Customer service capability improvement method and system based on user telephone traffic intention

Abstract

The invention discloses a customer service capability improvement method and system based on a user telephone traffic intention, and solves the problem of tedious processes of recording a user and filling a task work order by a manual seat. The user language text content is input into the coding and decoding combined prediction model to obtain a user intention prediction result and a user emotion prediction result, user question collection is performed according to the prediction results, in addition, seat allocation is performed based on prediction friendliness level target intention result recognition, the work order circulation process is further optimized, and the user experience is improved. According to the method, information transmission, task allocation and cooperative problem processing among business departments can be better assisted, and through the flow, the customer service center can efficiently and automatically allocate customer service resources according to user requirements and internal resource conditions, so that the service quality and the user satisfaction degree are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power grid customer service application, and in particular relates to a method and system for improving customer service capability based on user traffic intention. Background Art

[0002] IVR (Interactive Voice Response) service is a voice service value-added service. When users access the IVR system by phone, the IVR system will provide users with the corresponding IVR service process. For example, after the user dials the designated phone number 95598, the IVR system provides the user with a menu broadcast process. Through the corresponding selection operations, the user can enter the electricity fee inquiry process, fault repair process, power outage information inquiry process, account number inquiry process, electricity fee inquiry, complaint reporting process, new energy vehicle charging process, power purchase and power trading process, and manual transfer process. The current problems are:

[0003] 1. When the IVR system cannot solve the user's problem, the user's call needs to be transferred to a manual agent, who will collect and solve the user's problem. For some problems that need to be solved by business personnel, the manual agent will record the user's problem, fill out a task ticket, and transfer it to the business personnel through the ticket flow to solve the user's problem. However, in this process, after the terminal comes in, manual operation is required to complete the implementation of the background process engine, which is inefficient and inaccurate.

[0004] 2. In actual business, a problem may involve multiple departments such as the customer service department and provincial power companies, and multiple processing stages, which require transfer and collaborative processing between departments. Faced with the continuous changes and updates in user needs and internal organizations and the continuous expansion of business volume, this processing model can no longer serve customers well, and at the same time, it is also impossible to monitor the service quality and efficiency of the business well. Therefore, there is an urgent need for a method to assist various business departments in transmitting information, assigning tasks, and collaboratively processing problems. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present application provides a method and system for improving customer service capabilities based on user call intentions.

[0006] In a first aspect, the present application proposes a method for improving customer service capabilities based on user traffic intention, comprising the following steps:

[0007] Acquire voice data after interaction with the user based on the IVR connection established with the user terminal, perform voice recognition operation on the voice data through a voice recognition algorithm, and obtain user language text content corresponding to the voice data;

[0008] Constructing a codec joint prediction model, inputting the user language text content into the codec joint prediction model to obtain a user intention prediction result and a user emotion prediction result;

[0009] Converting the user intention prediction result and the user emotion prediction result into corresponding user intention prediction text and user emotion prediction text through reverse word embedding;

[0010] Establishing a friendliness prediction model based on deep learning, inputting the user emotion prediction text into the friendliness prediction model for friendliness recognition, and obtaining a predicted friendliness level;

[0011] Based on the predicted friendliness level, the customer service agents who answer the calls are assigned to the first category, and the first category assignment results are obtained;

[0012] Performing semantic recognition operation on the user intention prediction text based on the NLP language processing algorithm to obtain the corresponding target intention result;

[0013] An identification operation is performed on the target intention result to obtain a target agent service, and the call is automatically routed to the most suitable agent or department according to the first type of allocation result and the target agent service.

[0014] In some embodiments, the constructing of the codec joint prediction model and inputting the user language text content into the codec joint prediction model to obtain the user intention prediction result and the user emotion prediction result include:

[0015] The encoding and decoding joint prediction model is constructed using a recurrent neural network with BiLSTM as the core;

[0016] The architecture of the codec joint prediction model is a heterogeneous codec architecture, which includes an encoder part, a feature selection part and a decoder part. The encoder part uses RNN to abstract the input sequence data and outputs an abstract vector of a fixed dimension as an abstract representation of the sequence data. The decoder part uses a linear decoder to decode the abstract vector to obtain a sequence representation of the output sequence. The feature selection part selects one of the attention mechanism, PCA dimensionality reduction, regularization, convolution or pooling operations.

[0017] In some embodiments, the constructing of the codec joint prediction model, inputting the user language text content into the codec joint prediction model to obtain the user intention prediction result and the user emotion prediction result, further includes:

[0018] The encoder in the codec joint prediction model vectorizes the user language text content to obtain an abstract vector;

[0019] Performing a linear transformation on the abstract vector to obtain an intention prediction initial vector and an emotion prediction initial vector;

[0020] Decoding the intention prediction initial vector and the emotion prediction initial vector through a linear decoder to obtain a sequence representation of the vector;

[0021] The softmax function is used to predict the sequence representation of the vector to obtain the user intention prediction results and user emotion prediction results.

[0022] In some embodiments, converting the user intention prediction result and the user emotion prediction result into corresponding user intention prediction text and user emotion prediction text by reverse word embedding includes:

[0023] Selecting a pre-trained reverse word embedding model, wherein the reverse word embedding model is a neural network for mapping vectors in a high-dimensional space back to natural language vocabulary;

[0024] Inputting the numerical vector obtained from the user intention prediction result and the user emotion prediction result into the reverse word embedding model to output the words or phrases with the highest similarity to the user intention prediction result and the user emotion prediction result, and using the words or phrases as the output result of the reverse word embedding model;

[0025] Text conversion is performed according to the output result to obtain user intention prediction text and user emotion prediction text.

