Semantic understanding-based service fulfillment method and device, medium and product

Abstract

The embodiment of the invention relates to a business opening method and device based on semantic understanding, a medium and a product. The method comprises the steps of performing feature extraction on an input service opening text based on multi-source data to obtain a first semantic vector; performing field classification on the first semantic vector based on a classification model to obtain a target field; based on the target domain, obtaining structured information of the service opening text through a domain fine tuning large model; obtaining a second semantic vector according to the structured information, and matching an optimal operation path from the graph database based on the second semantic vector; and according to a preset template corresponding to the optimal operation path and in combination with the large model, outputting a service opening scheme to open the service.

Description

Semantic understanding-based service activation methods, equipment, media, and products Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a service activation method, device, medium, and product based on semantic understanding. Background Technology

[0002] Currently, automatic service activation mainly relies on rule-based methods, general natural language models, and static knowledge base matching methods. However, all of these methods suffer from low efficiency and poor accuracy. Summary of the Invention

[0003] To address the aforementioned issues, this disclosure provides a service activation method, device, medium, and product based on semantic understanding.

[0004] According to a first aspect of the present disclosure, a service activation method based on semantic understanding is provided. The method includes: extracting features from an input service activation text based on multi-source data to obtain a first semantic vector; classifying the first semantic vector into a domain based on a classification model to obtain a target domain; obtaining structured information of the service activation text by fine-tuning a large model based on the target domain; obtaining a second semantic vector based on the structured information, and matching an optimal operation path from a graph database based on the second semantic vector, wherein the graph database is a graphical knowledge base; and outputting a service activation scheme to activate the service based on a preset template corresponding to the optimal operation path and in combination with the large model.

[0005] In one implementation, the step of extracting features from the input service activation text based on multi-source data to obtain a first semantic vector includes: dividing the service activation text into multiple sentences, inputting the multiple sentences into a bidirectional pre-trained language model to generate multiple sub-vectors; and performing average pooling on the multiple sub-vectors to obtain the first semantic vector.

[0006] In one implementation, the step of classifying the first semantic vector based on a classification model to obtain a target domain includes: obtaining a model collaboration weight factor based on the classification confidence of a first classification model and the validation accuracy of a second classification model, wherein the first classification model is a lightweight classification model and the second classification model is a large-scale domain classification model; classifying the first semantic vector based on the model collaboration weight factor using the first classification model to obtain the target domain; or classifying the first semantic vector based on the model collaboration weight factor using both the first classification model and the second classification model to obtain the target domain.

[0007] In one implementation, obtaining the structured information of the service activation text based on the target domain and by fine-tuning a large model in the domain includes: obtaining a keyword mask matrix of the target domain; using the keyword mask matrix as an attention gating mechanism to enhance the large model with target domain attention, thereby obtaining the fine-tuned large model in the domain; and processing the service activation text based on the fine-tuned large model in the domain to obtain the structured information of the service activation text.

[0008] In one implementation, obtaining a second semantic vector based on the structured information and matching the optimal operation path from the graph database based on the second semantic vector includes: encoding the structured information into a second semantic vector using a bidirectional pre-trained language model; obtaining a comprehensive score based on the semantic similarity between the second semantic vector and node vectors in the graph database, and the historical success rate of the node paths; and matching the optimal operation path from the graph database based on the comprehensive score.

[0009] In one implementation, the step of outputting a service activation plan to activate the service based on a preset template corresponding to the optimal operation path and combined with a large model includes: automatically generating the service activation plan through a template engine based on the preset template, and dynamically adjusting the proportion of template-filled content and freely generated content from the large model in the service activation plan through a balance parameter, wherein the balance parameter is obtained based on the information entropy of the template-filled content and the user's rating of the accuracy of the plan.

[0010] In one embodiment, the method further includes: triggering an update step for the knowledge base if the difference between the user-reported correction operation step and the optimal operation path matching the knowledge base exceeds an update sensitivity threshold.

[0011] In one implementation, the method further includes: calculating feedback contribution weights based on user roles and historical correction behaviors; and incrementally training the large model using a loss function weighted and enhanced by the feedback contribution weights.

