A power customer service work order generation method and device

Abstract

The application discloses a power customer service work order generation method and device, relates to the field of work order identification and generation, and comprises the following steps: performing semantic feature extraction on user problem information to be processed to obtain a first feature vector; performing named entity recognition on the first feature vector in combination with a preset work order field mapping rule to obtain a first power customer service work order, and updating the first feature vector according to the first power customer service work order to obtain a second feature vector; performing identification on the second feature vector in combination with a preset fault description matching rule based on a preset fault feature identification model to obtain fault description information; performing intention prediction on the fault description information to obtain a first user intention and a corresponding first confidence; when the first confidence is lower than a preset confidence threshold, generating information supplement prompts according to the first power customer service work order, and updating the first power customer service work order according to first supplement information of the user to obtain a second power customer service work order.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of work order identification generation, and in particular to a power customer service work order generation method and device. BACKGROUND

[0002] With the improvement of the informatization degree of the power system, the processing efficiency of the traditional way of receiving customer feedback, recording and manually processing power customer service work orders by power customer service personnel has been unable to adapt to the situation of rapid increase in the number of work orders during peak periods, and manual processing may miss the power customer service work order generated by the user, which does not match the current situation. Therefore, how to automatically generate a power customer service work order according to user feedback and improve the accuracy of the generated power customer service work order is a hot issue that needs to be studied.

[0003] There are some automatic power customer service work order generation methods in the prior art. One is to automatically generate a power customer service work order by matching rules and keywords, which is better adapted to relatively simple power customer service work order scenarios, such as scenarios that only need to match rules and keywords to automatically fill in work order fields, but when the corresponding fault needs to be processed according to the user description, the recognition effect is poor, resulting in an inaccurate power customer service work order. The second is to generate a power customer service work order by matching a knowledge graph, which can match the corresponding fault by matching the user description through the knowledge graph, but when the user description is not clear or distorted, the matching error is likely to be large, resulting in an inaccurate power customer service work order. Therefore, how to improve the accuracy of automatic power customer service work order generation is still a problem to be solved in the prior art. SUMMARY

[0004] The present application provides a power customer service work order generation method and device to solve the technical problem of low accuracy of power customer service work order generation in the prior art.

[0005] According to a first aspect of the embodiments of the present application, a power customer service work order generation method is provided, comprising:

[0006] performing semantic feature extraction on the user problem information to be processed to obtain a first feature vector;

[0007] performing named entity recognition on the first feature vector in combination with a preset work order field mapping rule to obtain a first power customer service work order, and updating the first feature vector to obtain a second feature vector according to the first power customer service work order;

[0008] performing recognition on the second feature vector in combination with a preset fault description matching rule based on a preset fault feature recognition model to obtain fault description information;

[0009] performing intention prediction on the fault description information to obtain a first user intention and a corresponding first confidence;

[0010] when the first confidence is lower than a preset confidence threshold, generating an information supplement prompt according to the first power customer service work order, and updating the first power customer service work order to obtain a second power customer service work order according to first supplement information of the user; wherein the first supplement information is obtained by receiving feedback of the user on the information supplement prompt.

[0011] The application first extracts semantic features of the user problem information to be processed to obtain a first feature vector, then performs named entity recognition on the first feature vector in combination with a preset work order field mapping rule to obtain a first power customer service work order, and updates the first power customer service work order to obtain a second feature vector, which can further structure the characteristic problem information through named entity recognition, and can associate the structured problem information with the corresponding work order field in combination with the work order field mapping rule, thereby improving the matching degree of the first power customer service work order and the actual problem information of the user; and then the fault description information is recognized through a fault feature recognition model and a fault description matching rule, which can more accurately recognize and obtain the fault description information that is more consistent with the actual fault of the user through the cross recognition of the fault feature recognition model and the fault description matching rule, so that a more accurate first confidence is obtained when the intention prediction is performed on the fault description information, and when the first confidence is lower than a threshold and the user intention is not clear, an information supplement prompt is generated according to the first power customer service work order for the user to supplement to obtain a second power customer service work order, which can prevent the problem of incomplete work order caused by missing processing steps when the user intention is not clear, thereby improving the accuracy of work order generation.

[0012] In some embodiments of the application, the semantic feature extraction of the user problem information to be processed to obtain a first feature vector specifically includes:

[0013] performing word segmentation on the user problem information to be processed to obtain first segmented information;

[0014] performing feature representation on the first segmented information based on a pre-trained bidirectional encoder to obtain a first feature vector.

[0015] The application first performs word segmentation on the user problem information to be processed, and then performs feature representation on the first segmented information obtained through word segmentation based on a pre-trained bidirectional encoder to obtain a first feature vector, which can arrange the user problem information that is messy and has redundant expressions into simple segmented information that is easy to understand through word segmentation, and then obtain a feature vector that is easy to understand by the model after feature representation.

[0016] In some embodiments of the present application, the first feature vector is subjected to named entity recognition based on a preset work order field mapping rule to obtain a first power customer service work order, and the first feature vector is updated to obtain a second feature vector according to the first power customer service work order, specifically comprising:

[0017] The first feature vector is subjected to named entity recognition based on a bidirectional neural network to obtain a plurality of fault entities and corresponding fault entity types;

[0018] According to a preset work order field mapping rule, the plurality of fault entities are matched in combination with the fault entity types of the plurality of fault entities to obtain work order fields corresponding to the plurality of fault entities;

[0019] According to a preset work order template, the plurality of fault entities are embedded into corresponding work order fields to obtain a first power customer service work order;

[0020] The feature components in the first feature vector containing the plurality of fault entities are removed to obtain a second feature vector.

