Deep learning-driven order assignment management system, method and storage medium

Abstract

The present invention relates to a dispatching management system, method and storage medium driven by deep learning, and relates to the field of dispatching management, including: a dialogue vector acquisition module acquires a first dialogue vector between a customer and an intelligent customer service; a primary intention recognition module recognizes the first dialogue vector to obtain a first customer demand service; a text block optimization module optimizes and segments the vector according to the demand service to obtain a second dialogue vector; a secondary intention recognition module recognizes the second dialogue vector to obtain a second demand service, a demand quantity and a time limit; a dispatching management module constructs a work order dispatch according to the service, quantity and time limit when the two-level demand services are the same. This application solves the technical problem that in the existing intelligent customer service system, due to the lack of a verification mechanism in the dispatching link, the intention recognition accuracy rate is low and the dispatching effect of the intelligent customer service system is poor. By constructing a two-level intention recognition model and introducing a secondary verification mechanism, the technical effect of improving the intention recognition accuracy rate and optimizing the dispatching effect of the intelligent customer service system is achieved.

Description

Technical Field

[0001] The present invention relates to the field of dispatch management, and in particular to a deep learning driven dispatch management system, method and storage medium. Background Art

[0002] With the development and application of artificial intelligence technology, intelligent customer service systems have gradually become an important means for enterprises to improve customer service efficiency and service quality. Through natural language interaction, intelligent customer service systems can automatically respond to customer inquiries, identify customer intentions, and dispatch work orders according to the type of intention, thereby realizing the automated processing of customer service business. However, the existing intelligent customer service systems lack the necessary review and verification mechanisms for the intention recognition results based on a single round of dialogue in the work order dispatching link, and are easily affected by factors such as semantic misunderstanding errors and omissions of key information. When customer needs are complex and the expression is unclear, the accuracy of a single intention recognition is difficult to guarantee, resulting in a mismatch between the dispatching results and the actual needs of customers, affecting the effective resolution of customer problems and the improvement of service quality. Therefore, the existing intelligent customer service system has the problems of low accuracy of intention recognition in the dispatching link and poor dispatching effect of the intelligent customer service system. Summary of the invention

[0003] The present invention aims to solve the technical problems in existing intelligent customer service systems, such as low accuracy of intent recognition and poor order dispatching effect of the intelligent customer service system due to the lack of a verification mechanism in the order dispatching link, by providing a deep learning driven order dispatching management system, method and storage medium.

[0004] The technical solution of the present invention to solve the above technical problems is as follows:

[0005] In a first aspect, the present invention provides a deep learning-driven dispatch management system, including: a dialogue vector acquisition module, used to obtain a first dialogue vector between a customer and an intelligent customer service; a first-level intention recognition module, used to identify the customer intent of the first dialogue vector through a first-level deep learning semantic model, and obtain a first customer demand business; a text block optimization module, used to perform text block optimization on the dialogue vector according to the first customer demand business, obtain a first target text block size, and segment the first dialogue vector according to the first target text block size to obtain a second dialogue vector; a second-level intention recognition module, used to perform customer intent on the second dialogue vector through a second-level deep learning semantic model, and obtain a second customer demand business, customer demand quantity and customer demand time limit; an order management module, used to construct work order information for dispatch management according to the first customer demand business, customer demand quantity and customer demand time limit when the first customer demand business and the second customer demand business are the same.

[0006] In a second aspect, the present invention provides a deep learning-driven dispatch management method, including: obtaining a first conversation vector between a customer and an intelligent customer service; using a first-level deep learning semantic model, identifying customer intent on the first conversation vector to obtain a first customer demand business; optimizing the conversation vector according to the first customer demand business to obtain a first target text block size, and segmenting the first conversation vector according to the first target text block size to obtain a second conversation vector; using a second-level deep learning semantic model, identifying customer intent on the second conversation vector to obtain a second customer demand business, the number of customer demands, and a time limit for customer demands; when the first customer demand business and the second customer demand business are the same, constructing work order information according to the first customer demand business, the number of customer demands, and the time limit for customer demands to perform dispatch management.

[0007] In a third aspect, the present application provides a computer-readable storage medium storing a computer program, which is used to execute the deep learning-driven order dispatch management system provided by the present application.

[0008] The beneficial effects of the present invention are:

[0009] The first dialogue vector between the customer and the intelligent customer service is obtained through the dialogue vector acquisition module, providing a data basis for subsequent intent recognition; through the first-level intent recognition module, the first-level deep learning semantic model is used to identify the customer intent of the first dialogue vector, obtain the first customer demand business, realize the preliminary judgment of customer demand, and provide directional guidance for the second-level intent recognition; on the basis of the first-level intent recognition, the text block optimization module optimizes the text block of the first dialogue vector according to the first customer demand business, obtains the first target text block size, and divides the first dialogue vector according to the first target text block size to obtain the second dialogue vector. Through the feedback of the first-level intent recognition result, the granularity of the dialogue vector is dynamically adjusted, and a secondary judgment is made in the form of a more refined text block, thereby improving the accuracy of intent recognition; then, the second-level intent recognition module uses the second-level deep learning semantic model to identify the customer intent of the second dialogue vector, obtain the second customer demand business, customer demand quantity and customer demand time limit. On the basis of the first-level intent recognition and text block optimization, the second-level intent recognition obtains the detailed type, quantity and time limit requirements of customer demand through more professional and fine-grained semantic analysis, providing a key basis for accurate dispatching. When the order management module determines that the business demanded by the first customer is consistent with the business demanded by the second customer, it constructs the work order information for order management based on the business demanded by the first customer, the number of customer demands and the time limit of customer demands. By comprehensively utilizing the results of two-level intent recognition, it improves the accuracy of intent recognition, thereby optimizing the order dispatching effect of the intelligent customer service system. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1A schematic diagram of the structure of the deep learning-driven dispatch management system provided by the present invention;

[0011] Figure 2 A schematic diagram of the process of the deep learning-driven dispatch management method provided by the present invention;

[0012] Figure 3 A schematic diagram of the structure of a computer-readable storage medium provided by the present invention.