[0012] According to a second aspect of the present disclosure, a service activation device based on semantic understanding is provided. The device includes: a first acquisition module, configured to extract features from an input service activation text based on multi-source data to obtain a first semantic vector; a second acquisition module, configured to classify the first semantic vector into a domain based on a classification model to obtain a target domain; a third acquisition module, configured to obtain structured information of the service activation text based on the target domain by fine-tuning a large model within the domain; a fourth acquisition module, configured to obtain a second semantic vector based on the structured information and match an optimal operation path from a graph database based on the second semantic vector, wherein the graph database is a graphical knowledge base; and an output module, configured to output a service activation scheme to activate the service based on a preset template corresponding to the optimal operation path and in combination with the large model.

[0013] According to a third aspect of the present disclosure, an electronic device is provided, comprising: a memory having a computer program stored thereon; and a processor for executing the computer program in the memory to implement the steps of the method according to any one of the first aspects.

[0014] According to a fourth aspect of the present disclosure, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the method described in any of the first aspects.

[0015] According to a fifth aspect of the present disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method described in any one of the first aspects.

[0016] The embodiments disclosed herein can achieve the following beneficial effects: The embodiments disclosed herein determine the business domain through a classification model and match the optimal operation path for business activation based on a dynamic knowledge base, thereby realizing intelligent and automated operation of the business activation process, effectively improving the accuracy of business operation step generation and business activation efficiency.

[0017] Other features and advantages of this disclosure will be described in detail in the following detailed description section. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 is a flowchart of a service activation method based on semantic understanding provided by an embodiment of this disclosure.

[0019] Figure 2 is a flowchart of a service activation method based on semantic understanding provided in an embodiment of this disclosure.

[0020] Figure 3 is a flowchart of a service activation method based on semantic understanding provided in an embodiment of this disclosure.

[0021] Figure 4 is a flowchart of a service activation method based on semantic understanding provided in an embodiment of this disclosure.

[0022] Figure 5 is a flowchart of a service activation method based on semantic understanding provided in an embodiment of this disclosure.

[0023] Figure 6 is a flowchart of a service activation method based on semantic understanding provided in an embodiment of this disclosure.

[0024] Figure 7 is a flowchart of a service activation method based on semantic understanding provided in an embodiment of this disclosure.

[0025] Figure 8 is a flowchart of a service activation method based on semantic understanding provided in an embodiment of this disclosure.

[0026] Figure 9 is a block diagram of a service activation device based on semantic understanding provided in an embodiment of this disclosure.

[0027] Figure 10 is a block diagram of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0028] The specific embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this disclosure.

[0029] It should be understood that the term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the following description.

[0030] It should be noted that the concepts of "first," "second," etc., mentioned in this disclosure are used only to distinguish different devices, modules, or units, and are not used to limit the order of functions performed by these devices, modules, or units or their interdependencies. The modifiers "a" and "a plurality of" mentioned in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless explicitly stated in the context, they should be understood as "one or more." In the description of this disclosure, unless otherwise stated, "a plurality of" means two or more, and other quantifiers are similar; "at least one," "one or more," or similar expressions refer to any combination of these items, including any combination of single or multiple items.

[0031] Although operations or steps are described in a specific order in the accompanying drawings in the embodiments of this disclosure, it should not be construed as requiring these operations or steps to be performed in the specific order or serial order shown, or requiring all of the shown operations or steps to be performed to obtain the desired result. In the embodiments of this disclosure, these operations or steps may be performed serially; they may be performed in parallel; or a portion of these operations or steps may be performed.

[0032] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of these messages or information. It is understood that before using the technical solutions disclosed in the embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0033] Figure 1 is a flowchart of a service activation method based on semantic understanding according to an embodiment of this disclosure. As shown in Figure 1, this embodiment of the disclosure provides a service activation method based on semantic understanding, which may include the following steps: In step S10, feature extraction is performed on the input service activation text based on multi-source data to obtain a first semantic vector.

[0034] For example, the business activation text can be divided into multiple sentences, which are then input into a bidirectional pre-trained language model to generate multiple sub-vectors. Then, average pooling is performed on the multiple sub-vectors to obtain the first semantic vector.

[0035] In step S20, the first semantic vector is classified into different domains based on a classification model to obtain the target domain.

