Intelligent quality inspection system and method based on multi-dimensional data analysis

Abstract

The invention relates to the technical field of intelligent telephone traffic, in particular to an intelligent quality inspection system and method based on multi-dimensional data analysis. The system comprises a data acquisition module used for acquiring a voice media stream in a voice call process; the voice recognition module is used for converting the received voice media stream into text data; the semantic understanding module is used for performing semantic analysis on the text data and extracting key information; the sentiment analysis module is used for performing sentiment analysis on the text data and judging the service quality of telephone traffic personnel; the quality inspection rule management module is used for customizing and editing quality inspection rules, forming a personalized quality inspection model and storing the personalized quality inspection model in a database; the quality inspection execution module is used for matching the semantic analysis result and the sentiment analysis result with a preset quality inspection rule to generate a quality inspection result; and the intelligent training module is used for generating personalized training courses according to the quality inspection result. According to the technical scheme, the telephone traffic quality inspection efficiency and accuracy can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent call services, and particularly relates to an intelligent quality inspection system and method based on multi-dimensional data analysis. Background Art

[0002] As one of the main means of communicating with customers, the quality of voice calls is directly related to the service image of an enterprise and customer satisfaction. Traditional voice call quality inspection methods mainly rely on manual sampling inspection, that is, quality inspection personnel randomly select a part of call recordings to listen to and evaluate the call quality according to preset quality inspection standards. However, this method has many deficiencies.

[0003] First of all, the efficiency of manual quality inspection is low. With the growth of enterprise business, the volume of customer service calls has increased sharply. Manual quality inspection requires a large amount of time and effort to listen to the content of each call, which is not only time-consuming and laborious, but also difficult to complete a large number of quality inspection tasks in a short time. In addition, manual quality inspection is also easily affected by the subjective factors of quality inspection personnel. Different quality inspection personnel may have different evaluations of the quality of the same call, resulting in inconsistencies in quality inspection results.

[0004] Secondly, the extraction of key information in manual quality inspection may not be accurate enough. In voice calls, key information is often hidden in a large amount of conversation content. Manual quality inspection personnel need to listen carefully and record, which is not only cumbersome, but also prone to omitting key information due to fatigue or negligence. At the same time, the evaluation of the service quality of call operators in manual quality inspection also often depends on the experience and subjective judgment of quality inspection personnel, and it is difficult to achieve objective and accurate evaluation. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent quality inspection system and method based on multi-dimensional data analysis, which can improve the efficiency and accuracy of call quality inspection.

[0006] To achieve the above purpose, in a first aspect, the present invention provides an intelligent quality inspection system based on multi-dimensional data analysis, including:

[0007] A data collection module, configured to collect the voice media stream generated in real time during a voice call and transmit it to the voice recognition module;

[0008] A voice recognition module, configured to convert the received voice media stream into corresponding text data;

[0009] A semantic understanding module, configured to perform semantic analysis on the text data and extract key information;

[0010] An emotion analysis module, configured to perform emotion analysis on the text data and judge the service quality of call operators;

[0011] The quality inspection rule management module is used to customize and edit quality inspection rules, form a personalized quality inspection model, and store the edited quality inspection rules in the database;

[0012] The quality inspection execution module is used to match the semantic analysis result and the sentiment analysis result with the preset quality inspection rules to generate a quality inspection result;

[0013] The intelligent training module is used to generate personalized training courses according to the quality inspection results.

[0014] Beneficial effects of the basic solution: By collecting the voice media stream during the voice call in real time and quickly converting it into text data, the system can perform semantic analysis and sentiment analysis immediately, thus greatly improving the timeliness of quality inspection. The quality inspection execution module accurately matches the analysis result with the preset quality inspection rules, reducing the subjectivity and error of human judgment and improving the accuracy of quality inspection.

[0015] Through multi-dimensional data analysis, including text semantic analysis and sentiment analysis after voice recognition, the service quality of operators can be evaluated more comprehensively. Compared with the previous method of relying solely on manual sampling of call records for quality inspection, it may lead to omission of key information or misjudgment of service quality due to manual fatigue, subjective factors, etc. The semantic understanding module of this system can accurately extract key information, and the sentiment analysis module can quantify the emotional state of operators. By integrating this information to judge the service quality, the accuracy of quality inspection is improved.

