Automatic real-time order dispatching method and system based on service classification

Abstract

The invention discloses an automatic real-time order dispatching method and system based on business classification, and the method is characterized in that the method comprises the steps: collecting complaint information, respectively extracting complaint information of different dimensions, and constructing a complaint data set based on the extracted complaint information; after the complaint data set is input into a deep neural network, complaint information of different dimensions is decomposed into a first decomposition matrix and a second decomposition matrix; calculating common semantic features in complaint information of different dimensions by using the first decomposition matrix and the second decomposition matrix; training the deep neural network based on the common semantic features, and obtaining the output of the deep neural network; and confirming service classification by using the output of the deep neural network so as to realize real-time order dispatching for complaints. According to the method, refined, personalized and intelligent automatic real-time order dispatching can be quickly and accurately carried out, the customer experience is optimized, and the order dispatching accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of power systems, and more specifically, to a method and system for automatically implementing dispatching orders based on business classification. Background Art

[0002] Power supply service complaints refer to power customers within the power company's business scope expressing dissatisfaction with the damage to their rights and interests caused by the power supply company's responsibility in terms of power supply services, business operations, power outages, power supply quality, and grid construction, and demanding that their rights and interests be protected.

[0003] Customer complaints mainly include service complaints, business complaints, power outage complaints, power supply quality complaints, power grid construction complaints and other types. In service complaints, users complain about the irregular service behavior of power supply company employees and the blocked service channels. In business complaints, power supply companies have problems such as overtime, negligence, and errors in business expansion and registration, electricity consumption changes, meter reading and collection, electricity charges and prices, and energy metering. Therefore, it is necessary to make business improvements based on the content of the complaints. In power outage complaints, there are service deviations in power outage information announcements, power outage plan execution, emergency repair quality, and the opening of value-added services. Therefore, it is necessary to improve the service content based on complaints. In addition, in power supply quality complaints, customers may raise opinions on power supply quality problems such as voltage deviation, frequency deviation, voltage imbalance, voltage fluctuation or flicker in the transmitted electricity. Therefore, in order to optimize the power supply capacity in response to power supply quality complaints, it is necessary to make substantial improvements to transmission parameters such as precise voltage deviation and frequency deviation.

[0004] At present, power companies mainly provide customer service through WeChat business halls, online business halls, 95598 call hotlines, physical business halls, etc. Customer service for complaints mainly relies on manual customer service for classification and dispatch, and the power industry has not yet truly realized intelligent power customer service.

[0005] In the traditional complaint handling process, operators usually answer user complaints over the phone, then compile them into text messages for recording, and then assign them to relevant departments and personnel for processing. However, due to the wide variety of complaints and the involvement of many departments and personnel, qualified operators often need a lot of training before they can take up their posts, and they are unable to handle emergency issues during special hours at night. Therefore, there is an urgent need for automated dispatching applications to improve the efficiency of customer complaint handling.

[0006] Secondly, the categories of complaint cases are stored and classified in the database in a tree-like hierarchical structure. When accepting a complaint, it is necessary to first correctly classify it at each level of the tree, and finally accurately trace the complaint. If a category in an upper layer is classified incorrectly, the classification of the next level will also be wrong, which leads to the accumulation of error rates layer by layer.

[0007] On the other hand, with the upgrade of the customer service platform, various forms of complaints such as voice, text, pictures and videos can be obtained during the complaint service process. The complaint contents of various modes are interrelated and complementary, but it is difficult to provide complete complaint information depending on the source and access method of the complaint. These problems lead to the accumulation of classification model error rates.

[0008] When the error rate of the classification model is high, the automatic real-time dispatching service cannot be implemented accurately and effectively.

[0009] In view of the above problems, there is an urgent need for a new automatic real-time dispatching method and system based on business classification. Summary of the invention

[0010] In order to address the deficiencies in the prior art, the present invention provides an automatic real-time dispatching method and system based on business classification, which collects complaint information, constructs a complaint data set, decomposes the matrix to obtain the common semantic features of different types of information, and finally realizes automatic real-time dispatching for complaints.

[0011] The present invention adopts the following technical solution.