[0021] The present application first obtains a plurality of fault entities and corresponding fault entity types based on a bidirectional neural network for named entity recognition, which can structure the information of features into a plurality of entities, and then matches corresponding work order fields in combination with a work order field mapping rule, which can associate entities with work order fields, thereby improving the matching degree of the first power customer service work order and the actual problem information of the user when embedding entities into work order fields. The first feature vector is updated to obtain a second feature vector by removing feature components containing a plurality of entities in the first feature vector, which can prepare for analyzing additional other information in the user problem information in the next step.

[0022] In some embodiments of the present application, the second feature vector is subjected to recognition based on a preset fault feature recognition model in combination with a preset fault description matching rule to obtain fault description information, specifically comprising:

[0023] According to a preset fault description matching rule, each feature component in the second feature vector is matched to obtain a plurality of first fault descriptions;

[0024] Each feature component in the second feature vector is subjected to recognition based on a preset fault feature recognition model to obtain a plurality of second fault descriptions and corresponding weights;

[0025] According to the plurality of first fault descriptions, the plurality of second fault descriptions and the corresponding weights, fault description information is obtained.

[0026] The application first matches each feature component of the second feature vector according to the fault description to obtain a plurality of first fault descriptions, and then identifies based on a fault feature identification model to obtain a plurality of second fault descriptions and corresponding weights, and finally obtains the fault description information, which can obtain the fault description information more consistent with the actual fault of the user through the dual identification matching of rules and models, and improve the accuracy of the fault description information.

[0027] In some embodiments of the application, the fault description information is obtained according to the plurality of first fault descriptions, the plurality of second fault descriptions and the corresponding weights, specifically comprising:

[0028] a plurality of third fault descriptions are obtained according to the intersection of the plurality of first fault descriptions and the plurality of second fault descriptions;

[0029] The plurality of third fault descriptions are filtered according to the corresponding weights of the plurality of third fault descriptions and the preset weight threshold to obtain a plurality of fourth fault descriptions;

[0030] The fault description information is obtained by fusing the plurality of fourth fault descriptions.

[0031] The application first obtains a plurality of third fault descriptions according to the intersection of a plurality of first fault descriptions and a plurality of second fault descriptions, and then filters a plurality of fourth fault descriptions according to the corresponding weights and a preset weight threshold, and finally fuses to obtain the fault description information, which can be based on the dual identification matching of rules and models for fault description based on the intersection, so that the final fault description information is more consistent with the actual fault of the user.

[0032] In some embodiments of the application, the information supplement prompt is generated according to the first power customer service work order, specifically comprising:

[0033] The information missing type is obtained according to the first power customer service work order;

[0034] The information missing type is semantically expanded by combining a large language model to obtain an information supplement prompt.

[0035] The application obtains the information missing type according to the first power customer service work order, and semantically expands the information supplement prompt by combining a large language model, which can intuitively obtain the missing information that needs to be supplemented from the current first power customer service work order, and can convert the missing information that needs to be supplemented into a prompt language that is easy for users to understand, thereby improving user experience.

[0036] In some embodiments of the application, the second power customer service work order is obtained by updating the first power customer service work order according to the first supplement information of the user, specifically comprising:

[0037] tokenize the first supplementary information of the user to obtain second tokenized information;

[0038] featureize the second tokenized information based on a pre-trained bidirectional encoder to obtain a second feature vector;

[0039] update the first power customer service work order based on the second feature vector to obtain a second power customer service work order.

[0040] The application tokenizes the first supplementary information of the user, features the second tokenized information based on a pre-trained bidirectional encoder to obtain a second feature vector, and updates the first power customer service work order to obtain a second power customer service work order. The user input with disordered and redundant expressions can be sorted into a simple feature vector, the structural degree of the user input information is improved, and the matching degree between the second power customer service work order and the actual user input is improved.

[0041] In some embodiments of the application, the method further comprises:

[0042] obtaining a priority of the second power customer service work order based on a decision tree model;

[0043] updating the second power customer service work order to generate a third work order according to the priority, and processing the third work order according to the corresponding processing personnel.

[0044] The application obtains a priority of the second power customer service work order based on a decision tree model, and then updates the second power customer service work order to generate a third work order for distribution processing. The priority of the work order can be reasonably arranged, and the user problems corresponding to the work order can be reasonably processed.

[0045] In some embodiments of the application, the method of obtaining a priority of the second power customer service work order based on a decision tree model specifically comprises:

[0046] identifying the priority of each work order field of the second power customer service work order based on a decision tree model to obtain a plurality of priorities;

[0047] calculating the priority of the second power customer service work order according to a preset field priority weight and in combination with the plurality of priorities.

[0048] The application identifies the priority of each work order field of the second power customer service work order based on a decision tree model to obtain a plurality of priorities, and then calculates the priority of the second power customer service work order based on the priority weight of each field. The second power customer service work order can be reasonably prioritized, and the user problems corresponding to the work order can be reasonably processed.

