Customer service quality detection method and related devices based on deep learning algorithms

Abstract

The present application provides a customer service quality detection method and related devices based on a deep learning algorithm. The method obtains an initial text from the conversation information between the customer service and the user; preprocesses the initial text to obtain a text to be detected; and detects the customer service quality based on the text to be detected, a pre-constructed quality inspection model set, and a preset quality inspection rule table to obtain a detection result. The method provided by the present application can accurately and quantitatively monitor the customer service in multiple dimensions, timely discover customer demands, conduct public opinion monitoring on potential risks, etc., effectively reduce labor costs, reduce customer complaints, and improve the service quality of the agent seats.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and particularly to a method for detecting the service quality of customer service based on a deep learning algorithm and related devices. Background Art

[0002] In the operation process of various enterprises and institutions providing services, the customer service phone is one of the important channels for providing services to users. In order to ensure that users receive good services, supervising the service quality of customer service staff is one of the important tasks of enterprises and institutions. The conventional supervision method is to conduct supervision and management based on service quality evaluation.

[0003] Currently, usually after the call between the user and the customer service staff ends, the user is prompted to rate the service quality of the customer service staff for this service, and the user's rating is used as the evaluation standard for service quality.

[0004] However, in the actual service process, the subjectivity of users' service quality evaluation is relatively strong. The service quality evaluation of customer service staff only depends on the subjective evaluation of users' selected ratings, resulting in low accuracy of service quality evaluation and being unfavorable for the supervision and management of service quality. Summary of the Invention

[0005] In view of this, the purpose of this application is to propose a method for detecting the service quality of customer service based on a deep learning algorithm and related devices.

[0006] Based on the above purpose, this application provides a method for detecting the service quality of customer service based on a deep learning algorithm, including:

[0007] Obtain the initial text from the conversation information between the customer service and the user;

[0008] Preprocess the initial text to obtain the text to be detected;

[0009] Detect the service quality of the customer service based on the text to be detected, the pre-constructed quality inspection model set, and the preset quality inspection rule table to obtain the detection result.

[0010] Optionally, obtaining the initial text from the conversation information between the customer service and the user includes:

[0011] In response to the conversation information being text information, use this text information as the initial text;

[0012] In response to the conversation information being voice information, convert the voice information into text information and use it as the initial text.

[0013] Optionally, the preprocessing of the initial text includes:

[0014] Desensitize the initial text to remove sensitive information in the initial text.

[0015] Optionally, detect the customer service quality based on the text to be detected, a pre-constructed set of quality inspection models, and a preset quality inspection rule table, and obtain a detection result, including:

[0016] Determine at least one candidate quality inspection model from the set of quality inspection models according to the preset quality inspection rule table;

[0017] Input the text to be detected into all the candidate quality inspection models and output the detection result.

[0018] Optionally, the set of quality inspection models includes at least one of the following:

[0019] A keyword model, configured to detect keywords in the text to be detected;

[0020] An emotion recognition model, configured to classify the text to be detected according to preset emotion categories;

[0021] A customer complaint model, configured to classify the text to be detected according to preset complaint categories;

[0022] A customer demand model, configured to detect the clustering category of the text to be detected through a clustering algorithm;

[0023] A similar text model, configured to calculate the similarity of the text to be detected according to the quality inspection rule table;

[0024] Among them, the emotion recognition model, the customer complaint model, the customer demand model, and the similar text model are pre-trained.

[0025] Optionally, the pre-training of the emotion recognition model and the customer complaint model includes:

[0026] Obtain an initial data set and desensitize it to obtain an initial training data set;

[0027] Encode the initial training data set through a Bert model to obtain the semantic vector representation corresponding to each data in the initial training data set;

[0028] Calculate the cosine similarity between any two data in the initial training data set based on the semantic vector representation through a cosine similarity algorithm;

[0029] Construct a similar sentence index service based on the cosine similarity through the Annoy algorithm to obtain a set of similar sentences;

[0030] Select a part of the data from the initial training dataset for manual annotation to obtain an annotated dataset;

[0031] Query in the set of similar sentences based on the annotated dataset to obtain a training dataset;

[0032] Perform the pre-training on the emotion recognition model and the customer complaint model respectively based on the training dataset.

