Method and device for evaluating maintenance service, computer device and storage medium

Abstract

The application relates to a maintenance service evaluation method and device, computer equipment and a storage medium. The maintenance service evaluation method comprises the following steps: obtaining a preset voice inquiry graph; obtaining an evaluation voice of a user, converting the evaluation voice into evaluation text, and inputting the evaluation text into an intention recognition model; obtaining a maintenance record corresponding to the evaluation voice, extracting a keyword in the evaluation text through the intention recognition model, fusing the keyword and a maintenance feature vector in the maintenance record based on an attention mechanism to obtain a fused feature vector; the intention recognition model performs similarity matching on the fused feature vector in a preset question node and an intention node at each level, and if a most matched intention node is obtained, the most matched intention node is taken as a most matched feature vector; and the most matched feature vector and a follow-up result node corresponding to the most matched intention node are fused to obtain an evaluation result. The evaluation method provided by the application provides an evaluation standard and solves the problems of accuracy and consistency.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus, computer equipment, and storage medium for evaluating maintenance services. Background Technology

[0002] In the repair industry, repair technicians typically follow up with customers by phone after completing the repair work to inquire about their satisfaction with the installation process.

[0003] Traditional technical solutions rely on manually recording user responses to score and evaluate repair services, but the accuracy and consistency of the data are subject to significant deviations. Summary of the Invention

[0004] Therefore, it is necessary to provide a maintenance service evaluation method, device, computer equipment, and storage medium that can achieve standardized evaluation to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for evaluating maintenance services, the method comprising: Obtain a preset voice inquiry graph; the voice inquiry graph includes preset question nodes, multi-level intent nodes, and follow-up result nodes. Each level of intent node is connected to a previous level intent node or preset question node, and the follow-up result node is connected to a last level intent node; the follow-up result is an evaluation score or evaluation level. The system acquires the user's voice evaluation, converts the voice evaluation into text evaluation, and inputs the text evaluation into the intent recognition model. Obtain the maintenance record corresponding to the evaluation voice, extract keywords from the evaluation text through an intent recognition model, and fuse the keywords with the maintenance feature vector in the maintenance record based on an attention mechanism to obtain the fused feature vector; The intent recognition model performs similarity matching on the preset question nodes and intent nodes at all levels based on the fused feature vector. If the most matching intent node is obtained, it is used as the most matching feature vector. The evaluation result is obtained by fusing the best matching feature vector with the return result node corresponding to the best matching intent node.

[0006] Furthermore, each level of intent node includes unknown question nodes and open question nodes. After the intent recognition model performs similarity matching on the preset question nodes and intent nodes at each level based on the fused feature vector, the method further includes: If no most matching intent node exists, the evaluation text is categorized based on its attributes, defined as either an unknown question category or an open question category. The evaluation text is labeled and input into the intent recognition model as a sample for training.

[0007] Further, the conversion of the evaluation speech into evaluation text includes: The evaluation speech is subjected to background noise reduction and speech enhancement processing to obtain preprocessed speech; The preprocessed speech is subjected to speech recognition using a speech recognition model to obtain the converted text. The converted text is corrected using a spelling correction algorithm to remove stop words, and punctuation marks are added after sentence segmentation to obtain the processed text. The processed text is subjected to sentiment recognition using a sentiment analysis model to obtain sentiment analysis results. The sentiment results are then semantically fused with the processed text to obtain evaluation text.

[0008] Furthermore, the process of fusing the keywords with the maintenance feature vector in the maintenance record based on the attention mechanism to obtain the fused feature vector includes: The intent recognition model uses a natural language processing algorithm to extract maintenance feature words from the maintenance records and represents them as maintenance feature vectors; the maintenance feature words include numerical features, category features, and text features. Using the keywords as the query vector and the maintenance feature vector as the key vector and value vector, weights are assigned by calculating the dot product similarity between the query vector and the key vector, and then the value vector is weighted and fused to obtain the fused feature vector.

[0009] Furthermore, the intent recognition model performs similarity matching on the preset question node and intent nodes at each level based on the fused feature vector. If the most matching intent node is obtained, it is used as the most matching feature vector, including: The intent recognition model calculates the similarity between the fused feature vector and each of the preset question nodes to determine the best matching question node; The intent recognition model calculates the similarity with the intent node downstream of the best matching question node to determine the best matching first-level intent node; The similarity of intent nodes is calculated sequentially according to the level until the last level. If the last level intent node has a best matching intent node, it is vectorized as the best matching feature vector.