[0026] In some embodiments, the establishing of a friendliness prediction model based on deep learning, inputting the user emotion prediction text into the friendliness prediction model for friendliness recognition, and obtaining a predicted friendliness level includes:

[0027] Determine the neural network structure, embedding layer, hidden layer and output layer according to the requirements, and establish a neural network model according to the neural network structure, embedding layer, hidden layer and output layer;

[0028] Collect text data with friendliness labels, clean the text data to unify the text format, segment the unified text, and annotate the words with emotional connotations to obtain training text;

[0029] The training text is input into the neural network model and a corresponding loss function is selected for training, until the loss reaches a preset number of iterations to obtain the friendliness prediction model.

[0030] In some embodiments, the establishing of a friendliness prediction model based on deep learning, inputting the user emotion prediction text into the friendliness prediction model for friendliness recognition, and obtaining a predicted friendliness level further includes:

[0031] The user emotion prediction text is subjected to the same preprocessing steps as during training;

[0032] Input the processed text into the embedding layer of the neural network model;

[0033] The embedded text vector is forward propagated through the neural network model, and the probability distribution of the friendliness level is given through the model output layer;

[0034] According to the probability distribution of the model output, the friendliness level with the highest probability is selected as the predicted friendliness level.

[0035] In some embodiments, performing a first-category allocation on the customer service agents who answer the incoming calls based on the predicted friendliness level to obtain a first-category allocation result includes:

[0036] Obtain historical customer service audios corresponding to each friendliness level, and select historical customer service audios having a preset level higher than or equal to the predicted friendliness level from the historical customer service audios according to the predicted friendliness level;

[0037] Confirming the gender of the customer service agent of the selected historical call audio, and determining the gender of the customer service agent preferred by the user of this level of friendliness based on the male-female ratio of the gender of the agent;

[0038] Based on the gender of the customer service agent preferred by the user and the friendliness level, a first type of allocation is performed on the customer service agent who answers the incoming call to obtain a first type of allocation result.

[0039] In some embodiments, the semantic recognition operation is performed on the user intention prediction text based on the NLP language processing algorithm to obtain the corresponding target intention result, including performing a similar question recognition operation:

[0040] Segment the user intention prediction text into intention words, identify the part of speech of each of the intention words, and analyze the dependency relationship between each of the intention words according to the part of speech to obtain a dependency syntactic analysis result;

[0041] Establishing an intention question sentence according to the dependency syntactic analysis result, converting the intention question sentence into a vector form to obtain an intention question sentence vector, and calculating the similarity of the intention question sentence by using the cosine fourth degree;

[0042] The intended questions are grouped according to the similarity calculation results, and similar questions to the user intended prediction text are found to obtain the target intent result.

[0043] In some embodiments, the target agent service includes at least one of the following: knowledge query results for service script invocation, process positioning results for consulting service processes, entity extraction results for filling out business work orders, label retrieval results for describing customer portraits, and service quality inspection results for verifying the quality of agent service.

[0044] In the second aspect, the present application proposes a customer service capability improvement system based on user traffic intention, including a user text acquisition module, an emotional intention prediction module, an emotional intention text acquisition module, a friendliness prediction module, a first allocation module, an intention recognition module and an agent allocation module;

[0045] The user text acquisition module is used to acquire voice data after interaction with the user based on the IVR connection established with the user terminal, and perform voice recognition operation on the voice data through a voice recognition algorithm to obtain user language text content corresponding to the voice data;

[0046] The emotion intention prediction module is used to construct a codec joint prediction model, and input the user language text content into the codec joint prediction model to obtain a user intention prediction result and a user emotion prediction result;

[0047] The emotion intention text acquisition module is used to convert the user intention prediction result and the user emotion prediction result into corresponding user intention prediction text and user emotion prediction text through reverse word embedding;

[0048] The friendliness prediction module is used to establish a friendliness prediction model based on deep learning, input the user emotion prediction text into the friendliness prediction model to perform friendliness recognition, and obtain a predicted friendliness level;

[0049] The first allocation module is used to perform a first type allocation of customer service agents who answer incoming calls based on the predicted friendliness level to obtain a first type allocation result;

[0050] The intention recognition module is used to perform semantic recognition operations on the user intention prediction text based on the NLP language processing algorithm to obtain the corresponding target intention result;

[0051] The agent allocation module is used to perform an identification operation on the target intention result to obtain a target agent service, and automatically route the call to the most suitable agent or department according to the first type of allocation result and the target agent service.

[0052] In a third aspect, the present application proposes an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0053] In a fourth aspect, the present application proposes a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0054] Beneficial effects of the present invention:

[0055] It simplifies the cumbersome process of manual agents recording users and filling out task tickets. By directly building a joint codec prediction model, the user language text content is input into the joint codec prediction model to obtain user intention prediction results and user emotion prediction results, and user problems are collected according to the prediction results. In addition, agents are assigned based on the prediction friendliness level and target intention result identification, which further optimizes the process of work order circulation and can better assist various business departments in transmitting information, assigning tasks, and collaboratively solving problems. Through the above process, the customer service center can efficiently and automatically allocate customer service resources according to user needs and internal resource conditions, thereby improving service quality and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is the overall flow chart of the present invention.

[0057] Figure 2 It is a system principle diagram of the present invention. DETAILED DESCRIPTION

[0058] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein; on the contrary, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0059] In the first aspect, the present application proposes a method for improving customer service capabilities based on user traffic intentions, such as Figure 1 As shown, the following steps are included:

[0060] S100: acquiring voice data after interaction with the user based on an IVR connection established with the user terminal, performing a voice recognition operation on the voice data through a voice recognition algorithm, and obtaining user language text content corresponding to the voice data;

[0061] After the user terminal establishes a connection with the IVR system, the IVR system may first instruct the user to explain the reason for initiating the request in a voice manner, and collect the user's voice through the user terminal. The IVR system performs voice recognition on the user's voice, obtains the voice recognition result, and converts the voice information into text information. The text information in the voice recognition result may contain colloquial expressions, and the voice recognition result can be normalized according to a preset preprocessing method, such as removing modal particles and pause words in the voice recognition result, and determining the preprocessed voice recognition result as the user's language text content.

