared with traditional voice calls, the 5G new call function can support the real-time generation and display of multimedia content, such as text, pictures, videos and other forms of content, which greatly improves the user experience. However, the current multimedia content generation technology still has significant shortcomings. On the one hand, the existing technology has limited understanding and processing capabilities of call semantics, and cannot accurately extract and associate key information in multiple rounds of conversations, resulting in the lack of pertinence and scenario-based generated multimedia content. On the other hand, the generation process usually takes a long time, which makes it difficult to meet the requirements of real-time and efficient response in 5G new calls. In addition, the generated multimedia content has a low correlation with actual business needs, which makes it difficult to meet the refined service needs of users and customer service scenarios. Summary of the invention

[0003] The present application provides a method, device and storage medium for intelligently generating multimedia content under 5G new calls, which solves the technical problems in the prior art of low efficiency in generating multimedia content and poor semantic correlation between generated content and calls.

[0004] In view of the above problems, the present application provides a method, device and storage medium for intelligently generating multimedia content under 5G new calls.

[0005] In a first aspect of the present application, a method for intelligently generating multimedia content under a 5G new call is provided, the method comprising:

[0006] Acquire Q rounds of dialogue information of a target call within a preset window, where Q is an integer greater than or equal to 1; traverse the Q rounds of dialogue information to perform semantic understanding analysis and determine Q semantic understanding feature sets; perform feature concentration screening on the Q semantic understanding feature sets to determine a concentrated semantic understanding feature set; perform business association identification based on the concentrated semantic understanding feature set to obtain a first customer service scenario, and use the first customer service scenario as a first content generation constraint; obtain a first business keyword set in combination with the first content generation constraint, perform iterative association identification on the Q rounds of dialogue information according to the sequence of the first business keyword set and the Q rounds of dialogue information, and determine an iterative association feature set; call a multimedia content generator to identify the iterative association feature set, determine the target multimedia content, and display the target multimedia content on the call page of the target call.

[0007] The second aspect of the present application provides an electronic device, including: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the method for intelligent generation of multimedia content under the 5G new call provided in the present application.

[0008] According to a third aspect of an embodiment of the present application, a computer-readable storage medium is provided, which stores a computer program. When the program is executed by a processor, the method for intelligently generating multimedia content under the 5G new call provided in the present application is implemented.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0010] First, obtain Q rounds of dialogue information of the target call within a preset window, where Q is an integer greater than or equal to 1. Then, traverse the Q rounds of dialogue information to perform semantic understanding analysis and determine Q semantic understanding feature sets. Further, perform feature concentration screening on the Q semantic understanding feature sets to determine a concentrated semantic understanding feature set. Next, perform business association identification based on the concentrated semantic understanding feature set to obtain a first customer service scenario, and use the first customer service scenario as a first content generation constraint. Then, obtain a first business keyword set in combination with the first content generation constraint, perform iterative association identification on the Q rounds of dialogue information according to the sequence of the first business keyword set and the Q rounds of dialogue information, and determine an iterative association feature set. Finally, call a multimedia content generator to identify the iterative association feature set, determine the target multimedia content, and display the target multimedia content on the call page of the target call. The technical problems of low multimedia content generation efficiency and poor association between generated content and call semantics in the prior art are solved, and the technical effects of improving multimedia content generation efficiency and enhancing the association between generated content and call semantics are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0012] Figure 1 A schematic diagram of a method for intelligently generating multimedia content under a 5G new call provided in an embodiment of the present application;

[0013] Figure 2 A schematic diagram of a process for iterative association identification in a method for intelligently generating multimedia content under a 5G new call provided in an embodiment of the present application;

[0014] Figure 3This is a schematic diagram of the structure of an exemplary electronic device of the present application.

[0015] Description of the reference numerals: processor 31 , memory 32 , input device 33 , output device 34 . DETAILED DESCRIPTION

[0016] The present application solves the technical problems in the prior art of low efficiency in multimedia content generation and poor semantic correlation between generated content and call by providing a method, device and storage medium for intelligent generation of multimedia content under 5G new calls.