[0013] In the accompanying drawings, the components represented by the reference numerals are as follows:

[0014] Dialogue vector acquisition module 11, primary intention recognition module 12, text block optimization module 13, secondary intention recognition module 14, dispatch management module 15, computer readable storage medium 200, computer program 211. DETAILED DESCRIPTION

[0015] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0016] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0017] In the description of the present invention, the term "for example" is used to mean "used as an example, illustration or explanation". Any embodiment described as "for example" in the present invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is given to enable any technician in the field to implement and use the present invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes will not be elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed in the present invention.

[0018] Embodiment 1:

[0019] like Figure 1As shown, an embodiment of the present invention provides a deep learning driven dispatch management system, including:

[0020] The dialogue vector acquisition module 11 is used to obtain a first dialogue vector between the customer and the intelligent customer service.

[0021] Specifically, the conversation vector acquisition module 11 captures and vectorizes the conversation content between the customer and the intelligent customer service to obtain the first conversation vector. For example, the customer enters through the online chat window: "I want to check my phone bill last month because the total cost is much higher than usual. In addition, I would like to ask how to apply for a 5G package service? It is best to complete it within this week." The conversation vector acquisition module 11 uses natural language processing technology, such as word embedding or BERT model, to convert this text into a numerical vector, namely the first conversation vector. The first conversation vector represents the semantic features of this round of conversation in a digital form, preparing for subsequent customer intent recognition.

[0022] The first-level intention recognition module 12 is used to recognize the customer intention of the first dialogue vector through a first-level deep learning semantic model to obtain the first customer demand business.

[0023] Specifically, after obtaining the first conversation vector, the conversation vector acquisition module 11 sends the first conversation vector to the first-level intention recognition module 12. The first-level intention recognition module 12 analyzes the first conversation vector through a first-level deep learning semantic model to identify the customer's intention and obtain the first customer demand business.

[0024] The first-level intention recognition module 12 pre-trains a deep learning model specifically for customer intention recognition, which is a first-level deep learning semantic model. The model uses deep learning algorithms, such as convolutional neural networks, recurrent neural networks or attention mechanisms, to effectively extract semantic features from the first dialogue vector and accurately judge customer intentions. When the first dialogue vector is input into the first-level intention recognition module 12, the first-level deep learning semantic model automatically extracts features and understands semantics, identifies the core demands expressed by the customer in this round of dialogue, and classifies them into predefined business categories to form the first customer demand business. For example, for "I want to check my phone bill last month because the total cost is much higher than usual. In addition, I would like to ask how to apply for the 5G package service? It is best to complete it within this week.", the first-level intention recognition module 12 identifies it as a business demand for "bill inquiry" and "business handling".

[0025] Through the first-level intention recognition module 12, the business field to which the customer's needs belong can be quickly and accurately determined, providing important clues for subsequent fine-grained intention recognition and task dispatching. The introduction of the first-level deep learning semantic model improves the ability of the intelligent customer service system to understand customer intentions, reduces dependence on manual customer service, and improves response efficiency.

[0026] The text block optimization module 13 is used to optimize the text block according to the first customer demand business, obtain a first target text block size, and segment the first dialogue vector according to the first target text block size to obtain a second dialogue vector.

[0027] Specifically, the first-level intention recognition module 12 sends the obtained first customer demand business to the text block optimization module 13. The text block optimization module 13 optimizes the first dialogue vector according to the first customer demand business to extract more accurate and targeted semantic information, thereby obtaining the second dialogue vector.

[0028] First, the text block optimization module 13 determines the optimal text block size, that is, the first target text block size, according to the first customer demand business. The text block size refers to the length of each semantic unit in the dialogue vector. By adjusting the text block size, the granularity and richness of the semantic information can be controlled. Next, the text block optimization module 13 divides and reorganizes the first dialogue vector according to the first target text block size, and divides the first dialogue vector into a number of semantic segments with a length of the first target text block size according to the principles of semantic relevance and integrity. Each semantic segment is a relatively independent and complete sentence or phrase. In this way, the original dialogue vector is converted into a series of smaller and more fine-grained semantic units to form a second dialogue vector. Compared with the first dialogue vector, the second dialogue vector focuses more on the semantic information closely related to the first customer demand business and ignores irrelevant or redundant content. This optimized dialogue vector representation helps the subsequent secondary intention recognition module 14 to more accurately judge customer intentions and extract key information elements.

[0029] For example, a conversation of a customer is: "I want to check my phone bill last month, because the total cost is much higher than usual. In addition, I would like to ask how to apply for the 5G package service? It is best to complete it within this week." After processing by the conversation vector acquisition module 11 and the first-level intention recognition module 12, it is obtained that the first customer demand service is "bill inquiry" and "service handling". Subsequently, the text block optimization module 13 determines the optimal text block size, that is, the first target text block size, according to these two types of customer demand services. For "bill inquiry", a larger text block may be required to include the description details of the abnormal cost; while for "service handling", a smaller text block may be required to highlight key information such as the handling time limit. Then, the text block optimization module 13 divides and reorganizes the first conversation vector according to the first target text block size. For example, the following semantic fragments are obtained: "I want to check my phone bill last month, because the total cost is much higher than usual." "In addition, I would like to ask how to apply for the 5G package service?" "It is best to complete it within this week." These semantic fragments constitute the second conversation vector.

[0030] Through the text block optimization module 13, the first dialogue vector is optimized and reorganized to generate a more refined and targeted second dialogue vector, laying a solid foundation for subsequent deep-level intent recognition.

[0031] The secondary intention recognition module 14 is used to recognize the customer intention of the second dialogue vector through a secondary deep learning semantic model to obtain the second customer demand service, customer demand quantity and customer demand time limit.

[0032] Specifically, through the secondary intention recognition module 14, on the basis of text block optimization, the secondary deep learning semantic model is used to perform more refined and comprehensive semantic understanding and intention recognition on the second dialogue vector, thereby obtaining multiple key elements of customer needs, including the second customer demand business, customer demand quantity and customer demand time limit.