[0062] S200: constructing a codec joint prediction model, inputting the user language text content into the codec joint prediction model to obtain a user intention prediction result and a user emotion prediction result;

[0063] In some embodiments, the constructing of the codec joint prediction model and inputting the user language text content into the codec joint prediction model to obtain the user intention prediction result and the user emotion prediction result include:

[0064] The encoding and decoding joint prediction model is constructed using a recurrent neural network with BiLSTM as the core;

[0065] The architecture of the codec joint prediction model is a heterogeneous codec architecture, which includes an encoder part, a feature selection part and a decoder part. The encoder part uses RNN to abstract the input sequence data and outputs an abstract vector of a fixed dimension as an abstract representation of the sequence data. The decoder part uses a linear decoder to decode the abstract vector to obtain a sequence representation of the output sequence. The feature selection part selects one of the attention mechanism, PCA dimensionality reduction, regularization, convolution or pooling operations.

[0066] Among them, for recurrent neural networks, adding an attention mechanism can effectively alleviate the problems of gradient explosion and gradient vanishing of recurrent neural networks. By assigning different weights to context information, the model can focus on certain specific values ​​during training, realizing focused modeling of the model. The principle of the encoder-decoder architecture can be abstractly expressed in mathematics as the encoder compressing the model input, using feature extraction to focus on the temporal information of the abstract vector, and using the decoder to decompress and restore the abstract vector;

[0067] The abstract vector formed by the encoder has a fixed dimension, which is prone to information loss. If a certain feature extraction method is added at the abstract vector level, it can effectively alleviate the information loss of the encoder. The attention mechanism is the most effective feature extraction method at present. The addition of the attention mechanism can focus the contextual timing information of the abstract vector abstracted by the model, which is beneficial to the restoration of the contextual timing information and the performance optimization of the decoder for the two subtasks of intent recognition and emotion recognition. For the two subtasks of intent recognition and emotion recognition, both are classification tasks. The input of the model is the word vector, and the output is the classification task. There is a natural distribution gap between the two. The use of isomorphic encoder-decoder architecture faces serious model defects and easily leads to model overfitting. Therefore, this embodiment adopts a heterogeneous encoder-decoder architecture. For the decoder part of the model, an ordinary linear model is used to decode the task, thereby alleviating the problems of overfitting of the model and the huge difference in input and output distribution.

[0068] In some embodiments, the constructing of the codec joint prediction model, inputting the user language text content into the codec joint prediction model to obtain the user intention prediction result and the user emotion prediction result, further includes:

[0069] The encoder in the codec joint prediction model vectorizes the user language text content to obtain an abstract vector;

[0070] Performing a linear transformation on the abstract vector to obtain an intention prediction initial vector and an emotion prediction initial vector;

[0071] Decoding the intention prediction initial vector and the emotion prediction initial vector through a linear decoder to obtain a sequence representation of the vector;

[0072] The softmax function is used to predict the sequence representation of the vector to obtain the user intention prediction results and user emotion prediction results.

[0073] The sentence set of each sentence in the user language text content is represented as , represents the tth sentence in the sentence set, , Indicates the number of words in a sentence (usually the words contained in the longest sentence in the dataset are taken as ,insufficient In the case of multiple words, empty characters are used to fill the space.) Represents the dimension of the word vector in the sentence; The corresponding abstract vector expression obtained by inputting it into the encoder is: , , n represents the total number of sentences in the user text, and the abstract vector The n two-dimensional matrices are identified. The maximum value of each column in each two-dimensional matrix is ​​taken as the current column abstract representation. Each two-dimensional matrix is ​​converted into a one-dimensional vector. The one-dimensional vector is regarded as a sentence vector, thereby reducing the dimension of the abstract vector from Down to , and then the abstract vector is linearly transformed as the initial vector of intention prediction and the sentiment prediction initial vector ,but , ; Input intent recognition and emotion recognition into the fully connected attention neural network for attention calculation. By explicitly learning the attention of intention to emotion and the attention of emotion to intention, the explicit learning purpose of joint modeling is achieved. Finally, the softmax function is used to predict the model category of the output of the linear decoder.

[0074] Furthermore, the model prediction process is as follows: given user text ,in Indicates the first Sentences, , then , . Indicates the current conversation In the sentence words, . Indicates the current conversation The number of words in a sentence, generally Take the maximum number of words contained in a sentence in the dataset. Indicates the number of sentences in the conversation contained in the dataset, generally taking the maximum number of sentences contained in a conversation in the dataset. , they are all input into the BiLSTM encoder composed of a forward LSTM and a backward LSTM for encoding. So the overall operation flow of the model is as follows: the forward LSTM and the backward LSTM are:

[0075]

[0076] in, is the input gate, is the hidden layer vector, For input, is the weight matrix, is the bias value, is a non-linear activation function. For the Gate of Forgetfulness, , is the weight matrix, , is the bias value, is the forgetting coefficient, is the retention coefficient, is a non-linear activation function. is the output gate, is the matrix coefficient, is the bias value, is the current moment memory state, is the output value. For BiLSTM, the forward output is Indicates that the backward output adopts ,pass The encoded abstract vector can be obtained by concatenating the vectors. The abstract vector is recorded as , then:

[0077]

[0078] in, Indicates taking the maximum value in the matrix column as its column abstract vector. When the number of encoder layers of the model is greater than or equal to 2, define for Moment layer hidden layer vector, then there is a forward output , backward output , No. Abstract vector after layer encoding By taking the top-level vector of BiLSTM and concatenating them, we can get the abstract vector of the model stack:

[0079]

[0080] The abstract vectors are linearly changed to obtain the vector expression of the initial vector for intention prediction and the vector expression of the initial vector for emotion prediction. The vector expression of the initial vector for intention prediction is as follows:

[0081]

[0082] The vector expression of the initial vector of sentiment prediction is as follows:

[0083]

[0084] Among them, there are , . Then, the abstract vectors of emotion recognition and intention recognition are calculated by mutual attention mechanism to focus the information. The commonly used attention calculation method is used to calculate the attention of all positions. By superimposing the attention calculation network multiple times, the purpose of learning common knowledge is achieved, thereby enhancing the interpretability and robustness of the model.