[0010] Further, the step of fusing the most matching feature vector with the return result node corresponding to the most matching intent node to obtain the evaluation result includes: According to the preset standardized result template, the most matching feature vector is concatenated with the return result node to generate a comprehensive semantic vector; The comprehensive semantic vector is input into a pre-trained text generation model, which performs text fusion and outputs structured natural language text as the final evaluation result.

[0011] Secondly, this application also provides a maintenance service evaluation device, the device comprising: The acquisition module is used to acquire a preset voice inquiry graph; the voice inquiry graph includes preset question nodes, multi-level intent nodes and follow-up result nodes, each level of intent node is connected to a previous level intent node or preset question node, and the follow-up result node is connected to a last level intent node; the follow-up result is an evaluation score or evaluation level. The conversion module is used to acquire the user's evaluation voice, convert the evaluation voice into evaluation text, and input the evaluation text into the intent recognition model; The extraction module is used to obtain the maintenance record corresponding to the evaluation voice, extract keywords from the evaluation text through an intent recognition model, and fuse the keywords with the maintenance feature vector in the maintenance record based on an attention mechanism to obtain the fused feature vector. The matching module is used by the intent recognition model to perform similarity matching on the preset question nodes and intent nodes at all levels based on the fused feature vector. If the most matching intent node is obtained, it is used as the most matching feature vector. The fusion module is used to fuse the most matching feature vector with the return result node corresponding to the most matching intent node to obtain the evaluation result.

[0012] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any of the above claims.

[0013] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the preceding claims.

[0014] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in any of the above claims.

[0015] In the aforementioned maintenance service evaluation methods, devices, computer equipment, and storage media, a pre-set, hierarchical voice query map maps ambiguous, unstructured user speech to clear, pre-set business questions and intent nodes. This provides clear guidance and scope for intent recognition, avoiding misunderstandings caused by semantic ambiguity in traditional methods. The maintenance service evaluation method provided in this application innovatively integrates user evaluation text with background maintenance records, utilizing an attention mechanism to focus on the maintenance features most relevant to the current evaluation, thus improving the accuracy of intent recognition. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a diagram illustrating the application environment of a maintenance service evaluation method in one embodiment. Figure 2 This is a flowchart illustrating a maintenance service evaluation method in one embodiment; Figure 3 This is a structural block diagram of a maintenance service evaluation device in one embodiment; Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] The maintenance service evaluation method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or located in the cloud or on other network servers. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0020] In one exemplary embodiment, such as Figure 2 As shown, a maintenance service evaluation method is provided, which is applied to... Figure 1 Taking the server in the example, the explanation includes the following steps 202 to 210. Wherein: Step 202: Obtain a preset voice inquiry graph; the voice inquiry graph includes preset question nodes, multi-level intent nodes, and follow-up result nodes. Each level of intent node is connected to a higher-level intent node or preset question node, and the follow-up result node is connected to a last-level intent node; the follow-up result is the evaluation score or evaluation level.

[0021] For example, a voice query graph is a knowledge graph, which is a structured graphical model that represents knowledge through nodes (entities) and edges (relationships). As a knowledge graph, a voice query graph includes nodes (entities), edges (relationships), and a hierarchical structure.

[0022] For example, the preset question nodes represent questions the system asks the user during the follow-up process, such as "Are you satisfied with the repair service?" or "Was the installation process smooth?". Intent nodes represent the user's feedback intent; each intent node represents a different user response or answer to a question. The follow-up result node represents the final result of the follow-up, which determines the conclusion based on the user's feedback, such as "Satisfaction rating is 4 points" or "Evaluation level is satisfactory".

[0023] For example, the voice query graph is structured in a hierarchical manner. The first layer is the "preset question node", the second to N-1 layers are the "multi-level intent nodes" based on the user's answers, and the Nth and last layers are the "follow-up result nodes", which ultimately determine the follow-up result.

[0024] For example, the rating score can be from 1 to 10, and the rating level can be unsatisfactory, fairly satisfied, and very satisfied.