[0036] For example, the model collaboration weight factor can be obtained first based on the classification confidence of the first classification model and the validation accuracy of the second classification model, where the first classification model is a lightweight classification model and the second classification model is a large domain classification model. Then, based on the model collaboration weight factor, the first classification model is used to classify the first semantic vector into a domain to obtain the target domain. Alternatively, based on the model collaboration weight factor, the first classification model and the second classification model are used to classify the first semantic vector into a domain to obtain the target domain.

[0037] In step S30, based on the target domain, the structured information of the service activation text is obtained by fine-tuning the large model in the domain.

[0038] For example, a keyword mask matrix for the target domain can be obtained first, and then the keyword mask matrix can be used as an attention gating mechanism to enhance the target domain attention of the large model, resulting in a domain-fine-tuned large model. Then, the business activation text can be processed based on the domain-fine-tuned large model to obtain the structured information of the business activation text.

[0039] In step S40, a second semantic vector is obtained based on the structured information, and the optimal operation path is matched from the graph database based on the second semantic vector. The graph database is a graphical knowledge base.

[0040] For example, a bidirectional pre-trained language model can be used to encode structured information into a second semantic vector. Then, based on the semantic similarity between the second semantic vector and the node vectors in the graph database, as well as the historical success rate of the node path, a comprehensive score can be obtained. Finally, based on the comprehensive score, the best operation path can be matched from the graph database.

[0041] In step S50, based on the preset template corresponding to the optimal operation path and combined with the large model, a service activation scheme is output to activate the service.

[0042] For example, a business activation plan can be automatically generated by a template engine based on a preset template. The proportion of template-filled content and freely generated content from a large model in the business activation plan can be dynamically adjusted by a balance parameter. The balance parameter is obtained based on the information entropy of the template-filled content and the user's rating of the accuracy of the plan.

[0043] The entire process achieves automation and intelligence in business activation through multi-model collaboration and knowledge base linkage, which not only improves activation efficiency, but also ensures the accuracy of business activation solutions through domain adaptation and precise matching.

[0044] Figure 2 is a flowchart of a service activation method based on semantic understanding provided in an embodiment of this disclosure. As shown in Figure 2, the step of extracting features from the input service activation text based on multi-source data to obtain a first semantic vector may include the following steps: In step S101, the service activation text is divided into multiple sentences, and the multiple sentences are input into a bidirectional pre-trained language model to generate multiple sub-vectors.

[0045] For example, multi-source data related to the required business can be collected first, and data parsing, data cleaning and preprocessing can be performed to extract data features for small model training and classification. Multi-source data includes all business types of the requirement from the perspective of business type; from the perspective of data type, it includes unstructured and structured data, where unstructured data includes user manuals, technical manuals, business descriptions, business processes and other documents, images, audio and video, web page content, etc., and structured data includes relevant business data tables, business work order data, etc.; from the perspective of data source, it includes various related business systems and offline collected data.

[0046] Data is parsed using tools corresponding to various data formats, such as PyPDF2 and BeautifulSoup for parsing PDF / HTML documents, Textraction for extracting key information from images, and Tabula for extracting table data, converting various types of data into text. Data cleaning is then performed, removing irrelevant content such as headers, footers, page numbers, table of contents, and watermarks; removing special identifiers such as \n\t; and removing redundant information such as duplicate content.

[0047] The business activation text is divided into multiple sentences, and these sentences are input into the bidirectional pre-trained language model Sentence-BERT for text feature extraction, generating multiple sub-vectors. .

[0048] In step S102, average pooling is performed on the plurality of sub-vectors to obtain the first semantic vector.

[0049] In this step, multiple subvectors Perform average pooling to obtain the first semantic vector. For example, the first semantic vector It can be obtained from the following formula: Formula 1, where, The first semantic vector is the dense vector of the entire document. This is an average pooling operation used to merge multiple vectors into one vector. If the text exceeds the model's maximum length (512 tokens), a sliding window segmentation is used.

[0050] This process can accurately capture the semantic information of each part of the text. By integrating global features through average pooling, even when dealing with extremely long texts, it can ensure the integrity and accuracy of semantic vectors through sliding window segmentation, providing reliable feature support for subsequent domain classification.