[0033] Optionally, the pre-training on the customer demand model includes:

[0034] Calculate all the semantic vector representations through the k-means clustering algorithm to obtain the customer demand model with a preset number of clustering categories.

[0035] Optionally, the pre-training on the similar text model includes:

[0036] Perform the pre-training on the ROFORMER pre-training model through a preset similarity dataset to obtain the similar text model.

[0037] Based on the same inventive concept, the present application also provides a customer service quality detection device based on a deep learning algorithm, including:

[0038] An acquisition module configured to acquire an initial text from the conversation information between the customer service and the user;

[0039] A preprocessing module configured to preprocess the initial text to obtain a text to be detected;

[0040] A detection module configured to detect the customer service quality based on the text to be detected, a pre-constructed quality inspection model set, and a preset quality inspection rule table to obtain a detection result.

[0041] Based on the same inventive concept, the present application also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable by the processor, and the processor implements the method as described above when executing the computer program.

[0042] Based on the same inventive concept, the present application also provides a non-transitory computer-readable storage medium, and the non-transitory computer-readable storage medium stores computer instructions for causing a computer to execute the method as described above.

[0043] As can be seen from the above, a customer service quality detection method and related devices based on deep learning algorithms provided by this application obtain initial text from the conversation information between the customer service and the user; preprocess the initial text to obtain the text to be detected; and detect the customer service quality based on the text to be detected, the pre-constructed quality inspection model set, and the preset quality inspection rule table to obtain the detection result. The method provided by this application can achieve multi-dimensional and accurate quantitative monitoring of customer service, timely discovery of customer demands, public opinion monitoring of potential risks, etc., effectively reducing labor costs, reducing customer complaints, and improving the service quality of the operator seat. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in this application or related technologies, the following will briefly introduce the drawings required for use in the embodiments or related technology descriptions. Obviously, the drawings in the following descriptions are only the embodiments of this application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0045] Figure 1 It is a schematic flowchart of the customer service quality detection method based on deep learning algorithms in the embodiments of this application;

[0046] Figure 2 It is a schematic flowchart of a model pre-training method in the embodiments of this application;

[0047] Figure 3 It is a schematic structural diagram of the customer service quality detection device based on deep learning algorithms in the embodiments of this application;

[0048] Figure 4 It is a schematic diagram of the hardware structure of the electronic device in the embodiments of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following further describes this application in detail with reference to specific embodiments and the accompanying drawings.

[0050] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of this application should have the ordinary meanings understood by those of ordinary skill in the field to which this application belongs. The "first", "second" and similar terms used in the embodiments of this application do not denote any order, quantity or importance, but are only used to distinguish different components. Words such as "including" or "comprising" mean that the elements or objects appearing before this word cover the elements or objects listed after this word and their equivalents, without excluding other elements or objects. Words such as "connected" or "linked" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right", etc. are only used to indicate relative position relationships. When the absolute position of the object being described changes, the relative position relationship may also change accordingly.

[0051] As described in the background art, there is currently no effective and objective method for evaluating and supervising the quality of customer service calls.

[0052] Currently, the customer service channels in the insurance field include manual telephone customer service, intelligent customer service, etc. The main purpose of the customer service is to solve users' problems. Since the insurance industry has relatively high requirements for the level of customer service and the intelligent customer service in the insurance industry has relatively poor dialogue capabilities, currently, it mainly relies on manual customer service supplemented by intelligent customer service. Insurance companies have historical voice call and text dialogue data accumulated over the years. During the conversation between customers and customer service, how to accurately and quantitatively monitor manual customer service, timely discover customers' demands, and quickly discover potential risks for public opinion monitoring, etc., are all issues that have been concerned in the field of insurance customer service.

