95598 customer appeal identification method and system

Abstract

The invention relates to the technical field of digital data processing, in particular to a 95598 customer appeal identification method and system. The invention provides a 95598 customer demand identification method. The method comprises the following steps: model training; the method specifically comprises the steps of establishing a power industry lexicon and an appeal corpus; a hierarchical classification system is constructed by taking a customer perspective as a core, and full-scene appeals of the power industry are covered; modeling a hybrid neural network; model deployment and reasoning; and inputting the text into the trained model, and outputting the appeal type, the potential risk and the emotional tendency. Different from a traditional business-dominated classification mode, a client-centered classification system is established, events are objectively described from the perspective of clients, and it is ensured that the classification dimension covers the real appeals of the clients.

Description

Technical Field

[0001] This patent relates to the field of digital data processing technology, and specifically to a 95598 customer demand identification method and system. Background Art

[0002] With the digital transformation of the power industry, voice customer service has become a crucial channel for collecting customer information, but traditional analytical methods struggle to effectively unlock its value. Customer demands, emotions, and expectations are concentrated in voice interactions, containing a wealth of valuable information. However, traditional data mining methods face numerous challenges when processing this unstructured data. Traditional voice customer service systems primarily rely on manual transcription of work orders or simple keyword matching techniques. For example, work order text analysis uses regular expressions or rule engines to extract keywords, but cannot handle complex semantics, such as distinguishing between the intent of "sudden increase in electricity bills" and "problems with the meter."

[0003] Existing customer demand classification systems are often business-driven and lack the flexibility to adapt to rapidly changing customer needs. For example, traditional customer service knowledge bases require manual maintenance, are often updated late, and fail to reflect the latest policy or technology changes.

[0004] Existing voice request recognition technology has the following shortcomings:

[0005] In existing customer service systems, emotion recognition and risk prevention primarily rely on manual quality control, which is time-sensitive and unable to identify issues in real time. Furthermore, existing systems often use simple keyword matching methods for intent recognition, which cannot handle complex semantics and long-distance dependencies, resulting in insufficient recognition accuracy.

[0006] The traditional classification system is business-oriented. For example, "power outage" is classified as "fault repair" but ignores the customer's specific demands, such as "power outage compensation" and "power outage warning", resulting in a deviation between the classification results and the customer's actual needs.

[0007] The classification system is static. Fixed classification systems cannot be dynamically expanded and lack self-improvement mechanisms, making them unable to adapt to emerging customer demands or policy changes. This results in low classification accuracy. For example, a power company must manually update its knowledge base, which takes an average of 3-5 days to reflect the latest policy adjustments. Fixed classification systems cannot be dynamically expanded and cannot identify emerging demand types, such as "new energy vehicle charging station registration," resulting in low classification accuracy.

[0008] Key information such as customer service upgrade trends and public opinion risks mainly relies on manual identification and reporting by customer service specialists, which cannot comprehensively prevent and control risks. For example, collective complaints may evolve into a public opinion crisis due to delayed manual identification.

[0009] In summary, there is an urgent need for a method and system for automatically identifying 95598 customer demands. Summary of the Invention

[0010] The present invention combines artificial intelligence technology with classification system construction to achieve automated, high-precision identification of customer demands and risk warning.

[0011] In a first aspect, the present invention provides a 95598 customer demand identification method, comprising the following process:

[0012] Model training; specifically includes:

[0013] S100: Establish a vocabulary and appeal corpus for the power industry;

[0014] S200: With the customer perspective as the core, it builds a hierarchical classification system to cover all scenarios in the power industry;

[0015] S300: Hybrid neural network modeling; specifically includes the following processes:

[0016] S310: Training data preparation: dividing the annotated corpus into training set, validation set, and test set;

[0017] S320: Design a neural network with the task of appeal recognition and auxiliary tasks of sentiment polarity recognition and risk warning. The specific process includes the following:

[0018] S321: BiGRU text feature extraction; obtain word embedding sequences of speech transcription text and work order text through pre-trained word vectors, using two layers of BiGRU, each with 128 units. The forward GRU captures forward semantic dependencies, and the reverse GRU captures backward context associations. The forward and reverse hidden states are spliced ​​into a 256-dimensional context-aware feature vector, and the text feature matrix H is output. t×256 , where t represents the sequence length;

[0019] S322: Through multi-task output, in-depth analysis of the requirements of a single work order is achieved; the specific process is as follows:

[0020] S322a: Main Task - Hierarchical Classification: Use the self-attention mechanism to assign different weights to each word in the text, highlighting important information and ignoring unimportant information. The weighted features are then input into the softmax layer to convert the output into a probability distribution to obtain the final classification result.