[0017] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0018] It should be noted that the terms "including" and "having" are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules that are not explicitly listed or are inherent to these processes, methods, products or devices.

[0019] Embodiment 1, as Figure 1 As shown, the embodiment of the present application provides a method for intelligently generating multimedia content under a 5G new call, wherein the method includes:

[0020] Get Q rounds of conversation information of the target call within a preset window, where Q is an integer greater than or equal to 1.

[0021] The target call refers to the call instance that currently needs to generate multimedia content, including the interactive information between the two parties of the call (such as voice, text, video, etc.); the preset window is a fixed time range used to limit the conversation information that needs to be analyzed. The preset window can be flexibly set according to business needs, such as in minutes (such as the conversation content in the past 5 minutes).

[0022] Within the preset time window, obtain Q rounds of dialogue information of the target call. Each round of dialogue is defined as a complete interaction between the user and the system. For example, a user asks a question and the system returns a response, which is counted as one round of dialogue.

[0023] The Q rounds of dialogue information are traversed to perform semantic understanding analysis and determine Q semantic understanding feature sets.

[0024] Traverse the Q rounds of dialogue information obtained and perform semantic understanding analysis on each round of dialogue. Through natural language processing technology, extract the semantic features in the dialogue and form Q semantic understanding feature sets, which are used to describe the core semantics of the dialogue content.

[0025] Furthermore, the Q rounds of dialogue information are traversed to perform semantic understanding analysis, and Q semantic understanding feature sets are determined, including:

[0026] Acquire multiple sample conversation information and multiple sample semantic understanding feature sets as training sample sets; divide the training sample set into a training set and a verification set according to a preset division ratio; use the training set to train a framework built on a convolutional neural network, and use the verification set to supervise and verify the training results, update network parameters according to the supervised verification results until the training converges, and obtain a trained semantic understanding network layer; use the semantic understanding network layer to perform semantic understanding analysis on the Q rounds of conversation information respectively, and obtain the Q semantic understanding feature sets.

[0027] Collect multiple sample conversation information from historical conversation data, including text data (such as content after speech-to-text conversion), conversation turns, and context-related information; manually annotate the sample conversation information to generate a sample semantic understanding feature set, which includes core keywords, sentiment classification labels, context-related labels, etc. The sample semantic understanding feature set can accurately reflect the key information in the conversation, such as entities, events, relationships, etc. According to the preset ratio (for example, 80:20 or 70:30), the sample data is divided into a training set and a validation set. The training set is used for model training of the semantic understanding network, and the validation set is used to supervise and verify the model effect during the training process and evaluate the generalization ability and accuracy of the model.

[0028] The framework of the semantic understanding model is built based on the convolutional neural network (CNN). When processing text data, CNN usually converts the text into a sequence of word vectors, and then extracts text features through structures such as convolutional layers and pooling layers. Specifically, the framework built based on the convolutional neural network includes input layer, convolutional layer, pooling layer, fully connected layer, and output layer; among them, the input layer: the conversation text data is first converted into a fixed-length semantic vector through an embedding layer (such as Word2Vec, BERT); the convolutional layer: extracts local semantic features in the conversation text; the pooling layer: dimension reduction and feature selection, retaining key information; the fully connected layer: maps the extracted features to the target output space (such as feature vectors, sentiment classification); the output layer: generates a specific set of semantic understanding features (such as semantic embedding vectors, keyword weights, etc.). The convolutional neural network (CNN) model is trained using the training set, and the local semantic features of the text are extracted through forward propagation, and the network parameters are updated through back propagation combined with the loss function. At the same time, the training effect is supervised and verified using the validation set, and the hyperparameters are continuously adjusted until the model converges, and finally the trained semantic understanding network layer is obtained.

[0029] For each round of dialogue in the Q rounds of dialogue information, the trained semantic understanding network layer is used to perform semantic understanding analysis. The semantic understanding network layer will output the corresponding semantic understanding feature sets, which reflect the key information and semantic content in the dialogue.