[0051] Figure 3 is a flowchart of a service activation method based on semantic understanding provided in an embodiment of this disclosure. As shown in Figure 3, the step of classifying the first semantic vector based on a classification model to obtain the target domain may include the following steps: In step S201, a model collaboration weight factor is obtained based on the classification confidence of the first classification model and the validation accuracy of the second classification model, wherein the first classification model is a lightweight classification model and the second classification model is a large domain classification model.

[0052] In this step, the classification confidence is based on the first classification model. Validation accuracy of the second classification model To obtain the model collaboration weight factor (CHF), where the first classification model is a lightweight classification model and the second classification model is a large domain classification model. For example, the model collaboration weight factor (CHF) can be obtained using the following formula: Formula 2, where, This represents the classification confidence score for a lightweight classification model, with a value ranging from 0 to 1. The validation accuracy of the domain classification model information extraction based on historical data statistics within one year is measured, with a value ranging from 0 to 1. This is a dynamically adjustable coefficient, with a default value of 0.6, which is optimized through feedback learning.

[0053] In step S202a, based on the model collaboration weight factor, the first classification model is used to perform domain classification on the first semantic vector to obtain the target domain.

[0054] In this step, when the model collaboration weight factor CHF is greater than a first threshold (e.g., 0.75), the first classification model is used to classify the first semantic vector into its domain to obtain the probability that the target domain of the input document x is y. For example, the probability that the target domain of input document x is y. It can be obtained from the following formula: Formula 3, where, For activation function, and For classification layer parameters, The first semantic vector The corresponding BERT [CLS] label classification vector is used to aggregate semantics.

[0055] In step S202b, based on the model collaboration weight factor, the first classification model and the second classification model are used to perform domain classification on the first semantic vector to obtain the target domain.

[0056] In this step, when the model collaboration weight factor CHF is less than or equal to a first threshold (e.g., 0.75), the first classification model and the second classification model are used to classify the first semantic vector into its target domain to obtain the probability that the target domain of the input document x is y. For example, the probability that the target domain of input document x is y. It can be obtained from the following formula: Formula 4, where, The classification probability output by the first classification model. This represents the classification probability output by the second classification model.

[0057] This flexible classification mechanism can improve processing speed through lightweight models, ensure classification accuracy by leveraging large domain models, and simultaneously achieve reasonable allocation of tasks among different computing resources, thus optimizing resource utilization.

[0058] Figure 4 is a flowchart of a service activation method based on semantic understanding provided in an embodiment of this disclosure. As shown in Figure 4, obtaining the structured information of the service activation text based on the target domain by fine-tuning a large model of the domain may include the following steps: In step S301, obtaining the keyword mask matrix of the target domain.

[0059] In this step, a keyword mask matrix for the target domain is obtained. For example, a keyword mask matrix. Keyword position weight is +1, others are 0.

[0060] In step S302, the keyword mask matrix is ​​used as an attention gating mechanism to enhance the target domain attention of the large model, thereby obtaining the domain-fine-tuned large model.

[0061] In this step, the keyword mask matrix As an attention gating mechanism, it enhances the target domain attention of a large model, resulting in a domain-fine-tuned large model. For example, the keyword mask matrix... As an attention gating mechanism, enhancing the target domain attention of a large model to obtain a domain-fine-tuned large model can be expressed by the following formula: Formula 5, where, For domain-specific enhanced attention output, Let Q be the normalized exponential function, Q be the query matrix (a vector representation of the currently processed "question" or "focus"), and K be the key matrix (a vector representation of the "identifier" or "content" of all words / tokens in the input sequence). T For the dot product of the query and the key value, d k Let V be the dimension of the key vector, and let V be the value matrix, which is a vector representation of the "actual content" or "details" of all words / tokens in the input sequence. This is the domain enhancement factor, with a default value of 0.8, which can be adjusted and optimized through practice.

[0062] In step S303, the service activation text is processed based on the domain fine-tuning model to obtain the structured information of the service activation text.