[0021] S322b: Sentiment analysis; fine-tune classification based on the BERT model, freeze the parameters of the first 12 layers, and only train the last two fully connected layers. Define the optimizer and learning rate respectively, and start training with a small learning rate.

[0022] S322c: Risk identification and level prediction: Risk identification and detection are performed using rule features combined with semantic features and sentiment analysis results, and risks are graded on a scale of [0-5];

[0023] Model deployment and inference: Input text into the trained model and output the appeal type, potential risks, and emotional tendencies.

[0024] Preferably, after model deployment and inference, the following is also included:

[0025] S400: Self-improvement of the demand classification system; dynamic updating of the classification system through new word discovery and mining of highly relevant word pairs to adapt to the emerging demands of the power industry; specifically including:

[0026] S410: Text feature mining: Based on a weekly frequency, the newly added speech transcription text is grouped into [2, 3, ..., k] tuples as candidate words. The word frequency (TF), average mutual information (AMI), and left and right adjacency of each candidate word are calculated, and finally new words that meet the threshold requirements are selected.

[0027] S420: Mining related word pairs: Filtering words with a co-occurrence probability greater than 0.8 with the new word through co-occurrence matrix analysis;

[0028] S430: Classification system expansion: automatically expand the final classification based on the semantics of the new word;

[0029] S440: Expert review mechanism; newly expanded categories must be voted on and confirmed by 3-5 field experts, and the approval rate must be ≥ 70% before they can take effect;

[0030] S450: Historical data backtracking; after updating the classification system, historical data is automatically reclassified and the knowledge base is updated.

[0031] Preferably, the S100 includes the following process:

[0032] S110: Data collection: collect historical customer service voice data and work order texts to build a basic corpus;

[0033] S120: Data cleaning: segmenting and noise reduction processing of the original voice data, filtering invalid characters and blank characters using regular expressions;

[0034] S130: Using the Delphi method to organize power industry experts to conduct multiple rounds of anonymous feedback to gradually converge the scope of the vocabulary;

[0035] S140: Label the appeal type, sentiment label, and risk point of each sample, and store the cleaned text in the format of "appeal type-scenario-example" in JSON format. The fields include classification label, text, sentiment value, and risk point.

[0036] Preferably, the S130 includes the following process:

[0037] S131: First round of expert survey: a questionnaire containing a certain number of power industry terms was distributed to several experts, requiring them to screen core terms and add new terms;

[0038] S132: Count expert opinions, mark words with a support rate below 60% as "items for discussion", and feed them back to experts for a second round of evaluation;

[0039] S133: Experts score candidate terms on a scale of 1-5, retaining terms with a score of 4 or higher. After 3-5 rounds of iterations, a standardized vocabulary of 2,300 power industry terms is finally formed. At the same time, voting is conducted on controversial terms to finally determine the classification boundaries.

[0040] Preferably, the S200 includes the following process:

[0041] S210: Customer perspective demand research: Collect typical demand cases through interviews with several electricity users;

[0042] S220: Classification hierarchy design; construct a two-level classification system based on the Delphi vocabulary and user cases;

[0043] S230: Invite five customer service experts to review the classification logic and correct overlapping or ambiguous classification boundaries;

[0044] S240: Verification of the classification system. Invite the customers who participated in the interview to participate in the test, select a certain number of labeled samples, and test the coverage and accuracy of the classification system. The final classification coverage is required to reach 95%, and the misclassification rate is ≤5%. The classification accuracy of random appeal text is required to be ≥95%. If the user recognition rate of a certain category is lower than 80%, it will be renamed or merged into the parent category.

