Client interaction data analysis method and device, equipment and medium

Abstract

The invention discloses a client interaction data analysis method and device, equipment and a medium. By automatically extracting semantic features and emotional features in customer interaction data, accurately analyzing customer appeals and emotional states, intelligently associating a target business entity with historical cases in combination with a preset domain business knowledge base, and generating associated features with business interpretability, risk analysis is deeply rooted to a professional knowledge system. And performing comprehensive intelligent research and judgment by fusing the multi-dimensional features, and automatically outputting a structured decision result containing a risk type, a risk level and a processing priority. Traditional low-efficiency and lagged risk research and judgment depending on manpower are converted into a real-time and accurate intelligent decision-making process. According to the invention, intelligent discovery and early warning of service risks are realized, and service personnel are assisted to make efficient responses according to clear priorities, so that the service efficiency and the service quality of online customer services are comprehensively improved, and the customer satisfaction is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of customer service technology, and in particular to a method, apparatus, device, and medium for analyzing customer interaction data. Background Technology

[0002] In online customer service, customer service personnel need to handle a variety of complex tasks such as business inquiries, repair requests, and complaints simultaneously. The integration of relevant information and the retrieval of knowledge mainly rely on manual work, resulting in limited service efficiency and difficulty in guaranteeing response time. At the same time, there is a general lack of ability to capture and structure customer emotional states in real time, making it difficult to have objective evidence for service quality assessment and to achieve dynamic, accurate, personalized service recommendations and business process optimization. Summary of the Invention

[0003] In view of this, the present invention provides a method, apparatus, electronic device and medium for analyzing customer interaction data, in order to solve the technical problem that existing online customer service relies on manual information integration and cannot effectively analyze customer emotions, resulting in low efficiency and difficulty in accurately optimizing service quality.

[0004] Firstly, a method for analyzing customer interaction data is provided, the method comprising: Acquire customer interaction data, which includes call logs and text records between customers and service personnel; Extract semantic and sentiment features from customer interaction data. Semantic features include customer intent and business type, while sentiment features include customer sentiment polarity and sentiment intensity. Based on a pre-set business knowledge base, the business association features corresponding to customer interaction data are determined. These business association features include the target business entity and target case type corresponding to the customer interaction data. Based on semantic features, sentiment features, and business relevance features, risk analysis is performed on customer interaction data to generate risk analysis results, which include risk type, risk level, and processing priority.

[0005] Secondly, a customer interaction data analysis device is provided, the device comprising: The acquisition module is used to acquire customer interaction data, which includes call records and text records between customers and service personnel. The feature extraction module is used to extract semantic and sentiment features from customer interaction data. Semantic features include customer intent and business type, while sentiment features include customer sentiment polarity and sentiment intensity. The determination module is used to determine the business association features corresponding to customer interaction data based on a preset business knowledge base. The business association features include the target business entity and target case type corresponding to the customer interaction data. The generation module is used to perform risk analysis on customer interaction data based on semantic features, sentiment features, and business relevance features, and generate risk analysis results, which include risk type, risk level, and processing priority.

[0006] Thirdly, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for analyzing customer interaction data.

[0007] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described method for analyzing customer interaction data.

[0008] The aforementioned solutions, implemented using methods, devices, electronic equipment, and storage media for analyzing customer interaction data, automatically extract semantic and emotional features from customer interaction data to accurately analyze customer needs and emotional states. By combining this with a pre-set domain business knowledge base, they intelligently link target business entities with historical cases, generating business-interpretive relational features that deeply embed risk analysis within a professional knowledge system. Furthermore, by integrating these multi-dimensional features for comprehensive intelligent judgment, they automatically output structured decision results containing risk type, risk level, and processing priority. This transforms traditional, inefficient, and delayed risk assessment methods that rely on manual intervention into a real-time, accurate intelligent decision-making process. This method enables intelligent discovery and early warning of service risks, assisting service personnel in responding efficiently based on clear priorities. It fundamentally solves the problems of ambiguous intentions in manual identification, missing emotional records, and strong subjectivity in analysis, thereby comprehensively improving the efficiency and quality of online customer service and effectively enhancing customer satisfaction. Attached Figure Description

[0009] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart illustrating a method for analyzing customer interaction data according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a customer interaction data analysis device according to an embodiment of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in the present invention are only for illustrative and descriptive purposes and are not intended to limit the scope of protection of the present invention.

[0011] Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this invention illustrate operations implemented according to some embodiments of the invention. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or performed simultaneously. Moreover, those skilled in the art, guided by the content of this invention, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0012] Furthermore, the embodiments described herein are merely some, not all, of the embodiments of the invention. The components of the embodiments of the invention described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0013] It should be noted that the term "comprising" will be used in the embodiments of the present invention to indicate the presence of a feature subsequently declared, but does not exclude the addition of other features. It should also be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0014] The following is a detailed description of this case, in conjunction with the relevant accompanying drawings in the instruction manual.

[0015] Please see Figure 1 This description and embodiment provide a method for analyzing customer interaction data, specifically including the following steps: S10: Obtain customer interaction data; Customer interaction data includes call logs and text records between customers and service personnel; It is understood that the executing entity of this invention can be a device for analyzing customer interaction data, a terminal, or a server; no specific limitation is made here. This embodiment of the invention will be described using a server as an example.