[0053] Currently, the quality inspection of customer service in the insurance industry mainly relies on manual sampling. For telephone customer service, there are about 1 million calls per day. If you want to know whether there is a customer complaint in a call or whether the customer service attitude meets the requirements, manual sampling can only select a few hundred recordings for inspection every day. The method of manual quality inspection is time-consuming and laborious, with low accuracy, and it is impossible to achieve full inspection and rapid response. Automatic methods include keyword filtering, regular expression matching, and some single classification models that rely on word segmentation features. Although the quality inspection method of keyword filtering can respond quickly, it is prone to relatively large misidentifications and it is difficult to distinguish between customers and customer service. For example, there is a keyword "complaint". When it appears in a customer's conversation, it may be "I want to complain about you" or "I don't want to complain"; when it appears in a customer service's conversation, it may be "Your complaint is being processed", so it is impossible to judge whether there is a problem with the conversation solely based on the keyword. Keyword quality inspection can only be used as an auxiliary method. Classification models that simply rely on word segmentation features do not fully consider semantic information and can only be used singly, which is not convenient for general expansion.

[0054] With the development of deep learning technology, NLP (Natural Language Processing Technology) has also made breakthroughs. Natural language is the language people use in daily life. Natural language processing technology can be used for text analysis and mining, providing a technical foundation for intelligent quality inspection in the insurance customer service field.

[0055] Therefore, a customer service quality detection method and related devices based on deep learning algorithms provided in this application can detect the text to be detected through a quality inspection model constructed based on deep learning algorithms, enabling multi-dimensional and accurate quantitative monitoring of customer service, timely discovery of customer demands, public opinion monitoring of potential risks, etc., effectively reducing labor costs, reducing customer complaints, and improving the service quality of the agent.

[0056] The following will detail the embodiments of this application with reference to the accompanying drawings.

[0057] The embodiments of this application provide a customer service quality detection method based on deep learning algorithms, referring to Figure 1 , including the following steps:

[0058] Step S101: Obtain the initial text from the conversation information between the customer service and the user.

[0059] Specifically, currently, most of the conversation information between the customer and the user is voice information, accompanied by a small amount of text information. By obtaining the audio file or call text record of a conversation between the customer and the user, the initial text for customer service quality detection is obtained.

[0060] Step S102: Preprocess the initial text to obtain the text to be detected.

[0061] In this step, desensitize the initial text to remove sensitive information in the initial text. Specifically, the initial text may involve user sensitive information, and user sensitive information such as the customer's name, mobile phone number, and ID number needs to be filtered.

[0062] Step S103: Detect the customer service quality based on the text to be detected, a pre-constructed set of quality inspection models, and a preset quality inspection rule table to obtain the detection result.

[0063] Specifically, the specific detection items for the text to be detected can be adjusted according to the actual situation. For example, quality inspection personnel may want to detect whether certain keywords appear in the text to be detected, or whether there are fixed opening and closing remarks in the text to be detected, or whether there is some extreme emotion or complaint-related content in the text to be detected, etc. Set a quality inspection rule table according to the quality inspection intention of the quality inspection personnel, select relevant quality inspection models according to the quality inspection rule table to detect the text to be detected, so as to obtain the detection result. The customer service staff can be scored according to the detection result, and the scoring rule can be set according to the requirements. After the score is lower than a certain threshold, the administrator can be notified that there is a potential risk in this call, and the customer service can be focused on and monitored.

[0064] In some embodiments, an initial text is obtained from the conversation information between the customer service and the user, including:

[0065] In response to the conversation information being text information, using the text information as the initial text;

[0066] In response to the conversation information being voice information, converting the voice information into text information and using it as the initial text.

[0067] Specifically, if the obtained conversation information is text information, that is, the customer and the user communicate through text, then directly use the text information as the initial text. If the obtained conversation information is voice information, the voice information needs to be converted into text information and then input into the relevant quality inspection models for detection.