[0049] According to a second aspect of the embodiments of the present application, an electric power customer service order generation device is provided, comprising a semantic feature extraction module, a first electric power customer service order acquisition module, a user description acquisition module, a user intention prediction module, and a second electric power customer service order generation module;

[0050] The semantic feature extraction module is configured to perform semantic feature extraction on the user question information to be processed to obtain a first feature vector;

[0051] The first electric power customer service order acquisition module is configured to perform named entity recognition on the first feature vector in combination with a preset order field mapping rule to obtain a first electric power customer service order, and update the first feature vector to obtain a second feature vector according to the first electric power customer service order;

[0052] The user description acquisition module is configured to perform recognition on the second feature vector in combination with a preset fault description matching rule based on a preset fault feature recognition model to obtain fault description information;

[0053] The user intention prediction module is configured to perform intention prediction on the fault description information to obtain a first user intention and a corresponding first confidence;

[0054] The second electric power customer service order generation module is configured to generate information supplement prompts according to the first electric power customer service order when the first confidence is lower than a preset confidence threshold, and update the first electric power customer service order to obtain a second electric power customer service order according to first supplementary information of the user; wherein the first supplementary information is obtained by receiving feedback of the user on the information supplement prompts.

[0055] In some embodiments of the present application, the semantic feature extraction module comprises a first word segmentation sub-module and a first feature extraction sub-module;

[0056] The first word segmentation sub-module is configured to perform word segmentation on the user question information to be processed to obtain first word segmentation information;

[0057] The first feature extraction sub-module is configured to perform feature extraction representation on the first word segmentation information based on a pre-trained bidirectional encoder to obtain a first feature vector.

[0058] In some embodiments of the present application, the first electric power customer service order acquisition module comprises an entity recognition sub-module, a field matching sub-module, an order embedding sub-module, and a vector updating sub-module;

[0059] The entity recognition sub-module is configured to perform named entity recognition on the first feature vector based on a bidirectional neural network to obtain a plurality of fault entities and corresponding fault entity types;

[0060] The field matching submodule is configured to match the plurality of fault entities according to a preset work order field mapping rule and in combination with fault entity types of the plurality of fault entities, to obtain work order fields corresponding to the plurality of fault entities.

[0061] The work order embedding submodule is configured to embed the plurality of fault entities into corresponding work order fields according to a preset work order template, to obtain a first power customer service work order.

[0062] The vector updating submodule is configured to eliminate feature components in the first feature vector that contain the plurality of fault entities, to obtain a second feature vector.

[0063] In some embodiments of the present application, the user description obtaining module includes a rule matching submodule, a feature recognition submodule, and a description obtaining submodule.

[0064] The rule matching submodule is configured to match each feature component in the second feature vector according to a preset fault description matching rule, to obtain a plurality of first fault descriptions.

[0065] The feature recognition submodule is configured to recognize each feature component in the second feature vector based on a preset fault feature recognition model, to obtain a plurality of second fault descriptions and corresponding weights.

[0066] The description obtaining submodule is configured to obtain fault description information according to the plurality of first fault descriptions, the plurality of second fault descriptions, and the corresponding weights.

[0067] In some embodiments of the present application, the description obtaining submodule includes a description intersection unit, a description filtering unit, and a description fusion unit.

[0068] The description intersection unit is configured to obtain a plurality of third fault descriptions according to an intersection of the plurality of first fault descriptions and the plurality of second fault descriptions.

[0069] The description filtering unit is configured to filter the plurality of third fault descriptions according to corresponding weights of the plurality of third fault descriptions and a preset weight threshold, to obtain a plurality of fourth fault descriptions.

[0070] The description fusion unit is configured to fuse the plurality of fourth fault descriptions, to obtain fault description information.

[0071] In some embodiments of the present application, the second power customer service work order generation module includes a missing obtaining submodule and a semantic expansion submodule.

[0072] The missing obtaining submodule is configured to obtain an information missing type according to the first power customer service work order.

[0073] The semantic expansion submodule is configured to perform semantic expansion on the information missing type in combination with the large language model to obtain information supplement prompts.

[0074] In some embodiments of the present application, the second power customer service order generation module comprises a second word segmentation submodule, a second feature extraction submodule, and an order updating submodule.

[0075] The second word segmentation submodule is configured to perform word segmentation on the first supplementary information of the user to obtain second word segmentation information.

[0076] The second feature extraction submodule is configured to perform feature extraction on the second word segmentation information based on a pre-trained bidirectional encoder to obtain a second feature vector.

[0077] The order updating submodule is configured to update the first power customer service order based on the second feature vector to obtain a second power customer service order.

[0078] In some embodiments of the present application, an order allocation module is further included, and the order allocation module comprises a priority obtaining submodule and an order allocation submodule.

[0079] The priority obtaining submodule is configured to obtain a priority of the second power customer service order based on a decision tree model.

[0080] The order allocation submodule is configured to update the second power customer service order to generate a third order according to the priority, and to allocate a processing personnel to process the third order.

[0081] In some embodiments of the present application, the priority obtaining submodule comprises a priority identification unit and a priority calculation unit.

[0082] The priority identification unit is configured to perform priority identification on each order field of the second power customer service order based on a decision tree model to obtain a plurality of priorities.

[0083] The priority calculation unit is configured to calculate the priority of the second power customer service order based on a preset field priority weight in combination with the plurality of priorities.