[0063] In this step, the business activation text is processed based on the domain-fine-tuned large model to obtain structured information of the business activation text. For example, the fine-tuned model is used to perform deep processing on the business activation text to accurately identify key information such as the operation object, parameters, and dependencies, and finally output standardized JSON format structured information, such as {"Operation Target": "Activate VPN", "Dependency Service": ["Firewall Configuration"], "Parameters": {"IP Address": "192.168.1.1"}}.

[0064] This process enhances attention focus by using domain keywords, enabling the model to more accurately capture technical terms and core information, and significantly improve the accuracy and relevance of structured information extraction.

[0065] Figure 5 is a flowchart of a service activation method based on semantic understanding according to an embodiment of this disclosure. As shown in Figure 5, the step of obtaining a second semantic vector based on the structured information and matching the optimal operation path from the graph database based on the second semantic vector may include the following steps: In step S401, the structured information is encoded into a second semantic vector using a bidirectional pre-trained language model.

[0066] In this step, the bidirectional pre-trained language model BERT is used to encode structured information into a second semantic vector. Among them, the bidirectional pre-trained language model BERT is an existing technology, and will not be described in detail here.

[0067] In step S402, a comprehensive score is obtained based on the semantic similarity between the second semantic vector and the node vectors in the graph database, and the historical success rate of the node path.

[0068] In this step, based on the second semantic vector Node vectors in graph database semantic similarity A comprehensive score is obtained by analyzing the historical success rate of the node paths. For example, a comprehensive score. It can be obtained from the following formula: Formula 6, where, To find the mean.

[0069] Formula 7, where, To calculate cosine similarity, the cosine of the angle between two vectors in space is used to measure their similarity. The value ranges from [-1, 1], but in the semantic vectors of the document, it is usually taken as [0, 1] (because vectors are usually non-negative). The closer the value is to 1, the more similar they are. For the second semantic vector Node vectors in graph database The semantic similarity is a set (or array) containing the query vector (second semantic vector). ) and each node vector The cosine similarity value between them The second semantic vector captures the semantic information of the query (service activation text). Node vectors in a graph database are dense vectors, obtained by encoding the text (such as documents, sentences, or keywords) corresponding to the nodes using a language model, capturing the semantic information of the nodes.

[0070] In step S403, the optimal operation path is matched from the graph database based on the comprehensive score.

[0071] In this step, the optimal operation path is matched from the graph database based on the overall score. For example, the node path with the highest score can be selected from the graph database as the optimal operation path based on the overall score.

[0072] This process combines the relevance of semantic matching with the reliability of historical execution to ensure that the matched operation path not only meets the current business needs but also has a high success rate, providing a strong guarantee for business activation.

[0073] Figure 6 is a flowchart of a service activation method based on semantic understanding according to an embodiment of this disclosure. As shown in Figure 6, the step of outputting a service activation scheme to activate the service based on a preset template corresponding to the optimal operation path and combined with a large model may include the following steps: In step S501, based on the preset template, the service activation scheme is automatically generated by a template engine, and the proportion of template-filled content and freely generated content from the large model in the service activation scheme is dynamically adjusted by a balance parameter. The balance parameter is obtained based on the information entropy of the template-filled content and the user's rating of the accuracy of the scheme.

[0074] In this step, a business activation plan is automatically generated using a template engine based on a preset template. The weighting of template-filled content and freely generated content from a large model is dynamically adjusted using balancing parameters. These balancing parameters are obtained based on the information entropy of the template-filled content and user ratings of the plan's accuracy. For example, operation steps can be generated based on the template engine, i.e., a standardized basic template (preset template) can be set according to business requirements, such as: Step {Number}: {Action} {Parameters}; Expected Result: {Expected Output}; Timeout: {Default Value} seconds.

[0075] The equilibrium parameter DF can be obtained by the following formula: Formula 8, where, Information entropy for filling content into the template. The user rates the accuracy of the solution, with a value ranging from 0 to 1. The preset balance coefficient is 0.7 by default.

[0076] when When filling in the form, strictly follow the preset template, such as "Step 1: Configure the IP address as 192.168.1.1"); when At that time, large model optimization descriptions are allowed (such as "First, please log in to the management console and find the IP address configuration item in the top navigation bar of the 'Network Settings' page...").

