Call collection customer communication style analysis system and method based on large model

Abstract

The invention discloses a call collection customer communication style analysis system and method based on a large model. According to the method, a multi-dimensional customer communication style analysis index system is constructed, collection call voice data is collected and preprocessed in real time, deep analysis is performed on the data by using a large model based on a Transform architecture, and a customer communication style quantitative score and portrait are accurately generated. Compared with the prior art, the method has obvious advantages in the aspects of accuracy, real-time performance and adaptability of customer communication style analysis. The system covers an index system construction module, a data acquisition and preprocessing module, a large model analysis module, a result storage and application module and the like, and all the modules work cooperatively and have a perfect exception handling mechanism. In addition to greatly improving the collection efficiency and success rate in the collection industry and improving the customer experience, the method also has wide potential application value in similar scenes such as customer service, sales communication and the like, and provides an innovative solution for making an intelligent communication strategy in related fields.

Description

Technical Field

[0001] The present invention relates to the cross - field of artificial intelligence and financial collection, and particularly to a system and method for analyzing the customer communication style in collection calls based on a large model, which is used to analyze the customer communication style in collection calls, generate quantitative scores and customer portraits, so as to optimize the collection communication strategy. Background Art

[0002] In the financial collection industry, phone collection is a widely used means. At present, the analysis of collection call content mainly focuses on the compliance level, such as detecting whether there are illegal words in the call to ensure that the collection process complies with laws, regulations and industry norms. However, these traditional analysis methods have many limitations. On the one hand, the efficiency is low when dealing with a large amount of data. As the volume of collection business increases, traditional methods are difficult to quickly and accurately process a large amount of call data. On the other hand, the accuracy of analyzing the customer communication style is insufficient. It mainly relies on simple keyword matching or limited rule judgment, and cannot deeply explore the real communication style of customers, resulting in collection staff lacking targeted communication strategies during communication with customers and low collection efficiency.

[0003] With the rapid development of natural language processing and speech recognition technologies, new opportunities and technical bases have been brought to the analysis of collection calls. For example, the wide application of the Transformer architecture in the field of natural language processing has greatly improved the ability of text understanding and generation; advanced speech recognition models can convert speech into text more accurately. However, there is currently no mature solution that effectively integrates these advanced technologies to comprehensively and deeply analyze the customer communication style in collection calls. Based on the development of these technologies, the present invention innovatively proposes a set of systems and methods for analyzing the customer communication style in collection calls based on a large model to solve the deficiencies of the prior art. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide a system and method for analyzing the customer communication style in collection calls based on a large model. By constructing a scientific index system and using the large model to deeply analyze the collection call voice, the communication style of customers can be accurately identified, and quantitative scores and customer portraits can be generated. This not only helps to improve the collection efficiency and success rate, but also enhances the customer experience during the collection process, reduces the complaint rate. At the same time, it provides strong support for the improvement of collection staff skills and the optimization of team management, and promotes the development of the collection industry towards intelligence and precision. Technical Solution

[0005] Construct an analysis index system: Construct an analysis index system for customer communication style from multiple dimensions.

[0006] 3.2.1 Taking the language expression dimension as an example, the vocabulary richness index is calculated by counting the ratio of the number of different words used by the customer to the total number of words, and a similarity algorithm based on the word vector model is used to identify different words. In the actual debt collection call scenario, if the customer uses a large number of diverse words during the call, it indicates that their language expression ability is strong, the vocabulary richness ratio may be relatively high, and the corresponding score will also be high. The sentence complexity is comprehensively evaluated based on the nesting level of the sentence structure, grammar difficulty, and the usage frequency of rare words, and a syntactic analyzer in natural language processing is used to analyze the sentence structure. For example, when the customer utters a sentence with a complex structure, containing multiple clauses and using rare words, the sentence complexity score will be high.