[0045] In a second aspect, the present invention provides a 95598 customer demand identification method system, comprising:

[0046] The appeal corpus unit is used to build the electric power industry vocabulary and appeal corpus;

[0047] The classification system unit is used to build a hierarchical classification system with the customer perspective as the core, covering the full range of scenarios in the power industry;

[0048] Appeal recognition unit, used for automatic recognition of speech to appeal;

[0049] The self-improvement unit of the demand classification system dynamically updates the classification system through new word discovery and high-correlation word pair mining to adapt to the demands of the emerging power industry.

[0050] Preferably, the appeal corpus unit specifically includes:

[0051] The data collection module is used to collect historical customer service voice data and work order texts to build a basic corpus;

[0052] The data cleaning module is used to segment and reduce noise of the original voice data, and filter invalid characters and empty characters through regular expressions;

[0053] The corpus standardization and storage module is used to store the cleaned text in the "appeal type-scenario-example" format, using JSON format. The fields include classification labels, text, sentiment values, and risk points.

[0054] The advantages of the present invention over the prior art are as follows:

[0055] (1) Different from the traditional business-oriented classification method, the present invention adopts the Delphi method to establish a "customer-centric" classification system (two-level category system) through multiple rounds of expert consultation and statistical feedback, objectively describing events from the customer's perspective, and ensuring that the classification dimensions cover the real demands of customers.

[0056] (2) Solve the serious problem of information attenuation from voice conversation content to work order text. Conduct in-depth mining of customer service voice resources, extract key semantic features through speech-to-text and semantic analysis, combine sentiment analysis with association rule mining, and detect negative emotions and group demands in real time. For example, when multiple users complain about the same problem, automatic identification of customer demands, service risk points and other information in the recording can be achieved, reducing manual intervention and improving processing efficiency. For example, "Why did my bill suddenly increase?" is classified as "Bill dispute → Billing error". The BiGRU model is used to capture long-distance dependencies in text sequences, accurately understand the contextual semantics of customer demands, and reduce information loss. The model can handle multiple rounds of conversations between customer service and customers, and fully understand customer intentions, rather than just focusing on a single voice clip.

[0057] (3) In the embodiment, a dynamic customer demand classification system is constructed. Through the self-improvement mechanism of the classification system, the system can continuously learn new customer demands and policy changes without the need for manual frequent updates of the knowledge base. Customer service voice content can perceive changes in customer demands the fastest. Existing demand recognition technology is generally based on a solidified classification system from a business perspective. Fixed classification systems are updated with delays, and classification model updates rely on manual annotation of new samples. The response cycle is as long as 2-3 months, resulting in emerging customers' "new energy vehicle charging pile registration demands" not being included in monitoring in a timely manner, and there are blind spots in service responses. Through new word discovery algorithms and high-correlation word pair mining technology, new demands and potential change trends in text features are automatically identified, and customer feedback on the latest marketing management measures of the company or provincial company is tracked in a timely manner, so that the classification system can dynamically adapt to changes in customer needs and ensure the adaptability of the technical solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a schematic diagram of a 95598 customer demand identification method system and its operating principle in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] Explanation of relevant terms:

[0060] Delphi method: A forecasting method that reaches consensus through multiple rounds of anonymous expert feedback. It is used to build a scientific and objective classification system. For example, domain experts are invited to conduct multiple rounds of discussions on the classification dimensions of customer demands, ultimately forming a unified classification standard.

[0061] Long Short-Term Memory (LSTM) network: A special recurrent neural network (RNN) structure that addresses the vanishing gradient problem common in traditional RNNs when processing long sequences of data by introducing gating mechanisms such as a "forget gate," an "input gate," and an "output gate." LSTM excels at processing sequential data, capturing long-term dependencies through memory units and improving the accuracy of semantic analysis.

[0062] Gated Recurrent Unit (BiGRU): Bidirectional Gated Recurrent Unit (BiGRU) is an extension of the GRU. It is designed to capture richer contextual information by processing both the forward and reverse sequences of data simultaneously. The GRU is a neural network structure designed for processing and predicting sequential data, serving as a simplified variant of the LSTM.