[0016] In this step, customer interaction data is acquired in real time during online customer service. This includes real-time call recordings between customers and service personnel, as well as various text records such as online consultation chat logs, electronic complaint forms, and repair work orders submitted by customers. These data collectively form the basis for multi-dimensional analysis of customer needs.

[0017] For example, in an electricity marketing customer service scenario, customer service call records are obtained through voice recording devices and stored as audio files, containing telephone conversations between customers and service personnel (such as electricity bill inquiries or equipment repair requests). Text interaction records are collected through online platforms or mobile applications, containing customer-submitted inquiry texts, complaint forms, or work order notes, and stored as structured text data. All data is anonymized to remove customer identification elements (such as names and phone numbers) and stored in an encrypted database to ensure compliance.

[0018] S20: Extract semantic and sentiment features from customer interaction data; Among them, semantic features include customer intent and business type, while sentiment features include customer sentiment polarity and sentiment intensity.

[0019] In this step, deep feature extraction is performed on the raw interaction data. Using natural language processing and sentiment computing technologies, unstructured dialogues and texts are transformed into quantifiable semantic and sentiment features. Semantic features are used to analyze the customer's fundamental purpose, identifying their intent (such as inquiries, complaints, repair requests, or suggestions) and the associated business type (such as abnormal electricity bills, service delays, or equipment malfunctions). Sentiment features are used to assess the customer's service experience and emotional state, specifically quantified as sentiment polarity classification (such as positive, neutral, negative) and sentiment intensity level (such as high, medium, low), thus simultaneously completing the analysis of the customer's problem content and emotional state.

[0020] In one embodiment of this application, a specific feature extraction scheme is provided. In S20, that is, extracting semantic features and sentiment features from customer interaction data, the specific steps include the following steps S21-S24: S21: Convert the call log into speech and generate the corresponding text data.

[0021] In this step, the real-time call data is processed by an automatic speech recognition engine, converting continuous speech signals into corresponding text transcriptions to generate the text data corresponding to the call record.

[0022] S22: Merge the text data and text records to generate a text sequence.

[0023] In this step, to construct a complete customer interaction context, the converted text data is merged with various text records such as online chat logs, complaint form texts, and work order descriptions submitted by the customer, and arranged according to time or logical relationship to form a complete text sequence, so as to integrate all the text information expressed by the customer across channels and time periods.

[0024] S23: By using a pre-trained multimodal large model, the text sequence is processed through word segmentation, word embedding, and attention mechanisms to determine customer intent and business type as semantic features.

[0025] In this step, the generated integrated text sequence is input into a pre-trained multimodal large model. The model's natural language processing module segments the sequence into meaningful lexical units. Then, word embedding technology is used to map each word into a high-dimensional dense vector to capture semantic information. Finally, the model's core attention mechanism is activated to dynamically weigh the importance of different words in the sequence for overall intent judgment, focusing on key information. Through the above processing, the model accurately determines the customer's intent for this interaction (such as "inquiry," "complaint," "repair request," or "suggestion") and its business type (such as "abnormal electricity bill," "service delay," or "equipment malfunction"), and outputs the judgment result as a structured semantic feature vector.

[0026] Optionally, the pre-trained multimodal large model is based on a pre-trained language model and fine-tuned using data from the power marketing field, integrating text and speech processing capabilities to meet the needs of intelligent question answering and risk assessment.

[0027] S24: Perform sentiment analysis on text sequences and call logs to generate the target sentiment polarity and target sentiment intensity of customers as sentiment features.

[0028] In this step, sentiment analysis is performed on the same integrated text sequence. By analyzing sentiment words in the text and combining these words with the tone, speed, and volume characteristics of the call log, the emotional tendency expressed by the customer during the interaction is quantitatively determined. The analysis results are structured and output in two dimensions: sentiment polarity, i.e., whether the emotion is positive, neutral, or negative; and sentiment intensity, i.e., the degree of intensity of the emotion, such as high, medium, or low. These two dimensions together constitute the sentiment feature vector of this interaction.

[0029] The above methods systematically transform raw, difficult-to-calculate call and text data into standardized feature representations that include clear intent, business affiliation, emotional tendency, and emotional intensity, providing clear and reliable input for subsequent intelligent association and risk assessment.

[0030] In one embodiment of this application, a specific emotional polarity and intensity analysis scheme is provided. In S24, that is, performing sentiment analysis on the text sequence and call records to generate the customer's target emotional polarity and target emotional intensity, the specific steps include the following S241-S244: S241: Based on a preset sentiment keyword library, obtain at least one target sentiment keyword included in the text sequence.

[0031] In this step, the text sequence is matched against a pre-built, preset sentiment keyword library. This library is extracted from historical data for specific business scenarios and contains various sentiment keywords with clear emotional tendencies, such as positive words like "satisfied" and "thank you"; negative words like "slow" and "failure"; and neutral words like "consultation" and "query". Based on the matching results, at least one target sentiment keyword is identified within the text sequence.

[0032] S242: Determine the speech features of each target emotional keyword in the call record by pre-training a multimodal large model.