[0085] It is worth noting that the purpose of the model is to enhance the joint modeling effect of intent recognition and emotion recognition. The addition of the attention mechanism allows intent recognition and emotion recognition to have an explicit learning process, thereby increasing the robustness and interpretability of the model. Based on this, for the traditional soft attention mechanism, it is necessary to transform the query and key in the attention mechanism accordingly to achieve the purpose of explicit learning. The specific attention formula is:

[0086]

[0087]

[0088]

[0089] in, Indicates Layer network intention recognition is an explicit learning process for emotion recognition. Indicates Layer network emotion recognition is an explicit learning process for intention recognition, and the superposition of explicit processes can not only achieve the purpose of mutual learning, but also alleviate the gradient vanishing dilemma caused by network forgetting, thereby achieving the purpose of improving joint modeling. Indicates layer intent vector, It means the layer sentiment vector, , , These are the query weight matrix, key weight matrix, and value weight matrix of emotional attention. , , It is the query weight matrix, key weight matrix and valu weight matrix of the intention attention. is the regularization parameter, which is generally taken as the word vector dimension.

[0090] The decoder model uses the most common linear model for decoding, and the overall output is as follows:

[0091]

[0092] in, represents the user intent prediction result of the model, Represents the user sentiment prediction results of the model.

[0093] S300: converting the user intention prediction result and the user emotion prediction result into corresponding user intention prediction text and user emotion prediction text through reverse word embedding;

[0094] In some embodiments, converting the user intention prediction result and the user emotion prediction result into corresponding user intention prediction text and user emotion prediction text by reverse word embedding includes:

[0095] Selecting a pre-trained reverse word embedding model, wherein the reverse word embedding model is a neural network for mapping vectors in a high-dimensional space back to natural language vocabulary;

[0096] Inputting the numerical vector obtained from the user intention prediction result and the user emotion prediction result into the reverse word embedding model to output the words or phrases with the highest similarity to the user intention prediction result and the user emotion prediction result, and using the words or phrases as the output result of the reverse word embedding model;

[0097] Text conversion is performed according to the output result to obtain user intention prediction text and user emotion prediction text.

[0098] 1. Reverse Word Embedding

[0099] 1. Model selection: Choose a reverse word embedding model that matches the word embedding space used when training the intent and sentiment prediction model. This model is usually a neural network that can map vectors in a high-dimensional space back to natural language vocabulary.

[0100] 2. Vector mapping: The numerical vectors (usually points in multidimensional space) obtained by intent and sentiment prediction are input into the reverse word embedding model.

[0101] 3. Text generation: The model outputs the word or phrase that is closest to the input vector.

[0102] For example: Suppose the prediction model identifies a numerical vector of user intent, which is mapped to the word "shopping" through the reverse word embedding model. For another example, the numerical vector of sentiment prediction may be mapped to "happy" or "satisfied".

[0103] 2. Result Conversion

[0104] 1. User Intent Prediction Text

[0105] Intent category mapping: According to the intent category output by the prediction model, a predefined intent description library is searched to convert the category into descriptive text. Text generation: Combined with the results of reverse word embedding, a sentence or paragraph of text describing the user's intent is generated.

[0106] For example: The prediction model identifies the intent category as "search for products", and the reverse word embedding maps the relevant vector to "electronic products". The resulting conversion generates text: "The user's intent is to search for electronic products."

[0107] 2. User sentiment prediction text

[0108] Emotion label conversion: Match the predicted emotion labels (such as "happy", "angry", etc.) with the emotion description library. Text description: Combine the results of reverse word embedding to generate text describing the user's emotional state.

[0109] For example, the prediction model determines that the user's emotion is "positive", and the reverse word embedding maps the emotion vector to "excited". The resulting conversion generates text: "The user's emotional state is excited and positive."

[0110] S400: Establishing a friendliness prediction model based on deep learning, inputting the user emotion prediction text into the friendliness prediction model for friendliness recognition, and obtaining a predicted friendliness level;

[0111] In some embodiments, the establishing of a friendliness prediction model based on deep learning, inputting the user emotion prediction text into the friendliness prediction model for friendliness recognition, and obtaining a predicted friendliness level includes:

[0112] Determine the neural network structure, embedding layer, hidden layer and output layer according to the requirements, and establish a neural network model according to the neural network structure, embedding layer, hidden layer and output layer;

[0113] Collect text data with friendliness labels, clean the text data to unify the text format, segment the unified text, and annotate the words with emotional connotations to obtain training text;

[0114] The training text is input into the neural network model and a corresponding loss function is selected for training, until the loss reaches a preset number of iterations to obtain the friendliness prediction model.

[0115] In some embodiments, the establishing of a friendliness prediction model based on deep learning, inputting the user emotion prediction text into the friendliness prediction model for friendliness recognition, and obtaining a predicted friendliness level further includes:

[0116] The user emotion prediction text is subjected to the same preprocessing steps as during training;

[0117] Input the processed text into the embedding layer of the neural network model;

[0118] The embedded text vector is forward propagated through the neural network model, and the probability distribution of the friendliness level is given through the model output layer;

[0119] According to the probability distribution of the model output, the friendliness level with the highest probability is selected as the predicted friendliness level.

[0120] Specific implementation process:

[0121] 1. Model training: Use deep learning techniques (such as neural networks) to train a model that can identify the friendliness of text.