[0012] A first aspect of the present invention relates to an automatic real-time dispatching method based on business classification, the method comprising the following steps: collecting complaint voice, complaint text, complaint pictures and complaint videos, respectively extracting complaint information of different dimensions and constructing a complaint data set based on the complaint information; after inputting the complaint data set into a deep neural network, decomposing the complaint information of different dimensions into a first decomposition matrix and a second decomposition matrix; using the first decomposition matrix and the second decomposition matrix, calculating the common semantic features in the complaint information of different dimensions; based on the common semantic features, training the deep neural network and obtaining the output of the deep neural network; using the output of the deep neural network, confirming the business classification to achieve real-time dispatching for complaints.

[0013] Preferably, respectively extracting complaint information of different dimensions and constructing a complaint dataset based on the complaint information, including: the complaint dataset is

[0014]

[0015] m is the dimension number of complaint voice, complaint text, complaint picture and complaint video,

[0016] is the feature matrix of complaint information under the current dimension m,

[0017] is the difference between the current dimension m and The matrix composed of other data instances of .

[0018] Preferably, inputting the complaint data set into a deep neural network comprises: using the deep neural network to convert the complaint data set into a feature data set matrix In the formula, and They are and The deep feature transformation matrix of .

[0019] Preferably, the complaint information of different dimensions is decomposed into a first decomposition matrix and a second decomposition matrix, including: the product of the feature data set matrix V and the first decomposition matrix and the second decomposition matrix is ​​equal; and the first decomposition matrix W is transformed into:

[0020]

[0021] The second decomposition matrix H is transformed into:

[0022]

[0023] Where ij is the i-th row and j-th column of the matrix,

[0024] The iteration goal is min∑ ij (V ij -(W.H) ij ) 2 .

[0025] Preferably, the first decomposition matrix and the second decomposition matrix are used to calculate the common semantic features in the complaint information of different dimensions, including: the common semantic features are:

[0026]

[0027] In the formula, H C is the shared feature matrix of complaint information under all dimensions, They are different from the shared feature matrix H in different dimensions. C The matrix composed of other data instances of .

[0028] Preferably, the shared feature matrix H of complaint information under all dimensions C and other data instances Obtained through the objective function; the objective function is:

[0029]

[0030]

[0031] Where Wm is the first decomposition matrix under the current dimension m,

[0032] Tr(·) is the trace of the matrix,

[0033] α m is the dimension parameter,

[0034] L m is a Laplace matrix.

[0035] Preferably, training the deep neural network further includes: constructing a deep neural network parameter θ m , based on the deep neural network parameters θ m Adjust the lth layer of the deep neural network for the The network weights of the dimensions and the bias vector The deep neural network parameters θ m Initialized at the beginning of the iteration, using the function in each iteration Solve to improve.

[0036] Preferably, the first layer of the deep neural network is for The network weights of the dimensions and the bias vector The change is performed according to the product of the descent gradient and the learning rate of the deep neural network; the network weights The descent gradient is:

[0037]

[0038] The bias vector The descent gradient is:

[0039]

[0040] The deep neural network learning rate is set for the current dimension m.

[0041] A second aspect of the present invention relates to an automatic real-time dispatching system based on business classification, the system comprising an acquisition module, a decomposition module, a calculation module, a training module and a classification module; wherein the acquisition module is used to collect complaint voice, complaint text, complaint pictures and complaint videos, respectively extract complaint information of different dimensions and construct a complaint data set based on the complaint information; the decomposition module is used to input the complaint data set into a deep neural network, and then decompose the complaint information of different dimensions into a first decomposition matrix and a second decomposition matrix; the calculation module is used to calculate the common semantic features in the complaint information of different dimensions using the first decomposition matrix and the second decomposition matrix; the training module is used to train the deep neural network based on the common semantic features and obtain the output of the deep neural network; the classification module is used to confirm the business classification using the output of the deep neural network to achieve real-time dispatching for complaints.

[0042] Preferably, the collection module is used to extract complaint information of different dimensions and construct a complaint data set based on the complaint information, including:

[0043] The complaint dataset is

[0044]

[0045] m is the dimension number of complaint voice, complaint text, complaint picture and complaint video,

[0046] is the feature matrix of complaint information under the current dimension m,

[0047] is the difference between the current dimension m and The matrix composed of other data instances of .

[0048] Preferably, the decomposition module, used to input the complaint data set into a deep neural network, comprises:

[0049] The deep neural network is used to convert the complaint dataset into a feature dataset matrix.

[0050] In the formula, and They are and The deep feature transformation matrix of .