[0030] The Q semantic understanding feature sets are subjected to feature centralized screening to determine a centralized semantic understanding feature set.

[0031] Through methods such as feature weight calculation or feature clustering, Q semantic understanding feature sets are screened to remove redundant or noise information and determine a concentrated semantic understanding feature set. The concentrated semantic understanding feature set retains the keywords, semantic vectors and sentiment labels most relevant to the target call and has a higher feature density and correlation.

[0032] Furthermore, performing feature centralized screening on the Q semantic understanding feature sets to determine a centralized semantic understanding feature set includes:

[0033] The Q semantic understanding feature sets are clustered according to preset feature similarities to obtain multiple semantic understanding feature similarity sets; the number of features in the multiple semantic understanding feature similarity sets are respectively counted to obtain multiple feature quantities; the M semantic understanding feature similarity sets that are in the top M positions among the multiple feature quantities are extracted, and the M semantic understanding feature similarity sets are unioned to obtain the concentrated semantic understanding feature set, where M is an integer greater than or equal to 1 and less than or equal to Q.

[0034] Preferably, a similarity calculation method (such as cosine similarity and Euclidean distance) is used to quantify the similarity between features, and the features in Q semantic understanding feature sets are clustered according to a preset feature similarity threshold to generate multiple semantic understanding feature similarity sets, each semantic understanding feature similarity set represents a feature group whose similarity is higher than the threshold; the number of features in the multiple semantic understanding feature similarity sets obtained after clustering is counted separately to obtain multiple feature quantities; the first M semantic understanding feature similarity sets containing the largest number of features are selected, where M is an integer greater than or equal to 1 and less than or equal to Q, and can be dynamically set according to specific scenario requirements; the selected M semantic understanding feature similarity sets are unioned to integrate all high-value features to form a concentrated semantic understanding feature set.

[0035] Furthermore, multiple semantic understanding feature similarity sets are obtained, including:

[0036] A plurality of semantic understanding feature sets are randomly extracted from the Q semantic understanding feature sets, and clustering head particles are authenticated for the plurality of semantic understanding feature sets; when the authentication is passed, the plurality of semantic understanding feature sets are used as a plurality of clustering head particles, and the Q semantic understanding feature sets are clustered according to the preset feature similarity to obtain a plurality of clustering tail particle sets, wherein each clustering head particle corresponds to a clustering tail particle set; the plurality of clustering head particles are aggregated with the corresponding plurality of clustering tail particle sets to obtain a plurality of semantic understanding feature similarity sets.

[0037] A certain number of semantic understanding feature sets are randomly selected from Q semantic understanding feature sets as potential clustering head particles. Clustering head particles are feature sets with strong representativeness and discrimination in the semantic understanding feature set, and are used as candidates for the center points of subsequent clustering; these potential clustering head particles are authenticated to ensure that the selected clustering head particles can represent different parts of the entire feature set space. When the clustering head particles are authenticated, the similarity between each semantic understanding feature set and multiple clustering head particles is calculated using feature similarity indicators (such as cosine similarity). According to the preset similarity threshold, the features similar to a certain clustering head particle are classified into the corresponding clustering tail particle set. Each clustering head particle eventually forms a clustering tail particle set, which contains all similar features; all clustering head particles and their corresponding clustering tail particle sets are integrated to form multiple semantic understanding feature similarity sets.

[0038] Furthermore, clustering head particle authentication is performed on the plurality of semantic understanding feature sets, including:

[0039] The multiple semantic understanding feature sets are enumerated in pairs to obtain multiple enumeration combinations; it is determined whether there is an enumeration combination whose combination similarity among the multiple enumeration combinations is greater than a preset combination similarity threshold; if so, the clustering head particle authentication fails, and multiple semantic understanding feature sets are randomly extracted from the Q semantic understanding feature sets for authentication again; if not, the authentication passes.