[0033] In this step, the call log is input into a pre-trained multimodal large model. The intonation analysis module processes the continuous speech signal, first extracting objective acoustic parameters, mainly including speech rate (number of syllables per unit time) and volume (energy intensity of sound). These basic features can directly reflect the speaker's emotional state; for example, a rapid speech rate and a high volume are often associated with emotions such as excitement and anger. The intonation analysis module analyzes the speech segments when each target emotional keyword is uttered, outputting the specific speech features corresponding to them at the speech level. For example, if the speech segment corresponding to the target emotional keyword "dissatisfaction" is detected to have a significantly faster speech rate, a sudden increase in pitch, and an increase in volume, it can be determined that its speech features are characterized by "excitement," and this judgment is quantified as a high-intensity negative emotional signal.

[0034] The above methods enable the capture of more nuanced emotional signals from call recordings rich in audio information, thereby enhancing the accuracy and robustness of emotional feature extraction.

[0035] In practical applications, the model's temporal attention module analyzes the time-series characteristics of customer interaction data through a long short-term memory network to capture dynamic changes in emotion. For example, if a customer's tone changes from calm to agitated during a call, the temporal attention module generates a dynamic emotion trajectory, recording the change in emotion intensity from low to high. The module outputs a dynamic emotion vector, containing emotion intensity values ​​at each time step, which is stored in the customer relationship management system database. For instance, when a customer complains about an "incorrect meter reading," their tone may be neutral in the first half of the call and turn negative in the second half. The temporal attention module captures this dynamic change, supporting real-time service quality evaluation.

[0036] S243: Based on a preset mapping table and speech features, determine the emotional polarity and emotional intensity corresponding to each target emotional keyword; The preset mapping table includes the preset emotional polarity and preset emotional intensity corresponding to each preset emotional keyword.

[0037] In this step, the emotional polarity and intensity corresponding to all emotional keywords are pre-defined, and a preset mapping table is constructed. For example, the keyword "anger" is defined as "polarity: negative, intensity: high"; "a little slow" is defined as "polarity: negative, intensity: medium", and so on. When at least one target emotional keyword is matched in the text sequence, the preset emotional polarity and intensity corresponding to each target emotional keyword are determined by querying the mapping table.

[0038] Furthermore, by combining the phonological features of the target sentiment keyword with its preset sentiment polarity, for example, when the sentiment polarity corresponding to the target sentiment keyword "okay" is neutral, but the phonological features are "low-pitched and slow," it can be comprehensively determined that the customer may be saying one thing but meaning another, and their actual emotion is negative. Therefore, the sentiment polarity of the target sentiment keyword "okay" in this interaction was changed from "neutral" to "negative." This multimodal fusion of sentiment evidence greatly enhances the depth and reliability of judging the customer's true emotional state, providing a more solid basis for subsequent accurate risk level assessment.

[0039] S244: Based on the emotional polarity and emotional intensity corresponding to each target emotional keyword, generate the customer's target emotional polarity and target emotional intensity through preset priority rules.

[0040] In this step, during sentiment analysis, a single customer interaction text (i.e., the integrated text sequence) may contain multiple sentiment keywords. These words may point to different sentiment polarities (such as the simultaneous appearance of positive and negative words) or different sentiment intensities. To extract a deterministic feature vector representing the overall sentiment tendency of this interaction from complex and even contradictory sentiment signals, the system pre-sets priority rules for sentiment polarity and sentiment intensity. After confirming the sentiment polarity and intensity of all target sentiment keywords in this interaction, all matching results are comprehensively evaluated according to the pre-set priority rules to ultimately determine a unique target sentiment polarity and target sentiment intensity as the sentiment feature characterizing the overall sentiment state of this interaction, for use in subsequent risk assessment.

[0041] Optionally, the preset priority rules are: high-intensity emotion words take precedence over medium and low intensity words; negative polarity words take precedence over positive and neutral words. In most customer service scenarios, especially in sectors like electricity that involve people's livelihoods and safety, identifying and prioritizing negative emotions is a primary task. This is because negative emotions (such as dissatisfaction and anger) are directly associated with high-risk business consequences such as escalating complaints, customer churn, and safety incidents. Therefore, rules typically assign higher priority to "negative" polarity. Furthermore, emotion intensity (high, medium, low) directly reflects the intensity of the customer's emotions and the urgency of the problem. High-intensity emotions usually mean a more serious problem or that the customer's tolerance has reached its limit, requiring immediate intervention. Therefore, high-intensity emotions are prioritized over low-intensity emotions.

[0042] For example, a customer says, "The technician arrived too late (emotional keyword: 'too late' corresponds to 'negative, high intensity'), but the technical skills were still very professional (emotional keyword: 'professional' corresponds to 'positive, medium intensity'), and the problem is finally solved." Analysis yields two emotional signals: "negative, high" and "positive, medium." According to the preset priority rule, the high-intensity "negative, high" has a higher priority than the medium-intensity "positive, medium." Therefore, the target emotional polarity is negative, and the target emotional intensity is high. This indicates that although the customer acknowledges the technical skills, their core dissatisfaction with the service delay is stronger. Therefore, this interaction is judged as a high-intensity negative emotion, potentially triggering a high-priority service delay risk ticket, rather than a positive service praise record.