[0122] 2. Friendliness identification: Input the user sentiment prediction text into the model to obtain the friendliness level.

[0123] For example: Input: "The user's emotional state is shown as angry." Predicting friendliness level: The model may output "low friendliness".

[0124] Further refinement of model training and friendliness identification as well as specific implementation processes and examples are provided:

[0125] 1. Model training:

[0126] a. Data collection and preprocessing:

[0127] Collect data: Collect text data with friendliness labels from social media, forums, comments, etc. For example, collect comments on Weibo and their corresponding friendliness labels (such as friendly, unfriendly).

[0128] Data cleaning: remove useless information, such as HTML tags, user tags, etc., and unify the text format.

[0129] Word segmentation and annotation: Segment the text and annotate words with emotional connotations, such as "happy", "angry", etc.

[0130] Build the vocabulary: Create a list of all the words in the training text and assign a unique index to each word.

[0131] b. Build a neural network model:

[0132] Choose the model structure: Decide on the neural network structure to use, such as RNN, LSTM, GRU or Transformer.

[0133] Embedding layer: Use pre-trained word vectors (such as Word2Vec, GloVe, or BERT) as weight initialization for the embedding layer.

[0134] Hidden layer: Design multiple hidden layers and use activation functions such as ReLU and tanh.

[0135] Output layer: According to the number of friendliness levels, softmax is selected as the activation function of the output layer to output the probability distribution of each level.

[0136] c. Training model:

[0137] Loss Function: Cross entropy loss is chosen as the loss function to evaluate the difference between the model prediction and the true label.

[0138] Optimization algorithm: Use optimization algorithms such as Adam and RMSprop to update model weights.

[0139] Hyperparameter tuning: Adjust hyperparameters such as learning rate, batch size, number of hidden layer units, etc. to improve model performance.

[0140] Training process: Iterate the model until the loss drops to an acceptable level or the preset number of iterations is reached.

[0141] d. Model evaluation:

[0142] Validation set: Use the validation set to evaluate model performance and avoid overfitting.

[0143] Evaluation metrics: Use metrics such as accuracy, recall, and F1 score to evaluate the model.

[0144] 2. Friendliness identification:

[0145] a. Text input:

[0146] Data formatting: The user sentiment prediction text is subjected to the same preprocessing steps as during training, such as word segmentation and removal of stop words.

[0147] Embedding processing: The processed text is input into the embedding layer of the neural network model and converted into a vector of fixed length.

[0148] b. Model prediction:

[0149] Forward propagation: The embedded text vector is forward propagated through the neural network model.

[0150] Output results: The model output layer gives the probability distribution of friendliness levels.

[0151] c. Interpretation of results:

[0152] Determine the friendliness level: Based on the probability distribution of the model output, select the friendliness level with the highest probability as the prediction result.

[0153] For example, if the model outputs a text with a friendliness probability distribution of [0.1, 0.3, 0.6] (corresponding to unfriendly, neutral, and friendly levels, respectively), the predicted friendliness level is “friendly”.

[0154] S500: performing a first-category allocation of customer service agents who answer incoming calls based on the predicted friendliness level, and obtaining a first-category allocation result;

[0155] In some embodiments, performing a first-category allocation on the customer service agents who answer the incoming calls based on the predicted friendliness level to obtain a first-category allocation result includes:

[0156] Obtain historical customer service audios corresponding to each friendliness level, and select historical customer service audios having a preset level higher than or equal to the predicted friendliness level from the historical customer service audios according to the predicted friendliness level;

[0157] Confirming the gender of the customer service agent of the selected historical call audio, and determining the gender of the customer service agent preferred by the user of this level of friendliness based on the male-female ratio of the gender of the agent;

[0158] Based on the gender of the customer service agent preferred by the user and the friendliness level, a first type of allocation is performed on the customer service agent who answers the incoming call to obtain a first type of allocation result.

[0159] Specific implementation process:

[0160] 1. Historical audio acquisition: Collect historical customer service call audio and mark the friendliness level.

[0161] Collect data: Collect past call recordings from the customer service center’s recording system.

[0162] Label the data: manually or semi-automatically listen to the recordings and label them with a friendliness rating (e.g., 1 to 5)

[0163] Create a database: Create a database to store audio files and their corresponding friendliness levels.

[0164] For example, 1,000 recordings of calls between customer service and customers were collected and, after annotation, were divided into five levels: ,in Representative The friendliness level of each recording. Specifically, the records in the database can be expressed as: ,in It is audio files, is the corresponding friendship level.

[0165] 2. Audio selection: Based on the predicted friendliness level, select customer service audio of the corresponding level or higher.

[0166] Predicted rating: Use machine learning models to predict the friendliness rating of the current caller.

[0167] Retrieve data: Retrieve matching audio from the database based on the prediction level.

[0168] For example: predicting the current user The friendliness level of the user is level 3. All recordings with a friendliness level of level 3 or higher are retrieved from the database. The formula for predicting the level can be expressed as: ,in is a function that predicts the friendliness level.

[0169] The search conditions can be expressed as: .

[0170] 3. Gender ratio analysis: Analyze the gender ratio of customer service personnel in the selected audio.

[0171] Gender analysis: Use speech recognition and gender recognition technology to determine the gender of the customer service representative in each recording.

[0172] Calculate the ratio: Count the number of customer service representatives of different genders and calculate the ratio.

[0173] For example, among 100 recordings analyzed, there are 60 male customer service representatives and 40 female customer service representatives.

[0174] The gender ratio can be expressed as: and , where M is the number of male customer service representatives and F is the number of female customer service representatives.

[0175] 4. Determination of gender preference: Determine the gender of the customer service agent based on the gender ratio of the user preference.

[0176] Analyze the gender ratio: Analyze the results according to the gender ratio;

[0177] Determine bias: Combine historical data and user feedback to determine the user's gender preference.