[0077] By dynamically adjusting the ratio of template-filled content to freely generated content from the large model based on balance parameters, the standardization of the solution is ensured while allowing for flexible optimization of the description method. This approach effectively avoids the rigidity of pure templates and the uncontrollability of purely free generation, resulting in a business activation solution that is both standardized and tailored to real-world application scenarios.

[0078] Figure 7 is a flowchart of a service activation method based on semantic understanding provided in an embodiment of this disclosure. As shown in Figure 7, the method may further include the following steps: In step S60, if the difference between the user-reported correction operation step and the optimal operation path matching the knowledge base exceeds an update sensitivity threshold, an update step of the knowledge base is triggered.

[0079] In this step, if the difference between the user-reported correction steps and the optimal operation path matching the knowledge base exceeds an update sensitivity threshold, an update step for the knowledge base is triggered. For example, the difference... It can be obtained from the following formula: Formula 9, where cos is used to calculate the cosine similarity. This is the vector representation of the optimal operation path. A vector representation of the corrective action steps based on user feedback.

[0080] The updated sensitivity threshold UST can be obtained using the following formula: Formula 10, where, For similarity The historical minimum value, This represents the error rate over the past three months, based on sliding window statistics. This is an empirical decay factor, with an initial value of 0.5, which decreases over time, decaying by 0.05 per month, with a minimum of 0.2.

[0081] The system collects user feedback on corrected service activation plans, encodes these corrected steps and the best matching operation path in the knowledge base into vectors using a model, calculates the cosine similarity between the two vectors, and then determines their difference. Simultaneously, it calculates an update sensitivity threshold by combining the historical minimum similarity value and the operation error rate over the past three months. When the difference exceeds this threshold, the knowledge base update process is automatically triggered, integrating the corrected operation path into the knowledge base. This mechanism avoids stability issues caused by frequent knowledge base updates while promptly incorporating high-quality corrected content, ensuring the timeliness and accuracy of the knowledge base.

[0082] Figure 8 is a flowchart of a service activation method based on semantic understanding according to an embodiment of this disclosure. As shown in Figure 8, the method may further include the following steps: In step S70, a feedback contribution weight is calculated based on the user role and historical correction behavior.

[0083] In this step, feedback contribution weights are calculated based on user roles and historical correction behaviors. For example, feedback contribution weights. It can be obtained from the following formula: Formula 11, where, Contribute weight to user i's feedback. User role weights are assigned as follows: administrators = 1.0, regular users = 0.3. Let i be the number of times a historical correction has been adopted, k be the total number of users, and j represent each user from 1 to k. Not zero.

[0084] In step S80, the large model is incrementally trained using the loss function enhanced by the feedback contribution weights.

[0085] In this step, the large model is incrementally trained using the loss function weighted by the feedback contribution weights. For example, the loss function weighted by the feedback contribution weights is used. It can be obtained from the following formula: Formula 12, where, The original loss function, These are the predictions currently generated by the large model. The corrected feedback value provided to the user. To find the norm.

[0086] User roles are assigned corresponding weights, and the number of times a user's historical corrections were adopted is counted. These weights are then used in Formula 11 to calculate the feedback contribution weight for each user. This weight is incorporated into the loss function of the incremental model training. By weighting the feedback, the impact of high-quality feedback on the training of the large model is amplified, enabling continuous incremental training of the large model. This training method based on feedback contribution weights can accurately absorb high-value user feedback, continuously optimize the performance of the large model, and improve the accuracy and adaptability of subsequent business activation plans.

[0087] Figure 9 is a block diagram of a service activation device based on semantic understanding according to an embodiment of this disclosure. As shown in Figure 9, this embodiment of the disclosure provides a service activation device 900 based on semantic understanding. The device 900 may include the following modules: a first acquisition module 910, used to extract features from the input service activation text based on multi-source data to obtain a first semantic vector.

[0088] The second acquisition module 920 is used to perform domain classification on the first semantic vector based on the classification model to obtain the target domain.

[0089] The third acquisition module 930 is used to obtain the structured information of the service activation text based on the target domain by fine-tuning the large model in the domain.

[0090] The fourth acquisition module 940 is used to acquire a second semantic vector based on the structured information, and to match the best operation path from the graph database based on the second semantic vector, wherein the graph database is a graphical knowledge base.