[0051] Preferably, the calculation module is used to decompose complaint information of different dimensions into a first decomposition matrix and a second decomposition matrix, including:

[0052] The feature data set matrix V is equal to the product of the first decomposition matrix and the second decomposition matrix; and,

[0053] The first decomposition matrix W is transformed into:

[0054]

[0055] The second decomposition matrix H is transformed into:

[0056]

[0057] Where ij is the i-th row and j-th column of the matrix,

[0058] The iteration goal is min∑ ij (V ij -(W.H) ij ) 2 .

[0059] Preferably, the calculation module is used to calculate the common semantic features in the complaint information of different dimensions by using the first decomposition matrix and the second decomposition matrix, including:

[0060] The common semantic features are:

[0061]

[0062] In the formula, H C is the shared feature matrix of complaint information under all dimensions,

[0063] They are different from the shared feature matrix H in different dimensions. C The matrix composed of other data instances of .

[0064] Preferably, in the calculation module, the shared feature matrix H of complaint information under all dimensions C and other data instances Obtained through the objective function;

[0065] The objective function is:

[0066]

[0067]

[0068] Where W m is the first decomposition matrix under the current dimension m,

[0069] Tr(·) is the trace of the matrix,

[0070] α m is the dimension parameter,

[0071] Lm is a Laplace matrix.

[0072] Preferably, the training module, used for training the deep neural network, comprises:

[0073] Construct deep neural network parameters θ m , based on the deep neural network parameters θ m Adjust the lth layer of the deep neural network for the The network weights of the dimensions and the bias vector

[0074] The deep neural network parameters θ m Initialized at the beginning of the iteration, using the function in each iteration Solve to improve.

[0075] Preferably, in the training module,

[0076] The first layer of the deep neural network targets The network weights of the dimensions and the bias vector The changes are performed by the product of the descent gradient and the deep neural network learning rate;

[0077] The network weights The descent gradient is:

[0078]

[0079] The bias vector The descent gradient is:

[0080]

[0081] The deep neural network learning rate is set for the current dimension m.

[0082] The third aspect of the present invention relates to an electronic device, comprising a processor and a storage medium; the characteristics are: the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method in the first aspect of the present invention.

[0083] The beneficial effect of the present invention is that, compared with the prior art, the automatic real-time dispatching method and system based on business classification in the present invention collects complaint information, constructs a complaint data set, decomposes the matrix to obtain the common semantic features of different types of information, and finally realizes automatic real-time dispatching for complaints.

[0084] The beneficial effects of the present invention also include:

[0085] 1. Considering that in practice, case assignment is often not a single level, but a level-by-level assignment, each level of case assignment can be regarded as a classification task, so the error rate of the final classification will be accumulated level by level. To this end, the present invention realizes the joint optimization of multi-level classification labels. The method can quickly and accurately perform refined, personalized, intelligent and automatic real-time dispatch through historical customer service information and knowledge graphs, optimize customer experience, and improve the accuracy of dispatch.

[0086] 2. Methods Based on the real-time semantic understanding of customer service in multiple business scenarios, a full-scenario perception pre-training model is constructed. Multimodal features are extracted using models such as Robert, Wav2cev, and ViT. Combined with historical customer service information and knowledge graphs, a hierarchical classification evaluation index based on a hierarchical structure is designed. Multiple classifiers in the hierarchy are jointly optimized by minimizing the global loss function to improve accuracy and timeliness.

[0087] 3. In machine learning, the loss function is used to measure the gap between the output predicted by the model and the true label. The smaller the value of the loss function, the closer the model's prediction result is to the true label. Therefore, by minimizing the loss function, the optimal model parameters can be obtained, thereby improving the model's prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 A schematic diagram of a process flow of an automatic real-time dispatching method based on business classification according to the present invention;

[0089] Figure 2 A schematic diagram of the execution process of the perception scheduling decision in an automatic real-time dispatching method based on business classification of the present invention;

[0090] Figure 3 The present invention is a schematic diagram of the module structure of an automatic real-time dispatching system based on business classification. DETAILED DESCRIPTION

[0091] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in the present invention are only embodiments of a part of the present invention, not all embodiments. Based on the spirit of the present invention, all other embodiments not recorded in the present invention obtained by ordinary technicians in this field according to the embodiments recorded in the present invention without making creative work should belong to the protection scope of the present invention.