[0025] Step 204: Obtain the user's voice evaluation, convert the voice evaluation into text evaluation, and input the text evaluation into the intent recognition model.

[0026] For example, the user's evaluation voice is a voice recording formed from a follow-up phone call with the user, including the user's evaluation of the repair service.

[0027] For example, evaluation speech can be converted into evaluation text using a model such as speech-to-text.

[0028] Step 206: Obtain the maintenance record corresponding to the evaluation voice, extract keywords from the evaluation text through the intent recognition model, and fuse the keywords with the maintenance feature vector in the maintenance record based on the attention mechanism to obtain the fused feature vector.

[0029] For example, a repair record is a record of a repair order that has been executed, i.e., a repair order for which a telephone follow-up is made.

[0030] For example, the intent recognition model processes the evaluation text, such as by identifying keywords in the evaluation text based on natural language processing algorithms, and by vectorizing these keywords to obtain keyword vectors.

[0031] For example, the maintenance feature vector can be obtained by vectorizing maintenance features. For example, maintenance features can be structured or unstructured data such as maintenance type, maintenance item, maintenance work order number, fault description, maintenance measures, replaced parts, and maintenance time.

[0032] For example, the keyword vector and maintenance feature vector are input together into the attention-based fusion module. The attention mechanism dynamically calculates the relevance weights between the keyword vector and the maintenance feature vector, strengthens the maintenance features that are highly relevant to the evaluation intent, and generates a unified feature vector that integrates the evaluation keywords and maintenance features, i.e., the fused vector.

[0033] Step 208: The intent recognition model performs similarity matching on the preset question nodes and intent nodes at all levels based on the fused feature vector. If the most matching intent node is obtained, it is used as the most matching feature vector.

[0034] For example, the intent recognition model performs similarity matching calculations at each preset question node and each level of intent node based on the fused feature vector. Optionally, the calculation uses the cosine similarity method.

[0035] For example, the intent recognition model traverses intent nodes at all levels and selects the intent node with the highest similarity score. The feature vector corresponding to this intent node is then output as the "best-matching feature vector" representing the core intent of the current maintenance evaluation speech. Step 210: Fuse the best matching feature vector with the return result node corresponding to the best matching intent node to obtain the evaluation result.

[0036] For example, the follow-up result node can be a satisfaction rating, a score rating, or a level rating.

[0037] In the aforementioned maintenance service evaluation method, a pre-defined, hierarchical voice query graph maps ambiguous, unstructured user speech to clear, pre-defined business questions and intent nodes. This provides clear guidance and scope for intent recognition, avoiding misunderstandings caused by semantic ambiguity in traditional methods. The maintenance service evaluation method provided in this application innovatively integrates user evaluation text with background maintenance records, utilizing an attention mechanism to focus on the maintenance features most relevant to the current evaluation, thus improving the accuracy of intent recognition.

[0038] Furthermore, each level of intent node includes unknown question nodes and open question nodes. After the intent recognition model performs similarity matching between the preset question nodes and intent nodes at each level based on the fused feature vector, the method also includes: if there is no best matching intent node, classifying it based on the attributes of the evaluation text and defining it as an unknown question category or an open question category; labeling the evaluation text and inputting it into the intent recognition model as a sample for training.

[0039] For example, to handle user reviews that cannot be matched with a language query graph node, an "Unknown Problem" node and an "Open Problem" node are added to each level of intent node. An "Unknown Problem" node indicates that the repair service problem corresponding to the user's review voice does not exist; the user's review intent is completely outside the service scope of the current voice query graph, belonging to a new problem type. An "Open Problem" node indicates that the repair service problem corresponding to the user's review voice exists, but there is no solution, i.e., there is no downstream intent node.

[0040] For example, when the intent recognition model performs global similarity matching in the entire speech question graph based on the "fused feature vector", if the similarity scores of all preset nodes (including preset question nodes and intent nodes at all levels) are lower than the preset confidence threshold, the system determines that "there is no best matching intent node".

[0041] For samples identified as belonging to the "Unknown Question" and "Open Question" categories, these unmatched "evaluation texts" are temporarily labeled "Unknown Question" or "Open Question," and stored along with their corresponding "fused feature vectors" in a special sample library for processing. The samples are then manually calibrated to assign them the correct intent labels. Subsequently, these calibrated, high-quality samples are input into the intent recognition model as new training data for incremental training or fine-tuning.