[0016] The data collection module can collect the voice media stream in real time, and the whole system has a high degree of automation. Compared with the traditional manual quality inspection method, manual quality inspection takes a lot of time to listen to call content, record information, etc. This intelligent quality inspection system can quickly convert voice into text, perform semantic and sentiment analysis, and match with quality inspection rules, greatly shortening the quality inspection cycle, being able to process a large number of call quality inspection tasks in a short time, and improving the overall quality inspection efficiency.

[0017] The quality inspection rule management module allows users to customize and edit quality inspection rules to form a personalized quality inspection model. Different enterprises and different business departments may have different requirements for the service quality of operators. They can flexibly formulate quality inspection rules suitable for themselves according to their own business characteristics and quality requirements to better meet the service requirements of different enterprises, improving the practicability and applicability of the system.

[0018] The intelligent training module generates personalized training courses based on the quality inspection results and the skill levels of the call center agents. This helps to provide targeted training for each agent's specific problems and improve the effectiveness of training. For example, for agents who are lacking in emotional communication skills, training courses on emotion management and communication skills can be arranged; for agents who have an incomplete understanding of business knowledge, focused training on business knowledge reinforcement can be provided, thereby enhancing the overall service level of the call center agents.

[0019] As an implementable preferred solution, the data acquisition module uses a linear microphone array and, through beamforming technology, focuses the acquisition direction on the location of the call.

[0020] As an implementable preferred solution, the speech recognition module is used to perform denoising, filtering, and enhancement operations on the collected voice media stream. The denoising operation uses a multi-layer CNN model for denoising; the filtering operation uses a band-pass filter to process the voice media stream; the enhancement operation performs frame segmentation on the voice media stream, conducts STFT transformation on each frame to convert the time-domain signal into a frequency-domain signal, enhances the spectrum of the voice signal in the frequency domain, and then converts the enhanced frequency-domain signal back into a time-domain signal through inverse STFT transformation.

[0021] As an implementable preferred solution, the semantic understanding module uses a conditional random field model to annotate the text to construct a part-of-speech tagging corpus, and trains using the corpus conditional random field model. The model learns the relationship between each word and its part of speech in the text, and inputs the text data output by the speech recognition module into the trained model. The model tags each word with the corresponding part of speech.

[0022] As an implementable preferred solution, the sentiment analysis module constructs a sentiment dictionary, including positive words, negative words, and neutral words, and judges the sentiment tendency of the text data based on the sentiment dictionary to obtain the service attitude and service quality scores of the call center agents.

[0023] As an implementable preferred solution, the intelligent training module determines the training needs of the call center agents based on the violation items and / or reasons for low scores, selects appropriate course content from the training resource library, and constructs an online learning platform.

[0024] As an implementable preferred solution, it further includes a multi-modal data acquisition module and a multi-modal feature fusion module. The multi-modal data acquisition module is used to track the facial area of the operator in real time, extract expression features, and capture the operation logs of the operator; the multi-modal feature fusion module is used to input the speech emotion score, expression emotion score, and operation efficiency score into the attention mechanism network, dynamically allocate the weights of each modality, and vote on the multi-modal analysis results through a random forest model to output the comprehensive service quality score.

[0025] As an implementable preferred solution, the quality inspection rule management module includes a rule evolution sub-module and a priority scheduling sub-module. The rule evolution sub-module encodes the quality inspection rule parameters into vectors, combines the context features, and guides the policy network to generate the target quality inspection rule through a reward function; the priority scheduling sub-module is used to dynamically adjust the rule execution order according to the real-time traffic volume and service type.

[0026] In a second aspect, the present invention also provides an intelligent quality inspection method based on multi-dimensional data analysis, which uses the above-mentioned intelligent quality inspection system based on multi-dimensional data analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of the architecture of an intelligent quality inspection system based on multi-dimensional data analysis.