[0043] In practical applications, pre-set scores for sentiment polarity and sentiment intensity are used. Sentiment polarity scores are: positive 1, neutral 0, negative -1; sentiment intensity scores are: high 3, medium 2, low 1. When there are multiple target sentiment keywords in customer interactions, a weighted summation method can be used to calculate the sentiment polarity and sentiment intensity scores for all target sentiment keywords. The scores are then stored in the customer relationship management system, supporting visual display and assisting service personnel in real-time risk warnings and quality assessments.

[0044] S30: Based on a preset business knowledge base, determine the business association features corresponding to the customer interaction data; Among them, business association characteristics include the target business entity and target case type corresponding to customer interaction data.

[0045] In this step, to enhance the accuracy and business relevance of the analysis, the current customer interaction content is intelligently associated with the structured business entities, service processes, and historical cases stored in the knowledge base by querying a preset business knowledge base. This identifies the target business entity and target case type most relevant to the current interaction data and quantifies the degree of association as a business association feature. This ensures that the analysis is rooted in specific business scenarios and historical experience, rather than relying solely on general semantics.

[0046] For example, in the scenario of electricity marketing customer service, a business knowledge base is preset, which includes structured business entities (such as customer files and electricity meter equipment), service processes (such as electricity bill inquiry specifications and fault handling processes) and historical cases (such as electricity usage abnormality records and complaint cases) to support knowledge integration and identification of potential electricity usage hazards.

[0047] In one embodiment of this application, a specific scheme for determining business association features is provided. In S30, that is, based on a preset business knowledge base, the business association features corresponding to the customer interaction data are determined, which specifically includes the following steps S31-S33: S31: Business types based on customer interaction data, determine the target business entity in the preset business knowledge base.

[0048] In this step, the identified business type (such as abnormal electricity bills or service delays) is used as the query index to search a pre-defined business knowledge base. This knowledge base structures the company's core business objects, such as electricity meters, transformers, billing packages, and service processes. Through semantic matching, the target business entity most relevant to the current customer problem is identified from the knowledge base, thus anchoring the abstract customer problem to a specific business responsibility object.

[0049] For example, for the business type "electricity bill anomaly", it can be associated with the "electricity meter" entity through semantic matching; for "service delay", it can be associated with the "fault reporting process" entity through semantic matching.

[0050] S32: Obtain at least one historical case type for the target business entity.

[0051] In this step, after identifying the target business entity, at least one historical case type associated with that entity is further retrieved from a pre-defined business knowledge base. These case types are pre-summarized from the historical case database. For example, historical case types associated with the electricity meter entity include: electricity meter data freeze, sudden increase or decrease in electricity consumption, electricity meter offline and loss of connection, etc.; historical case types associated with the fault reporting process include: dispatch delay, unprocessed after timeout, multiple reminders from users, etc.

[0052] S33: Based on the matching degree between the text sequence and each historical case type, determine the target case type corresponding to the customer interaction data.

[0053] In this step, the detailed text sequence of the current customer interaction is compared with the descriptions of various historical case types to calculate the matching degree. Then, the case type with the highest matching degree is selected as the target case type corresponding to this customer interaction. For example, if a customer describes, "My electricity meter screen isn't displaying anything, but it seems to be still running," the matching calculation shows the highest matching degree with the description of the "electricity meter data frozen" event, thus identifying it as the target case type. Ultimately, the target business entity and the target case type together constitute key business association features, enabling the system not only to understand what the customer is saying but also what their problem means in terms of business and how it typically occurs historically. This achieves intelligent association from the original customer problem to structured business knowledge, providing a solid domain knowledge basis for subsequent comprehensive risk assessment and effectively avoiding misjudgments caused by a lack of professional knowledge in general models.

[0054] In one embodiment of this application, a specific target case type determination scheme is provided. In S33, the target case type corresponding to customer interaction data is determined based on the matching degree between the text sequence and each historical case type. This specifically includes the following steps S331-S333: S331: Obtain the description text for each historical case type.

[0055] In this step, all relevant historical case description texts are filtered and retrieved from a pre-defined business knowledge base. These description texts are not original case records, but standardized problem descriptions that have been summarized and abstracted, such as "electricity meter data freeze case: the screen shows no display but the pulse indicator light flashes, and the data collected in the background remains unchanged."

[0056] S332: Calculate the semantic similarity between the text sequence and each descriptive text using a semantic similarity model, and use this as the matching degree.

[0057] In this step, the text sequence of the current customer interaction (containing the customer's specific description of the problem, such as "My electricity meter screen is not displaying, but it seems to be running numbers") and the description text of each acquired historical case are compared with each other. The semantic similarity between the two is calculated using a semantic similarity model (such as a sentence vector model based on Sentence-BERT). The similarity score is the matching score of this comparison.

[0058] In practical applications, semantic similarity is calculated using the cosine similarity algorithm. The formula is the cosine of the angle between two vectors; a higher value indicates a higher similarity. For example, the vector calculation for the customer interaction data "electricity meter reading error" and the knowledge base case "electricity meter malfunction" yields a high similarity value (e.g., 0.85), indicating a high correlation between the two. Furthermore, the calculation process can incorporate specific weights from the electricity marketing field (e.g., keywords like "electricity meter" and "malfunction" have higher weights) to make the results more relevant to the business scenario.