[0033] First, the secondary deep learning semantic model uses neural networks such as bidirectional long short-term memory networks to encode each semantic segment in the second dialogue vector and extract the contextual information and key semantic information contained in the segment. Subsequently, based on the encoding, the secondary deep learning semantic model introduces the attention mechanism to automatically learn the importance weights of different semantic segments and keywords for customer intent recognition. Through the attention mechanism, the model can focus more on key information that is highly relevant to customer needs, while reducing the interference of irrelevant information, and achieve a multi-angle and multi-level understanding of customer intentions, thereby obtaining the second customer demand business, customer demand quantity and customer demand time limit. Among them, the second customer demand business is to determine the specific business to which the customer demand belongs, such as bill inquiry, fault repair, business handling, etc., by performing keyword matching and semantic similarity calculation in each semantic segment; the customer demand quantity is to extract the specific quantity information mentioned in the customer dialogue by performing number word recognition and keyword matching in the semantic segment, such as "applying for 5 mobile phone numbers"; the customer demand time limit is to determine the customer's time requirements for demand resolution by performing time word recognition and keyword matching in the semantic segment, such as "must be completed today" and "handled before next week".

[0034] For example, the second dialogue vector obtained by the text block optimization module 13 includes the following semantic segments: "I want to check my phone bill last month, because the total cost is much higher than usual." "In addition, I would like to ask how to apply for the 5G package service?" "It is best to complete it within this week." According to the secondary deep learning semantic model, "I want to check my phone bill last month, because the total cost is much higher than usual." is analyzed to obtain the second customer demand service: bill query, customer demand quantity: not mentioned, customer demand time limit: not mentioned; "In addition, I would like to ask how to apply for the 5G package service?" is analyzed to obtain the second customer demand service: service handling-5G package, customer demand quantity: not mentioned, customer demand time limit: not mentioned; "It is best to complete it within this week." is analyzed to obtain the second customer demand service: not mentioned, customer demand quantity: not mentioned, customer demand time limit: within this week.

[0035] Through the more refined semantic understanding of the secondary intention recognition module 14, a refined understanding and multi-dimensional analysis of customer needs are achieved, which improves the intelligence level and service quality of dispatch management.

[0036] The dispatch management module 15 is used to construct work order information for dispatch management according to the first customer demand business, the customer demand quantity and the customer demand time limit when the first customer demand business and the second customer demand business are the same.

[0037] Specifically, after obtaining the second customer demand business, the dispatch management module 15 compares the second customer demand business with the first customer demand business to verify the consistency of the two-level intent recognition and improve the accuracy of business judgment. If the two are consistent, it means that the customer's core demands have been accurately understood. When the first customer demand business is consistent with the second customer demand business, the dispatch management module 15 extracts the first customer demand business, the customer demand quantity and the customer demand time limit, and builds structured work order information based on this. Among them, the first customer demand business indicates the business category to which the work order belongs; the customer demand quantity clarifies the amount of business that needs to be processed; and the customer demand time limit indicates the time the customer expects to resolve.

[0038] The order management module 15 integrates the first customer demand business, customer demand quantity and customer demand time limit into work order information, and then automatically dispatches the work order information to the most suitable customer service representative for processing based on the work order information, combined with factors such as the customer service personnel's business expertise, service quality score, current workload, etc., so as to fully match customer service expertise with customer needs, reasonably allocate human resources, and thus improve service efficiency and customer satisfaction.

[0039] The accuracy of customer intent recognition is improved through a two-level intent recognition mechanism. Compared with traditional single intent recognition, the introduction of a two-level deep learning semantic model at the primary and secondary levels, through two rounds of iterative semantic understanding, analyzes customer needs from coarse-grained to fine-grained, effectively reducing the probability of misidentification. By combining clear demand information with a task allocation mechanism, it is possible to quickly respond to customer demands and match the best service resources for each demand, thereby significantly improving the dispatch effect and service level of the intelligent customer service system.

[0040] Furthermore, the embodiment of the present application further includes an intention correction processing module, and the execution steps of the intention correction processing module include:

[0041] According to the first customer demand business, sort the business type set to obtain a business type sorting result;

[0042] When the first customer demand service and the second customer demand service are different, extracting the first serial number service type of the service type sorting result and setting it as the third customer demand service;

[0043] Optimize the text block according to the third customer demand business to obtain a second target text block size, and segment the first dialogue vector according to the second target text block size to obtain a third dialogue vector;

[0044] By using the secondary deep learning semantic model, the third dialogue vector is used to identify the customer's intention, and a fourth customer demand service, customer demand quantity, and customer demand time limit are obtained;

[0045] When the third customer demand business and the fourth customer demand business are the same, constructing work order information for dispatch management according to the third customer demand business, the customer demand quantity and the customer demand time limit;

[0046] When the third customer demand service and the fourth customer demand service are different, the second serial number service type of the service type sorting result is extracted for iterative analysis.

[0047] In a feasible implementation, the deep learning-driven dispatch management system also includes an intent correction processing module. The function of this module is to further optimize the intent recognition process by iteratively analyzing the business type ranking results when the first customer demand business and the second customer demand business are inconsistent, until the final customer demand business is determined.

[0048] First, the intent correction processing module sorts the preset business type set according to the first customer demand business obtained by the first-level intent recognition. The sorting process comprehensively considers the relevance of different business types to the first customer demand business, and puts the business type that is most likely to match the customer's true intention in the front row to form a business type sorting result. When the first customer demand business and the second customer demand business are inconsistent, the intent correction processing module will extract the first-numbered business type in the business type sorting result and set it as the third customer demand business, aiming to select one from the most relevant alternative business types as a new intent recognition target and re-judge the customer's needs.

[0049] Next, the intent correction processing module optimizes the text block according to the third customer demand business. By dynamically adjusting the text block size, key information blocks that are more in line with the semantic characteristics of the third customer demand business can be extracted from the first dialogue vector. The optimized text block size is defined as the second target text block size. The intent correction processing module then uses the second target text block size to segment the first dialogue vector to obtain the third dialogue vector. Compared with the initial first dialogue vector, the semantic fragments contained in the third dialogue vector are more in line with the characteristics of the third customer demand business, creating favorable conditions for subsequent accurate intent recognition. Subsequently, the intent correction processing module inputs the third dialogue vector into the secondary deep learning semantic model to re-identify the customer intent. Based on the optimized dialogue vector, the secondary semantic model can more accurately judge the customer demand and derive the fourth customer demand business, the customer demand quantity and the customer demand time limit. If the third customer demand business and the fourth customer demand business are consistent, it means that the customer's real needs have been successfully identified through intent correction. At this time, the intent correction processing module will construct the work order information and perform dispatch management according to the consistent business type, demand quantity and time limit requirements to complete the processing of the service request. However, if there are still differences between the third customer demand service and the fourth customer demand service, the intention correction processing module will extract the second order service type from the service type sorting result and start the next round of iterative analysis. The new round of iteration will repeat the above process until a service type that fully matches the customer demand is determined or all sorting results are traversed.