[0028] Figure 2 It is a schematic diagram of the electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0029] To make the technical solutions and their advantages of the present application clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings. It can be understood that the specific embodiments described herein are only partial embodiments of the present invention, which are only used to explain the present application and are not intended to limit the present application. It should be noted that the technical features or combinations of technical features described in the following embodiments should not be considered as isolated, and they can be combined with each other to achieve better technical effects. The same reference numerals in the accompanying drawings of the following embodiments represent the same features or components, which can be applied to different embodiments.

[0030] In addition, unless otherwise defined, the technical terms or scientific terms used in the description of the present invention should have the ordinary meaning understood by those of ordinary skill in the art to which the present invention belongs.

[0031] The present invention will be further described in detail below with reference to the accompanying drawings:

[0032] Reference numerals: Electronic device 500, Processor 501, Communication interface 502, Memory 503, Bus 504.

[0033] Example 1

[0034] Refer to Figure 1 , an intelligent quality inspection system based on multi-dimensional data analysis provided by an embodiment of the present disclosure includes a data acquisition module, a speech recognition module, a semantic understanding module, an emotion analysis module, a quality inspection rule management module, a quality inspection execution module, and an intelligent training module.

[0035] The data acquisition module is used to collect the voice media stream generated in real time during a voice call and transmit it to the speech recognition module. A linear microphone array is adopted, and through beamforming technology, the collection direction is focused on the position of the call to reduce the interference of ambient noise.

[0036] At the beginning of a voice call, the data acquisition module is immediately started. By docking with the interface of the call system, the voice media stream generated during the voice call is obtained in real time. For example, for a call system based on an IP network, the data acquisition module uses protocols such as SIP (Session Initiation Protocol) or RTP (Real-time Transport Protocol) to extract the transmission address and port information of the voice media stream from the call signaling, and then establishes a connection for data acquisition.

[0037] The speech recognition module uses automatic speech recognition technology (ASR) to convert the voice media stream into corresponding text data. Specifically, preprocessing is performed on the collected voice media stream, including operations such as denoising, filtering, and enhancement, to improve the accuracy of speech recognition.

[0038] Specifically, a multi-layer CNN model is used for denoising. Its input layer receives the voice media stream data that has been sampled and quantized. The intermediate layer extracts the features of the voice and noise through convolutional kernels, and the output layer outputs the pure voice signal after removing the noise; a band-pass filter is used to process the voice media stream according to the frequency range of the voice signal; the voice media stream is frame-processed, and the length of each frame is preferably 20 - 30 milliseconds. The STFT transform is performed on each frame to convert the time-domain signal into a frequency-domain signal. In the frequency domain, according to the energy distribution characteristics of the voice signal and the noise signal, the spectrum of the voice signal is enhanced. By increasing the amplitude spectrum of the voice signal and suppressing the amplitude spectrum of the noise signal, the enhanced frequency-domain signal is converted back to the time-domain signal through the inverse STFT transform to obtain the enhanced voice media stream.

[0039] The preprocessed voice media stream is input into the speech recognition engine, and the engine converts it into corresponding text data and outputs it to the semantic understanding module for processing.

[0040] The semantic understanding module uses natural language processing technology (NLP) to perform semantic parsing on the text data and extract key information. For the event subject, through part-of-speech tagging and named entity recognition technology, entities such as person names, enterprise names, and tax matters in the text are recognized.

[0041] Specifically, the part-of-speech tagging technology uses a conditional random field (CRF) model to tag the text and construct a part-of-speech tagging corpus. The CRF model is trained using the corpus, and the model learns the relationship between each word and its part of speech in the text. In practical applications, the text data output by the speech recognition module is input into the trained CRF model, and the model tags each word with the corresponding part of speech, such as nouns, verbs, adjectives, etc. The named entity recognition technology includes constructing a tagging corpus containing entities such as person names, company names, and tax matters. Through the training of this corpus, the CRF model can learn the characteristics and patterns of these entities in the text, so as to accurately identify the taxpayer name, tax agency name, etc. in the tax return. Part-of-speech tagging can further improve the accuracy of named entity recognition. For example, for a noun phrase, if its part of speech is tagged as a noun and it is within the pattern of the taxpayer name learned by the model, it is determined to be a taxpayer name entity.