[0063] New word discovery: An unsupervised learning technology that analyzes word combination patterns in text data to identify words that do not appear in the existing vocabulary but have specific meanings, such as "flexible direct current transmission" and "green electricity" and other emerging terms in the power industry, supporting the dynamic expansion of the classification system.

[0064] Highly correlated word pairs: These are word combinations that frequently co-occur or have highly related semantics in text data, such as "electricity price" and "subsidy," "power outage" and "compensation," reflecting the potential correlation and changing trends of customer demands.

[0065] Customer Complaint Classification System: This system objectively describes customer call events from the customer's perspective. It consists of two levels of categories, such as electricity bill disputes, service processes, equipment failures, and other major categories, and a final level of categories such as electricity price calculation errors, power outage repair delays, and green electricity subsidy policies. It is used to structure the classification of customer complaints.

[0066] Compare with Figure 1 , a 95598 customer demand identification method system, comprising:

[0067] The appeal corpus unit is used to establish the electric power industry vocabulary and appeal corpus; specifically includes:

[0068] The data collection module is used to collect historical customer service voice data and work order texts to build a basic corpus;

[0069] The data cleaning module is used to segment and reduce noise of the original voice data, and filter invalid characters and empty characters through regular expressions;

[0070] The corpus standardization and storage module is used to store the cleaned text in the "appeal type-scenario-example" format in JSON format. The fields include classification label, text, sentiment value, and risk point.

[0071] The classification system unit is used to build a hierarchical classification system with the customer perspective as the core, covering the full range of scenarios in the power industry;

[0072] Appeal recognition unit, used for automatic recognition of speech to appeal;

[0073] The self-improvement unit of the demand classification system dynamically updates the classification system through new word discovery and high-correlation word pair mining to adapt to the demands of the emerging power industry.

[0074] A 95598 customer demand identification method includes the following steps:

[0075] Model training; specifically includes:

[0076] S100: Use the Delphi method and expert judgment method to establish a vocabulary and demand corpus for the power industry; the specific steps are as follows:

[0077] S110: Data collection: The data comes from voice recordings and work order texts from the 95598 customer service system and is stored in structured text file formats such as txt, JSON, and Oracle. Historical customer service voice data and work order texts are collected to build a basic corpus. Voice recordings and work order texts from the power customer service system over the past year are extracted, covering typical complaints such as "electricity bill disputes" and "service attitude complaints."

[0078] S120: Data cleaning: The original voice data contains a large number of modal particles and blank characters. The original voice data is segmented and noise-reduced, and invalid and blank characters are filtered using regular expressions. For example, modal particles such as "hmm" and "ah" are filtered out, as well as non-electricity-related words, while retaining core demand text such as "fault repair" and "electricity bill dispute";

[0079] S130: Use the Delphi method to organize power industry experts to conduct 3-4 rounds of anonymous feedback to gradually narrow the scope of the vocabulary. Power industry experts can be customer service managers, business analysts, and technical engineers. The process includes the following:

[0080] S131: First round of expert survey: A questionnaire containing 1,000 power industry terms was distributed to 10 experts, who were asked to screen core terms and add new ones. For example, experts reported that "power outage type" should be further subdivided into "planned outage" and "fault-related outage";

[0081] S132: Count expert opinions, mark words with a support rate below 60% as "items for discussion", and feed them back to experts for a second round of evaluation;

[0082] S133: Experts score candidate terms on a scale of 1-5, retaining those with a score of 4 or higher. After 3-5 rounds of iteration, a standardized vocabulary of 2,300 power industry terms is finally formed, covering high-frequency scenarios such as "electricity bill calculation" and "fault repair." Voting is also conducted on controversial terms to determine the classification boundaries, such as "electricity bill calculation dispute" and "electricity price policy consultation." The first-level categories and second-level subcategories are shown in the table below.