[0059] S333: Identify the historical case type with the highest matching degree as the target case type corresponding to the customer interaction data.

[0060] In this step, after all matching scores are calculated, all scores are compared and sorted, and the historical case type with the highest matching score is determined as the target case type corresponding to the current customer interaction data. For example, if the current text sequence has a matching score of 0.92 with the "electricity meter data frozen" case and a matching score of 0.45 with the "electricity meter completely damaged" case, then "electricity meter data frozen" is automatically selected as the target case type.

[0061] The above method enables intelligent mapping from specific, conversational customer descriptions to standardized, structured case types. The output target case type serves as a key basis for subsequent risk level assessments, processing priority rankings, and solution recommendations, allowing the system to provide professional and accurate judgments on current new problems based on historical experience.

[0062] In practical applications, to improve the accuracy of risk entity discovery, after determining the target case type corresponding to customer interactions, sentiment features, semantic features, and association features are input into a graph structure. A graph neural network (GNN) is used to calculate the association strength value between risk entities. The association strength is calculated through weighted summation, with sentiment features and semantic features assigned different weights (e.g., sentiment weight 0.4, semantic weight 0.6). The association strength value reflects the relevance between customer interaction data and knowledge base entities in the business scenario. For example, a strength value of 0.9 indicates high association, and 0.3 indicates low association; this strength value quantifies the degree of association between the two. The association strength value is converted into an association feature vector through a feature encoding module. This module uses a fully connected layer to map the association strength value to a fixed-dimensional vector space, forming a high-dimensional association feature vector. The encoding process, combined with the business characteristics of the power marketing scenario, normalizes the association strength value to ensure that the vector value is within the range [0, 1]. For example, the correlation strength value of 0.9 between customer interaction data and the "electricity meter malfunction" case is encoded into a vector containing multiple dimensions (such as malfunction type, customer type, and service process), with the dimension values ​​reflecting the weights of different business attributes. The encoded correlation feature vector is stored in the system as input for subsequent risk entity classification. For example, the correlation feature vector can be represented as [0.9, 0.7, 0.3], corresponding to malfunction correlation, customer history correlation, and service process correlation, respectively, to support the accuracy of risk entity discovery.

[0063] S40: Based on semantic features, sentiment features, and business-related features, perform risk analysis on customer interaction data and generate risk analysis results; The risk analysis results include risk type, risk level, and handling priority.

[0064] In this step, semantic features, sentiment features, and business-related features are integrated to conduct a comprehensive analysis of customer interaction data, outputting structured risk results, including risk type (such as customer complaint risk, service anomaly risk, or business dispute risk), risk level (such as high, medium, and low), and processing priority. This guides service personnel in arranging the processing order, thereby quickly focusing on core risks and implementing differentiated responses based on level and priority. This enables early detection, early assessment, and early handling of potential problems, comprehensively improving service efficiency and risk management capabilities.

[0065] In one embodiment of this application, a specific risk analysis scheme is provided. In S40, risk analysis is performed on customer interaction data based on semantic features, sentiment features, and business relevance features to generate risk analysis results. Specifically, this includes the following steps S41-S45: S41: Input semantic features, sentiment features, and business-related features into a pre-trained classifier to generate risk types from customer interaction data.

[0066] In this step, three sets of feature vectors—semantic features, sentiment features, and business-related features—are concatenated and input into a pre-trained classifier specifically designed for risk classification. The network's core attention mechanism dynamically assigns importance weights to different features. For example, for an interaction with a semantic element of complaint, a highly negative sentiment element, and a connection to historical meter malfunctions, the classifier will focus more on the strong risk signals of complaint intent and high-intensity negative sentiment. By performing nonlinear transformations and classifications on the weighted features, the risk type of the customer interaction data is ultimately determined, such as categorizing it as customer complaint risk, service anomaly risk, or business dispute risk, thus completing the qualitative assessment of the risk.

[0067] Optionally, a pre-trained classifier is trained on labeled data (including labels for customer complaints, service anomalies, business disputes, etc.) to identify risk types such as "customer complaint risk" and "service anomaly risk." This classifier employs a multilayer perceptron architecture, trained on labeled data from the electricity marketing domain (e.g., labels for complaints, service anomalies, disputes, etc.). It calculates the probability distribution of the comprehensive risk feature vector across each risk category using fully connected layers and a softmax function, outputting the risk entity category. For example, if customer interaction data includes "electricity bill error" with a high negative sentiment, the classifier might output a probability of 0.8 for "customer complaint risk," 0.15 for "service anomaly risk," and 0.05 for "business dispute risk," selecting the highest probability "customer complaint risk" as the risk entity category. The classifier supports dynamic adjustment, and its performance can be optimized by periodically updating training data (e.g., adding new complaint cases) to adapt to the diverse risks in electricity marketing scenarios.