[0042] Furthermore, the evaluation speech is converted into evaluation text, including: performing background noise reduction and speech enhancement on the evaluation speech to obtain preprocessed speech; performing speech recognition on the preprocessed speech using a speech recognition model to obtain converted text; performing text correction on the converted text based on a spelling correction algorithm to remove stop words, and adding punctuation marks after sentence segmentation to obtain processed text; performing sentiment recognition on the processed text using a sentiment analysis model to obtain sentiment analysis results, and semantically fusing the sentiment results with the processed text to obtain the evaluation text.

[0043] For example, the original evaluation speech may contain noise, requiring front-end signal preprocessing. By using a spectral subtraction-based noise reduction algorithm and a deep learning-based speech enhancement model, the human voice and background noise are effectively separated, improving the signal-to-noise ratio of the speech signal and obtaining preprocessed speech with clean sound quality.

[0044] For example, preprocessed speech is input into a speech recognition model to convert acoustic features into text. Optionally, the speech recognition model can be an encoder-decoder architecture, combined with an attention mechanism, to map the audio feature sequence of the preprocessed speech into a text sequence and output the converted text.

[0045] For example, the converted text may contain recognition errors, so the recognized text undergoes multiple rounds of refinement processing. Optionally, homophone and near-homophone errors are corrected based on spelling correction algorithms; stop words with small semantic contributions are removed using TF-IDF weighted algorithms; sentence segmentation is achieved by combining speech pause detection and semantic integrity analysis to identify sentence boundaries; punctuation marks are added according to sentence structure and semantic relationships to finally generate the processed text.

[0046] For example, the sentiment analysis model is used for deep semantic understanding to identify the sentiment polarity and intensity contained in the processed text, obtaining sentiment analysis results that include sentiment labels and confidence levels. The sentiment vector and the text semantic vector are concatenated and fused at the feature level to form an enhanced evaluation text that retains the original semantics while being rich in sentiment information, providing a richer feature representation for subsequent intent recognition.

[0047] Furthermore, based on the attention mechanism, keywords are fused with maintenance feature vectors in maintenance records to obtain a fused feature vector. This includes: the intent recognition model uses natural language processing algorithms to extract maintenance feature words from maintenance records and represents them as maintenance feature vectors; maintenance feature words include numerical features, category features, and text features; keywords are used as query vectors, and maintenance feature vectors are used as key vectors and value vectors. Weights are assigned by calculating the dot product similarity between the query vector and the key vector, and then the value vectors are weighted and fused to obtain the fused feature vector.

[0048] For example, numerical features can be maintenance time, maintenance cost, etc., categorical features can be maintenance type, fault level, maintenance personnel ID, etc., and text features can be unstructured text such as fault description and maintenance measures.

[0049] For example, the keyword vector is used as the query vector, and the entire set of maintenance feature vectors is used as both the key vector and the value vector. The key vector is used to calculate the relevance with the query vector, while the value vector is the final aggregated information ontology.

[0050] Weight Calculation: An attention score is obtained by calculating the dot product similarity between the query vector (keywords) and each key vector (maintenance features), serving as an indicator of association strength. This attention score is then normalized using the Softmax function to obtain weight coefficients. These weight coefficients are then used to weight and sum the corresponding value vectors (maintenance features) to obtain the fused feature vector.

[0051] Furthermore, the intent recognition model performs similarity matching on preset question nodes and intent nodes at all levels based on the fused feature vector. If the most matching intent node is obtained, it is used as the most matching feature vector. This includes: the intent recognition model calculates the similarity between the fused feature vector and each preset question node to determine the most matching question node; the intent recognition model calculates the similarity with the intent nodes downstream of the most matching question node to determine the most matching first-level intent node; the similarity of intent nodes is calculated sequentially according to the level until the last level. If the last-level intent node has a most matching intent node, it is vectorized as the most matching feature vector.

[0052] For example, the root node is first determined. The intent recognition model first calculates the similarity between the fused feature vector and the vector representations of all preset question nodes in the first layer of the graph (optionally, cosine similarity is used for calculation). By comparing all similarity scores, the system selects the node with the highest score and determines it as the best-matching question node.