[0178] 5. Allocate seats: Allocate seats based on friendliness level and gender preference.

[0179] Determine allocation criteria: Based on gender preference and friendliness level.

[0180] Automatic allocation: Use allocation algorithms to allocate incoming calls to the most suitable customer service agent;

[0181] For example, if the predicted friendliness level of user U is level 3, and the system determines that the user prefers female customer service, then priority is given to customer service staff with a friendliness level of level 3 or higher and who are female. The allocation algorithm can be expressed by the following mathematical expression: ,in is the jth customer service agent, is a function used to calculate customer service With users degree of matching.

[0182] S600: Performing a semantic recognition operation on the user intention prediction text based on an NLP language processing algorithm to obtain the corresponding target intention result;

[0183] In some embodiments, the semantic recognition operation is performed on the user intention prediction text based on the NLP language processing algorithm to obtain the corresponding target intention result, including performing a similar question recognition operation:

[0184] Segment the user intention prediction text into intention words, identify the part of speech of each of the intention words, and analyze the dependency relationship between each of the intention words according to the part of speech to obtain a dependency syntactic analysis result;

[0185] The user intention prediction text is segmented into intention words, expressed as ,in The word sequence of the intent word, It is Intent words, identify the part of speech of each word, the formula is expressed as: , represents a sequence of parts of speech, Is an intention word Part of speech, analyze the dependency between words. It is expressed as middle It is the headword. It is a dependent word. It is the relationship between them;

[0186] Specific examples include:

[0187] User input text: "How do I reset my password?

[0188] Word segmentation result: T={"how","reset","password"}

[0189] Part-of-speech tagging result: POS={"question word""verb""noun"}

[0190] Dependency parsing result: D = {("reset", "how", "question"), ("password", "reset", "object")};

[0191] Establishing an intention question sentence according to the dependency syntactic analysis result, converting the intention question sentence into a vector form to obtain an intention question sentence vector, and calculating the similarity of the intention question sentence by using the cosine fourth degree;

[0192] The intended questions are grouped according to the similarity calculation results, and similar questions to the user intended prediction text are found to obtain the target intent result.

[0193] Similar question identification:

[0194] 1. Vector representation: Convert text into vector form, formula: ,in is a sequence of word vectors, It is the intention word of the intention question The vector representation of .

[0195] 2. Similarity calculation: Use cosine similarity and other algorithms to calculate the similarity of the questions. Formula: cosine similarity , where A and B are the vector representations of the two questions of the intended question;

[0196] 3. Clustering or nearest neighbor search: Group questions based on similarity or find the most similar questions.

[0197] For K nearest neighbor search, finding the K most similar questions can be expressed as ,in, This is one of the questions that is most similar to the target question.

[0198] For example, suppose there are two versions of the intended question: "How do I reset my password?" and "How do I change my password?"

[0199] The vector is represented as:

[0200] and .

[0201] Similarity calculation:

[0202]

[0203] Clustering or nearest neighbor search: Based on the calculated similarity, two questions can be classified into the same category of target intent results, or "How do I change my password?" can be considered a similar question to "How do I reset my password?"

[0204] S700: performing an identification operation on the target intention result to obtain a target agent service, and automatically routing the call to the most suitable agent or department according to the first type of allocation result and the target agent service.

[0205] In some embodiments, the target agent service includes at least one of the following: knowledge query results for service script invocation, process positioning results for consulting service processes, entity extraction results for filling out business work orders, label retrieval results for describing customer portraits, and service quality inspection results for verifying the quality of agent service.

[0206] Specific implementation process:

[0207] a. Based on the intent results, different analysis modules are called to generate corresponding analysis results.

[0208] b. Each analysis result corresponds to a seat service type.

[0209] c. The following are several processes for generating analysis results:

[0210] Knowledge query results: Based on the keywords in the user's intent (such as "international roaming charges"), a search is performed in the knowledge base to generate query results containing information such as roaming charge descriptions and package details.

[0211] Process location results: Analyze the business processes involved in user intentions and locate the corresponding service process nodes, such as the "roaming service consultation" process.

[0212] Entity extraction results: Extract key business entities such as country, time, and cost from user intent and automatically fill them into relevant fields of the business ticket.

[0213] Tag search results: Through user intent analysis, retrieve user-related tags, such as "frequent overseas users" and "cost-sensitive users". Update or create customer profiles.

[0214] Service quality inspection results: Analyze the voice and text records during the agent service process, evaluate the service quality, generate a quality inspection report, and point out potential service deficiencies.

[0215] For example:

[0216] Knowledge query results: The analysis result may be a detailed information page containing roaming charges for the agent to refer to when answering the user's question.

[0217] Process positioning result: The system may indicate that the agent should enter the third step of the "International Roaming Service" process, which is "Provide Charge Details".

[0218] Entity extraction results: The system identifies "Country: United States" and "Duration: 1 month" from the user's question and automatically fills them in the business work order.

[0219] Tag search results: The system adds a tag of "frequently travels to the United States" to the user and updates the customer profile.

[0220] Service quality inspection results: The system analyzes the agent's responses and generates a quality inspection report to indicate whether the agent's explanation of roaming charges was accurate.

[0221] 2. Determine target agent service

[0222] Specific implementation process:

[0223] a. Match the generated analysis results with the preset agent service rules.

[0224] b. Determine the most suitable agent service type based on the matching results.

[0225] c. Transfer or assign the user to an agent who can provide the corresponding service.

[0226] For example:

[0227] If the analysis results show that the user needs to know the detailed international roaming charges, the system will match the "roaming charge consultation" service type and assign it to the agent who is specifically responsible for roaming consultation.

[0228] Through such a process, we can ensure that users' needs are accurately understood and efficiently served.