[0042] Through syntactic analysis and semantic understanding, combined with the knowledge base in the tax field, the problems and corresponding solutions in the text are classified and extracted. Syntactic analysis constructs a parse tree to analyze the structure and components of the sentence, providing a basis for semantic analysis. For example, for the sentence "Taxpayer Zhang San has filed a tax return", the parse tree can show the dependency relationships such as "Taxpayer Zhang San" being the subject of the sentence, "has filed" being the predicate, and "tax return" being the object. Semantic analysis combines the professional knowledge base in the tax field to understand the semantics of the words and sentences in the text. It identifies the semantic relationships and importance levels between the words, matches the words in the text with the terms in the knowledge base to determine the semantic meanings of the words. For example, for the term "tax return", the knowledge base stores information such as its definition, filing process, and relevant regulations. Then, based on the parse tree obtained from syntactic analysis, the semantic relationships between the words in the sentence are analyzed. For example, in the sentence "Taxpayer Zhang San has filed a tax return", through semantic analysis, it is determined that "Zhang San" is the subject of the "tax return" action, and "tax return" is a tax matter. In this way, the problems and corresponding solutions in the text are classified and extracted, providing a basis for subsequent quality inspection and analysis.

[0043] The sentiment analysis module conducts sentiment analysis on text data to judge the service quality of the operator. Specifically, a sentiment dictionary is constructed, including positive words, negative words, and neutral words, etc. Based on the sentiment dictionary, the sentiment tendency of the text data is judged, and the service attitude and service quality scores of the operator are obtained. Each word in the text data is matched with the words in the sentiment dictionary. If the word is in the positive word category, a certain weight is added to the positive sentiment score of the text; if it is in the negative word category, a certain weight is added to the negative sentiment score; if it is in the neutral word category, the sentiment score is not affected. Calculate the ratio of the positive sentiment score to the negative sentiment score, and judge the sentiment tendency of the text according to the preset threshold. If the ratio of the positive sentiment score to the negative sentiment score is greater than the threshold (such as 1.5), it is judged that the service attitude of the operator is positive; if the ratio is less than the threshold (such as 0.5), it is judged to be negative; if the ratio is within the threshold range, it is judged to be neutral. At the same time, according to the specific situation of the sentiment score, the service quality score of the operator is given. For example, if the positive sentiment score is high, the score is high; if the negative sentiment score is high, the score is low. The score range can be set from 0 to 10 points.

[0044] The quality inspection rule management module is used to customize and edit quality inspection rules to form a personalized quality inspection model, and store the edited quality inspection rules in the database for the quality inspection execution module to call. The database uses a relational database, such as MySQL or Oracle, to facilitate data management and query. In the database, a record is created for each quality inspection rule, and the record includes fields such as rule name, rule description, rule conditions, rule actions, and rule priorities. The rule priority is used to determine the execution order when multiple rules are matched, and the rule with a higher priority is executed first. For example, for some important violation behaviors, such as the rule of leaking customer information, a higher priority is set to ensure that such rules are checked first during the quality inspection process.

[0045] The quality inspection execution module is used to match the semantic parsing result and the sentiment analysis result with the preset quality inspection rules to generate a quality inspection result. The quality inspection result includes information such as quality inspection score, violation items, and recommended improvement measures. For example, for a quality inspection rule, the condition is that "the type of question raised by the customer is a tax declaration question and the service attitude of the operator is negative", and the rule action is to "generate a quality inspection score of 6 points, mark the violation item as poor service attitude, and the recommended improvement measure is to strengthen service attitude training". When the customer question is identified as a tax declaration question in the semantic parsing result and the sentiment analysis result is negative, the rule engine triggers this rule and executes the corresponding action.