[0053] For example, after determining the best matching question node, the intent recognition model matches the first-level intent node by calculating the similarity between the fused feature vector and all downstream first-level intent nodes directly connected to the best matching question node, and then determining the node with the highest similarity as the best matching first-level intent node.

[0054] Then, according to the preset levels of the graph, the similarity is calculated repeatedly multiple times. The intent node matched in the previous step is used as the new parent node, and the similarity with its downstream child nodes is calculated again to determine the best matching node at that level.

[0055] The top-down traversal process will continue until the intent node at the last level in the graph is reached. If a node is successfully matched at each level of this path, and the node similarity at the final level exceeds a preset confidence threshold, then the final level intent node is determined as the best matching intent node globally, and it is vectorized into the best matching feature vector.

[0056] Furthermore, the evaluation result is obtained by fusing the most matching feature vector with the corresponding return result node of the most matching intent node. This includes: concatenating the most matching feature vector with the return result node according to a preset standardized result template to generate a comprehensive semantic vector; inputting the comprehensive semantic vector into a pre-trained text generation model, which then performs text fusion and outputs structured natural language text as the final evaluation result.

[0057] For example, according to the preset standardized result template, the "most matching feature vector" representing the user's specific intent and the vector of the "follow-up result node" representing the standardized processing scheme are concatenated and fused to generate a comprehensive semantic vector.

[0058] For example, the text generation model performs conditional text generation based on the structured information encoded by the comprehensive semantic vector, understands and integrates all the key elements of the comprehensive semantic vector, and transforms it into fluent and coherent natural language.

[0059] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0060] Based on the same inventive concept, this application also provides a maintenance service evaluation device for implementing the maintenance service evaluation method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more maintenance service evaluation device embodiments provided below can be found in the limitations of the maintenance service evaluation method described above, and will not be repeated here.

[0061] In one exemplary embodiment, such as Figure 3As shown, a maintenance service evaluation device 300 is provided, including: an acquisition module 302, a conversion module 304, an extraction module 306, a matching module 308, and a fusion module 310. Wherein: The acquisition module 302 is used to acquire a preset voice inquiry graph; the voice inquiry graph includes preset question nodes, multi-level intent nodes and follow-up result nodes, each level of intent node is connected to a previous level intent node or preset question node, and the follow-up result node is connected to a last level intent node; the follow-up result is an evaluation score or evaluation level. The conversion module 304 is used to acquire the user's evaluation voice, convert the evaluation voice into evaluation text, and input the evaluation text into the intent recognition model; Extraction module 306 is used to obtain the maintenance record corresponding to the evaluation voice, extract keywords in the evaluation text through an intent recognition model, and fuse the keywords with the maintenance feature vector in the maintenance record based on an attention mechanism to obtain the fused feature vector; The matching module 308 is used by the intent recognition model to perform similarity matching on the preset question node and intent nodes at all levels based on the fused feature vector. If the most matching intent node is obtained, it is used as the most matching feature vector. The fusion module 310 is used to fuse the most matching feature vector with the return result node corresponding to the most matching intent node to obtain the evaluation result.

[0062] Each module in the aforementioned maintenance service evaluation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0063] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores maintenance data. The I / O interfaces allow the processor to exchange information with external devices. The communication interface allows communication with external terminals via a network connection. When executed by the processor, the computer program implements a maintenance evaluation method.

[0064] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0065] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0066] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0067] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0068] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0069] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0070] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for evaluating maintenance services, characterized in that, The method includes: Obtain a preset voice inquiry graph; the voice inquiry graph includes preset question nodes, multi-level intent nodes, and follow-up result nodes. Each level of intent node is connected to a higher-level intent node or preset question node, and the follow-up result node is connected to a last-level intent node; the follow-up result is an evaluation score or evaluation level; the intent node represents the user's feedback intent, and each intent node represents a different response or answer from the user to a question. The system acquires the user's voice evaluation, converts the voice evaluation into text evaluation, and inputs the text evaluation into the intent recognition model. The repair record corresponding to the evaluation voice is obtained, and keywords in the evaluation text are extracted through an intent recognition model. The keywords are then fused with the repair feature vector in the repair record based on an attention mechanism to obtain a fused feature vector. The repair feature vector is obtained by vectorizing the repair features, which include repair type, repair item, repair work order number, and fault description. The intent recognition model performs similarity matching on the preset question nodes and intent nodes at each level based on the fused feature vector. If the most matching intent node is obtained, it is used as the most matching feature vector. This includes: the intent recognition model calculates the similarity between the fused feature vector and each preset question node to determine the most matching question node; the intent recognition model calculates the similarity with the intent nodes downstream of the most matching question node to determine the most matching first-level intent node; the similarity of intent nodes is calculated sequentially according to the level until the last level. If the last-level intent node has a most matching intent node, it is vectorized as the most matching feature vector. The evaluation result is obtained by fusing the best matching feature vector with the return result node corresponding to the best matching intent node.