[0068] In some embodiments, the quality of the customer service is detected based on the text to be detected, a pre-constructed set of quality inspection models, and a preset quality inspection rule table, and the detection results are obtained, including:

[0069] According to the preset quality inspection rule table, determine at least one candidate quality inspection model from the set of quality inspection models;

[0070] Input the text to be detected into all the candidate quality inspection models and output the detection results.

[0071] Specifically, the quality inspection model set includes multiple quality inspection models for different quality inspection purposes. Before each quality inspection, it is necessary to identify the quality inspection intention according to the quality inspection rule table, and then call one or more quality inspection models in the quality inspection model set to detect the text to be detected. The quality inspection rule table includes different tables, and each table stipulates different quality inspection intentions. For example, [Similar Text Table 1] stipulates the opening remarks, including 10 similar texts; [Similar Text Table 2] stipulates the closing remarks, including 20 similar texts, and so on. If this quality inspection wants to identify whether the text to be detected contains the opening remarks, then [Similar Text Table 1] is input as the quality inspection rule table during the detection, and the relevant quality inspection models in the quality inspection model set are called to detect the text to be detected. The quality inspection model can calculate the similarity scores between the text to be detected and the 10 similar texts in [Similar Text Table 1]. The detection result is the returned similarity score. If the score exceeds the preset similarity threshold, it can be determined that the text to be detected contains the opening remarks. If the similarity score is less than the preset similarity threshold, it can be determined that the text to be detected does not contain the opening remarks, indicating that the customer service staff did not communicate with the user according to the call regulations.

[0072] By setting the quality inspection rule table, the quality inspection intention can be flexibly set, and the relevant quality inspection models can be automatically called to detect the text to be detected, effectively reducing the labor cost and time cost.

[0073] In some embodiments, the quality inspection model set includes at least one of the following:

[0074] A keyword model, configured to detect keywords in the text to be detected;

[0075] An emotion recognition model, configured to classify the text to be detected according to preset emotion categories;

[0076] A customer complaint model, configured to classify the text to be detected according to preset complaint categories;

[0077] A customer demand model, configured to detect the clustering category of the text to be detected through a clustering algorithm;

[0078] A similar text model, configured to calculate the similarity of the text to be detected according to the quality inspection rule table;

[0079] Among them, the emotion recognition model, the customer complaint model, the customer demand model, and the similar text model are pre-trained.

[0080] Specifically, the keyword model is implemented based on the Double Array Trie algorithm. The Double Array Trie is a data structure that balances query efficiency and space storage. It does not require tokenization of the query text and supports multi-pattern matching. The dictionary size is over 100,000, and the query time is within 10 ms. The dictionary is a commonly used keyword dictionary in the insurance field, and the keyword model does not require training. When performing quality inspection, the text to be inspected is input into the keyword model, and the model outputs all the keywords contained in the dictionary that appear in the text to be inspected or the keywords specified in the quality inspection rule table. For example, it outputs the keywords contained in [Keyword Table 1] in the quality inspection rule table that appear in the text to be inspected.

[0081] The emotion recognition model is a multi-classification model and requires pre-training. In this embodiment, the preset emotion categories at least include: angry, abusive, high-risk, and normal. The pre-trained model uses the ALBERT model, with model parameters maxlen = 128, batch_size = 32, epoch_num = 10. Among them, the number of transformer layers is 4, and the classification layer is a linear layer and a softmax layer. The ALBERT model is trained through a training dataset to obtain the emotion recognition model. When performing quality inspection, the text to be inspected is input into the emotion recognition model, and the model outputs the emotion category of the text to be inspected.

[0082] The customer complaint model is a multi-classification model, with the same model structure and parameters as the emotion recognition model, and requires pre-training. The complaint categories at least include: customer complains about the customer service, asks for the customer service supervisor, questions the poor attitude of the customer service, and normal. The pre-trained model uses the ALBERT model, with model parameters maxlen = 128, batch_size = 32, epoch_num = 10. Among them, the number of transformer layers is 4, and the classification layer is a linear layer and a softmax layer. The ALBERT model is trained through a training dataset to obtain the customer complaint model. When performing quality inspection, the text to be inspected is input into the customer complaint model, and the model outputs the complaint category of the text to be inspected.