[0046] A report generation module for generating various intuitive visualization charts based on quality inspection data. For example, generating trend charts of quality inspection pass rates by day, week, and month, and presenting the changing trend of quality inspection pass rates over time in the form of line charts to clearly reflect the fluctuations in service quality; generating pie charts of the proportion of problem types to visually show the proportion of various quality inspection problems in the overall problems and help quickly locate the main problem types.

[0047] It is also used to generate detailed quality inspection reports, including call record details, quality inspection problem descriptions, the time when problems occur, and operator information, etc. The reports are classified and summarized according to different dimensions, such as classified by business type, operator group, etc., to facilitate targeted analysis by management. At the same time, the reports also contain statistical analysis results of quality inspection data, such as statistical data of key indicators like average response time and first-time resolution rate, providing comprehensive data support for management decisions.

[0048] An intelligent training module for generating personalized training courses based on quality inspection results and the skill levels of operators to improve the service levels of operators. Specifically, determine the training needs of operators according to violation items and / or reasons for low scores, and generate personalized training courses, including information such as course content, learning methods, and learning time. Specifically, select appropriate course content from the training resource library according to the training needs. The training resource library stores various training materials such as tax business knowledge, service skills, and communication methods, including various forms such as documents, videos, and audios. For example, for the training of tax declaration business knowledge, relevant tax law interpretation documents, tax declaration process demonstration videos, etc. can be selected as course content. Build an online learning platform to provide functions such as course learning, online testing, and interactive communication to facilitate operators' self-study.

[0049] Embodiment 2

[0050] The distinguishing technical features of this embodiment from the above-mentioned embodiment are that it further includes a multi-modal data acquisition module and a multi-modal feature fusion module.

[0051] The multi-modal data acquisition module is used to continuously track the facial area of the operator in real time through a face monitoring model, extract expression features (such as smiling, frowning, surprised, etc.), and quantify the emotional state; it is also used to capture the operation logs of the operator (such as system interface switching, information query duration) by calling the system's underlying screen recording interface. During the capture process, the system focuses on recording operation logs, including key information such as the time points of system interface switching and the durations of different information queries. For example, when the operator is querying tax information for a customer, the duration from clicking the query button to obtaining the result, as well as the number of times and time intervals of switching different business system interfaces during the query process, etc., to judge the service response efficiency.

[0052] A multi-modal feature fusion module for combining the speech emotion score S 语音 (output of the emotion analysis module), the facial expression emotion score S 表情 (output of the facial feature analysis), and the operation efficiency score S 操作 (output of the screen operation analysis) are input into the attention mechanism network, and the weights of each modality are dynamically allocated according to the matching degree between the collected data and the scoring items. For example, when the collected speech emotion data has a low match with the scoring item, it indicates that the score of the speech emotion data is relatively fuzzy, and at this time, the weights of other analyses are automatically increased.

[0053] A random forest model is used to vote on the multi-modal analysis results to output the comprehensive service quality score. The formula is as follows:

[0054] Score = α·S 语音 +β·S 表情 +γ·S 操作

[0055] Among them, α, β, and γ are weight parameters that are dynamically adjusted through training with historical data.

[0056] When the speech emotion analysis is "positive" but the facial expression analysis is "negative", the arbitration process is triggered, and the semantic understanding result (such as whether sensitive words are mentioned) is called for secondary verification to avoid misjudgment of a single modality. Specifically, the semantic understanding module will re-examine the previously parsed text data, focusing on whether sensitive words are mentioned. For example, in a tax consulting scenario, if the operator shows a positive attitude in the voice during the communication with the customer, but the facial expression is negative, the semantic understanding module will check whether there are sensitive words such as "unable to handle" or "not clear" in the text. If there are sensitive words, it indicates that there may be problems with the operator's service quality.

[0057] The above technical solution solves the limitations of single speech / text analysis (such as the contradictory scenario of "smiling voice" and "negative facial expression").

[0058] Example Three

[0059] The distinguishing technical features of this example from the above example are that the quality inspection rule management module further includes a rule evolution sub-module and a priority scheduling sub-module, which can intelligently optimize the quality inspection rules.