[0229] 1. Comprehensive allocation results

[0230] Data collection: Collect the first type of allocation results (such as the customer service representative’s gender, professional skills, etc.) and target intent analysis results (the user’s specific needs or problem types).

[0231] Data integration: The collected data is integrated in the system to form a comprehensive information table for routing decisions.

[0232] 2. Routing Decision

[0233] Matching algorithm: Use the preset matching algorithm, combined with the comprehensive information table, to determine the most suitable customer service agent or department. Preset matching algorithm: Use different algorithms to determine the best match. The following are some common matching algorithms:

[0234] Longest matching algorithm: select the agent that best matches the user's needs.

[0235] Example: If a user needs technical support, the system will select the agent with the highest technical skills score.

[0236] Shortest processing time algorithm: select the agent with the smallest current workload.

[0237] Example: If customer service B has the shortest waiting queue, the system will route to customer service B.

[0238] Priority queue algorithm: Matching is performed according to preset priority rules.

[0239] Example: If the user is a VIP customer, the system may match customer service based on VIP priority.

[0240] Priority rules: The system sets priorities based on the following rules:

[0241] Skills match (such as account management expertise) is the highest priority.

[0242] Based on the current customer service workload, customers with lower workload will be given priority.

[0243] Gender is a secondary factor that meets user preferences but does not affect professional skill matching.

[0244] Automatic routing: The system automatically performs routing operations based on decision results.

[0245] Example: Assume that a call management system in a customer service center is processing the following user request:

[0246] The first type of assignment result: Based on user preference, the system recommends assigning a female customer service representative. Target intent analysis result: Through speech recognition and natural language processing, the system analyzes that the user's intention is to change the account password, which is a service request that requires professional knowledge.

[0247] Example result: Suppose customer service A is a female customer service who specializes in handling account issues and currently has only two requests being processed, while customer service B is also a female but handles general consultations and currently has four requests being processed. Based on the above process, the system ultimately decides to route the user's call to customer service A because she meets the best matching criteria in terms of professional skills and current workload.

[0248] In the second aspect, the present application proposes a customer service capability improvement system based on user traffic intention, such as Figure 2 As shown, it includes a user text acquisition module, an emotion intention prediction module, an emotion intention text acquisition module, a friendliness prediction module, a first allocation module, an intention recognition module and an agent allocation module;

[0249] The user text acquisition module is used to acquire voice data after interaction with the user based on the IVR connection established with the user terminal, and perform voice recognition operation on the voice data through a voice recognition algorithm to obtain user language text content corresponding to the voice data;

[0250] The emotion intention prediction module is used to construct a codec joint prediction model, and input the user language text content into the codec joint prediction model to obtain a user intention prediction result and a user emotion prediction result;

[0251] The emotion intention text acquisition module is used to convert the user intention prediction result and the user emotion prediction result into corresponding user intention prediction text and user emotion prediction text through reverse word embedding;

[0252] The friendliness prediction module is used to establish a friendliness prediction model based on deep learning, input the user emotion prediction text into the friendliness prediction model to perform friendliness recognition, and obtain a predicted friendliness level;

[0253] The first allocation module is used to perform a first type allocation of customer service agents who answer incoming calls based on the predicted friendliness level to obtain a first type allocation result;

[0254] The intention recognition module is used to perform semantic recognition operations on the user intention prediction text based on the NLP language processing algorithm to obtain the corresponding target intention result;

[0255] The agent allocation module is used to perform an identification operation on the target intention result to obtain a target agent service, and automatically route the call to the most suitable agent or department according to the first type of allocation result and the target agent service.

[0256] In a third aspect, the present application proposes an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0257] In a fourth aspect, the present application proposes a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0258] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0259] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0260] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this disclosure.

[0261] In the embodiments provided in the present disclosure, it should be understood that the disclosed apparatus / computer equipment and methods can be implemented in other ways. For example, the apparatus / computer equipment embodiments described above are merely schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation. Multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection of the apparatus or unit, which may be electrical, mechanical or other forms.

[0262] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0263] In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0264] If the integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present disclosure implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. The computer program may include computer program code, and the computer program code may be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electric carrier signals and telecommunication signals.

[0265] The above are only preferred implementations of the present invention. It should be pointed out that a number of modifications and improved technical solutions made by those skilled in the art without departing from the technical solution should also be deemed to fall within the scope of protection required by the claims.

Claims

1. A method for improving customer service capabilities based on user traffic intention, characterized by: The following steps are involved: Acquire voice data after interaction with the user based on the IVR connection established with the user terminal, perform voice recognition operation on the voice data through a voice recognition algorithm, and obtain user language text content corresponding to the voice data; Constructing a codec joint prediction model, inputting the user language text content into the codec joint prediction model to obtain a user intention prediction result and a user emotion prediction result; Converting the user intention prediction result and the user emotion prediction result into corresponding user intention prediction text and user emotion prediction text through reverse word embedding; Establishing a friendliness prediction model based on deep learning, inputting the user emotion prediction text into the friendliness prediction model for friendliness recognition, and obtaining a predicted friendliness level; Based on the predicted friendliness level, the customer service agents who answer the calls are assigned to the first category, and the first category assignment results are obtained; Performing semantic recognition operation on the user intention prediction text based on the NLP language processing algorithm to obtain the corresponding target intention result; An identification operation is performed on the target intention result to obtain a target agent service, and the call is automatically routed to the most suitable agent or department according to the first type of allocation result and the target agent service.

2. The method according to claim 1, characterized in that: The constructing of the codec joint prediction model, inputting the user language text content into the codec joint prediction model to obtain a user intention prediction result and a user emotion prediction result, includes: The encoding and decoding joint prediction model is constructed using a recurrent neural network with BiLSTM as the core; The architecture of the codec joint prediction model is a heterogeneous codec architecture, which includes an encoder part, a feature selection part and a decoder part. The encoder part uses RNN to abstract the input sequence data and outputs an abstract vector of a fixed dimension as an abstract representation of the sequence data. The decoder part uses a linear decoder to decode the abstract vector to obtain a sequence representation of the output sequence. The feature selection part selects one of the attention mechanism, PCA dimensionality reduction, regularization, convolution or pooling operations.