[0083] The customer demand model is a clustering model and requires pre-training. The clustering category of the text to be inspected can be calculated through a clustering algorithm. When performing quality inspection, the text to be inspected is input into the customer demand model, and the model outputs the clustering category of the text to be inspected.

[0084] The similar text model requires pre-training and can calculate the similarity score of the text to be inspected according to the quality inspection rule table.

[0085] In some embodiments, refer to Figure 2, the pre-training of the emotion recognition model and the customer complaint model includes the following steps:

[0086] Step S201, obtain an initial data set and desensitize it to obtain an initial training data set;

[0087] Step S202, encode the initial training data set through a Bert model to obtain the semantic vector representation corresponding to each data in the initial training data set;

[0088] Step S203, calculate the cosine similarity between any two data in the initial training data set based on the semantic vector representation through the cosine similarity algorithm;

[0089] Step S204, construct a similar sentence index service based on the cosine similarity through the Annoy algorithm to obtain a set of similar sentences;

[0090] Step S205, select some data from the initial training data set for manual annotation to obtain an annotated data set;

[0091] Step S206, query in the set of similar sentences based on the annotated data set to obtain a training data set;

[0092] Step S207, perform the pre-training on the emotion recognition model and the customer complaint model respectively based on the training data set.

[0093] Specifically, extract a certain number of call text conversations, split them according to the customer and customer service roles to obtain multiple sentences. Since the conversation may involve user sensitive information, user sensitive information such as names, mobile phone numbers, and ID card numbers need to be filtered out to obtain the initial training data set. Input the initial training data set into the BERT model trained on the public similar sentence data set, and the BERT model outputs the semantic vector representation corresponding to each data in the initial training data set. In order to find multiple groups of similar texts in all the data of the initial training data set, calculate the similarity between any two semantic vector representations through the cosine similarity algorithm. The cosine similarity algorithm is as follows:

[0094]

[0095] where A and B represent sentence vectors, A i and B i represent the components of sentence vectors A and B, 1 ≤ i ≤ n. The range of the cosine similarity value is between [-1, 1]. The closer the value is to 1, the closer the directions of the two vectors are; the closer it is to -1, the more opposite their directions are; close to 0, it means the two vectors are nearly orthogonal.

[0096] Build a similar sentence index service with the help of nearest neighbor search libraries such as Annoy (Approximate Nearest Neighbors Oh Yeah) or Faiss (Facebook AI Similarity Search). In this application, Annoy is used to build the index because the query speed of the Annoy index is relatively fast. When building the index, the input is a vector matrix, the number of index trees is 30, and the above-mentioned cosine similarity is used in the building process. The output is an index file, which can be stored on the hard disk. The entire building process only needs to be executed once. After the building is completed, load the index file and build a query service. The input is a sentence, which is converted into a sentence vector and then similar sentences are queried. The output is the TopN similar sentences, and the value of N can be adjusted according to actual needs. The query example is as follows:

[0097] Query sentence: Do I need to go to the counter to handle this?

[0098] Similar sentences (Top10): ['Do I need to go to the counter to handle this?', 'Do I need to go to the counter for this?', 'Do I need to go to the counter to handle it, right?', 'Oh, do I need to go to the counter to handle it?', 'Do I need to go to the counter to handle it?', 'Do I need to go to the counter to handle it, right?', 'Oh, that is to say, do I need to go to the counter to handle it?', 'Must I handle it at the counter?', 'Do I need to go to the counter to handle it?'].