[0092] Figure 1 FIG. 1 is a flow chart of an automatic real-time dispatching method based on business classification according to the present invention. Figure 1The first aspect of the present invention relates to an automatic real-time dispatching method based on business classification, and the method includes steps 1 to 5.

[0093] Step 1: Collect complaint voice, complaint text, complaint pictures and complaint videos, extract complaint information of different dimensions respectively and construct a complaint dataset based on the complaint information.

[0094] With the widespread application of multimedia technology and the enrichment of data description methods, a large amount of data content in different forms is widely available, and the description of a description object is jointly realized from multiple angles, in multiple languages, and in multiple visual forms. In the complaint information of the power system, the same complaint content may include multiple different forms such as audio, image, text, and video. Pictures and videos from different sources may have different pixel intensity, grayscale, and texture information. How to integrate various types of information and realize decision-making and prediction based on multi-modal data is the key issue considered in this invention.

[0095] To this end, the method collects complaint voice, text, picture and video information through a variety of different complaint channels, and constructs a complaint dataset by using the different dimensions of the collected data and the dimensions of the data content.

[0096] Preferably, the complaint information of different dimensions is extracted respectively and a complaint data set is constructed based on the complaint information, including: the complaint data set is

[0097]

[0098] m is the dimension number of complaint voice, complaint text, complaint picture and complaint video,

[0099] is the feature matrix of complaint information under the current dimension m,

[0100] is the difference between the current dimension m and The matrix composed of other data instances of .

[0101] In the above formula, complaint voice, text, picture and video are identified by dimensions 1, 2, 3 and 4 respectively. Considering that in the complaint data set, data content of different dimensions can represent essentially the same complaint issues, the data set includes 4 different data dimensions and multiple data instances belonging to each data dimension. Among them, the information of some data instances can be complete, including all information of the four dimensions and the correlation between the information in each dimension. For this reason, the present invention records it as That is, the matrix composed of instances of the complete data instance in the current dimension, that is, the feature matrix mentioned above.

[0102] Through the above method, the method can effectively extract instances with multiple different dimensional information, thereby obtaining the distance between instances based on the emphasis of these instances, constructing the multidimensional space where the instances are located, ensuring the validity of the data set, and providing a basis for the subsequent type classification of data instances.

[0103] On the other hand, other multiple data instances may not be complete data instances, and the number of data instances in each dimension is also different, so they are recorded as L, that is, for the current dimension, the number of all data instances is C+Lm. The number of L1, L2, L3 and L4 is not equal.

[0104] In this process, the method will not abandon other incomplete data instance contents. By constructing the integration of instance information of different dimensions, it ensures that all complaint information is taken into account, thereby providing a model data basis for subsequent data classification.

[0105] For a complete data instance, its content in different dimensions is different, but it actually has the same semantic description information. Therefore, the present invention uses a deep neural network to extract the semantic description.

[0106] Step 2: After the complaint data set is input into the deep neural network, the complaint information of different dimensions is decomposed into a first decomposition matrix and a second decomposition matrix.

[0107] Preferably, inputting the complaint data set into a deep neural network includes: using the deep neural network to convert the complaint data set into a feature data set matrix In the formula, and They are and The deep feature transformation matrix of .

[0108] In the present invention, the method can convert the matrix of data instances into a matrix of deep feature representation. Then, the method extracts common information of multiple dimensions in the data set matrix and feature information of each dimension by decomposing the data set matrix.

[0109] Preferably, the complaint information of different dimensions is decomposed into a first decomposition matrix and a second decomposition matrix, including: the product of the feature data set matrix V and the first decomposition matrix and the second decomposition matrix is ​​equal; and the first decomposition matrix W is transformed into:

[0110]

[0111] The second decomposition matrix H is transformed into:

[0112]

[0113] In the formula, ij is the i-th row and j-th column of the matrix, and the iteration target is min∑ ij (V ij -(W.H) ij ) 2 .

[0114] In the present invention, it is assumed that a parameter H can be used to represent the data because is the information content matrix of the complete data instance, so for different dimensions, have the same semantic information. Therefore, The potential meaning of the parameter H can be consistent in multiple dimensions.

[0115] According to this goal, the present invention adopts an iterative method to solve the parameter H. When solving the parameter H, another associated decomposition parameter is introduced. In the present invention, it is called the first decomposition matrix. The product of the first decomposition matrix and the second decomposition matrix can obtain a consistent feature data set matrix. And through iteration, the method can optimize the second decomposition matrix.