[0068] In practical applications, the system presets a probability threshold (e.g., 0.6) to determine the confidence level of the classification results. If the highest probability value is lower than this threshold, for example, if the risk type probability distribution is [0.4, 0.3, 0.2] (highest probability 0.4 < 0.6), a clear risk entity type cannot be determined, and the interaction record is marked as a "suspected risk." Suspected risk records are pushed to customer service personnel's terminals through the CRM system, generating a review task containing a summary of the interaction data (e.g., text content, call logs) to prompt human intervention. For example, if a customer mentions "the electricity bill may be incorrect" but the risk classification probability distribution is low, the system marks it as a suspected risk and pushes it for manual analysis to ensure the accuracy and reliability of risk identification.

[0069] S42: Determine the risk level of customer interaction data based on emotional intensity, target case type, and correlation strength value.

[0070] In this step, after determining the risk type, its severity needs to be further assessed. This is calculated based on the quantified emotional intensity, the business entity type to which the target case type belongs, and the correlation strength between customer interaction data and the target case type. For example, an interaction with high emotional intensity and a correlation strength value greater than 0.8 with the target case type that triggered the security complaint would be classified as high-risk.

[0071] S43: Using a multi-scale feature analyzer, semantic features, sentiment features, and business relevance features are decomposed into first-scale features and second-scale features.

[0072] S44: Calculate the priority score based on the first-scale feature and the second-scale feature.

[0073] For steps S43-S44, the multi-scale feature analyzer employs a convolutional neural network architecture, decomposing the comprehensive risk feature vector through convolutional kernels of different scales to support intelligent question answering, risk assessment, and electricity hazard identification. Semantic features, sentiment features, and business-related features are fused using a multi-head attention mechanism to generate comprehensive risk features. These comprehensive risk features are then input into the multi-scale feature analyzer for decomposition, yielding first-scale features and second-scale features. The first-scale features use smaller convolutional kernels (e.g., 3×3) to extract local features, capturing detailed information from the interactive data, such as the keyword "abnormal electricity bill" or changes in sentiment intensity, helping to accurately identify customer complaint tendencies. The second-scale features use larger convolutional kernels (e.g., 5×5) to extract global features, reflecting the overall correlation between the interactive data and the knowledge base (e.g., electricity anomaly cases), supporting the correlation analysis of electricity hazard. For example, if a customer mentions "meter reading error" in a heated tone during a call, the first-scale feature captures the keyword "error" and the agitated tone, while the second-scale feature integrates the complaint intent with the electricity meter malfunction case.

[0074] Subsequently, the importance score S1 of the first-scale feature is obtained by calculating the L2 norm of the first-scale feature vector, reflecting the saliency of local features (such as the strength of complaint keywords); the importance score S2 of the second-scale feature is calculated by calculating the L2 norm of the second-scale feature vector, reflecting the saliency of global features (such as the strength of association with the service process). Substituting S1 and S2 into the priority scoring formula, the priority score is calculated.

[0075] The priority score calculation formula is as follows:

[0076] Where P represents the priority score; The importance score for the first-scale feature; E represents the importance score of the second-scale feature; E is the sentiment intensity value, extracted from the sentiment features, used to quantify the sentiment level of customer interaction (e.g., high intensity is 3, medium intensity is 2, low intensity is 1) to support fine-grained sentiment recognition; α, β, and γ are preset weight coefficients, set to 0.4, 0.3, and 0.3 based on the needs of the power marketing scenario, to balance local features, global features, and sentiment intensity, and adapt to the needs of rapid processing in multiple scenarios.

[0077] For example, for customer interaction data "abnormal electricity bill and agitated tone", assuming S1=0.8, S2=0.6, E=3, the priority score P is calculated using the formula P=0.4×0.8+0.3×0.6+0.3×3=1.4.

[0078] S45: Determine the processing priority based on priority scores and multiple preset processing priority thresholds.

[0079] In this step, the preset processing priority thresholds are set through the rule engine in the customer relationship management system, determined based on business needs and historical data analysis. For example, three thresholds can be set: T1=0.5, T2=0.8, and T3=1.2, corresponding to low, medium, and high processing priorities, respectively. The calculated priority score P is compared with the thresholds T1, T2, and T3. For example, a customer complaint about "incorrect meter reading" has a priority score of P=1.4. After comparison, P>T3 (1.4>1.2), thus confirming this interaction as a high-priority issue, allowing service personnel to handle high-risk problems first.

[0080] In one embodiment of this application, a specific risk warning scheme is provided, which involves performing risk analysis on customer interaction data based on semantic features, sentiment features, and business relevance features. After generating the risk analysis results, the scheme further includes the following steps: Based on the risk analysis results and business-related characteristics, risk warning information is generated; Risk warning information will be sent to the service personnel's terminals.

[0081] In this embodiment, risk warning information is generated in a structured format based on risk analysis results and business-related characteristics, and sent in real time to service personnel's terminals (such as mobile applications or desktop clients) through the push interface of the customer relationship management system. The risk warning information is displayed through a visual interface, supporting pop-ups and sound prompts to ensure rapid response. Furthermore, the risk warning information may also include suggested handling measures generated based on historical cases from a preset business knowledge base. For example, for complaints about "abnormal electricity bills," a suggestion to "check the meter records and contact technical support" could be included. This facilitates rapid response by service personnel to urgent risks and meets the needs of intelligent service and knowledge recommendation.