[0099] Taking the emotion recognition model as an example, select thousands of data from the initial training dataset for manual annotation. In this embodiment, the emotion categories are set to 4 types: angry, abusive, high-risk, and normal, and the above-mentioned thousands of data are marked as these 4 categories. For example, the category of "I want to ask what health insurance there is" is "normal", and the category of "Are you a stupid salesman, may I ask?" is "abusive". In the customer complaint model, the data is manually annotated in the same way. In this embodiment, the complaint categories are set to 4 types: customer complains about the customer service, looks for the customer service supervisor, questions the poor attitude of the customer service, and normal. The data after manual annotation is used as the annotated dataset.

[0100] However, because the cost of manual annotation is very high, not all data in the initial training dataset can be annotated. Therefore, the annotated dataset can be used to expand the training dataset. Specifically, input each data in the annotated dataset into the above-mentioned similar sentence index service that has been built, and use the similar sentence index service to output a set of similar sentences similar to each data. Combine all the sets of similar sentences with the annotated dataset as the training dataset. In this way, it not only ensures that the training dataset all comes from the actual business scenario, but also realizes the rapid expansion of the training dataset, greatly saving the labor annotation cost.

[0101] After obtaining the training data set, pre-train the emotion recognition model and the customer complaint model respectively. Split the training data set into a training set, a validation set and a test set, train the emotion recognition model and the customer complaint model on the training set, verify the accuracy of the classification results through the validation set, and finally test on the test set. If the training standard is met, stop training the model. The training standard is to minimize the loss function, and the loss function is expressed as

[0102]

[0103] where output size represents the number of classification categories. For example, for emotion recognition, it is 4. The true category of the i-th data sample is y i , and the predicted category is

[0104] In some embodiments, the pre-training of the customer demand model includes:

[0105] Calculate all the semantic vector representations through the k-means clustering algorithm to obtain the customer demand model with a preset number of clustering categories.

[0106] The customer demand model is a clustering model. Since clustering can use unlabeled data, part of the data is extracted from the initial training data set. After being converted into sentence vectors by BERT, it is input into the k-means clustering model. The number of clustering categories in this embodiment is set to 50 after repeated experiments. The specific number of clustering categories can be adjusted according to different business requirements. When performing quality inspection, the text to be detected is input into the customer demand model, and the clustering category of the text to be detected is output by the model.

[0107] In some embodiments, the pre-training of the similar text model includes:

[0108] Pre-train the ROFORMER pre-trained model through a preset similarity data set to obtain the similar text model.

[0109] The similarity model is constructed based on the ROFOMER pre-trained model. Among them, the number of transformer layers is 12. It is pre-trained on the similarity data set in the insurance field. During training, a noisy sentence (that is, the data in the training set of the similarity data set) is input, and a similar sentence is output by the ROFOMER pre-trained model. The meaning of the loss function of this model is to maximize the probability of the next target word under the condition of the known input sequence and the previous generated sequence, and finally expect the probability of the entire output sequence to be the largest, that is, to output the expected similar sentence. The loss function formula is as follows

[0110]

[0111] Among them, θ is the parameter of the model, N is the number of samples in the training set. Xn is the input sequence, and Yn is the predicted sequence. When performing quality detection, the text to be detected and the similar text to be detected are respectively output as the sentence vector representations of these two texts by the similarity model, and then the similarity score between these two texts is calculated based on the sentence vector representations using cosine similarity.

[0112] It should be noted that the method of the embodiment of the present application can be executed by a single device, such as a computer or a server. The method of this embodiment can also be applied to a distributed scenario and completed by multiple devices cooperating with each other. In this case of a distributed scenario, one of the multiple devices can only execute one or more steps of the method of the embodiment of the present application, and these multiple devices will interact with each other to complete the described method.

[0113] It should be noted that some embodiments of the present application have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in a different order from that in the above embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require the specific order or sequential order shown to achieve the desired result. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0114] Based on the same inventive concept, corresponding to the method of any of the above embodiments, the present application also provides a customer service quality detection device based on a deep learning algorithm.