[0084] The application first extracts semantic features of the user question information to be processed to obtain a first feature vector, then performs named entity recognition on the first feature vector to obtain a first power customer service ticket according to a preset ticket field mapping rule, and updates the first feature vector to obtain a second feature vector. Through named entity recognition, the characteristic question information can be further structured, and by combining the ticket field mapping rule, the structured question information can be associated with the corresponding ticket field, thereby improving the matching degree of the first power customer service ticket and the actual question information of the user. Then, the fault description information is recognized through the fault feature recognition model and the fault description matching rule. Through the cross recognition of the fault feature recognition model and the fault description matching rule, more accurate fault description information that is more consistent with the actual fault of the user can be recognized and obtained, so that when the intent of the fault description information is predicted, a more accurate first confidence is obtained. When the first confidence is lower than a threshold value and it is determined that the user intent is unclear, an information supplement prompt is generated according to the first power customer service ticket to supplement the information of the user, and a second power customer service ticket is obtained. By determining the information supplement prompt according to the user intent and its confidence, the problem of incomplete ticket caused by missing processing steps when the user intent is unclear can be prevented, thereby improving the accuracy of ticket generation. BRIEF DESCRIPTION OF DRAWINGS

[0085] Figure 1 : a flowchart of a power customer service ticket generation method shown in some embodiments of the application;

[0086] Figure 2 : a module structure diagram of a power customer service ticket generation device shown in some embodiments of the application. DETAILED DESCRIPTION

[0087] The embodiments of the application are described in detail below, and examples of the embodiments are shown in the drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the drawings are exemplary and are only used to explain some embodiments of the application, and cannot be understood as limiting the embodiments of the application. Based on the embodiments shown in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0088] In the description of the application, it should be understood that the terms "first" and "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" can explicitly or implicitly include at least one of the features. In the description of the application, unless otherwise specifically limited, the meaning of "multiple" "several" is two or more.

[0089] In the prior art, the method of automatically generating power customer service work orders through rule and keyword matching is relatively effective for simple scenarios that only need to be automatically filled according to rule and keyword matching, but the recognition effect is poor for scenarios that need to be processed according to user description to obtain the corresponding fault, and the generated power customer service work order is inaccurate. Another method of generating power customer service work orders through knowledge graph matching can obtain the corresponding fault by matching the user description through the knowledge graph, but when the user description is not clear or the user description is distorted, matching errors are likely to occur, and the generated power customer service work order is also inaccurate. Therefore, how to improve the accuracy of automatic power customer service work order generation is still a difficult problem to be solved in the prior art.

[0090] Based on the above technical background, see Figure 1 The embodiments of the present application provide a power customer service work order generation method, which comprises steps S101 to S105, and each step is specifically as follows:

[0091] Step S101: performing semantic feature extraction on the user problem information to be processed to obtain a first feature vector.

[0092] In some embodiments of the present application, the semantic feature extraction on the user problem information to be processed to obtain the first feature vector specifically comprises:

[0093] performing word segmentation on the user problem information to be processed to obtain first segmented information;

[0094] performing feature representation on the first segmented information based on a pre-trained bidirectional encoder to obtain the first feature vector.

[0095] In some embodiments of the present application, the user problem information to be processed is obtained by processing the dialogue content with the user, and the form of the dialogue content includes but is not limited to voice, video or text content.

[0096] In some embodiments of the present application, the pre-trained bidirectional encoder is a BERT model and its improved model, and the preferred embodiment is a BERT model.

[0097] The present application first performs word segmentation on the user problem information to be processed, and then performs feature representation on the first segmented information obtained by word segmentation based on a pre-trained bidirectional encoder to obtain a first feature vector. The user problem information which is messy and has redundant expressions can be sorted into simple segmented information which is easy to understand based on word segmentation, and then the feature vector which is easy to understand by the model is obtained after feature representation.

[0098] Step S102: perform named entity recognition on the first feature vector according to a preset work order field mapping rule to obtain a first power customer service work order, and update the first feature vector according to the first power customer service work order to obtain a second feature vector.

[0099] In some embodiments of the present application, the named entity recognition on the first feature vector according to the preset work order field mapping rule to obtain the first power customer service work order, and the updating of the first feature vector according to the first power customer service work order to obtain the second feature vector specifically includes:

[0100] performing named entity recognition on the first feature vector based on a bidirectional neural network to obtain a plurality of fault entities and corresponding fault entity types;

[0101] matching the plurality of fault entities according to the fault entity types of the plurality of fault entities based on a preset work order field mapping rule to obtain work order fields corresponding to the plurality of fault entities;

[0102] embedding the plurality of fault entities into the corresponding work order fields according to a preset work order template to obtain the first power customer service work order;

[0103] eliminating feature components containing the plurality of fault entities in the first feature vector to obtain the second feature vector.

[0104] In some embodiments of the present application, the implementation of the bidirectional neural network includes but is not limited to a bidirectional RNNN model and a bidirectional LSTM model, and the preferred implementation is a bidirectional LSTM model.

[0105] In some embodiments of the present application, the preset work order field mapping rule is obtained based on a pre-constructed power customer service domain knowledge graph, and the pre-constructed power customer service domain knowledge graph is generated based on historical power work order data and historical power customer service reply data.

[0106] In some embodiments of the present application, the fault entity types include but are not limited to fault types, fault times, fault devices, fault durations and fault locations; and the work order fields include but are not limited to fault type fields, fault time fields, fault device fields, fault duration fields and fault location fields.

[0107] The application first performs named entity recognition based on a bidirectional neural network to obtain a plurality of fault entities and corresponding fault entity types, can structure the characteristic information into a plurality of entities, and then matches the corresponding work order field by combining the work order field mapping rule, can associate the entity with the work order field, and then embed the entity into the work order field to improve the matching degree of the first electric power customer service work order and the actual problem information of the user, and then update the first feature vector by removing the feature components containing a plurality of entities in the first feature vector to obtain a second feature vector, which can prepare for the next step of analyzing additional other information in the user problem information.