[0050] The intent correction processing module iteratively analyzes the business type ranking results and dynamically optimizes the conversation vector and intent recognition process, greatly improving the system's ability to adapt to the diverse expressions of customers, further enhancing the accuracy and robustness of order dispatch, and ensuring that every customer demand can be accurately identified and efficiently processed, providing a guarantee for providing a high-quality customer service experience.

[0051] Furthermore, the execution steps of the intention correction processing module also include:

[0052] Get a business type set;

[0053] Based on the first customer demand business, traverse the business type set to perform business text vector similarity analysis to obtain a business text vector similarity set;

[0054] The business type set is sorted from large to small according to the business text vector similarity set to obtain a business type sorting result.

[0055] In a preferred embodiment, when obtaining the result of the business type ranking, first, the intention correction processing module obtains a preset business type set. The set covers various common businesses in the intelligent customer service scenario, such as bill inquiry, fault repair, business handling, etc. Each business type corresponds to a text vector describing its characteristics. The business type set is the basis of intention correction and provides a reference standard for the selection of candidate intentions. Next, the intention correction processing module traverses the business type set based on the first customer demand business and performs business text vector similarity analysis. Specifically, the first customer demand business is compared with the text vector of each business type in the business type set, and the similarity between them is calculated. The higher the similarity, the closer the two businesses are in semantic features, and the greater the possibility that the customer demand is misjudged as the business. Through traversal analysis, the intention correction processing module obtains a business text vector similarity set. Each element in the set is composed of a business type and its similarity with the first customer demand business. These similarities quantify the degree of matching between different business types and actual customer needs, and provide data support for the reasonable adjustment of the intention recognition direction. Afterwards, the intent correction processing module re-sorts the original business type set according to the business text vector similarity set. The sorting principle is from large to small similarity, that is, the business type most similar to the first customer demand business is ranked first, and so on. Through sorting, the optimized business type sorting result is obtained. The business type ranked at the top represents the most likely correct attribution of customer demand and will be used for intent correction and re-identification first.

[0056] Through quantitative analysis and sorting of business relevance, the intent correction processing module constructs a business type sequence from similar to different for subsequent iterative identification, so that each correction focuses on the business direction that is most likely to match the customer's real needs. This dynamic sorting mechanism based on similarity greatly improves the efficiency and accuracy of intent correction.

[0057] Furthermore, the execution steps of the text block optimization module 13 include:

[0058] According to the first customer demand business, obtaining a historical recognition log of the first customer demand business by the secondary deep learning semantic model, wherein the historical recognition log includes a recognition accuracy identifier and a text block size identifier;

[0059] Extract the text block size identifier that is identified as accurate in the historical recognition log;

[0060] The length of the text block is the first coordinate axis, the width of the text block is the second coordinate axis, and the height of the text block is the third coordinate axis;

[0061] Constructing an optimization three-dimensional space according to the first coordinate axis, the second coordinate axis and the third coordinate axis;

[0062] Distributing the text block size identifiers in the optimization three-dimensional space to obtain text block distribution coordinates;

[0063] Deleting outlier coordinates based on the text block distribution coordinates to obtain concentrated text block distribution coordinates;

[0064] Performing density analysis on the concentrated text block distribution coordinates to obtain a density parameter of the concentrated text block distribution coordinates;

[0065] The maximum value coordinates of the concentrated text block distribution coordinate density parameters are extracted to generate the first target text block size.

[0066] In a preferred embodiment, the text block optimization module 13 uses three-dimensional optimization to determine the optimal text block size.

[0067] First, the text block optimization module 13 obtains the historical recognition log of the secondary deep learning semantic model for the first customer demand business type based on the first-level intent recognition. The historical recognition log records the performance of the secondary deep learning semantic model in actual applications, including the accuracy evaluation of each intent recognition (recognition accuracy mark) and the corresponding text block size setting (text block size mark). These log data provide an empirical reference for optimizing the text block size. Next, the text block optimization module 13 extracts those text block size marks with the recognition accuracy mark as "accurate" from the historical recognition log, thereby focusing on the samples predicted correctly by the secondary deep learning semantic model, because only under the premise of correct recognition can the setting of the text block size be considered effective and excellent.

[0068] In order to evaluate and optimize the text block size in multiple dimensions, the text block optimization module 13 introduces a three-dimensional coordinate system. Among them, the first coordinate axis (such as the x-axis) represents the length of the text block, the second coordinate axis (such as the y-axis) represents the width of the text block, and the third coordinate axis (such as the z-axis) represents the height of the text block. These three dimensions comprehensively characterize the spatial size characteristics of the text block. Based on the above three-dimensional coordinate axes, the text block optimization module 13 constructs a three-dimensional optimization space. In this space, each coordinate point corresponds to a possible text block size combination (length, width, height). By mapping the extracted text block size identifier marked as "accurate" to the three-dimensional optimization space, a series of text block distribution coordinates distributed in the space are obtained.

[0069] In order to further optimize the selection of text block size, the text block optimization module 13 removes outliers from the text block distribution coordinates. The so-called outliers refer to those coordinate points that are far away from the main distribution area and may represent abnormalities or noise. By eliminating outliers, a more concentrated and stable set of text block distribution coordinates is obtained, eliminating the interference of some accidental factors. After obtaining the concentrated text block distribution coordinates, the text block optimization module 13 carries out density analysis. The purpose of density analysis is to identify the areas where text blocks are most densely distributed, because in these areas, the comprehensive performance of different text block size settings is the closest and best. By calculating the density parameters around each coordinate point, the text block optimization module 13 obtains the density parameters of the concentrated text block distribution coordinates. Afterwards, the text block optimization module 13 extracts the coordinate point corresponding to the maximum density from the concentrated text block distribution coordinate density parameters, and outputs it as the first target text block size. The maximum value point represents the text block size with the best comprehensive performance in the historical data, and is also the most concentrated and representative optimization result, so it is selected as the first target text block size.