[0007] 3.2.2 In the emotional expression dimension, the emotional intensity is analyzed by using a pre-trained sentiment analysis model for the emotional words and intonation in the text, and weighted summation is used to measure it in combination with acoustic features such as the volume and speech rate of the voice. In a debt collection call, if the customer has a high volume, a fast speech rate, and frequently uses negative emotional words, the emotional intensity score will be high. The emotional stability is determined by calculating the fluctuation frequency and amplitude of the emotion during the call, and the sliding window algorithm is used to count the ratio of the number of emotion changes to the call duration. If the customer's emotion changes frequently during the call, the emotional stability score will be low.

[0008] 3.2.3 In the communication willingness dimension, the active speech frequency is obtained by counting the ratio of the number of times the customer actively initiates a topic to the total call duration; the response duration calculates the average duration of each response of the customer. In practical applications, customers with a high active speech frequency and a long response duration usually show a high communication willingness.

[0009] 3.2.4 In the attitude tendency dimension, the politeness level is scored according to the usage frequency of polite expressions and the analysis of their sincerity through semantic analysis, and a keyword matching combined with semantic understanding algorithm is used to identify polite expressions; the confrontation level is judged based on the usage frequency, intensity, and tone hardness of aggressive words, and an emotion dictionary and a machine learning classification model are used to identify aggressive words. For example, if the customer frequently uses polite expressions and has a sincere tone, the politeness level score is high; if the customer frequently uses aggressive words, the confrontation level score is high. Detailed quantitative calculation methods and scoring criteria are set for each index. For example, if the vocabulary richness ratio is higher than 70%, the score is 4 - 5 points; if it is 50% - 70%, the score is 2 - 3 points; if it is lower than 50%, the score is 1 point.

[0010] 3.3 Specific implementation part: 3.3.1 Voice data collection and preprocessing: When the collection call is connected, the system automatically activates the data collection and preprocessing module. The collection module is integrated with the phone system through a specific interface and has the function of adaptive sampling rate adjustment. When the network bandwidth is good and the call quality is stable, a higher sampling rate is adopted to obtain clearer voice data; when the network is poor, the sampling rate is automatically reduced to ensure the continuity of the data. The collected data is first denoised using the wavelet denoising algorithm to remove interference such as environmental noise and current noise, ensuring the data quality. Then, the hash algorithm is used for duplicate removal to avoid the impact of duplicate data on the analysis results. Next, a speech-to-text model based on the Transformer architecture is used to convert the speech into text format, and metadata such as call time, collector number, and customer number are marked. At the same time, the text data format is converted to JSON format for subsequent processing, and the AES encryption algorithm is used for encryption during data transmission and storage to ensure data security. Under different network environments and devices, the system dynamically adjusts the collection and processing strategies by real-time monitoring of the network status and device performance to ensure the quality and stability of data collection.

[0011] 3.3.2 Large Model Analysis and Scoring: A large model based on the Transformer architecture is pre-trained using a large amount of historical collection call voice data covering different customer types and collection scenarios. The training data is preprocessed and labeled, and the labeled content includes the customer's communication style type. During the training process, cross-entropy is used as the loss function, and the Adam optimizer is used to adjust the model parameters to minimize the difference between the prediction result and the label. When the accuracy of the model on the validation set reaches more than 90%, the model training is considered qualified. During the call, the large model analysis module obtains the preprocessed text data. According to the constructed index system, through the attention mechanism, it identifies and understands information such as vocabulary, sentence structure, and emotional tendency in the text, extracts features related to the customer's communication style, calculates the scores of each index, and synthesizes the scores of each index to generate a quantitative score and portrait of the customer's communication style. For example, the model judges whether the customer belongs to the friendly cooperation type, the negative perfunctory type, the emotional type, etc. by analyzing features such as vocabulary, sentence structure, and emotional expression.