[0108] Step S103: Based on the preset fault feature recognition model, the second feature vector is recognized in combination with the preset fault description matching rule to obtain fault description information.

[0109] In some embodiments of the application, the second feature vector is recognized based on the preset fault feature recognition model in combination with the preset fault description matching rule to obtain fault description information, specifically including:

[0110] According to the preset fault description matching rule, each feature component in the second feature vector is matched to obtain a plurality of first fault descriptions;

[0111] Based on the preset fault feature recognition model, each feature component in the second feature vector is recognized to obtain a plurality of second fault descriptions and corresponding weights;

[0112] According to the plurality of first fault descriptions, the plurality of second fault descriptions and the corresponding weights, the fault description information is obtained.

[0113] Exemplarily, the preset fault description matching rule, such as "fault description: broken | not good | black screen | power failure | restart", can match "my home has been power failure since last night, and the power switch is on but still no power" and "the TV at home has been broken since last night and has not been good, and has been black screen".

[0114] In some embodiments of the application, the implementation of the fault feature recognition model includes but is not limited to a support vector machine model (SVM) and its improved model, and the preferred implementation is an SVM model.

[0115] The application first matches each feature component of the second feature vector to obtain a plurality of first fault descriptions according to the fault description matching rule, and then recognizes a plurality of second fault descriptions and corresponding weights based on the fault feature recognition model, and then obtains the fault description information, which can obtain the fault description information more consistent with the actual fault of the user through the double recognition and matching of rules and models, and improve the accuracy of the fault description information.

[0116] In some embodiments of the present application, the obtaining of the fault description information according to the plurality of first fault descriptions, the plurality of second fault descriptions and the corresponding weights specifically comprises:

[0117] obtaining a plurality of third fault descriptions according to the intersection of the plurality of first fault descriptions and the plurality of second fault descriptions;

[0118] filtering the plurality of third fault descriptions according to the corresponding weights of the plurality of third fault descriptions and a preset weight threshold to obtain a plurality of fourth fault descriptions;

[0119] fusing the plurality of fourth fault descriptions to obtain the fault description information.

[0120] The present application first obtains a plurality of third fault descriptions according to the intersection of a plurality of first fault descriptions and a plurality of second fault descriptions, then filters a plurality of fourth fault descriptions according to the corresponding weights and a preset weight threshold, and finally fuses to obtain fault description information, which can be based on the intersection fusion rule and model to double-identify and match the fault description, so that the final fault description information is more consistent with the actual fault of the user.

[0121] Step S104: performing intent prediction on the fault description information to obtain a first user intent and a corresponding first confidence.

[0122] In some embodiments of the present application, the intent prediction is specifically an intent recognition and classification task, and the preferred embodiment is a BERT model.

[0123] Specifically, if the first confidence of the first user intent is lower than a preset confidence threshold, it indicates that the user's description of the fault is incomplete or the intent of the fault is not clear, and the user needs to supplement the fault information to make the description of the fault complete or the intent of the fault clear.

[0124] Step S105: when the first confidence is lower than the preset confidence threshold, generating an information supplement prompt according to the first power customer service work order, and updating the first power customer service work order to obtain a second power customer service work order according to the first supplement information of the user; wherein the first supplement information is obtained by receiving feedback of the user on the information supplement prompt.

[0125] In some embodiments of the present application, the generating of the information supplement prompt according to the first power customer service work order specifically comprises:

[0126] obtaining an information missing type according to the first power customer service work order;

[0127] performing semantic extension on the information missing type in combination with a large language model to obtain an information supplement prompt.

[0128] In some embodiments of the present application, the preferred implementation of the information missing type obtained according to the first power customer service work order is to determine the information missing type according to the missing work order field corresponding to the fault entity type in the first power customer service work order.

[0129] In some embodiments of the present application, the large language model includes but is not limited to WenBen YiYan, GPT, Xunfei Xinghuo, or TongYi QianWen.

[0130] Specifically, the information missing type is semantically expanded by combining the large language model to obtain an information supplement prompt, specifically:

[0131] The information missing type and the preset supplement prompt word are input into the large language model for semantic expansion to obtain an information supplement prompt.

[0132] For example, if the fault description information is "network problem", the missing work order fields in the first power customer service work order are the fault equipment field and the fault duration field, and the predicted first user intent and its first confidence are difficult to determine the user's intent for the fault due to the simplicity of the fault description information. At this time, the information missing type is determined to be "fault equipment" and "fault duration" according to the missing work order fields in the first power customer service work order, and the information missing type is input into the large language model for semantic expansion to obtain an information supplement prompt, such as "Please describe which network equipment has a problem? How long has the problem been? " for the user to supplement.

[0133] The present application obtains the information missing type according to the first power customer service work order and combines the large language model for semantic expansion to obtain the information supplement prompt, which can intuitively obtain the missing information that needs to be supplemented from the current first power customer service work order, and can convert the missing information that needs to be supplemented into a prompt language that is easy for users to understand, thereby improving the user experience.

[0134] In some embodiments of the present application, the second power customer service work order is obtained by updating the first power customer service work order according to the first supplement information of the user, specifically including:

[0135] The first supplement information of the user is segmented to obtain second segmented information;

[0136] The second segmented information is represented by features based on a pre-trained bidirectional encoder to obtain a second feature vector;

[0137] The first power customer service work order is updated according to the second feature vector to obtain a second power customer service work order.