[0091] The output module 950 is used to output a service activation scheme to activate the service based on the preset template corresponding to the optimal operation path and in combination with the large model.

[0092] In one possible implementation, the first acquisition module 910 is further configured to: divide the service activation text into multiple sentences, input the multiple sentences into a bidirectional pre-trained language model to generate multiple sub-vectors; and perform average pooling on the multiple sub-vectors to obtain the first semantic vector.

[0093] In one possible implementation, the second acquisition module 920 is further configured to: acquire a model collaboration weight factor based on the classification confidence of the first classification model and the validation accuracy of the second classification model, wherein the first classification model is a lightweight classification model and the second classification model is a large domain classification model; based on the model collaboration weight factor, perform domain classification on the first semantic vector using the first classification model to obtain the target domain; or based on the model collaboration weight factor, perform domain classification on the first semantic vector using the first classification model and the second classification model to obtain the target domain.

[0094] In one possible implementation, the third acquisition module 930 is further configured to: acquire the keyword mask matrix of the target domain; use the keyword mask matrix as an attention gating mechanism to perform target domain attention enhancement on the large model to obtain the domain fine-tuned large model; and process the service activation text based on the domain fine-tuned large model to obtain the structured information of the service activation text.

[0095] In one possible implementation, the fourth acquisition module 940 is further configured to: encode the structured information into a second semantic vector using a bidirectional pre-trained language model; obtain a comprehensive score based on the semantic similarity between the second semantic vector and the node vectors in the graph database, and the historical success rate of the node path; and match the optimal operation path from the graph database according to the comprehensive score.

[0096] In one possible implementation, the output module 950 is further configured to: automatically generate the service activation plan based on the preset template using a template engine, and dynamically adjust the proportion of template-filled content and freely generated content from the large model in the service activation plan using a balance parameter, wherein the balance parameter is obtained based on the information entropy of the template-filled content and the user's rating of the plan's accuracy.

[0097] In one possible implementation, the device 900 further includes an update module for triggering an update step of the knowledge base if the difference between the user-reported correction operation step and the optimal operation path matching the knowledge base exceeds an update sensitivity threshold.

[0098] In one possible implementation, the device 900 further includes a feedback module for: calculating feedback contribution weights based on user roles and historical correction behaviors; and incrementally training the large model using a loss function weighted and enhanced by the feedback contribution weights.

[0099] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0100] In summary, the first acquisition module collects multi-source data, parses, cleans, and extracts features from the input business activation text, and outputs a first semantic vector. The second acquisition module receives the first semantic vector, performs domain classification using a classification model, determines the target domain, and passes it to the third acquisition module. Based on the target domain, the third acquisition module uses a domain-fine-tuned large model to process the text, generates structured information, and sends it to the fourth acquisition module. The fourth acquisition module encodes the structured information into a second semantic vector, matches the optimal operation path in the graph database, and feeds it back to the output module. The output module generates and outputs a business activation scheme based on the optimal operation path and a preset template, combined with the large model. The modules work together to fully realize the entire process of automatic business activation, ensuring the efficient implementation of the overall technical solution and combining the advantages of automation, precision, and intelligence.

[0101] Figure 10 is a block diagram of an electronic device according to an embodiment of the present disclosure. As shown in Figure 10, the electronic device 1000 may include a processor 1001 and a memory 1002. The electronic device 1000 may also include one or more of a multimedia component 1003, an input / output (I / O) interface 1004, and a communication component 1005.

[0102] The processor 1001 controls the overall operation of the electronic device 1000 to complete all or part of the steps in the semantic understanding-based service activation method described above. The memory 1002 stores various types of data to support the operation of the electronic device 1000. This data may include, for example, instructions for any application or method operating on the electronic device 1000, and application-related data such as contact data, sent and received messages, pictures, audio, video, etc. The memory 1002 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. Multimedia component 1003 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in memory 1002 or transmitted via communication component 1005. The audio component also includes at least one speaker for outputting audio signals. I / O interface 1004 provides an interface between processor 1001 and other interface modules, such as a keyboard, mouse, buttons, etc. These buttons may be virtual or physical buttons. Communication component 1005 is used for wired or wireless communication between the electronic device 1000 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G, NB-IoT, eMTC, or other 5G technologies, or combinations thereof, is not limited here. Therefore, the corresponding communication component 1005 may include: a Wi-Fi module, a Bluetooth module, an NFC module, etc.