[0083]

[0084] S140: Corpus Standardization and Storage: Using a combination of manual and automatic annotation, label each sample with the appeal type, such as "bill dispute → billing error," sentiment labels, such as "negative" and "neutral," and risk points, such as "group risk." Label consistency must be ≥ 90%. The cleaned text is stored in the "appeal type-scenario-example" format in JSON format. Fields include classification label, text, sentiment value, and risk point. An example is as follows:

[0085] {

[0086] "category":"Electricity Bill Disputes→Billing Rules",

[0087] "text":"Why did my electricity bill suddenly increase by 200 yuan?",

[0088] "emotion":"negative",

[0089] "risk":"Increase in electricity bills"

[0090] };

[0091] S200: With the customer perspective at the core, we build a hierarchical classification system that covers all scenarios in the power industry. This includes the following processes:

[0092] S210: Customer Perspective Demand Research: Through interviews with 100 electricity users, we collected typical demand cases. For example, users reported frequent issues such as "no notification after power outage" and "lack of transparency in electricity fee calculation rules."

[0093] S220: Classification hierarchy design: Based on the Delphi vocabulary and user cases, a two-level classification system is constructed; 26 major categories, such as "Electricity Billing Service" and "Fault Repair"; 189 final categories, such as "Billing Error" and "Equipment Failure";

[0094] S230: Invite five customer service experts to review the classification logic and correct overlapping or ambiguous classification boundaries; for example, separate "service attitude complaints" from "business consultation" to the "service experience" category;

[0095] S240: Classification system verification: To ensure the classification system's suitability from a customer perspective, we invited interviewed customers to participate in testing, selecting 1,000 annotated samples to test the classification system's coverage and accuracy. The final classification coverage was required to be 95%, with a misclassification rate ≤ 5%. The classification accuracy for random appeal text was required to be ≥ 95%. If a category's user recognition rate fell below 80%, it would be renamed or merged into a parent category.

[0096] S300: Hybrid neural network modeling; specifically includes the following processes:

[0097] S310: Training data preparation: Divide the annotated corpus into a training set (70%), a validation set (15%), and a test set (15%).

[0098] S320: Design a neural network with the task of appeal recognition and auxiliary tasks of sentiment polarity recognition and risk warning. The specific process includes the following:

[0099] S321: BiGRU text feature extraction; obtain word embedding sequences of speech transcription text and work order text through pre-trained word vectors, dimension d = 300, such as: word "electricity fee" → [0.12, -0.23, ..., 0.45] 1×300 , using two layers of BiGRU, each layer has 128 units, the forward GRU captures the forward semantic dependency, the reverse GRU captures the backward context association, the forward and reverse hidden states are concatenated into a 256-dimensional context-aware feature vector, and the output text feature matrix H t×256 , where t represents the sequence length;

[0100] S322: Multi-task output header. Through multi-task output, in-depth analysis of the requirements of a single work order can be achieved;

[0101] Main Task - Hierarchical Classifier: Use the self-attention mechanism to assign different weights to each word in the text, highlighting important information and ignoring unimportant information. The weighted features are then input into the softmax layer to convert the output into a probability distribution, resulting in the final classification result. Since this is a two-level classification, dynamic sub-classifiers are configured under each major category, and the corresponding second-level classifier is selected based on the first-level results.

[0102] Sentiment analysis task: Classification fine-tuning is performed based on the BERT model. The parameters of the first 12 layers are frozen, and only the last two fully connected layers are trained. The optimizer and learning rate are defined respectively. The learning rate is trained from a small value. The learning rate of this invention is set to 5e-5. The number of samples selected for one training is 16, and a total of 3 rounds of training are performed.

[0103] Risk identification and level prediction: Risk identification and detection are performed using rule features combined with semantic features and sentiment analysis results. Risks are graded on a scale of [0-5], such as group words "our village | multiple households | everyone | collective", multiple cases, and threat expressions.

[0104] Model deployment and inference: Input text into the trained model and output the demand type, potential risk, and sentiment. Based on the above multi-task model training, the demand type, potential risk, and sentiment of the work order are realized; as shown in the following example:

[0105] User description: "The transformer at the east end of our village has been humming loudly these past few days during peak hours. The other day, Mr. Zhang's TV burned out due to unstable voltage. Shouldn't the power supply station give an explanation for this situation?"