[0138] In some embodiments of the present application, the implementation of the pre-trained bidirectional encoder is the BERT model and its improved models, and the preferred implementation is the BERT model.

[0139] The first supplementary information of the user is segmented first, and then represented as a second feature vector based on the pre-trained bidirectional encoder, and the first power customer service order is updated to obtain a second power customer service order. The user input with messy and redundant expressions can be sorted into a simple feature vector, the structural degree of the user input information is improved, and the matching degree of the second power customer service order and the actual input of the user is improved.

[0140] In some embodiments of the present application, it further comprises:

[0141] Based on the decision tree model, the priority of the second power customer service order is obtained;

[0142] According to the priority, the second power customer service order is updated to generate a third order, and a processing personnel is allocated and processed according to the third order.

[0143] In some embodiments of the present application, the implementation of the decision tree model includes but is not limited to the ID3-based decision tree model, the C4.5-based decision tree model and the CART-based decision tree model, and the preferred implementation is the CART-based decision tree model.

[0144] The priority of the second power customer service order is obtained based on the decision tree model first, and then the second power customer service order is updated to generate a third order for allocation and processing, which can reasonably arrange the priority of the order, and then reasonably process the user problems corresponding to the order.

[0145] In some embodiments of the present application, the priority of the second power customer service order is obtained based on the decision tree model, and specifically comprises:

[0146] Based on the decision tree model, the priority of each order field of the second power customer service order is identified to obtain a plurality of priorities.

[0147] According to the preset field priority weight, the priority of the second power customer service order is calculated in combination with the plurality of priorities.

[0148] In some embodiments of the present application, the priority of the second power customer service order is calculated in combination with the plurality of priorities according to the preset field priority weight, and specifically comprises:

[0149] The sum of the products of the plurality of priorities and the corresponding preset field priority weight is taken as the priority of the second power customer service order.

[0150] The application first identifies the priority of each work order field of the second power customer service work order based on a decision tree model to obtain multiple priorities, and then calculates the priority of the second power customer service work order based on the priority weight of each field, so that the second power customer service work order is reasonably assigned a priority, thereby reasonably processing the user problem corresponding to the work order.

[0151] Compared with the prior art, the application first extracts semantic features from the to-be-processed user problem information to obtain a first feature vector, then performs named entity recognition on the first feature vector in combination with a preset work order field mapping rule to obtain a first power customer service work order, and updates the first feature vector based on the first power customer service work order to obtain a second feature vector. Through named entity recognition, the feature-based problem information can be further structured, and in combination with the work order field mapping rule, the structured problem information can be associated with the corresponding work order field, thereby improving the matching degree of the first power customer service work order and the actual problem information of the user. Then, the fault description information is recognized through the fault feature recognition model and the fault description matching rule. Through the cross recognition of the fault feature recognition model and the fault description matching rule, more accurate fault description information that is more consistent with the actual fault of the user can be recognized and obtained, so that a more accurate first confidence is obtained when the intent of the fault description information is predicted, and when the first confidence is lower than a threshold value and the user intent is determined to be unclear, an information supplement prompt is generated based on the first power customer service work order for the user to supplement to obtain a second power customer service work order. By determining the information supplement prompt based on the user intent and the confidence, the problem of incomplete work order caused by missing processing steps when the user intent is unclear can be prevented, thereby improving the accuracy of work order generation.

[0152] Corresponding to the foregoing method, please refer to Figure 2 The embodiment of the application provides a power customer service work order generation device, which comprises a semantic feature extraction module 210, a first power customer service work order acquisition module 220, a user description acquisition module 230, a user intent prediction module 240 and a second power customer service work order generation module 250.

[0153] The semantic feature extraction module 210 is configured to extract semantic features from to-be-processed user problem information to obtain a first feature vector.

[0154] The first power customer service work order acquisition module 220 is configured to perform named entity recognition on the first feature vector in combination with a preset work order field mapping rule to obtain a first power customer service work order, and update the first feature vector based on the first power customer service work order to obtain a second feature vector.

[0155] The user description acquisition module 230 is configured to identify the second feature vector based on a preset fault feature recognition model in combination with a preset fault description matching rule to obtain fault description information.

[0156] The user intention prediction module 240 is configured to perform intention prediction on the fault description information to obtain a first user intention and a corresponding first confidence level.

[0157] The second power customer service order generation module 250 is configured to, when the first confidence level is lower than a preset confidence threshold, perform information supplement prompting based on the first power customer service order generation information, and update the first power customer service order based on first supplement information of the user to obtain a second power customer service order, wherein the first supplement information is obtained by receiving feedback of the user on the information supplement prompting.

[0158] In some embodiments of the present application, the semantic feature extraction module 210 includes a first word segmentation sub-module and a first feature extraction sub-module.

[0159] The first word segmentation sub-module is configured to perform word segmentation on the user question information to be processed to obtain first word segmentation information.

[0160] The first feature extraction sub-module is configured to perform feature extraction representation on the first word segmentation information based on a pre-trained bidirectional encoder to obtain a first feature vector.

[0161] In some embodiments of the present application, the first power customer service order acquisition module 220 includes an entity recognition sub-module, a field matching sub-module, an order embedding sub-module, and a vector updating sub-module.