[0060] The rule evolution sub-module encodes the quality inspection rule parameters into 128-dimensional vectors. The quality inspection rule parameters include sentiment thresholds, keyword weights, etc. At the same time, it combines context features, including information such as customer types and call periods. Customer types can be divided into enterprise customers and individual customers, and call periods can be divided into weekdays during the day, weekdays at night, weekends, etc. After digitizing and encoding the above information, it is integrated into the state vector, transforming complex rules and environmental information into a vector form that can be processed by the policy network.

[0061] A reward function is established. The reward function consists of two parts. One is the similarity score between the manual review result and the automatic quality inspection result, which is measured by calculating the matching degree between the two. The higher the similarity, the higher the score. The other is the rule complexity penalty term. Overly complex rules may lead to waste of computing resources and increased misjudgment. Therefore, the rule complexity is penalized. For example, metrics such as the number of conditions and actions in the rule and the depth of logical nesting are used to measure the rule complexity, and the higher the complexity, the greater the penalty. The reward function guides the policy network to generate quality inspection rules that are both accurate and concise.

[0062] Regularly use the trained policy network to guide the generation of new rules. During the new rule generation process, high-yield rule fragments are retained. Through knowledge distillation, new rules can draw on the excellent parts of existing successful rules, accelerating the process of rule evolution and improving the quality and effectiveness of the rules.

[0063] The priority scheduling sub-module is used to dynamically adjust the rule execution order according to the real-time traffic volume and service type, ensuring that the rules that have the greatest impact on service quality can be preferentially executed in different situations.

[0064] Specifically, the priority scheduling sub-module predicts future traffic peaks based on the long short-term memory network (LSTM). The LSTM network can effectively process time series data. By learning historical traffic volume data, it predicts the future traffic change trend. For example, during the tax filing period, the traffic volume usually increases significantly, and the LSTM network can accurately predict the arrival of the tax filing peak period. Adjust the rule execution priority according to the traffic volume change trend. For example, if it is predicted to be the tax filing peak period, the rules related to "declaration process errors" are preferentially executed. At the same time, the Q-learning algorithm is adopted, with the rule execution time and service impact degree as the state, to learn the optimal rule triggering path. The Q-learning algorithm updates the Q value according to the obtained rewards by continuously trying different rule execution orders, so as to find the optimal rule triggering strategy.

[0065] Through the above technical solutions, the quality inspection rule library can self-evolve, continuously optimize the rule logic according to historical quality inspection data and manual review feedback, dynamically adjust the rule execution order, and provide interpretability for the rule evolution process, improving the performance and reliability of the intelligent quality inspection system.

[0066] An embodiment of the present disclosure further provides an intelligent quality inspection method based on multi-dimensional data analysis, which employs the above-mentioned intelligent quality inspection system based on multi-dimensional data analysis.

[0067] Those of ordinary skill in the art can understand that all or part of the processes in implementing an intelligent quality inspection method based on multi-dimensional data analysis can be completed by instructing relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of various embodiments of an intelligent quality inspection method based on multi-dimensional data analysis. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in this application can include non-volatile and / or volatile memories. 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), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0068] An embodiment of the present application also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps of the above-mentioned intelligent quality inspection method based on multi-dimensional data analysis. In the embodiment of the present application, the processor is the control center of the computer method, which can be the processor of a physical machine or the processor of a virtual machine.

[0069] Referring to Figure 2 , the electronic device 500 includes: at least one processor 501, at least one communication interface 502, at least one memory 503, and at least one bus 504. Among them, the bus 504 is used to realize the connection and communication between these components. The communication interface 502 is used to communicate signaling or data with other node devices. The memory 503 stores machine-readable instructions executable by the processor 501. When the electronic device 500 runs, the processor 501 communicates with the memory 503 through the bus 504. When the machine-readable instructions are called by the processor 501, they execute the steps of an intelligent quality inspection method based on multi-dimensional data analysis.