[0040] Preferably, multiple randomly selected semantic understanding feature sets are combined in pairs to form all possible enumeration combination sets; for each enumeration combination, its combination similarity is calculated, and the combination similarity reflects the degree of semantic overlap between two semantic understanding feature sets. Low similarity indicates that the feature set has good discrimination and is suitable as a candidate for clustering head particle; a combination similarity threshold is set to determine whether there are too similar enumeration combinations; all enumeration combinations are traversed to check whether there is a combination similarity greater than a preset combination similarity threshold; if there is an enumeration combination whose similarity exceeds the threshold, the clustering head particle authentication fails. At this time, it is necessary to randomly re-extract a new feature set from the Q semantic understanding feature sets and re-perform the authentication process; if the similarities of all enumeration combinations are less than or equal to the preset combination similarity threshold, the authentication is passed, and the selected semantic understanding feature set can be used as a candidate set for clustering head particles.

[0041] Business association identification is performed based on the centralized semantic understanding feature set to obtain a first customer service scenario, and the first customer service scenario is used as a first content generation constraint.

[0042] Based on a centralized semantic understanding feature set, the specific business scenario to which the user conversation content belongs (such as product consultation, after-sales service, order inquiry, etc.) is identified, and the first customer service scenario is generated as a constraint condition for subsequent multimedia content generation. Specifically, a label library covering a variety of customer service business scenarios is created, such as product consultation, after-sales service, order inquiry, complaint handling and other business scenario labels; a set of predefined feature sets corresponding to each business scenario label is stored in the scenario feature library, including a keyword list, a semantic vector template, sentiment distribution, etc.; a classification model (such as support vector machine SVM, deep learning classification model, etc.) is used to classify the business scenarios of the centralized feature set, and a neural network (such as BERT classification model) can be used in combination with semantic understanding features to classify business scenarios; core keywords are extracted from the centralized semantic understanding feature set and matched with the keyword list in the scenario feature library, and the matching score can be calculated based on the frequency of keyword occurrence, weight, etc.; the similarity between the semantic vector of the centralized semantic understanding feature set and the semantic vector template in the scenario feature library is calculated; the sentiment distribution of the centralized semantic understanding feature set is analyzed to see if it is consistent with the sentiment characteristics of the target scenario, for example, complaint handling scenarios are usually accompanied by negative emotions, while product consultation scenarios are mainly neutral or positive emotions; the keyword matching score, semantic vector similarity and sentiment distribution score are weighted and fused to calculate the total business association score; according to the business association score, the scenario with the highest score is selected from the business scenario label library as the first customer service scenario.

[0043] A first business keyword set is obtained in combination with the first content generation constraint, and iterative association recognition is performed on the Q-round dialogue information according to the sequence of the first business keyword set and the Q-round dialogue information to determine an iterative association feature set.

[0044] The first customer service scenario obtained from business association identification is used as a constraint condition for content generation; based on the first content generation constraint, a relevant first business keyword set is extracted, and the first business keyword set represents the core concepts of the business, user intentions, question types, etc.; using the first business keyword set, Q rounds of dialogue information are iteratively processed; in each round of dialogue information, content related to business keywords is identified, including direct mentions of keywords, synonyms, antonyms, contextual associations, etc.; as the iteration proceeds, business-related features and information are continuously accumulated to form an iterative association feature set; after completing all rounds of dialogue information processing, all identified business-related features and information are summarized and organized to form a final iterative association feature set.

[0045] Furthermore, if Figure 2As shown, a first business keyword set is obtained in combination with the first content generation constraint, and the Q semantic understanding feature sets are iteratively associated with each other according to the sequence of Q rounds of dialogue information corresponding to the first business keyword set and the Q semantic understanding feature sets, to determine an iterative associated feature set, including:

[0046] Obtain a set of historical conversation information within a preset historical window constrained by the first content generation constraint, extract keywords whose appearance frequency exceeds a preset frequency in the historical conversation information, and obtain the first business keyword set; perform associated text capture on the first round of conversation information in the Q rounds of conversation information based on the first business keyword set to obtain a first memory vector; perform iterative association integration on the second round of conversation information in the Q rounds of conversation information based on the first memory vector and the first business keyword set to obtain a second memory vector; perform iterative association integration on the Q rounds of conversation information in sequence using the first memory vector, the second memory vector, and the first business keyword set to obtain a Qth memory vector; perform semantic feature recognition on the Qth memory vector to obtain the iterative association feature set.