[0012] 3.3.3 Result Storage and Application: The result storage and application module stores the quantified scores of the customer communication style, the corresponding call metadata, and the customer communication style portrait in JSON format in the database. Based on the customer communication style portrait, through a similarity matching algorithm with historical successful collection cases, targeted communication strategy suggestions are provided for the collection staff. For example, for friendly and cooperative customers, it is recommended to adopt a simple and clear communication method to quickly reach a repayment agreement; for negative and perfunctory customers, it is recommended to adjust the conversation skills to increase the interest and attraction of the communication; for emotionally excited customers, it is recommended to soothe the emotions first and then conduct rational communication. At the same time, the system provides data support for the collection team. Through the cluster analysis of a large number of customer communication style portraits, it assists in formulating the overall collection strategy and optimizing resource allocation. Through the data visualization function, data such as the distribution of customer communication styles and the comparison of collection effects are displayed in the form of charts, providing intuitive data support for team management.

[0013] 3.4 Beneficial effects: 3.4.1 Precise customer portrait: Through a multi-dimensional index system and in-depth analysis of the large model, the communication style of customers can be accurately identified, and a detailed customer portrait can be generated. In actual tests, compared with traditional analysis methods, the accuracy rate of the judgment of the customer communication style by the present invention has increased by 30%, providing comprehensive and accurate customer information for the collection staff.

[0014] 3.4.2 Improve collection efficiency: Targeted communication strategy suggestions are provided according to the customer communication style portrait, enabling the collection staff to adopt more appropriate communication methods. In the actual application of a certain collection team, after adopting the method of the present invention, the collection efficiency has increased by 40%, and the collection success rate has increased by 25%, significantly improving the collection effect.

[0015] 3.4.3 Enhance customer experience: Through targeted communication strategies, the resistance of customers during the collection process is reduced, and the complaint rate is lowered. According to statistics, the complaint rate has decreased by 50% compared with before, enhancing the customer experience.

[0016] 3.4.4 Data-driven decision-making: Data support is provided for the collection team. Through the analysis of customer communication style data, it assists in formulating the overall collection strategy and optimizing resource allocation. At the same time, the data visualization function facilitates the team manager to intuitively understand the business situation and promotes the scientific management of the collection business.

[0017] 3.4.5 Promote the improvement of personnel skills: The targeted communication strategy suggestions provided for the collection staff help the collection staff learn and master different communication skills, and improve their own communication ability and business level. Description of the drawings Figure 1 This is the system architecture diagram description: 4.1.1 Index System Construction Module: Responsible for constructing a multi-dimensional customer communication style analysis index system, including dimensions such as language expression, emotional expression, communication willingness, and attitude tendency.

[0019] 4.1.2 Data Collection and Preprocessing Module: Real-time collects call voice data and performs preprocessing operations such as noise reduction, duplicate removal, and speech-to-text conversion to ensure data quality and security.

[0020] 4.1.3 Large Model Analysis Module: Uses a pre-trained Transformer architecture large model to analyze the preprocessed text data, calculates the scores of each index, and generates a customer communication style portrait.

[0021] 4.1.4 Result Storage and Application Module: Stores the analysis results in the database and provides targeted communication strategy suggestions for debt collectors.

[0022] 4.1.5 Data Quality Management Module: Performs quality inspection and cleaning on the collected data to ensure data accuracy and integrity.

[0023] 4.1.6 Model Monitoring Module: Real-time monitors the performance indicators of the large model, such as accuracy and recall rate. When the performance indicators are below the threshold, it triggers the model optimization process.

[0024] Figure 2 This is the business process diagram description: 4.2.1 Start Debt Collection Call: When the debt collection call is connected, the system starts the data collection and preprocessing module.

[0025] 4.2.2 Data Collection and Preprocessing Module Starts: Real-time collects call voice data and performs preprocessing.

[0026] 4.2.3 Is Data Collection Successful: Checks whether the data collection is successful. If successful, continue with the large model analysis; if failed, switch to the backup network or re-collect the data.

[0027] 4.2.4 Large Model Analysis Module Obtains Preprocessed Data: The large model obtains the preprocessed text data and performs analysis 4.2.5 Is Model Analysis Successful: Checks whether the model analysis is successful. If successful, store the results and generate communication strategy suggestions; if failed, switch to the backup model or re-train.

[0028] 4.2.6 Result Storage and Application Module Stores the Results: Stores the analysis results in the database.