[0070] By adopting three-dimensional optimization and making full use of existing recognition logs, the optimal text blocks are dynamically evaluated and screened from multiple dimensions to improve the adaptability and accuracy of text block division.

[0071] Furthermore, the steps of constructing the first-level deep learning semantic model and the second-level deep learning semantic model include:

[0072] Construct semantic analysis loss function:

[0073] ,

[0074] ,

[0075] ,

[0076] in, Characterize business type identification loss, Characterizes the predicted business type label of the i-th training, Characterizes the supervised business type label of the i-th training, Represents the classification loss, Represent the intent recognition sentence of the i-th training dialogue sequence, mark the j-th sequence number text, The previous text representing the intent recognition sentence text of the i-th training dialogue, Represents the predicted business type label identified based on the previous text, i.e. the first j-1 sentences, Characterize the number of business categories;

[0077] According to the semantic analysis loss function, the first-level deep learning semantic model and the second-level deep learning semantic model are trained.

[0078] In a preferred embodiment, when constructing the first-level deep learning semantic model and the second-level deep learning semantic model, first, a semantic analysis loss function is constructed as:

[0079] ,

[0080] ,

[0081] ,

[0082] The semantic analysis loss function consists of two parts: cross entropy loss and semantic consistency loss, which are used to evaluate the performance of the model in business type recognition and context consistency. It is composed of the average loss of n trainings. For the i-th training, the loss is the sum of two items. The first item is the cross entropy loss , used to measure the business type label predicted by the model With the true label The difference between. is the prediction result of the model at the i-th training. is the corresponding manually annotated supervision label, and M represents the total number of business categories. It describes the degree of deviation between the probability distribution of the predicted label and the true distribution. The second term is the semantic consistency loss , used to measure the model's ability to identify the current intent recognition sentence When, as in the previous The semantic consistency of . represents the jth sentence in the dialogue sequence during the i-th training, express Previous The context information of the sentence. In the semantic analysis loss function, is the weight coefficient of the semantic consistency loss, which is used to balance the contribution of the two losses to the total loss. By jointly optimizing these two losses, we can improve the accuracy of business type recognition while taking into account the coherence of the conversation context, and learn a more comprehensive semantic representation that is more in line with the true intention.

[0083] In the process of building the first-level deep learning semantic model and the second-level deep learning semantic model, the semantic analysis loss function is used as the optimization target to back-propagate and iteratively update the parameters of the first-level deep learning semantic model and the second-level deep learning semantic model. By minimizing LOSS, we gradually learn to accurately understand and express customer intentions, which improves the overall effect of business type identification and dispatch management.

[0084] By introducing the constructed semantic analysis loss function when building the semantic understanding model, integrating the classification loss and semantic consistency loss, the model is prompted to pay attention to the coherence of the context while recognizing the intent. This multi-objective optimization approach can effectively improve the semantic understanding ability and actual effect of the dispatching link.

[0085] Furthermore, the embodiment of the present application also includes:

[0086] Obtaining a first conversation vector record data set and a business type identification data set, a customer quantity identification data set, and a customer time limit identification data set;

[0087] Segmenting the first conversation vector record data set according to the text block identifier size set of the business type identifier data set to obtain a second conversation vector record data set;

[0088] According to the semantic analysis loss function, calling the first conversation vector record data set and the business type identification data set to train the first-level deep learning semantic model;

[0089] According to the semantic analysis loss function, the second conversation vector record dataset, the business type identification dataset, the customer quantity identification dataset and the customer time limit identification dataset are retrieved to train the secondary deep learning semantic model.

[0090] In a preferred embodiment, when training the first-level deep learning semantic model and the second-level deep learning semantic model according to the semantic analysis loss function, first, four types of data sets are obtained, namely, the first conversation vector record data set, the business type identification data set, the customer quantity identification data set, and the customer time limit identification data set. Among them, the first conversation vector record data set contains a large number of customer conversation texts and their corresponding vector representations; the business type identification data set marks the business type to which each conversation belongs; the customer quantity identification data set and the customer time limit identification data set respectively record the specific business quantity and time limit requirements involved in the conversation. Next, the first conversation vector record data set is further processed. According to the text block size information provided by the business type identification data set, each conversation vector is segmented according to the corresponding text block size to obtain the second conversation vector record data set. Compared with the first conversation vector record data set, each sample of the second conversation vector record data set contains more fine-grained semantic fragments, providing more refined input for the training of the second-level deep learning semantic model.

[0091] Subsequently, the first-level deep learning semantic model is trained. The first conversation vector record dataset is used as input, the business type identification dataset is used as the supervision label, and the first-level deep learning semantic model is trained and optimized according to the semantic analysis loss function. By minimizing the semantic analysis loss, the first-level deep learning semantic model learns the global understanding of the entire conversation, and can grasp the customer's intention from a macro perspective, laying the foundation for the next step of refined analysis. After the first-level deep learning semantic model is trained, the second-level deep learning semantic model is trained. Different from the first-level deep learning semantic model, the training data of the second-level deep learning semantic model is richer and more diverse. At the same time, the second conversation vector record dataset, the business type identification dataset, the customer quantity identification dataset, and the customer time limit identification dataset are used to comprehensively optimize the performance of the model in multiple dimensions. Similarly, the semantic analysis loss function is used as the optimization target to drive the second-level deep learning semantic model to learn the deep semantic features of the conversation fragment, so as to achieve accurate understanding and judgment of the details of customer needs.

[0092] Through a phased and granular training process, the first-level deep learning semantic model and the second-level deep learning semantic model complement each other and work together. The first-level deep learning semantic model is responsible for controlling the global intent, and the second-level deep learning semantic model is responsible for mining local details, jointly building a complete and efficient semantic understanding framework.