3. The method according to claim 2, characterized in that: The constructing of the codec joint prediction model, inputting the user language text content into the codec joint prediction model to obtain the user intention prediction result and the user emotion prediction result, further includes: The encoder in the codec joint prediction model vectorizes the user language text content to obtain an abstract vector; Performing a linear transformation on the abstract vector to obtain an intention prediction initial vector and an emotion prediction initial vector; Decoding the intention prediction initial vector and the emotion prediction initial vector through a linear decoder to obtain a sequence representation of the vector; The softmax function is used to predict the sequence representation of the vector to obtain the user intention prediction results and user emotion prediction results.

4. The method according to claim 3, characterized in that: The converting the user intention prediction result and the user emotion prediction result into corresponding user intention prediction text and user emotion prediction text by reverse word embedding includes: Selecting a pre-trained reverse word embedding model, wherein the reverse word embedding model is a neural network for mapping vectors in a high-dimensional space back to natural language vocabulary; Inputting the numerical vector obtained from the user intention prediction result and the user emotion prediction result into the reverse word embedding model to output the words or phrases with the highest similarity to the user intention prediction result and the user emotion prediction result, and using the words or phrases as the output result of the reverse word embedding model; Text conversion is performed according to the output result to obtain user intention prediction text and user emotion prediction text.

5. The method according to claim 4, characterized in that: The establishing of a friendliness prediction model based on deep learning, inputting the user emotion prediction text into the friendliness prediction model for friendliness recognition, and obtaining a predicted friendliness level includes: Determine the neural network structure, embedding layer, hidden layer and output layer according to the requirements, and establish a neural network model according to the neural network structure, embedding layer, hidden layer and output layer; Collect text data with friendliness labels, clean the text data to unify the text format, segment the unified text, and annotate the words with emotional connotations to obtain training text; The training text is input into the neural network model and a corresponding loss function is selected for training, until the loss reaches a preset number of iterations to obtain the friendliness prediction model.

6. The method according to claim 5, characterized in that: The method of establishing a friendliness prediction model based on deep learning, inputting the user emotion prediction text into the friendliness prediction model for friendliness recognition, and obtaining a predicted friendliness level, further includes: The user emotion prediction text is subjected to the same preprocessing steps as during training; Input the processed text into the embedding layer of the neural network model; The embedded text vector is forward propagated through the neural network model, and the probability distribution of the friendliness level is given through the model output layer; According to the probability distribution of the model output, the friendliness level with the highest probability is selected as the predicted friendliness level.

7. The method according to claim 6, characterized in that: The first type of allocation is performed on the customer service agents who answer the incoming calls based on the predicted friendliness level, and the first type of allocation result is obtained, including: Obtain historical customer service audios corresponding to each friendliness level, and select historical customer service audios having a preset level higher than or equal to the predicted friendliness level from the historical customer service audios according to the predicted friendliness level; Confirming the gender of the customer service agent of the selected historical call audio, and determining the gender of the customer service agent preferred by the user of this level of friendliness based on the male-female ratio of the gender of the agent; Based on the gender of the customer service agent preferred by the user and the friendliness level, a first type of allocation is performed on the customer service agent who answers the incoming call to obtain a first type of allocation result.

8. The method according to claim 7, characterized in that: The semantic recognition operation is performed on the user intention prediction text based on the NLP language processing algorithm to obtain the corresponding target intention result, including performing a similar question recognition operation: Segment the user intention prediction text into intention words, identify the part of speech of each of the intention words, and analyze the dependency relationship between each of the intention words according to the part of speech to obtain a dependency syntactic analysis result; Establishing an intention question sentence according to the dependency syntactic analysis result, converting the intention question sentence into a vector form to obtain an intention question sentence vector, and calculating the similarity of the intention question sentence by using the cosine fourth degree; The intended questions are grouped according to the similarity calculation results, and similar questions to the user intended prediction text are found to obtain the target intent result.

9. The method according to claim 7, characterized in that: The target agent service includes at least one of the following: knowledge query results for service script invocation, process positioning results for consulting service processes, entity extraction results for filling out business work orders, label retrieval results for describing customer portraits, and service quality inspection results for verifying agent service quality.

10. A customer service capability improvement system based on user traffic intention, characterized by: It includes a user text acquisition module, an emotion intention prediction module, an emotion intention text acquisition module, a friendliness prediction module, a first allocation module, an intention recognition module and an agent allocation module; The user text acquisition module is used to acquire voice data after interaction with the user based on the IVR connection established with the user terminal, and perform voice recognition operation on the voice data through a voice recognition algorithm to obtain user language text content corresponding to the voice data; The emotion intention prediction module is used to construct a codec joint prediction model, and input the user language text content into the codec joint prediction model to obtain a user intention prediction result and a user emotion prediction result; The emotion intention text acquisition module is used to convert the user intention prediction result and the user emotion prediction result into corresponding user intention prediction text and user emotion prediction text through reverse word embedding; The friendliness prediction module is used to establish a friendliness prediction model based on deep learning, input the user emotion prediction text into the friendliness prediction model to perform friendliness recognition, and obtain a predicted friendliness level; The first allocation module is used to perform a first type allocation of customer service agents who answer incoming calls based on the predicted friendliness level to obtain a first type allocation result; The intention recognition module is used to perform semantic recognition operations on the user intention prediction text based on the NLP language processing algorithm to obtain the corresponding target intention result; The agent allocation module is used to perform an identification operation on the target intention result to obtain a target agent service, and automatically route the call to the most suitable agent or department according to the first type of allocation result and the target agent service.

Images