[0082] As can be seen, the above solution, by automatically extracting semantic and emotional features from customer interaction data, accurately analyzes customer needs and emotional states. Combined with a pre-set domain business knowledge base, it intelligently associates target business entities with historical cases, generating business-interpretive relational features that deeply embed risk analysis within a professional knowledge system. Furthermore, by integrating these multi-dimensional features for comprehensive intelligent judgment, it automatically outputs structured decision results containing risk type, risk level, and processing priority, transforming traditional, inefficient, and delayed risk assessment relying on manual intervention into a real-time, accurate intelligent decision-making process. This method achieves intelligent discovery and early warning of service risks, assisting service personnel in responding efficiently based on clear priorities. It fundamentally solves the problems of ambiguous intentions in manual identification, missing emotional records, and strong subjectivity in analysis, thereby comprehensively improving the service efficiency and quality of online customer service and effectively enhancing customer satisfaction.

[0083] In one embodiment, a customer interaction data analysis device is provided, which corresponds one-to-one with the customer interaction data analysis method described in the above embodiments. For example... Figure 2 As shown, the customer interaction data analysis device 100 includes: an acquisition module 101, a feature extraction module 102, a determination module 103, and a generation module 104. Detailed descriptions of each functional module are as follows: The acquisition module 101 is used to acquire customer interaction data, which includes call records and text records between customers and service personnel. The feature extraction module 102 is used to extract semantic features and sentiment features from customer interaction data. The semantic features include customer intent and business type, and the sentiment features include customer sentiment polarity and sentiment intensity. The determination module 103 is used to determine the business association features corresponding to customer interaction data based on a preset business knowledge base. The business association features include the target business entity and target case type corresponding to the customer interaction data. The generation module 104 is used to perform risk analysis on customer interaction data based on semantic features, sentiment features and business relevance features, and generate risk analysis results, which include risk type, risk level and processing priority.

[0084] In one embodiment, the feature extraction module 102 is specifically used for: The call log is converted into speech, generating corresponding text data. The text data and text records are merged to generate a text sequence; By using a pre-trained multimodal large model to process text sequences through word segmentation, word embedding, and attention mechanisms, customer intent and business type can be determined as semantic features. Sentiment analysis is performed on text sequences and call logs to generate the target sentiment polarity and target sentiment intensity of customers as sentiment features.

[0085] In one embodiment, the feature extraction module 102 is further configured to: Based on a pre-defined sentiment keyword library, at least one target sentiment keyword is obtained from the text sequence. By pre-training a multimodal large model, the speech features of each target emotional keyword in the call record are determined; Based on a preset mapping table and speech features, the emotional polarity and emotional intensity corresponding to each target emotional keyword are determined. The preset mapping table includes the preset emotional polarity and preset emotional intensity corresponding to each preset emotional keyword. Based on the emotional polarity and intensity corresponding to each target emotional keyword, the target emotional polarity and intensity of the customer are generated through preset priority rules.

[0086] In one embodiment, the determining module 103 is specifically used for: Based on the business type of customer interaction data, the target business entity is determined in the preset business knowledge base through semantic matching. Obtain at least one historical case type for the target business entity; Based on the matching degree between the text sequence and each historical case type, the target case type corresponding to the customer interaction data is determined.

[0087] In one embodiment, the determining module 103 is further configured to: Retrieve the descriptive text for each historical case type; The semantic similarity model is used to calculate the semantic similarity between the text sequence and each descriptive text, which is then used as the matching degree. The historical case type with the highest matching degree is identified as the target case type corresponding to the customer interaction data.

[0088] In one embodiment, the generation module 104 is specifically used for: Input semantic features, sentiment features, and business-related features into a pre-trained classifier to generate risk types from customer interaction data; The risk level of customer interaction data is determined based on the correlation strength values ​​of emotional intensity and target case type. The semantic features, sentiment features, and business relevance features are decomposed into first-scale features and second-scale features by using a multi-scale feature analyzer. Calculate priority scores based on first-scale features and second-scale features; Processing priorities are determined based on priority scores and multiple preset processing priority thresholds.

[0089] In one embodiment, the generation module 104 is further configured to generate risk warning information based on the risk analysis results and business-related characteristics.

[0090] In one embodiment, the device further includes: The sending module is used to send risk warning information to the service personnel's terminal.

[0091] This invention provides a customer interaction data analysis device 100. By automatically extracting semantic and emotional features from customer interaction data, it accurately analyzes customer needs and emotional states. Combined with a pre-set domain business knowledge base, it intelligently associates target business entities with historical cases, generating business-interpretive relational features that deeply embed risk analysis within a professional knowledge system. Furthermore, it integrates these multi-dimensional features for comprehensive intelligent judgment, automatically outputting structured decision results including risk type, risk level, and processing priority. This transforms traditional, inefficient, and delayed risk assessment, which relies on manual intervention, into a real-time, accurate intelligent decision-making process. This method enables intelligent discovery and early warning of service risks, assisting service personnel in responding efficiently based on clear priorities. It fundamentally solves the problems of ambiguous intentions in manual identification, missing emotional records, and strong subjectivity in analysis, thereby comprehensively improving the service efficiency and quality of online customer service and effectively enhancing customer satisfaction.