[0116] Through the decomposed second decomposition matrix, the method extracts all the information of the most simplified and lossless complete instance without difference and defines it as the second decomposition matrix. The second decomposition matrix is ​​solved in an iterative manner to prevent local optimality and excessive errors in the solution results. These common information in multiple dimensions supports the construction of instance space and ensures spatial scalability, thereby providing redundancy for the multi-level classification process.

[0117] In one embodiment of the present invention, the iteration target is min∑ ij (V ij -(W.H) ij ) 2 That is, after iteration, the error between the value of each element in the product of the decomposition matrix and the value of each element in the original matrix is ​​the smallest.

[0118] In the following, the common semantic features can be used to describe the direction of iteration in this step.

[0119] Step 3: Calculate the common semantic features in complaint information of different dimensions using the first decomposition matrix and the second decomposition matrix.

[0120] After obtaining the first decomposition matrix and the second decomposition matrix, the method can calculate the common semantic features in complaint information of different dimensions.

[0121] Preferably, the common semantic features are:

[0122]

[0123] In the formula, H C is the shared feature matrix of complaint information under all dimensions,

[0124] They are different from the shared feature matrix H in different dimensions. C The matrix composed of other data instances of .

[0125] As mentioned above, the common semantic features here are actually the common features H of the parameter H in multiple dimensions. C , and separate features in each dimension By calculating the common semantic features, the method can obtain the basis for complaint classification. By inputting the common semantic features into the pre-constructed knowledge graph related to power business complaints, the corresponding business improvement methods for complaints should be obtained, so as to automatically dispatch orders in real time.

[0126] Through this method, semantic features under multiple different dimensions can ensure that classification will not be inaccurate due to the additional features of a certain dimension of data. Since the information of multiple dimensions constrains and cooperates with each other, it overcomes the classification errors caused by incompleteness in the data collection process, greatly improves the effectiveness of classification, and provides a semantic basis for automatic dispatching.

[0127] The method for calculating the shared semantic features is as follows. Preferably, the shared feature matrix H of complaint information in all dimensions is C and other data instances Obtained through the objective function; the objective function is:

[0128]

[0129]

[0130] Where W m is the first decomposition matrix under the current dimension m,

[0131] Tr(·) is the trace of the matrix,

[0132] α m is the dimension parameter,

[0133] L m is a Laplace matrix.

[0134] The method solves the minimization target by constructing an objective function. The solved minimization target includes the parameter H to be solved. C and The value of .

[0135] The above objective function includes two aspects: the first term decomposition error and the second term error in the spatial query positioning process of the data instance. By minimizing the sum of the errors in the two aspects, it is fully ensured that the result of each iteration converges accurately enough.

[0136] In the present invention, H C and are all matrices greater than or equal to 0. According to the matrix compilation method, the information it carries, etc., the method can set the capacity of the matrix, that is, the number of rows and columns. The objective function also includes some constraints, such as W m It is also a matrix greater than or equal to 0, and the parameters can be pre-defined according to the actual situation. During multiple iterations of the deep learning network, the value of the parameter does not change.

[0137] Laplace matrix L m =D m -S m , where D m is a diagonal matrix, and the value of each element on the diagonal is S m The sum of each row or column in a matrix. S m is a similarity matrix, which is used to characterize the similarity between two data instances. For example, two complaint information are for the same power business, the problems are exactly the same or similar, and the customers who made the complaints are close to each other. In this case, the similarity between the two instances is high. If the locations, actual problems, business types, etc. of the two complaint information are different, the distance between the two instances is far.

[0138] Therefore, through the similarity matrix, the method can not only fully define the construction method of the initial data set matrix where the complaint information is located, but also consider various factors to ensure the accuracy of instance distance calculation.

[0139] By constructing the spatial relationship between any two complaint information, we can obtain accurate basis for complaint classification and ensure the accuracy of information classification results. Moreover, this classification method takes into account different information in multiple dimensions and will not cause low classification accuracy due to incomplete information in different dimensions, which greatly improves classification accuracy.

[0140] In one embodiment, the method uses the Euclidean distance as a basis to characterize the similarity between two instances. In the similarity matrix, the distance between any two instances p and q is an element in a matrix. Therefore, the number of rows and columns of the matrix is ​​equal and equal to the total number of all data instances in the data set.