[0106] {

[0107] "Appeal Type":[

[0108] {"type":"Equipment Maintenance","entity":"Village East Head Transformer"},

[0109] {"type":"Loss compensation","entity":"Zhang San's TV"}

[0110] ],

[0111] "Potential Risk":[

[0112] {"type":"mass incident risk","level":4},

[0113] ],

[0114] "Sentiment tendency":{"score":0.87,"type":"negative"},

[0115] };

[0116] S400: Self-improvement of the demand classification system; dynamic updating of the classification system through new word discovery and mining of highly relevant word pairs to adapt to the emerging demands of the power industry; specifically including:

[0117] S410: Text feature mining: Based on a weekly frequency, the newly added speech transcription texts are grouped into [2, 3, ..., k] tuples as candidate words (k is 5 in the present invention). The word frequency (TF), average mutual information (AMI), and left and right adjacency entropy of each candidate word are then calculated. Finally, new words that meet the threshold requirements are selected.

[0118] S420: Mining related word pairs: Through co-occurrence matrix analysis, filter out words with a co-occurrence probability greater than 0.8 with new words; for example, "virtual power plant" often co-occurs with "electricity subsidy", and it is inferred that it belongs to the "policy consultation" category;

[0119] S430: Classification system expansion: Automatically expand the final classification based on the semantics of new words; for example, add a "virtual power plant policy" subcategory under the "new energy services" category and trigger the expert review process;

[0120] S440: Expert review mechanism; newly expanded categories must be voted on and confirmed by 3-5 field experts, and the approval rate must be ≥ 70% before they can take effect;

[0121] S450: Historical data backtracking: After updating the classification system, historical data is automatically reclassified and the knowledge base is updated. For example, work orders mentioning "virtual power plants" in 2024 are reclassified into a new subcategory.

[0122] This embodiment achieves indiscriminate extraction of customer demand data, achieving a demand recognition accuracy rate of 80%. For example, a payment channel analysis based solely on work order content would not extract all related work orders for that demand, nor would it analyze the problems existing in each channel, such as Alipay, WeChat, and the website. The demand recognition results of the present invention can fully match relevant work orders, greatly improving the breadth of customer service thematic analysis.

[0123] The efficiency of data analysis is greatly improved. It takes about 30 seconds to manually identify a demand, while the demand identification method of the present invention only takes 10ms. The demand identification efficiency is improved by 3000 times, and the processing time of special analysis data can be shortened by more than 50%.

Claims

1. A 95598 customer demand identification method, comprising the following steps: Model training; specifically includes: S100: Establish a vocabulary and appeal corpus for the power industry; S200: With the customer perspective as the core, it builds a hierarchical classification system to cover all scenarios in the power industry; S300: Hybrid neural network modeling; specifically includes the following processes: S310: Training data preparation: dividing the annotated corpus into training set, validation set, and test set; S320: Design a neural network with the task of appeal recognition and auxiliary tasks of sentiment polarity recognition and risk warning. The specific process includes the following: S321: BiGRU text feature extraction; obtain word embedding sequences of speech transcription text and work order text through pre-trained word vectors, using two layers of BiGRU, each with 128 units. The forward GRU captures forward semantic dependencies, and the reverse GRU captures backward context associations. The forward and reverse hidden states are spliced ​​into a 256-dimensional context-aware feature vector, and the text feature matrix H is output. t×256 , where t represents the sequence length; S322: Through multi-task output, in-depth analysis of the requirements of a single work order is achieved; the specific process is as follows: S322a: Main Task - Hierarchical Classification: Use the self-attention mechanism to assign different weights to each word in the text, highlighting important information and ignoring unimportant information. The weighted features are then input into the softmax layer to convert the output into a probability distribution to obtain the final classification result. S322b: Sentiment analysis; fine-tune classification based on the BERT model, freeze the parameters of the first 12 layers, and only train the last two fully connected layers. Define the optimizer and learning rate respectively, and start training with a small learning rate. S322c: Risk identification and level prediction; Risk identification and detection are performed using rule features combined with semantic features and sentiment analysis results, and risks are graded from [0-5]; Model deployment and inference: Input text into the trained model and output the appeal type, potential risks, and emotional tendencies.