[0093] Furthermore, the embodiment of the present application also includes:

[0094] According to the semantic analysis loss function, calling the second conversation vector record data set and the business type identification data set to train the business type analysis channel;

[0095] According to the mean square error loss function, the second conversation vector record data set and the customer quantity identification data set are retrieved to train the customer quantity analysis channel;

[0096] According to the mean square error loss function, the second conversation vector record data set and the customer time limit identification data set are retrieved to train the customer time limit analysis channel;

[0097] The business type analysis channel, the customer quantity analysis channel and the customer time limit analysis channel are combined to generate the secondary deep learning semantic model.

[0098] In a preferred embodiment, a multi-channel parallel learning mechanism is introduced when training the secondary deep learning semantic model.

[0099] First, according to the semantic analysis loss function, the second dialogue vector record dataset and the business type identification dataset are used to train a sub-model specifically used for business type analysis, called the business type analysis channel. This channel focuses on the business type information contained in the dialogue segment, and learns to identify the specific business field to which the customer needs belong by optimizing the semantic analysis loss, such as bill inquiry, fault repair, etc. The training of the business type analysis channel enables the secondary deep learning semantic model to have the ability to accurately understand the business context. At the same time, according to the mean square error loss function, the second dialogue vector record dataset and the customer quantity identification dataset are used to train a sub-model specifically used for customer quantity analysis, called the customer quantity analysis channel. This channel focuses on the specific business quantity information expressed in the dialogue segment, such as "I need to apply for 3 mobile phone numbers". By minimizing the mean square error between the output value and the label value, the customer quantity analysis channel has mastered the skills of accurately extracting quantity details from the dialogue, so that the secondary deep learning semantic model can accurately perceive the quantity level of customer needs. At the same time, according to the mean square error loss function, the second dialogue vector record dataset and the customer time limit identification dataset are used to train a sub-model specifically used for customer time limit analysis, called the customer time limit analysis channel. This channel focuses on the time requirement information contained in the conversation fragment, such as "complete it within today". By learning and fitting the time limit labels, the customer time limit analysis channel enables the secondary deep learning semantic model to understand the customer's time requirements, providing an important basis for subsequent task dispatch.

[0100] After the business type analysis channel, customer quantity analysis channel, and customer time limit analysis channel are trained, the three analysis channels are merged to form a complete secondary deep learning semantic model. The merged model structurally contains three parallel feature extraction branches, corresponding to the three dimensions of business type, customer quantity, and customer time limit. This multi-channel design enables the model to focus on different semantic elements in the conversation at the same time and fully understand all aspects of customer intent.

[0101] Through the above training, the secondary deep learning semantic model has achieved all-round and multi-angle semantic understanding capabilities. The business type analysis channel provides an accurate grasp of the conversation context, while the customer quantity analysis channel and customer time limit analysis channel supplement the extraction and quantification of key details. The synergy of the three channels enables the model to fully deconstruct customer intentions and provide sufficient information support for dispatch management.

[0102] The deep learning-driven dispatch management system provided by the embodiment of the present invention has at least the following technical effects:

[0103] The dialogue vector acquisition module is used to obtain the first dialogue vector between the customer and the intelligent customer service, providing a data basis for subsequent intent recognition. The first-level intent recognition module is used to perform customer intent recognition on the first dialogue vector through a first-level deep learning semantic model, obtain the first customer demand business, use the pre-trained first-level deep learning semantic model to perform preliminary customer intent judgment on the first dialogue vector, quickly lock the approximate business scope to which the customer demand belongs, and provide directional guidance for the next step of in-depth analysis. The text block optimization module is used to optimize the text block according to the first customer demand business, obtain the first target text block size, segment the first dialogue vector according to the first target text block size, obtain the second dialogue vector, and dynamically adjust the length and content of the original dialogue vector according to the first customer demand business given by the first-level intent recognition, and perform secondary semantic judgment with a more refined text block granularity. The second-level intent recognition module is used to perform customer intent recognition on the second dialogue vector through a second-level deep learning semantic model, obtain the second customer demand business, customer demand quantity and customer demand time limit, and achieve more refined semantic understanding and intent judgment of customer demand. The order dispatch management module is used to construct work order information for order dispatch management when the first customer demand business and the second customer demand business are the same, based on the first customer demand business, the customer demand quantity and the customer demand time limit, thereby achieving the technical effect of improving the accuracy of intent recognition and optimizing the order dispatch effect of the intelligent customer service system.

[0104] Embodiment 2:

[0105] like Figure 2 As shown, based on the same inventive concept of the deep learning driven dispatch management system provided in the first embodiment, the embodiment of the present invention also provides a deep learning driven dispatch management method, including:

[0106] Obtain the first conversation vector between the customer and the intelligent customer service;

[0107] Using a first-level deep learning semantic model, the customer intent is identified on the first conversation vector to obtain the first customer demand service;

[0108] Optimize the text block according to the first customer demand business to obtain a first target text block size, and segment the first dialogue vector according to the first target text block size to obtain a second dialogue vector;

[0109] Using a secondary deep learning semantic model, the second conversation vector is used to identify the customer's intention, and the second customer's required business, customer required quantity, and customer required time limit are obtained;

[0110] When the first customer demand business and the second customer demand business are the same, work order information is constructed according to the first customer demand business, the customer demand quantity and the customer demand time limit to perform order management.

[0111] Furthermore, the embodiment of the present application also includes:

[0112] According to the first customer demand business, sort the business type set to obtain a business type sorting result;

[0113] When the first customer demand service and the second customer demand service are different, extracting the first serial number service type of the service type sorting result and setting it as the third customer demand service;

[0114] Optimize the text block according to the third customer demand business to obtain a second target text block size, and segment the first dialogue vector according to the second target text block size to obtain a third dialogue vector;

[0115] By using the secondary deep learning semantic model, the third dialogue vector is used to identify the customer's intention, and a fourth customer demand service, customer demand quantity, and customer demand time limit are obtained;

[0116] When the third customer demand business and the fourth customer demand business are the same, constructing work order information for dispatch management according to the third customer demand business, the customer demand quantity and the customer demand time limit;

[0117] When the third customer demand service and the fourth customer demand service are different, the second serial number service type of the service type sorting result is extracted for iterative analysis.