[0162] The entity recognition sub-module is configured to perform named entity recognition on the first feature vector based on a bidirectional neural network to obtain a plurality of fault entities and corresponding fault entity types.

[0163] The field matching sub-module is configured to match the plurality of fault entities according to a preset order field mapping rule in combination with the fault entity types of the plurality of fault entities to obtain order fields corresponding to the plurality of fault entities.

[0164] The order embedding sub-module is configured to embed the plurality of fault entities into corresponding order fields according to a preset order template to obtain a first power customer service order.

[0165] The vector updating sub-module is configured to eliminate feature components containing the plurality of fault entities in the first feature vector to obtain a second feature vector.

[0166] In some embodiments of the present application, the user description acquisition module 230 includes a rule matching sub-module, a feature recognition sub-module, and a description acquisition sub-module.

[0167] The rule matching submodule is configured to match each feature component in the second feature vector according to a preset fault description matching rule to obtain a plurality of first fault descriptions.

[0168] The feature identification submodule is configured to identify each feature component in the second feature vector based on a preset fault feature identification model to obtain a plurality of second fault descriptions and corresponding weights.

[0169] The description obtaining submodule is configured to obtain fault description information according to the plurality of first fault descriptions, the plurality of second fault descriptions, and the corresponding weights.

[0170] In some embodiments of the present application, the description obtaining submodule includes a description intersection unit, a description filtering unit, and a description fusion unit.

[0171] The description intersection unit is configured to obtain a plurality of third fault descriptions according to an intersection of the plurality of first fault descriptions and the plurality of second fault descriptions.

[0172] The description filtering unit is configured to filter the plurality of third fault descriptions according to corresponding weights of the plurality of third fault descriptions and a preset weight threshold to obtain a plurality of fourth fault descriptions.

[0173] The description fusion unit is configured to fuse the plurality of fourth fault descriptions to obtain fault description information.

[0174] In some embodiments of the present application, the second power customer service order generation module 250 includes a missing obtaining submodule and a semantic expansion submodule.

[0175] The missing obtaining submodule is configured to obtain an information missing type according to the first power customer service order.

[0176] The semantic expansion submodule is configured to perform semantic expansion on the information missing type in combination with a large language model to obtain an information supplement prompt.

[0177] In some embodiments of the present application, the second power customer service order generation module includes a second tokenization submodule, a second featureization submodule, and an order updating submodule.

[0178] The second tokenization submodule is configured to tokenize the first supplement information of the user to obtain second tokenization information.

[0179] The second featureization submodule is configured to perform featureization representation on the second tokenization information based on a pre-trained bidirectional encoder to obtain a second feature vector.

[0180] The work order updating submodule is configured for updating the first power customer service work order according to the second feature vector to obtain a second power customer service work order.

[0181] In some embodiments of the present application, a work order distribution module is further included, which comprises a priority obtaining submodule and a work order distribution submodule.

[0182] The priority obtaining submodule is configured for obtaining the priority of the second power customer service work order based on a decision tree model.

[0183] The work order distribution submodule is configured for generating a third work order according to the priority of the second power customer service work order, and distributing a processing personnel to process the third work order.

[0184] In some embodiments of the present application, the priority obtaining submodule comprises a priority identifying unit and a priority calculating unit.

[0185] The priority identifying unit is configured for identifying the priority of each work order field of the second power customer service work order based on a decision tree model to obtain a plurality of priorities.

[0186] The priority calculating unit is configured for calculating the priority of the second power customer service work order according to a preset field priority weight in combination with the plurality of priorities.

[0187] The present application first extracts the semantic features of the user problem information to be processed to obtain a first feature vector, then performs named entity recognition on the first feature vector in combination with a preset work order field mapping rule to obtain a first power customer service work order, and updates the first power customer service work order to obtain a second feature vector. The named entity recognition can further structure the feature information, and the combination of the work order field mapping rule can associate the structured problem information with the corresponding work order field, thereby improving the matching degree of the first power customer service work order and the actual problem information of the user. Then, the fault description information is recognized through a fault feature recognition model and a fault description matching rule. The cross recognition of the fault feature recognition model and the fault description matching rule can more accurately recognize and obtain the fault description information that is more consistent with the actual fault of the user. Therefore, when predicting the intention of the fault description information, a more accurate first confidence is obtained. When the first confidence is lower than a threshold value and the user intention is determined to be unclear, an information supplement prompt is generated according to the first power customer service work order for the user to supplement to obtain a second power customer service work order. The information supplement prompt is determined by judging the user intention and the confidence, which can prevent the problem of incomplete work order caused by missing processing steps when the user intention is unclear, thereby improving the accuracy of work order generation.

[0188] It should be understood that the device provided by the embodiments of the present application corresponds to the foregoing method, and the device for generating a power customer service work order provided by the embodiments of the present application can implement the method for generating a power customer service work order provided by any of the embodiments of the present application.

[0189] Adaptively, the embodiments of the present application further provide a computer device and a computer readable storage medium.

[0190] The computer device comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor;

[0191] When the processor executes the computer program, the method for generating a power customer service work order is implemented.

[0192] The computer readable storage medium stores a plurality of instructions adapted to be loaded by the processor to execute the method for generating a power customer service work order.