[0115] Refer to Figure 3 , the customer service quality detection device based on a deep learning algorithm includes:

[0116] An acquisition module 301, configured to acquire an initial text from the conversation information between the customer service and the user;

[0117] A preprocessing module 302, configured to preprocess the initial text to obtain a text to be detected;

[0118] A detection module 303, configured to detect the customer service quality based on the text to be detected, a pre-constructed quality inspection model set, and a preset quality inspection rule table, and obtain a detection result.

[0119] For the convenience of description, when describing the above device, it is divided into various modules according to functions for separate description. Of course, when implementing the present application, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0120] The device in the above embodiments is used to implement the corresponding customer service quality detection method based on the deep learning algorithm in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be elaborated herein.

[0121] Based on the same inventive concept, corresponding to the method in any of the above embodiments, the present application further 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 customer service quality detection method based on the deep learning algorithm described in any of the above embodiments.

[0122] As Figure 4 FIG. shows a more specific schematic diagram of the hardware structure of the electronic device provided in this embodiment. The device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. Among them, the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are communicatively connected to each other inside the device through the bus 1050.

[0123] The processor 1010 may be implemented in a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0124] The memory 1020 may be implemented in the form of a ROM (Read Only Memory), a RAM (Random Access Memory), a static storage device, a dynamic storage device, etc. The memory 1020 may store an operating system and other application programs. When implementing the technical solutions provided in the embodiments of this specification through software or firmware, the relevant program codes are stored in the memory 1020 and are called and executed by the processor 1010.

[0125] The input / output interface 1030 is used to connect to an input / output module to implement information input and output. The input / output module may be configured as a component in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Among them, the input device may include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device may include a display, a speaker, a vibrator, an indicator light, etc.

[0126] The communication interface 1040 is used to connect to a communication module (not shown in the figure) to enable communication and interaction between this device and other devices. The communication module can achieve communication through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0127] The bus 1050 includes a path for transmitting information between various components of the device (such as the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040).

[0128] It should be noted that although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040, and the bus 1050, in the specific implementation process, the device may also include other components necessary for normal operation. In addition, those skilled in the art can understand that the above device may also only include the components necessary to implement the solution of the embodiments of this specification, and does not necessarily include all the components shown in the figure.

[0129] The electronic device of the above embodiment is used to implement the corresponding customer service quality detection method based on the deep learning algorithm in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be elaborated here.

[0130] Based on the same inventive concept, corresponding to the method of any of the above embodiments, the present application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the customer service quality detection method based on the deep learning algorithm as described in any of the foregoing embodiments.

[0131] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device.

[0132] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the customer service quality detection method based on the deep learning algorithm described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be elaborated here.

[0133] Those of ordinary skill in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the present application (including the claims) is limited to these examples; within the concept of the present application, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations in different aspects of the embodiments of the present application as described above, which are not provided in detail for the sake of brevity.

[0134] In addition, to simplify the description and discussion, and to avoid making the embodiments of the present application difficult to understand, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Further, the devices may be shown in block diagram form to avoid making the embodiments of the present application difficult to understand, and this also takes into account the fact that the details of the implementation of these block diagram devices are highly dependent on the platform on which the embodiments of the present application are to be implemented (i.e., these details should be fully within the understanding of those skilled in the art). In cases where specific details (such as circuits) are set forth to describe the exemplary embodiments of the present application, it will be apparent to those skilled in the art that the embodiments of the present application can be implemented without these specific details or with variations of these specific details. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0135] Although the present application has been described in connection with specific embodiments of the present application, many alternatives, modifications, and variations of these embodiments will be apparent to those of ordinary skill in the art based on the foregoing description. For example, other memory architectures (such as dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0136] The embodiments of the present application are intended to cover all such alternatives, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present application shall be included within the protection scope of the present application.