[0047] Specifically, according to the first content generation constraint, a set of historical dialogue information within a preset historical window is obtained, wherein the historical window can be based on a time range; keywords whose occurrence frequency exceeds a preset frequency are extracted from the historical dialogue information to form a first business keyword set; the first business keyword set is used to capture associated text of the first round of dialogue information in the Q rounds of dialogue information, and keyword matching and semantic similarity calculation (such as semantic embedding based on BERT or Word2Vec) can be used to capture the first round of dialogue text, identify semantic fragments related to the business keyword set, and generate a first memory vector to represent the feature representation of the first round of dialogue information; the first memory vector and the first business keyword set are used to capture the second round of dialogue information Perform semantic matching and association expansion; capture the associated content based on keyword matching and context semantic similarity, and merge the key features in the first memory vector with the associated features of the second round to generate a second memory vector; repeat the above steps, using the previous memory vector, the current round of dialogue information and the first business keyword set to iteratively associate and integrate the subsequent rounds of dialogue information in turn; after each round of iteration, a new memory vector is formed, which contains all the key information and semantics from the first round to the current round; after completing the iterative association integration of all rounds, the Qth memory vector is obtained; use the semantic understanding network layer to perform semantic feature recognition on the Qth memory vector, extract features closely related to the business, and form an iterative association feature set.

[0048] Furthermore, obtaining the first memory vector includes:

[0049] Using the first business keyword set as an index, keyword matching is performed on the first round of dialogue information in the Q rounds of dialogue information to obtain a first round of matching keyword set; using the first round of matching keyword set as a capture target, text content capture is performed on the first round of dialogue information to obtain a first round of capture result, and the first round of capture result is added to an initially empty memory vector to obtain a first memory vector.

[0050] Specifically, taking the first business keyword set as an index, keyword matching is performed on the first round of dialogue information in the Q rounds of dialogue information, and a semantic matching method (such as BERT-based word embedding similarity) can be used to identify words semantically related to the business keywords to obtain a first round of matching keyword set, which represents keywords in the first round of dialogue that are directly or indirectly related to the business keyword set; taking the first round of matching keyword set as a capture target, text content capture is performed on the first round of dialogue information, and the first round of capture results can be obtained by extracting contextual content within a certain range around the keywords (such as sentences or text fragments of fixed length), which are a set of text content fragments related to the first round of matching keyword set; the first round of capture results are added to the initially empty memory vector to form a first memory vector.

[0051] Further, including:

[0052] Using the first business keyword set as an index, keyword matching is performed on the second round of dialogue information in the Q rounds of dialogue information to obtain a second round of matching keyword set; using the second round of matching keyword set and the first memory vector as capture targets, text capture is performed on the second round of dialogue information to obtain a second round of capture result, and the second round of capture result is added to the second memory vector.

[0053] Specifically, using the first business keyword set as an index, keyword matching is performed on the second round of dialogue information in the Q rounds of dialogue information. Similar to the processing of the first round of dialogue information, the match can be a complete match or a partial match. Using the second round matching keyword set and the first memory vector as capture targets, text capture is performed on the second round of dialogue information. The captured content includes not only text directly related to the second round matching keywords, but also text associated with the information in the first memory vector. The second round capture results are added to the second memory vector.

[0054] A multimedia content generator is called to identify the iterative association feature set, determine target multimedia content, and display the target multimedia content on a call page of the target call.