[0193] The above describes some embodiments of the present application, which further details the purposes, technical solutions and beneficial effects of the present application. It should be understood that the above describes some embodiments of the present application and cannot be understood as a limitation of the present application. It is particularly pointed out that any changes, modifications, equivalent replacements and variations made by those skilled in the art within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for generating power customer service work orders, characterized in that, include: Semantic feature extraction is performed on the user question information to be processed to obtain the first feature vector; Combined with the preset work order field mapping rules, the first feature vector is subjected to named entity recognition to obtain the first power customer service work order, and the first feature vector is updated according to the first power customer service work order to obtain the second feature vector. Based on the preset fault feature recognition model and combined with the preset fault description matching rules, the second feature vector is identified to obtain fault description information; The fault description information is used to predict intent, thereby obtaining the first user intent and the corresponding first confidence level; When the first confidence level is lower than a preset confidence threshold, a supplementary information prompt is generated based on the first power customer service work order, and the first power customer service work order is updated based on the user's first supplementary information to obtain a second power customer service work order; wherein, the first supplementary information is obtained by receiving feedback from the user on the supplementary information prompt.

2. The method for generating power customer service work orders according to claim 1, characterized in that, The semantic feature extraction of the user question information to be processed to obtain the first feature vector specifically includes: The user question information to be processed is segmented into words to obtain the first segmentation information; Based on a pre-trained bidirectional encoder, the first word segmentation information is represented by features to obtain a first feature vector.

3. The method for generating power customer service work orders according to claim 1, characterized in that, The step of combining the preset work order field mapping rules to perform named entity recognition on the first feature vector to obtain the first power customer service work order, and updating the first feature vector according to the first power customer service work order to obtain the second feature vector, specifically includes: Based on a bidirectional neural network, named entity recognition is performed on the first feature vector to obtain multiple fault entities and their corresponding fault entity types. According to the preset work order field mapping rules, and combined with the fault entity type of the multiple fault entities, the multiple fault entities are matched to obtain the work order fields corresponding to the multiple fault entities. According to the preset work order template, the multiple fault entities are embedded into the corresponding work order fields to obtain the first power customer service work order; The second feature vector is obtained by removing the feature components of the multiple faulty entities contained in the first feature vector.

4. The method for generating power customer service work orders according to claim 1, characterized in that, The method based on a preset fault feature recognition model, combined with preset fault description matching rules, identifies the second feature vector to obtain fault description information, specifically including: According to the preset fault description matching rules, each feature component in the second feature vector is matched to obtain multiple first fault descriptions; Based on the preset fault feature recognition model, each feature component in the second feature vector is identified to obtain multiple second fault descriptions and their corresponding weights. Based on the plurality of first fault descriptions, the plurality of second fault descriptions, and their corresponding weights, fault description information is obtained.

5. The method for generating power customer service work orders according to claim 4, characterized in that, The step of obtaining fault description information based on the plurality of first fault descriptions, the plurality of second fault descriptions, and their corresponding weights specifically includes: Based on the intersection of the plurality of first fault descriptions and the plurality of second fault descriptions, a plurality of third fault descriptions are obtained; Based on the weights corresponding to the multiple third fault descriptions and the preset weight threshold, the multiple third fault descriptions are filtered to obtain multiple fourth fault descriptions; By integrating the multiple fourth fault descriptions, fault description information is obtained.

6. The method for generating power customer service work orders according to claim 1, characterized in that, The supplementary prompts generated based on the information from the first power customer service work order specifically include: Based on the first power customer service work order, the type of missing information is obtained; By combining the large language model, the semantic expansion of the information missing types is performed to obtain information supplementation prompts.

7. The method for generating power customer service work orders according to claim 1, characterized in that, The step of updating the first power customer service work order based on the user's first supplementary information to obtain the second power customer service work order specifically includes: The user's first supplementary information is segmented into words to obtain the second segmentation information; Based on the pre-trained bidirectional encoder, the second word segmentation information is represented by features to obtain the second feature vector; Based on the second feature vector, update the first power customer service work order to obtain the second power customer service work order.

8. A method for generating power customer service work orders according to claim 1, characterized in that, Also includes: Based on the decision tree model, the priority of the second power customer service work order is obtained; The second power customer service work order is updated according to the priority to generate a third work order, and the corresponding personnel are assigned to process the work order.

9. A method for generating power customer service work orders according to claim 8, characterized in that, The process of prioritizing the second power customer service work order based on the decision tree model specifically includes: Based on the decision tree model, priority is identified for each work order field of the second power customer service work order to obtain multiple priorities; The priority of the second power customer service work order is calculated based on the preset field priority weights and the multiple priorities.

10. A power customer service work order generation device, characterized in that, It includes a semantic feature extraction module, a first power customer service work order acquisition module, a user description acquisition module, a user intent prediction module, and a second power customer service work order generation module; The semantic feature extraction module is used to extract semantic features from the user question information to be processed, and obtain a first feature vector. The first power customer service work order acquisition module is used to combine a preset work order field mapping rule to perform named entity recognition on the first feature vector to obtain the first power customer service work order, and update the first feature vector according to the first power customer service work order to obtain the second feature vector. The user description acquisition module is used to identify the second feature vector based on a preset fault feature recognition model and a preset fault description matching rule to obtain fault description information. The user intent prediction module is used to predict the intent of the fault description information to obtain a first user intent and a corresponding first confidence level. The second power customer service work order generation module is used to supplement the first power customer service work order generation information with a prompt when the first confidence level is lower than a preset confidence level threshold, and update the first power customer service work order with the first supplementary information of the user to obtain a second power customer service work order; wherein, the first supplementary information is obtained by receiving feedback from the user on the information supplement prompt.

Images