2. The method according to claim 1, characterized in that, Each level of intent node includes unknown question nodes and open question nodes. After the intent recognition model performs similarity matching between the preset question nodes and intent nodes at each level based on the fused feature vector, the method further includes: If no most matching intent node exists, the evaluation text is categorized based on its attributes, defined as either an unknown question category or an open question category. The evaluation text is labeled and input into the intent recognition model as a sample for training.

3. The method according to claim 1, characterized in that, The step of converting the evaluation speech into evaluation text includes: The evaluation speech is subjected to background noise reduction and speech enhancement processing to obtain preprocessed speech; The preprocessed speech is subjected to speech recognition using a speech recognition model to obtain the converted text. The converted text is corrected using a spelling correction algorithm to remove stop words, and punctuation marks are added after sentence segmentation to obtain the processed text. Sentiment recognition is performed on the processed text using a sentiment analysis model to obtain sentiment analysis results. The sentiment analysis results are then semantically fused with the processed text to obtain evaluation text.

4. The method according to claim 1, characterized in that, The process of fusing the keywords with the maintenance feature vector in the maintenance record based on an attention mechanism to obtain a fused feature vector includes: The intent recognition model uses a natural language processing algorithm to extract maintenance feature words from the maintenance records and represents them as maintenance feature vectors; the maintenance feature words include numerical features, category features, and text features. Using the keywords as the query vector and the maintenance feature vector as the key vector and value vector, weights are assigned by calculating the dot product similarity between the query vector and the key vector, and then the value vector is weighted and fused to obtain the fused feature vector.

5. The method according to claim 1, characterized in that, The step of fusing the best-matching feature vector with the return result node corresponding to the best-matching intent node to obtain the evaluation result includes: According to the preset standardized result template, the most matching feature vector is concatenated with the return result node to generate a comprehensive semantic vector; The comprehensive semantic vector is input into a pre-trained text generation model, which performs text fusion and outputs structured natural language text as the final evaluation result.

6. A maintenance service evaluation device, characterized in that, The device includes: The acquisition module is used to acquire a preset voice inquiry graph; the voice inquiry graph includes preset question nodes, multi-level intent nodes, and follow-up result nodes. Each level of intent node is connected to a higher-level intent node or preset question node, and the follow-up result node is connected to a last-level intent node; the follow-up result is an evaluation score or evaluation level; the intent node represents the user's feedback intent, and each intent node represents a different response or answer from the user to a question. The conversion module is used to acquire the user's evaluation voice, convert the evaluation voice into evaluation text, and input the evaluation text into the intent recognition model; The extraction module is used to obtain the maintenance record corresponding to the evaluation voice, extract keywords from the evaluation text through an intent recognition model, and fuse the keywords with the maintenance feature vector in the maintenance record based on an attention mechanism to obtain the fused feature vector; the maintenance feature vector is obtained by vectorizing the maintenance features, which include maintenance type, maintenance item, maintenance work order number, and fault description. The matching module is used by the intent recognition model to perform similarity matching on the preset question nodes and intent nodes at all levels based on the fused feature vector. If the most matching intent node is obtained, it is used as the most matching feature vector. The module includes: the intent recognition model calculating the similarity between the fused feature vector and each preset question node to determine the most matching question node; the intent recognition model calculating the similarity with the intent nodes downstream of the most matching question node to determine the most matching first-level intent node; and calculating the similarity of intent nodes sequentially according to the level until the last level. If the last-level intent node has a most matching intent node, it is vectorized as the most matching feature vector. The fusion module is used to fuse the most matching feature vector with the return result node corresponding to the most matching intent node to obtain the evaluation result.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images