[0029] 4.2.7 Generate communication strategy suggestions: Based on the customer communication style profile, generate targeted communication strategy suggestions.

[0030] 4.2.8 Feedback the suggestions to the debt collectors: Feedback the communication strategy suggestions to the debt collectors to end the process.

Claims

1. A method for analyzing the communication style of customers in collection calls based on a large model, characterized in that: The following steps are involved: (1) Constructing an analysis index system: constructing a customer communication style analysis index system from four dimensions: language expression, emotional expression, communication willingness, and attitude tendency. The language expression dimension includes vocabulary richness and sentence complexity, the emotional expression dimension includes emotional intensity and emotional stability, the communication willingness dimension includes active speaking frequency and response time, and the attitude tendency dimension includes politeness and confrontation. (2) Voice data collection and preprocessing: collect call voice data in real time through the collection interface integrated with the telephone system, perform noise reduction and deduplication processing on the collected data in turn, convert it into text data using a speech-to-text model, and annotate call metadata. (3) Large model analysis and scoring: adopt a large model based on the Transformer architecture to analyze the preprocessed text data according to the analysis index system, extract language expression, emotional expression, communication willingness, and attitude tendency related features, calculate the scores of each indicator, and generate a quantitative score and profile of the customer communication style. (4) Result storage and application: store the quantitative score, call metadata, and customer profile in a database, and generate targeted communication strategy recommendations based on the customer profile.

2. The analysis method according to claim 1, characterized in that In the step of constructing the analysis indicator system, the principal component analysis (PCA) method is used to screen key features from the historical collection call data, and the experience of collection business experts is combined to determine the indicators of each dimension and their quantitative calculation methods.

3. The analysis method according to claim 1, characterized in that The vocabulary richness is calculated by counting the ratio of the number of different words used by the customer to the total number of words, and the sentence complexity is evaluated by analyzing the nesting level of the sentence structure, grammatical difficulty and the frequency of use of rare words.

4. The analysis method according to claim 1, characterized in that The emotion intensity is calculated by combining the emotional vocabulary analysis and acoustic features (volume, speaking speed) in the speech, and the emotion stability is calculated by using a sliding window algorithm to calculate the frequency and amplitude of emotion fluctuations during the call.

5. The analysis method according to claim 1, characterized in that The training process of the large model includes: collecting historical collection call voice data and corresponding customer communication style annotation information, preprocessing and annotating the data and inputting it into the model, and optimizing the model parameters through supervised learning to minimize the difference between the predicted results and the annotations.

6. The analysis method according to claim 1, characterized in that The targeted communication strategy recommendation is generated through a similarity matching algorithm between customer portraits and historical successful collection cases. The algorithm uses a distance measurement method to calculate the portrait similarity and recommend corresponding strategies.

7. A collection call customer communication style analysis system based on a large model, characterized in that: include: (1) Index system construction module: used to construct a customer communication style analysis index system including language expression, emotional expression, communication willingness, and attitude tendency dimensions. The module uses principal component analysis (PCA) to screen key features and determines indicators based on expert experience; (2) Data collection and preprocessing module: collects call voice data in real time through an integrated interface, performs noise reduction, deduplication, speech-to-text processing, and annotates metadata. The module includes an adaptive sampling rate adjustment unit and a data encryption unit; (3) Large model analysis module: uses a large model based on the Transformer architecture to extract features and score indicators for the preprocessed text data, and generates a quantitative score and profile of the customer's communication style. The module includes a feature extraction submodule and a score calculation submodule; (4) Result storage and application module: stores analysis results and generates targeted communication strategy recommendations based on customer profiles. The module includes a strategy matching unit and a data visualization unit.

8. The analysis system according to claim 7, characterized in that The data encryption unit uses the AES encryption algorithm to encrypt the data transmission and storage process.

9. The analysis system according to claim 7, characterized in that The strategy matching unit calculates the similarity between the customer portrait and the historical case through the Euclidean distance algorithm, and recommends corresponding strategies for adjusting the speech and controlling the communication rhythm.

Images