Claims

1. A method for detecting the quality of customer service based on deep learning algorithms, characterized in that, Including: Obtain the initial text from the conversation information between the customer service and the user; Preprocess the initial text to obtain the text to be detected; Detect the service quality of the customer service based on the text to be detected, the pre-constructed quality inspection model set and the preset quality inspection rule table, and obtain the detection result; The quality inspection model set includes the following models: The keyword model is configured to detect keywords in the text to be detected; The emotion recognition model is configured to classify the text to be detected according to the preset emotion categories; The customer complaint model is configured to classify the text to be detected according to the preset complaint categories; The customer demand model is configured to detect the clustering category of the text to be detected through a clustering algorithm; The similar text model is configured to calculate the similarity of the text to be detected according to the quality inspection rule table; Among them, the emotion recognition model, the customer complaint model, the customer demand model and the similar text model are pre-trained; The pre-training of the emotion recognition model and the customer complaint model includes: Obtain the initial data set and perform desensitization processing on it to obtain the initial training data set; Encode the initial training data set through the Bert model to obtain the semantic vector representation corresponding to each data in the initial training data set; Calculate the cosine similarity of any two data in the initial training data set based on the semantic vector representation through the cosine similarity algorithm; Construct a similar sentence index service based on the cosine similarity through the Annoy algorithm to obtain a set of similar sentences; Select some data from the initial training data set for manual annotation to obtain an annotated data set; Query in the set of similar sentences based on the annotated data set to obtain a training data set; Based on the training data set, perform the pre-training on the emotion recognition model and the customer complaint model respectively.

2. The method according to claim 1, wherein Obtain the initial text from the conversation information between the customer service and the user, including: In response to the conversation information being text information, use the text information as the initial text; In response to the conversation information being voice information, convert the voice information into text information and use it as the initial text.

3. The method according to claim 1, wherein The preprocessing of the initial text includes: Perform desensitization processing on the initial text to remove sensitive information in the initial text.

4. The method according to claim 1, wherein The detection of the service quality of the customer service based on the text to be detected, the pre-constructed quality inspection model set and the preset quality inspection rule table to obtain the detection result includes: According to the preset quality inspection rule table, determine at least one candidate quality inspection model from the quality inspection model set; Input the text to be detected into all the candidate quality inspection models and output the detection result.

5. The method according to claim 1, characterized in that The pre-training of the customer demand model includes: Calculate all the semantic vector representations through the k-means clustering algorithm to obtain the customer demand model with a preset number of clustering categories.

6. The method according to claim 1, wherein The pre-training of the similar text model includes: Perform pre-training on the ROFORMER pre-training model through a preset similarity data set to obtain the similar text model.

7. A customer service quality detection device based on a deep learning algorithm, characterized in that, Including: An acquisition module, configured to acquire initial text from the conversation information between the customer service and the user; A preprocessing module, configured to preprocess the initial text to obtain a text to be detected; A detection module, configured to detect the service quality of the customer service based on the text to be detected, a pre-constructed quality inspection model set, and a preset quality inspection rule table, and obtain a detection result; The quality inspection model set includes the following models: A keyword model, configured to detect keywords in the text to be detected; An emotion recognition model, configured to classify the text to be detected according to preset emotion categories; A customer complaint model, configured to classify the text to be detected according to preset complaint categories; A customer demand model, configured to detect the clustering category of the text to be detected through a clustering algorithm; A similar text model, configured to calculate the similarity of the text to be detected according to the quality inspection rule table; Wherein, the emotion recognition model, the customer complaint model, the customer demand model, and the similar text model are pre-trained; Performing the pre-training on the emotion recognition model and the customer complaint model includes: Acquiring an initial data set and performing desensitization processing on it to obtain an initial training data set; Encoding the initial training data set through a Bert model to obtain a semantic vector representation corresponding to each data in the initial training data set; Calculating the cosine similarity between any two data in the initial training data set based on the semantic vector representation through a cosine similarity algorithm; Constructing a similar sentence index service based on the cosine similarity through an Annoy algorithm to obtain a set of similar sentences; Selecting part of the data in the initial training data set for manual annotation to obtain an annotated data set; Querying in the set of similar sentences based on the annotated data set to obtain a training data set; Performing the pre-training on the emotion recognition model and the customer complaint model respectively based on the training data set.

8. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

Images