[0070] The above content is only an embodiment of the present invention. Specific structures and common knowledge such as characteristics that are well-known in the solution are not described in detail here. Those of ordinary skill in the art know all the common general technical knowledge in the technical field to which the invention belongs before the filing date or the priority date, are able to obtain all the prior art in this field, and have the ability to apply the conventional experimental means before this date. Those of ordinary skill in the art can, under the inspiration given in this application, combine their own abilities to improve and implement this solution. Some typical well-known structures or well-known methods should not become an obstacle for those of ordinary skill in the art to implement this application. It should be noted that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can still be made, and these should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicality of the patent. The protection scope required by this application should be based on the content of its claims, and the specific implementation manners and the like recorded in the specification can be used to explain the content of the claims.

Claims

1. An intelligent quality inspection system based on multi-dimensional data analysis, characterized by: include: The data acquisition module is used to collect the voice media stream generated in real time during the voice call and transmit it to the voice recognition module; A speech recognition module, used for converting a received speech media stream into text data; Semantic understanding module, used to perform semantic analysis on text data and extract key information; Sentiment analysis module, used to perform sentiment analysis on text data and judge the service quality of the operators; The quality inspection rule management module is used to customize and edit quality inspection rules, form a personalized quality inspection model, and store the edited quality inspection rules in the database; The quality inspection execution module is used to match the semantic parsing results and sentiment analysis results with the preset quality inspection rules to generate quality inspection results; Intelligent training module, used to generate personalized training courses based on quality inspection results.

2. The intelligent quality inspection system based on multi-dimensional data analysis according to claim 1, characterized in that: The data collection module adopts a linear microphone array and focuses the collection direction on the location of the call through beam forming technology.

3. The intelligent quality inspection system based on multi-dimensional data analysis according to claim 1, characterized in that: The speech recognition module is used to perform denoising, filtering and enhancement operations on the collected speech media stream; the denoising operation uses a multi-layer CNN model for denoising; the filtering operation uses a bandpass filter to process the speech media stream; the enhancement operation divides the speech media stream into frames, performs STFT transformation on each frame, converts the time domain signal into a frequency domain signal, enhances the spectrum of the speech signal in the frequency domain, and then converts the enhanced frequency domain signal back to the time domain signal through an inverse STFT transformation.

4. The intelligent quality inspection system based on multi-dimensional data analysis according to claim 1, characterized in that: The semantic understanding module uses the conditional random field model to annotate the text and build a part-of-speech tagging corpus. The corpus conditional random field model is used for training. The model learns the relationship between each word and the part of speech in the text. The text data output by the speech recognition module is input into the trained model, and the model tags each word with the corresponding part of speech.

5. The intelligent quality inspection system based on multi-dimensional data analysis according to claim 1, characterized in that: The sentiment analysis module constructs a sentiment dictionary including positive words, negative words and neutral words, and makes sentiment tendency judgment on text data based on the sentiment dictionary to obtain the service attitude and service quality score of the operator.

6. The intelligent quality inspection system based on multi-dimensional data analysis according to claim 1, characterized in that: The intelligent training module determines the training needs of the operators according to the violations and / or reasons for the low scores, selects appropriate course content from the training resource library, and builds an online learning platform.

7. The intelligent quality inspection system based on multi-dimensional data analysis according to claim 1, characterized in that: It also includes a multimodal data acquisition module and a multimodal feature fusion module. The multimodal data acquisition module is used to track the facial area of ​​the operator in real time, extract expression features, and capture the operation log of the operator; the multimodal feature fusion module is used to input the voice emotion score, expression emotion score and operation efficiency score into the attention mechanism network, dynamically allocate the weight of each modality, and vote on the multimodal analysis results through the random forest model to output a comprehensive service quality score.

8. The intelligent quality inspection system based on multi-dimensional data analysis according to claim 1, characterized in that: The quality inspection rule management module includes a rule evolution submodule and a priority scheduling submodule. The rule evolution submodule encodes the quality inspection rule parameters into vectors, combines context features, and generates target quality inspection rules through a reward function to guide the strategy network; the priority scheduling submodule is used to dynamically adjust the rule execution order according to real-time traffic volume and service type.

9. An intelligent quality inspection method based on multi-dimensional data analysis, characterized in that: An intelligent quality inspection system based on multi-dimensional data analysis as described in any one of claims 1 to 8 is used.

Images