2. A 95598 customer demand identification method according to claim 1, characterized in that: After model deployment and inference, it also includes: S400: Self-improvement of the demand classification system; dynamic updating of the classification system through new word discovery and mining of highly relevant word pairs to adapt to the emerging demands of the power industry; specifically including: S410: Text feature mining: Based on a weekly frequency, the newly added speech transcription text is grouped into [2, 3, ..., k] tuples as candidate words. The word frequency (TF), average mutual information (AMI), and left and right adjacency of each candidate word are calculated, and finally new words that meet the threshold requirements are selected. S420: Mining related word pairs: Filtering words with a co-occurrence probability greater than 0.8 with the new word through co-occurrence matrix analysis; S430: Classification system expansion: automatically expand the final classification based on the semantics of the new word; S440: Expert review mechanism; newly expanded categories must be voted on and confirmed by 3-5 field experts, and the approval rate must be ≥ 70% before they can take effect; S450: Historical data backtracking; after updating the classification system, historical data is automatically reclassified and the knowledge base is updated.

3. A 95598 customer demand identification method according to claim 1, characterized in that: The S100 includes the following process: S110: Data collection: collect historical customer service voice data and work order texts to build a basic corpus; S120: Data cleaning: segmenting and noise reduction processing of the original voice data, filtering invalid characters and blank characters using regular expressions; S130: Using the Delphi method to organize power industry experts to conduct multiple rounds of anonymous feedback to gradually converge the scope of the vocabulary; S140: Label the appeal type, sentiment label, and risk points of each sample, and store the cleaned text in the "appeal type-scenario-example" format using JSON. Fields include classification label, text, sentiment value, and risk point.

4. A 95598 customer demand identification method according to claim 3, characterized in that: The S130 includes the following process: S131: First round of expert survey: a questionnaire containing a certain number of power industry terms was distributed to several experts, requiring them to screen core terms and add new terms; S132: Expert opinions are collected and words with a support rate below 60% are marked as "for discussion" and fed back to the experts for a second round of evaluation. S133: Experts score candidate terms on a scale of 1-5, retaining terms with a score of 4 or higher. After 3-5 rounds of iterations, a standardized vocabulary of 2,300 power industry terms is finally formed. At the same time, voting is conducted on controversial terms to finally determine the classification boundaries.

5. A 95598 customer demand identification method according to claim 1, characterized in that: The S200 includes the following process: S210: Customer perspective demand research: Collect typical demand cases through interviews with several electricity users; S220: Classification hierarchy design; construct a two-level classification system based on the Delphi vocabulary and user cases; S230: Invite five customer service experts to review the classification logic and correct overlapping or ambiguous classification boundaries; S240: Classification system verification: Invite interviewed customers to participate in the test, select a certain number of labeled samples, and test the coverage and accuracy of the classification system. The final classification coverage is required to reach 95%, and the misclassification rate is ≤5%. The classification accuracy of random appeal text is required to be ≥95%. If the user recognition rate of a certain category is lower than 80%, it will be renamed or merged into the parent category.

6. A 95598 customer demand identification method system, characterized in that: include: The appeal corpus unit is used to build the electric power industry vocabulary and appeal corpus; The classification system unit is used to build a hierarchical classification system with the customer perspective as the core, covering the full range of scenarios in the power industry; Appeal recognition unit, used for automatic recognition of speech to appeal; The self-improvement unit of the demand classification system dynamically updates the classification system through new word discovery and high-correlation word pair mining to adapt to the demands of the emerging power industry.

7. A 95598 customer demand identification method system according to claim 6, characterized in that: The appeal corpus unit specifically includes: The data collection module is used to collect historical customer service voice data and work order texts to build a basic corpus; The data cleaning module is used to segment and reduce noise of the original voice data, and filter invalid characters and empty characters through regular expressions; The corpus standardization and storage module is used to store the cleaned text in the "appeal type-scenario-example" format, using JSON format. Fields include classification labels, text, sentiment values, and risk points.

Images