[0092] Specific limitations regarding the customer interaction data analysis device can be found in the limitations of the customer interaction data analysis method described above, and will not be repeated here. Each module in the aforementioned customer interaction data analysis 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 the electronic device, or stored in the memory of the electronic device as software, so that the processor can call and execute the corresponding operations of each module.

[0093] In one embodiment, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned method for analyzing customer interaction data.

[0094] In one embodiment, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the aforementioned method for analyzing customer interaction data.

[0095] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or electronic device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0096] Those skilled in the art will understand that all or part of the processes in the methods of 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, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0097] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0098] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for analyzing customer interaction data, characterized in that, include: Acquire customer interaction data, wherein the customer interaction data includes call records and text records between customers and service personnel; Extract semantic and sentiment features from the customer interaction data, wherein the semantic features include customer intent and business type, and the sentiment features include customer sentiment polarity and sentiment intensity; Based on a pre-set business knowledge base, the business association features corresponding to the customer interaction data are determined, wherein the business association features include the target business entity and the target case type corresponding to the customer interaction data; Based on the semantic features, the sentiment features, and the business-related features, risk analysis is performed on the customer interaction data to generate risk analysis results, wherein the risk analysis results include risk type, risk level, and processing priority.

2. The method for analyzing customer interaction data according to claim 1, characterized in that, The step of extracting semantic and sentiment features from the customer interaction data specifically includes: The call log is converted into speech to generate text data corresponding to the call log; The text data and the text records are merged to generate a text sequence; The text sequence is processed by word segmentation, word embedding, and attention mechanism through a pre-trained multimodal large model to determine customer intent and business type as semantic features; Sentiment analysis is performed on the text sequence and the call records to generate the customer's target sentiment polarity and target sentiment intensity as sentiment features.

3. The method for analyzing customer interaction data according to claim 2, characterized in that, The step of performing sentiment analysis on the text sequence and the call record to generate the customer's target sentiment polarity and target sentiment intensity specifically includes: Based on a preset sentiment keyword library, at least one target sentiment keyword is obtained from the text sequence; By pre-training a multimodal large model, the speech features of each target emotional keyword in the call record are determined; Based on the preset mapping table and the speech features, the emotional polarity and emotional intensity corresponding to each target emotional keyword are determined. The preset mapping table includes the preset emotional polarity and preset emotional intensity corresponding to each preset emotional keyword. Based on the emotional polarity and intensity corresponding to each target emotional keyword, the target emotional polarity and intensity of the customer are generated through a preset priority rule.

4. The method for analyzing customer interaction data according to claim 1, characterized in that, The step of determining the business association features corresponding to the customer interaction data based on a preset business knowledge base specifically includes: Based on the business type of the customer interaction data, the target business entity is determined in the preset business knowledge base through semantic matching. Obtain at least one historical case type for the target business entity; Based on the matching degree between the text sequence and each historical case type, the target case type corresponding to the customer interaction data is determined.

5. The method for analyzing customer interaction data according to claim 4, characterized in that, The step of determining the target case type corresponding to the customer interaction data based on the matching degree between the text sequence and each historical case type specifically includes: Retrieve the descriptive text for each historical case type; The semantic similarity between the text sequence and each descriptive text is calculated using a semantic similarity model, and this is used as the matching degree. The historical case type with the highest matching degree is determined as the target case type corresponding to the customer interaction data.

6. The method for analyzing customer interaction data according to claim 1, characterized in that, The step of performing risk analysis on the customer interaction data based on the semantic features, the sentiment features, and the business relevance features, and generating risk analysis results, specifically includes: The semantic features, the sentiment features, and the business-related features are input into a pre-trained classifier to generate the risk type of the customer interaction data; The risk level of the customer interaction data is determined based on the correlation strength values ​​of emotional intensity and target case type. The semantic features, the sentiment features, and the business-related features are decomposed into first-scale features and second-scale features using a multi-scale feature analyzer. Calculate the priority score based on the first scale feature and the second scale feature; The processing priority is determined based on the priority score and multiple preset processing priority thresholds.

7. The method for analyzing customer interaction data according to claim 1, characterized in that, After performing risk analysis on the customer interaction data based on the semantic features, the sentiment features, and the business relevance features, the method further includes: Based on the risk analysis results and business-related characteristics, a risk warning message is generated; The risk warning information is sent to the service personnel's terminal.

8. A device for analyzing customer interaction data, characterized in that, include: The acquisition module is used to acquire customer interaction data, wherein the customer interaction data includes call records and text records between customers and service personnel; The feature extraction module is used to extract semantic features and sentiment features from the customer interaction data. The semantic features include customer intent and business type, and the sentiment features include customer sentiment polarity and sentiment intensity. The determination module is used to determine the business association features corresponding to the customer interaction data based on a preset business knowledge base, wherein the business association features include the target business entity and the target case type corresponding to the customer interaction data; The generation module is used to perform risk analysis on the customer interaction data based on the semantic features, the sentiment features, and the business association features, and generate risk analysis results, wherein the risk analysis results include risk type, risk level, and processing priority.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for analyzing customer interaction data as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method for analyzing customer interaction data as described in any one of claims 1 to 7.

Images