[0118] Further, when the first customer demand service and the second customer demand service are different, the first customer demand service is updated to obtain a third customer demand service, including:

[0119] Get a business type set;

[0120] Based on the first customer demand business, traverse the business type set to perform business text vector similarity analysis to obtain a business text vector similarity set;

[0121] The business type set is sorted from large to small according to the business text vector similarity set to obtain a business type sorting result.

[0122] Further, optimizing the text block according to the first customer demand business to obtain a first target text block size includes:

[0123] According to the first customer demand business, obtaining a historical recognition log of the first customer demand business by the secondary deep learning semantic model, wherein the historical recognition log includes a recognition accuracy identifier and a text block size identifier;

[0124] Extract the text block size identifier that is identified as accurate in the historical recognition log;

[0125] The length of the text block is the first coordinate axis, the width of the text block is the second coordinate axis, and the height of the text block is the third coordinate axis;

[0126] Constructing an optimization three-dimensional space according to the first coordinate axis, the second coordinate axis and the third coordinate axis;

[0127] Distributing the text block size identifiers in the optimization three-dimensional space to obtain text block distribution coordinates;

[0128] Deleting outlier coordinates based on the text block distribution coordinates to obtain concentrated text block distribution coordinates;

[0129] Performing density analysis on the concentrated text block distribution coordinates to obtain a density parameter of the concentrated text block distribution coordinates;

[0130] The maximum value coordinates of the concentrated text block distribution coordinate density parameters are extracted to generate the first target text block size.

[0131] Furthermore, the steps of constructing the first-level deep learning semantic model and the second-level deep learning semantic model include:

[0132] Construct semantic analysis loss function:

[0133] ,

[0134] ,

[0135] ,

[0136] in, Characterize business type identification loss, Characterizes the predicted business type label of the i-th training, Characterizes the supervised business type label of the i-th training, Represents the classification loss, Represent the intent recognition sentence of the i-th training dialogue sequence, mark the j-th sequence number text, The previous text representing the intent recognition sentence text of the i-th training dialogue, Represents the predicted business type label identified based on the previous text, i.e. the first j-1 sentences, Characterize the number of business categories;

[0137] According to the semantic analysis loss function, the first-level deep learning semantic model and the second-level deep learning semantic model are trained.

[0138] Further, according to the semantic analysis loss function, training the first-level deep learning semantic model and the second-level deep learning semantic model includes:

[0139] Obtaining a first conversation vector record data set and a business type identification data set, a customer quantity identification data set, and a customer time limit identification data set;

[0140] Segmenting the first conversation vector record data set according to the text block identifier size set of the business type identifier data set to obtain a second conversation vector record data set;

[0141] According to the semantic analysis loss function, calling the first conversation vector record data set and the business type identification data set to train the first-level deep learning semantic model;

[0142] According to the semantic analysis loss function, the second conversation vector record dataset, the business type identification dataset, the customer quantity identification dataset and the customer time limit identification dataset are retrieved to train the secondary deep learning semantic model.

[0143] Further, according to the semantic analysis loss function, the second conversation vector record data set, the business type identification data set, the customer quantity identification data set and the customer time limit identification data set are retrieved to train the secondary deep learning semantic model, including:

[0144] According to the semantic analysis loss function, calling the second conversation vector record data set and the business type identification data set to train the business type analysis channel;

[0145] According to the mean square error loss function, the second conversation vector record data set and the customer quantity identification data set are retrieved to train the customer quantity analysis channel;

[0146] According to the mean square error loss function, the second conversation vector record data set and the customer time limit identification data set are retrieved to train the customer time limit analysis channel;

[0147] The business type analysis channel, the customer quantity analysis channel and the customer time limit analysis channel are combined to generate the secondary deep learning semantic model.

[0148] Embodiment three:

[0149] See also Figure 3 , Figure 3A schematic diagram of an embodiment of a computer-readable storage medium provided in an embodiment of the present invention. Figure 3 As shown, this embodiment provides a computer-readable storage medium 200 on which a computer program 211 is stored. When the computer program 211 is executed by a processor, a deep learning-driven dispatch management system is implemented.

[0150] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and for parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0151] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0152] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0153] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0154] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0155] Although preferred embodiments of the present invention have been described, additional changes and modifications may occur to these embodiments once those skilled in the art understand the basic inventive concepts.

[0156] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention belong to the scope of the present invention and its equivalent technology, the present invention is also intended to include these changes and variations.

Claims

1. A deep learning-driven dispatch management system, characterized by: include: A dialogue vector acquisition module, wherein the dialogue vector acquisition module is used to obtain a first dialogue vector between the customer and the intelligent customer service; A primary intention recognition module, the primary intention recognition module is used to recognize the customer intention of the first dialogue vector through a primary deep learning semantic model to obtain a first customer demand business; a text block optimization module, the text block optimization module being used to optimize the text block according to the first customer demand business, obtain a first target text block size, and segment the first dialogue vector according to the first target text block size to obtain a second dialogue vector; A secondary intention recognition module, the secondary intention recognition module is used to recognize the customer intention of the second dialogue vector through a secondary deep learning semantic model, and obtain the second customer demand business, customer demand quantity and customer demand time limit; An order management module, wherein when the first customer demand business and the second customer demand business are the same, the order management module is used to construct work order information for order management according to the first customer demand business, the customer demand quantity and the customer demand time limit; The execution steps of the text block optimization module include: According to the first customer demand business, obtaining a historical recognition log of the first customer demand business by the secondary deep learning semantic model, wherein the historical recognition log includes a recognition accuracy identifier and a text block size identifier; Extract the text block size identifier that is identified as accurate in the historical recognition log; The length of the text block is the first coordinate axis, the width of the text block is the second coordinate axis, and the height of the text block is the third coordinate axis; Constructing an optimization three-dimensional space according to the first coordinate axis, the second coordinate axis and the third coordinate axis; Distributing the text block size identifiers in the optimization three-dimensional space to obtain text block distribution coordinates; Deleting outlier coordinates based on the text block distribution coordinates to obtain concentrated text block distribution coordinates; Performing density analysis on the concentrated text block distribution coordinates to obtain a density parameter of the concentrated text block distribution coordinates; Extracting the maximum value coordinates of the concentrated text block distribution coordinate density parameters to generate the first target text block size; The steps of constructing the first-level deep learning semantic model and the second-level deep learning semantic model include: Construct semantic analysis loss function: ， ， ， in, Characterize business type identification loss, Characterizes the predicted business type label of the i-th training, Characterizes the supervised business type label of the i-th training, Represents the classification loss, Represent the intent recognition sentence of the i-th training dialogue sequence, mark the j-th sequence number text, The previous text representing the intent recognition sentence text of the i-th training dialogue, Represents the predicted business type label identified based on the previous text, i.e. the first j-1 sentences, represents the number of business categories, and n represents the preset number of training times for calculating the loss; According to the semantic analysis loss function, the first-level deep learning semantic model and the second-level deep learning semantic model are trained.