[0141] With the above method as the goal, the method obtains the common semantic features in complaint information of different dimensions.

[0142] Step 4: Based on the common semantic features, train the deep neural network and obtain the output of the deep neural network.

[0143] In the present invention, training a deep neural network also includes: constructing a deep neural network parameter θ m , based on the deep neural network parameter θ m Adjust the lth layer of the deep neural network for the The network weights of the dimensions and the bias vector Deep neural network parameters θ m Initialized at the beginning of the iteration, using the function in each iteration Solve to improve.

[0144] In the function Middle, H m Based on the last iteration, the deep neural network parameters θ are obtained by calculating the function. m of adjustment.

[0145] Specifically, in each iteration of the present invention, the W of the previous iteration is first m and parameter θ m , update H m Then, through the H of the previous iteration m and parameter θ m , update W m . The updated W m and H m Substitute into the function and solve for θ for the next iteration m , through this method, the update is achieved.

[0146] Preferably, the first layer of the deep neural network is for The network weights of the dimensions and the bias vector The changes are performed according to the product of the descent gradient and the learning rate of the deep neural network; the network weights The descent gradient is:

[0147]

[0148] Bias Vector The descent gradient is:

[0149]

[0150] The deep neural network learning rate is set for the current dimension m.

[0151] Through the above method, the network weights and bias vectors are improved, and finally the parameters θ are fine-tuned. m, to achieve iterative calculation of the first decomposition matrix and the second decomposition matrix. In the above process, the trained deep neural network can converge quickly and accurately obtain the output of the neural network.

[0152] Step 5: Use the output of the deep neural network to confirm the business classification to achieve real-time dispatch of complaints.

[0153] It is easy to think that the way to confirm the business classification can be to learn the knowledge of the complaint information content and construct a knowledge graph. In the present invention, the current complaint is input into the deep neural network, and the business type is found and located through the pre-constructed knowledge graph.

[0154] In the knowledge graph, we can not only learn the business type of the complaint, but also learn the solution and technical solution for the complaint. By finding the corresponding node on the knowledge graph, the method can implement accurate dispatch and provide technical support.

[0155] Figure 2 FIG. 1 is a schematic diagram of the process of perceiving the scheduling decision execution in an automatic real-time dispatching method based on business classification according to the present invention. Figure 2 , the method identifies various types of information in the perception center, inputs them into the dispatch center, implements the analysis process mentioned above, and then obtains decisions based on the input. The decision center deploys a knowledge graph to guide the semantic output to the specific node location of the knowledge graph, and provides services such as dispatching, event prediction, performance evaluation, and process optimization to support the handling and tracking of complaints. Finally, in the execution center, the dispatched operation and maintenance personnel can improve the power business and handle complaints.

[0156] Figure 3 FIG. 1 is a schematic diagram of the module structure of an automatic real-time dispatching system based on business classification according to the present invention. Figure 3 The second aspect of the present invention relates to an automatic real-time dispatching system based on business classification, the system comprising an acquisition module, a decomposition module, a calculation module, a training module and a classification module; wherein the acquisition module is used to collect complaint voice, complaint text, complaint picture and complaint video, respectively extract complaint information of different dimensions and construct a complaint data set based on the complaint information; the decomposition module is used to input the complaint data set into a deep neural network, and then decompose the complaint information of different dimensions into a first decomposition matrix and a second decomposition matrix; the calculation module is used to calculate the common semantic features in the complaint information of different dimensions using the first decomposition matrix and the second decomposition matrix; the training module is used to train the deep neural network based on the common semantic features and obtain the output of the deep neural network; the classification module is used to confirm the business classification using the output of the deep neural network to realize real-time dispatching for complaints.

[0157] Preferably, the acquisition module is used to extract complaint information of different dimensions and construct a complaint data set based on the complaint information, including: the complaint data set is

[0158]

[0159] m is the dimension number of complaint voice, complaint text, complaint picture and complaint video,

[0160] is the feature matrix of complaint information under the current dimension m,

[0161] is the difference between the current dimension m and The matrix composed of other data instances of .

[0162] Preferably, the decomposition module is used to input the complaint data set into a deep neural network, including: using the deep neural network to convert the complaint data set into a feature data set matrix

[0163] In the formula, and They are and The deep feature transformation matrix of .