[0055] The multimedia content generator is a system that can generate corresponding multimedia content based on input features. After receiving the iterative associated feature set, the multimedia content generator performs internal processing and analysis. The generator identifies the multimedia content that is most relevant or most in line with the requirements of the target call, including pictures, videos, audio clips or other forms of multimedia elements; once the target multimedia content is determined, it will be sent to the call page of the target call. The call page is the interface that the user sees when making a call, usually including call information, control buttons, etc.; on the call page, the target multimedia content will be displayed to the user in an appropriate manner so that the user can view or interact during the call.

[0056] In summary, the embodiments of the present application have at least the following technical effects:

[0057] First, obtain Q rounds of dialogue information of the target call within a preset window, where Q is an integer greater than or equal to 1. Then, traverse the Q rounds of dialogue information to perform semantic understanding analysis and determine Q semantic understanding feature sets. Further, perform feature concentration screening on the Q semantic understanding feature sets to determine a concentrated semantic understanding feature set. Next, perform business association identification based on the concentrated semantic understanding feature set to obtain a first customer service scenario, and use the first customer service scenario as a first content generation constraint. Then, obtain a first business keyword set in combination with the first content generation constraint, perform iterative association identification on the Q rounds of dialogue information according to the sequence of the first business keyword set and the Q rounds of dialogue information, and determine an iterative association feature set. Finally, call a multimedia content generator to identify the iterative association feature set, determine the target multimedia content, and display the target multimedia content on the call page of the target call. The technical problems of low multimedia content generation efficiency and poor association between generated content and call semantics in the prior art are solved, and the technical effects of improving multimedia content generation efficiency and enhancing the association between generated content and call semantics are achieved.

[0058] Embodiment 2: Figure 3 The schematic structural diagram of the electronic device provided in the second embodiment of the present invention shows a block diagram of an exemplary electronic device suitable for implementing the implementation mode of the present invention. Figure 3 The electronic device shown is only an example and should not limit the functions and scope of use of the embodiments of the present invention. Figure 3 As shown, the electronic device includes a processor 31, a memory 32, an input device 33 and an output device 34; the number of processors 31 in the electronic device can be one or more. Figure 3 Taking a processor 31 as an example, the processor 31, memory 32, input device 33 and output device 34 in the electronic device can be connected through a bus or other means. Figure 3 The example of connecting through bus is taken in the following.

[0059] Embodiment 3, based on the same inventive concept as the method for intelligently generating multimedia content under 5G new calls in the above-mentioned embodiments, this embodiment provides a computer-readable storage medium, which can be used to store software programs, computer executable programs and modules, such as program instructions / modules corresponding to the method for intelligently generating multimedia content under 5G new calls in the embodiments of this application. The processor executes various functional applications and data processing of the computer device by running the software programs, instructions and modules stored in the memory, that is, the method for intelligently generating multimedia content under 5G new calls is implemented.

[0060] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned specific embodiments of this specification are described. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0061] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0062] This specification and drawings are merely exemplary illustrations of the present application and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, a person skilled in the art may make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application intends to include these modifications and variations.