2. The system according to claim 1, characterized in that The system further includes an intention correction processing module, and the execution steps of the intention correction processing module include: According to the first customer demand business, sort the business type set to obtain a business type sorting result; When the first customer demand service and the second customer demand service are different, extracting the first serial number service type of the service type sorting result and setting it as the third customer demand service; Optimize the text block according to the third customer demand business to obtain a second target text block size, and segment the first dialogue vector according to the second target text block size to obtain a third dialogue vector; By using the secondary deep learning semantic model, the third dialogue vector is used to identify the customer's intention, and a fourth customer demand service, customer demand quantity, and customer demand time limit are obtained; When the third customer demand business and the fourth customer demand business are the same, constructing work order information for dispatch management according to the third customer demand business, the customer demand quantity and the customer demand time limit; When the third customer demand service and the fourth customer demand service are different, the second serial number service type of the service type sorting result is extracted for iterative analysis.

3. The system according to claim 2, characterized in that The execution steps of the intention correction processing module also include: Get a business type set; Based on the first customer demand business, traverse the business type set to perform business text vector similarity analysis to obtain a business text vector similarity set; The business type set is sorted from large to small according to the business text vector similarity set to obtain a business type sorting result.

4. The system according to claim 1, characterized in that According to the semantic analysis loss function, training the first-level deep learning semantic model and the second-level deep learning semantic model includes: Obtaining a first conversation vector record data set and a business type identification data set, a customer quantity identification data set, and a customer time limit identification data set; Segmenting the first conversation vector record data set according to the text block identifier size set of the business type identifier data set to obtain a second conversation vector record data set; According to the semantic analysis loss function, calling the first conversation vector record data set and the business type identification data set to train the first-level deep learning semantic model; According to the semantic analysis loss function, the second conversation vector record dataset, the business type identification dataset, the customer quantity identification dataset and the customer time limit identification dataset are retrieved to train the secondary deep learning semantic model.

5. The system according to claim 4, characterized in that According to the semantic analysis loss function, calling the second conversation vector record data set, the business type identification data set, the customer quantity identification data set, and the customer time limit identification data set to train the secondary deep learning semantic model includes: According to the semantic analysis loss function, calling the second conversation vector record data set and the business type identification data set to train the business type analysis channel; According to the mean square error loss function, the second conversation vector record data set and the customer quantity identification data set are retrieved to train the customer quantity analysis channel; According to the mean square error loss function, the second conversation vector record data set and the customer time limit identification data set are retrieved to train the customer time limit analysis channel; The business type analysis channel, the customer quantity analysis channel and the customer time limit analysis channel are combined to generate the secondary deep learning semantic model.

6. A deep learning-driven dispatch management method, characterized in that: include: The acquisition module is used to obtain the first conversation vector between the customer and the intelligent customer service; Using a first-level deep learning semantic model, the customer intent is identified on the first conversation vector to obtain the first customer demand service; Optimize the text block according to the first customer demand business to obtain a first target text block size, and segment the first dialogue vector according to the first target text block size to obtain a second dialogue vector; Using a secondary deep learning semantic model, the second conversation vector is used to identify the customer's intention, and the second customer's required business, customer required quantity, and customer required time limit are obtained; When the first customer demand business and the second customer demand business are the same, constructing work order information for dispatch management according to the first customer demand business, the customer demand quantity and the customer demand time limit; The step of optimizing the text block according to the first customer demand business to obtain a first target text block size, and segmenting the first dialogue vector according to the first target text block size to obtain a second dialogue vector includes: According to the first customer demand business, obtaining a historical recognition log of the first customer demand business by the secondary deep learning semantic model, wherein the historical recognition log includes a recognition accuracy identifier and a text block size identifier; Extract the text block size identifier that is identified as accurate in the historical recognition log; The length of the text block is the first coordinate axis, the width of the text block is the second coordinate axis, and the height of the text block is the third coordinate axis; Constructing an optimization three-dimensional space according to the first coordinate axis, the second coordinate axis and the third coordinate axis; Distributing the text block size identifiers in the optimization three-dimensional space to obtain text block distribution coordinates; Deleting outlier coordinates based on the text block distribution coordinates to obtain concentrated text block distribution coordinates; Performing density analysis on the concentrated text block distribution coordinates to obtain a density parameter of the concentrated text block distribution coordinates; Extracting the maximum value coordinates of the concentrated text block distribution coordinate density parameters to generate the first target text block size; The steps of constructing the first-level deep learning semantic model and the second-level deep learning semantic model include: Construct semantic analysis loss function: ， ， ， in, Characterize business type identification loss, Characterizes the predicted business type label of the i-th training, Characterizes the supervised business type label of the i-th training, Represents the classification loss, Represent the intent recognition sentence of the i-th training dialogue sequence, mark the j-th sequence number text, The previous text representing the intent recognition sentence text of the i-th training dialogue, Represents the predicted business type label identified based on the previous text, i.e. the first j-1 sentences, represents the number of business categories, and n represents the preset number of training times for calculating the loss; According to the semantic analysis loss function, the first-level deep learning semantic model and the second-level deep learning semantic model are trained.

7. A non-transitory computer-readable storage medium, characterized in that: The storage medium stores a computer software program, which, when executed by a processor, implements the deep learning-driven dispatch management system as described in any one of claims 1 to 5.

Images