[0164] Preferably, the decomposition module is used to decompose complaint information of different dimensions into a first decomposition matrix and a second decomposition matrix, including: the product of the feature data set matrix V and the first decomposition matrix and the second decomposition matrix is ​​equal; and,

[0165] The first decomposition matrix W is transformed into:

[0166]

[0167] The second decomposition matrix H is transformed into:

[0168]

[0169] Where ij is the i-th row and j-th column of the matrix,

[0170] The iteration goal is min∑ ij (V ij -(W.H) ij ) 2 .

[0171] Preferably, the calculation module uses the first decomposition matrix and the second decomposition matrix to calculate the common semantic features in the complaint information of different dimensions, including: the common semantic features are:

[0172]

[0173] In the formula, H C is the shared feature matrix of complaint information under all dimensions,

[0174] They are different from the shared feature matrix H in different dimensions. C The matrix composed of other data instances of .

[0175] Preferably, in the calculation module, the shared feature matrix H of complaint information in all dimensions C and other data instances Obtained through the objective function;

[0176] The objective function is:

[0177]

[0178]

[0179] Where W m is the first decomposition matrix under the current dimension m,

[0180] Tr(·) is the trace of the matrix,

[0181] α m is the dimension parameter,

[0182] L m is a Laplace matrix.

[0183] Preferably, the training module, used for training the deep neural network, further comprises: constructing a deep neural network parameter θ m , based on the deep neural network parameters θ m Adjust the lth layer of the deep neural network for the The network weights of the dimensions and the bias vector The deep neural network parameters θ m Initialized at the beginning of the iteration, using the function in each iteration Solve to improve.

[0184] Preferably, in the training module, the first layer of the deep neural network is for the first The network weights of the dimensions and the bias vector The changes are performed by the product of the descent gradient and the deep neural network learning rate;

[0185] The network weights The descent gradient is:

[0186]

[0187] The bias vector The descent gradient is:

[0188]

[0189] The deep neural network learning rate is set for the current dimension m.

[0190] A third aspect of the present invention relates to an electronic device, comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method according to the first aspect of the present invention.

[0191] A fourth aspect of the present invention relates to a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of the method in the first aspect of the present invention.

[0192] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An automatic real-time dispatching method based on business classification, characterized in that: include: Collect complaint information, extract complaint information of different dimensions and construct a complaint data set based on the extracted complaint information; After inputting the complaint data set into the deep neural network, the complaint information of different dimensions is decomposed into a first decomposition matrix and a second decomposition matrix; Calculate the common semantic features in complaint information of different dimensions by using the first decomposition matrix and the second decomposition matrix; Based on the common semantic features, training the deep neural network and obtaining an output of the deep neural network; The output of the deep neural network is used to confirm business classification to achieve real-time dispatch of complaints.

2. The automatic real-time dispatching method based on business classification according to claim 1 is characterized by: The extracting complaint information of different dimensions respectively and constructing a complaint data set based on the complaint information includes: The complaint dataset is m is the dimension number of complaint voice, complaint text, complaint picture and complaint video, is the feature matrix of complaint information under the current dimension m, is the difference between the current dimension m and The matrix composed of other data instances of .

3. The automatic real-time dispatching method based on business classification according to claim 2 is characterized by: The step of inputting the complaint data set into a deep neural network comprises: The deep neural network is used to convert the complaint dataset into a feature dataset matrix. In the formula, and They are and The deep feature transformation matrix of .

4. The automatic real-time dispatching method based on business classification according to claim 3 is characterized by: Decomposing the complaint information of different dimensions into a first decomposition matrix and a second decomposition matrix includes: The feature data set matrix V is equal to the product of the first decomposition matrix and the second decomposition matrix; and, The first decomposition matrix W is transformed into: The second decomposition matrix H is transformed into: Where ij is the i-th row and j-th column of the matrix, The iteration goal is min∑ ij (V ij -(W.H) ij ) 2 .

5. The automatic real-time dispatching method based on business classification according to claim 4 is characterized by: The using the first decomposition matrix and the second decomposition matrix to calculate the common semantic features in the complaint information of different dimensions includes: The common semantic features are: In the formula, H C is the shared feature matrix of complaint information under all dimensions, They are different from the shared feature matrix H in different dimensions. C The matrix composed of other data instances of .