Claims

1. A method for intelligently generating multimedia content under a 5G new call, characterized in that: The method comprises: Obtaining Q rounds of conversation information of the target call within a preset window, where Q is an integer greater than or equal to 1; Traversing the Q rounds of dialogue information to perform semantic understanding analysis and determine Q semantic understanding feature sets; Clustering the Q semantic understanding feature sets according to preset feature similarities to obtain multiple semantic understanding feature similarity sets; Counting the number of features in the plurality of semantic understanding feature similarity sets respectively to obtain a plurality of feature quantities; Extracting M semantic understanding feature similarity sets that are located in the first M positions among the multiple feature quantities, and performing a union of the M semantic understanding feature similarity sets to obtain a concentrated semantic understanding feature set, where M is an integer greater than or equal to 1 and less than or equal to Q; Performing business association identification based on the centralized semantic understanding feature set to obtain a first customer service scenario, and using the first customer service scenario as a first content generation constraint; Acquiring a first business keyword set in combination with the first content generation constraint, iteratively identifying associations of the Q semantic understanding feature sets according to the sequence of Q rounds of dialogue information corresponding to the first business keyword set and the Q semantic understanding feature sets, and determining an iterative association feature set, including: Acquire a set of historical conversation information within a preset historical window constrained by the first content generation, extract keywords whose appearance frequency exceeds a preset frequency in the historical conversation information, and obtain the first business keyword set; Capturing the associated text of the first round of dialogue information in the Q rounds of dialogue information based on the first business keyword set to obtain a first memory vector; Iteratively associating and integrating the second round of dialogue information in the Q rounds of dialogue information based on the first memory vector and the first business keyword set to obtain a second memory vector; Iteratively associating and integrating the Q rounds of dialogue information in sequence using the first memory vector, the second memory vector, and the first business keyword set to obtain a Qth memory vector; Performing semantic feature recognition on the Qth memory vector to obtain the iterative association feature set; Calling a multimedia content generator to identify the iterative association feature set, determine target multimedia content, and display the target multimedia content on a call page of the target call; Wherein, obtaining a plurality of semantic understanding feature similarity sets includes: randomly extracting a plurality of semantic understanding feature sets from the Q semantic understanding feature sets, and performing clustering head particle authentication on the plurality of semantic understanding feature sets; when the authentication is passed, taking the plurality of semantic understanding feature sets as a plurality of clustering head particles, clustering the Q semantic understanding feature sets according to the preset feature similarity, and obtaining a plurality of clustering tail particle sets, wherein each clustering head particle corresponds to a clustering tail particle set; and summarizing the plurality of clustering head particles and the corresponding plurality of clustering tail particle sets to obtain a plurality of semantic understanding feature similarity sets; Among them, clustering head particle authentication is performed on the multiple semantic understanding feature sets, including: enumerating the multiple semantic understanding feature sets in pairs to obtain multiple enumeration combinations; judging whether there is an enumeration combination whose combination similarity of the multiple enumeration combinations is greater than a preset combination similarity threshold; if so, the clustering head particle authentication fails, and multiple semantic understanding feature sets are randomly extracted from the Q semantic understanding feature sets for authentication again; if not, the authentication passes.

2. The method for intelligently generating multimedia content under the 5G new call according to claim 1, characterized in that: include: Using the first business keyword set as an index, performing keyword matching on the first round of dialogue information in the Q rounds of dialogue information to obtain a first round of matching keyword set; The first-round matching keyword set is used as a capture target, the text content of the first-round dialogue information is captured to obtain a first-round capture result, and the first-round capture result is added to an initially empty memory vector to obtain a first memory vector.

3. The method for intelligently generating multimedia content under the 5G new call according to claim 2, characterized in that: include: Using the first business keyword set as an index, performing keyword matching on the second round of dialogue information in the Q rounds of dialogue information to obtain a second round of matching keyword set; Taking the second-round matching keyword set and the first memory vector as capture targets, text capture is performed on the second-round dialogue information to obtain a second-round capture result, and the second-round capture result is added to the second memory vector.

4. The method for intelligently generating multimedia content under the 5G new call according to claim 1, characterized in that: The Q rounds of dialogue information are traversed to perform semantic understanding analysis and determine Q semantic understanding feature sets, including: Acquire multiple sample conversation information and multiple sample semantic understanding feature sets as training sample sets; Dividing the training sample set into a training set and a validation set according to a preset division ratio; Use the training set to train the framework based on the convolutional neural network, and use the verification set to supervise and verify the training results. Update the network parameters according to the supervised verification results until the training converges to obtain the trained semantic understanding network layer. The semantic understanding network layer is used to perform semantic understanding analysis on the Q rounds of dialogue information to obtain the Q semantic understanding feature sets.

5. An electronic device, characterized in that: The electronic device comprises: A memory for storing executable instructions; The processor is used to implement the method for intelligently generating multimedia content under the 5G new call described in any one of claims 1 to 4 when executing the executable instructions stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, it implements the method for intelligently generating multimedia content under a 5G new call as described in any one of claims 1-4.

Images