[0103] In an exemplary embodiment, the electronic device 1000 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described semantic understanding-based service activation method.

[0104] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, which, when executed by a processor, implement the steps of the semantic understanding-based service activation method described above. For example, the computer-readable storage medium may be the memory 1002 including the program instructions described above, which may be executed by the processor 1001 of the electronic device 1000 to complete the semantic understanding-based service activation method described above.

[0105] In another exemplary embodiment, a computer program product is also provided, the computer program product comprising a computer program executable by a programmable device, the computer program having a code portion for performing the above-described semantic understanding-based service activation method when executed by the programmable device.

[0106] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0107] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0108] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

Claims

1. A service activation method based on semantic understanding, characterized in that, The method includes: extracting features from the input service activation text based on multi-source data to obtain a first semantic vector; classifying the first semantic vector into a domain based on a classification model to obtain a target domain; obtaining structured information of the service activation text by fine-tuning a large model based on the target domain; obtaining a second semantic vector based on the structured information, and matching the optimal operation path from a graph database, wherein the graph database is a graphical knowledge base; and outputting a service activation scheme to activate the service based on a preset template corresponding to the optimal operation path and in combination with the large model.

2. The method according to claim 1, characterized in that, The step of extracting features from the input service activation text based on multi-source data to obtain a first semantic vector includes: dividing the service activation text into multiple sentences, inputting the multiple sentences into a bidirectional pre-trained language model to generate multiple sub-vectors; and performing average pooling on the multiple sub-vectors to obtain the first semantic vector.

3. The method according to claim 1, characterized in that, The step of classifying the first semantic vector based on a classification model to obtain the target domain includes: obtaining a model collaboration weight factor based on the classification confidence of the first classification model and the validation accuracy of the second classification model, wherein the first classification model is a lightweight classification model and the second classification model is a large-scale domain classification model; classifying the first semantic vector based on the model collaboration weight factor using the first classification model to obtain the target domain; or classifying the first semantic vector based on the model collaboration weight factor using both the first classification model and the second classification model to obtain the target domain.

4. The method according to claim 1, characterized in that, The step of obtaining the structured information of the service activation text based on the target domain and by fine-tuning the large model in the domain includes: obtaining a keyword mask matrix of the target domain; using the keyword mask matrix as an attention gating mechanism to enhance the large model with target domain attention, thereby obtaining the fine-tuned large model in the domain; and processing the service activation text based on the fine-tuned large model in the domain to obtain the structured information of the service activation text.

5. The method according to claim 1, characterized in that, The step of obtaining a second semantic vector based on the structured information and matching the optimal operation path from the graph database based on the second semantic vector includes: encoding the structured information into a second semantic vector using a bidirectional pre-trained language model; obtaining a comprehensive score based on the semantic similarity between the second semantic vector and the node vectors in the graph database, and the historical success rate of the node path; and matching the optimal operation path from the graph database based on the comprehensive score.

6. The method according to claim 1, characterized in that, The step of outputting a service activation plan to activate the service based on the preset template corresponding to the optimal operation path and in combination with the large model includes: automatically generating the service activation plan through a template engine based on the preset template, and dynamically adjusting the proportion of template-filled content and freely generated content from the large model in the service activation plan through a balance parameter, wherein the balance parameter is obtained based on the information entropy of the template-filled content and the user's rating of the accuracy of the plan.

7. The method according to claim 1, characterized in that, The method further includes: triggering an update step for the knowledge base if the difference between the user-reported correction operation steps and the optimal operation path matching the knowledge base exceeds an update sensitivity threshold.

8. The method according to claim 1, characterized in that, The method further includes: calculating feedback contribution weights based on user roles and historical correction behaviors; and using the loss function weighted and enhanced by the feedback contribution weights to perform incremental training on the large model.

9. An electronic device, characterized in that, include: A memory on which computer programs are stored; A processor for executing the computer program in the memory to implement the steps of the method according to any one of claims 1-8.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method described in any one of claims 1-8.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-8.

Images