6. The automatic real-time dispatching method based on business classification according to claim 5 is characterized by: The shared feature matrix H of the complaint information under all dimensions C and other data instances Obtained through the objective function; The objective function is: Where W m is the first decomposition matrix under the current dimension m, Tr(·) is the trace of the matrix, α m is the dimension parameter, L m is a Laplace matrix.

7. The automatic real-time dispatching method based on business classification according to claim 4 is characterized by: The training of the deep neural network further includes: Construct deep neural network parameters θ m , based on the deep neural network parameters θ m Adjust the lth layer of the deep neural network for the The network weights of the dimensions and the bias vector The deep neural network parameters θ m Initialized at the beginning of the iteration, using the function in each iteration Solve to improve.

8. The automatic real-time dispatching method based on business classification according to claim 7 is characterized by: The first layer of the deep neural network is for The network weights of the dimensions and the bias vector The changes are performed by the product of the descent gradient and the deep neural network learning rate; The network weights The descent gradient is: The bias vector The descent gradient is: The deep neural network learning rate is set for the current dimension m.

9. An automatic real-time dispatching system based on business classification, characterized by: The system includes an acquisition module, a decomposition module, a calculation module, a training module and a classification module; wherein, The collection module is used to collect complaint information, extract complaint information of different dimensions respectively, and construct a complaint data set based on the complaint information; The decomposition module is used to decompose the complaint information of different dimensions into a first decomposition matrix and a second decomposition matrix after inputting the complaint data set into the deep neural network; The calculation module is used to calculate the common semantic features in the complaint information of different dimensions by using the first decomposition matrix and the second decomposition matrix; The training module is used to train the deep neural network based on the common semantic features and obtain the output of the deep neural network; The classification module is used to use the output of the deep neural network to confirm the business classification so as to realize real-time dispatching of complaints.

10. The automatic real-time dispatching system based on business classification according to claim 9, characterized in that: The acquisition module is used to extract complaint information of different dimensions and construct a complaint data set based on the complaint information, including: The complaint dataset is m is the dimension number of complaint voice, complaint text, complaint picture and complaint video, is the feature matrix of complaint information under the current dimension m, is the difference between the current dimension m and The matrix composed of other data instances of .

11. The automatic real-time dispatching system based on business classification according to claim 10, characterized in that: The decomposition module is used to input the complaint data set into a deep neural network, including: The deep neural network is used to convert the complaint dataset into a feature dataset matrix. In the formula, and They are and The deep feature transformation matrix of .

12. The automatic real-time dispatching system based on business classification according to claim 11, characterized in that: The calculation module is used to decompose complaint information of different dimensions into a first decomposition matrix and a second decomposition matrix, including: The feature data set matrix V is equal to the product of the first decomposition matrix and the second decomposition matrix; and, The first decomposition matrix W is transformed into: The second decomposition matrix H is transformed into: Where ij is the i-th row and j-th column of the matrix, The iteration goal is min∑ ij (V ij -(W.H) ij ) 2 .

13. The automatic real-time dispatching system based on business classification according to claim 12, characterized in that: The calculation module is used to calculate the common semantic features in the complaint information of different dimensions by using the first decomposition matrix and the second decomposition matrix, including: The common semantic features are: In the formula, H C is the shared feature matrix of complaint information under all dimensions, They are different from the shared feature matrix H in different dimensions. C The matrix composed of other data instances of .

14. The automatic real-time dispatching system based on business classification according to claim 13, characterized in that: In the calculation module, the shared feature matrix H of complaint information under all dimensions C and other data instances Obtained through the objective function; The objective function is: Where W m is the first decomposition matrix under the current dimension m, Tr(·) is the trace of the matrix, α m is the dimension parameter, L m is a Laplace matrix.

15. The automatic real-time dispatching system based on business classification according to claim 14, characterized in that: The training module is used to train the deep neural network, including: Construct deep neural network parameters θ m , based on the deep neural network parameters θ m Adjust the lth layer of the deep neural network for the The network weights of the dimensions and the bias vector The deep neural network parameters θ m Initialized at the beginning of the iteration, using the function in each iteration Solve to improve.

16. The automatic real-time dispatching system based on business classification according to claim 15, characterized in that: In the training module, The first layer of the deep neural network targets The network weights of the dimensions and the bias vector The changes are performed by the product of the descent gradient and the deep neural network learning rate; The network weights The descent gradient is: The bias vector The descent gradient is: The deep neural network learning rate is set for the current dimension m.

17. An electronic device, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-8.

Images