Seat unstructured label generation method and device, equipment and storage medium

Abstract

The invention relates to a seat unstructured label generation method, which comprises the following steps of: acquiring voice text data of a seat, coding the voice text data into a sentence vector, and forming a position embedded sentence vector in combination with a position code; processing the position embedded sentence vector through a deep learning model, and identifying and classifying the objection type of each sentence; performing processing intention and processing capability evaluation on the sentences belonging to the objection type, and generating an objection processing intention score and a processing capability score of the seat; and finally, generating an unstructured label reflecting the comprehensive performance of the seat according to the scores. According to the invention, the performance of the seat in processing customer objection can be automatically identified and evaluated, the generated unstructured label can comprehensively and accurately reflect the comprehensive capability of the seat, the management efficiency of a customer service system is improved, the customer resource allocation is optimized, and the accuracy and satisfaction of customer service are enhanced.

Description

Technical Field

[0001] The present invention relates to the fields of natural language processing and machine learning technology and financial technology, and in particular to a method, device, equipment and storage medium for generating unstructured labels for seats. Background Art

[0002] With the continuous development of information technology and the popularization of the Internet, more and more companies are beginning to use call centers and customer service systems to manage customer relationships and provide service support. In these systems, the work quality of agents (customer service personnel) directly affects customer satisfaction and the company's service efficiency. However, due to the diversity and complexity of customer needs, there are significant differences in the performance of agents when handling different customer objections. How to effectively evaluate the work performance of agents and optimize customer resource allocation has become a problem that companies need to solve urgently.

[0003] Traditional customer service systems usually rely on manual management of agents and customer resources, mainly through manual evaluation and experience judgment to determine customer allocation and agent training. However, this method is often inefficient, unable to accurately reflect the performance of agents in specific business scenarios, and difficult to cope with complex changes in customer needs. At the same time, with the rapid development of artificial intelligence technology, more and more companies hope to use natural language processing (NLP) and deep learning technology to improve the intelligence level of customer service systems, so as to automatically and accurately process the interaction data between customers and agents.

[0004] In recent years, natural language processing technology, especially text embedding and semantic analysis methods based on deep learning, has made significant progress. These technologies can extract useful information from a large amount of customer-agent conversation data and generate labels that reflect agent performance, thereby supporting the intelligent allocation of customer resources.

[0005] At present, although there are some intelligent customer service systems based on NLP, most of them are still limited to simple text analysis and keyword matching, unable to deeply understand the specific content of customer objections, and difficult to achieve targeted customer resource allocation. In addition, how to quickly and accurately evaluate the performance of agents in a large-scale call center environment in real time and use the evaluation results to optimize customer allocation and agent training is still an urgent problem to be solved. Therefore, the development of an intelligent agent label generation and customer allocation system based on natural language processing and machine learning technology has important application value and market demand. Summary of the invention

[0006] The main purpose of the present invention is to provide a method, device, equipment and storage medium for generating unstructured labels for agents, aiming to solve the technical problem of how to automatically identify and evaluate the performance of agents in handling customer objections and generate unstructured labels for optimized management.

[0007] To achieve the above object, the present invention provides a method for generating unstructured seat tags, the method comprising the following steps:

[0008] Acquire the voice text data of the seat, encode each sentence in the voice text data into a sentence vector, and combine the sentence vector with the position encoding of the sentence to form a position embedded sentence vector;

[0009] Processing the position embedding sentence vectors through a deep learning model to identify and classify the objection type of each sentence;

[0010] The willingness and ability to handle each objection sentence belonging to the objection type are evaluated to generate an objection handling willingness score and an objection handling ability score of the agent, and an unstructured label of the agent is generated according to the objection handling willingness score and the objection handling ability score.

[0011] Optionally, obtain the agent's voice text data, including:

[0012] Acquire the call voice data of the seat that meets the preset conditions, and convert the call voice data of the seat into the voice text data of the seat through automatic speech recognition technology.

[0013] Optionally, the step of processing the position embedding sentence vector by a deep learning model to identify and classify the objection type of each sentence includes:

[0014] The position embedding sentence vector is input into the Transformer encoder to obtain sentence vector information, and then the sentence vector information is predicted through the fully connected layer and the softmax network layer to identify and classify the objection type of each sentence.

[0015] Optionally, the willingness and ability to handle each objection sentence belonging to the objection type are evaluated to generate the objection handling willingness score and objection handling ability score of the agent, including:

[0016] Input the objection sentences into the preset large model, and identify whether each objection sentence has been processed and whether it has been processed in accordance with the requirements in combination with the preset prompt words;

[0017] Count the total number of objections for all objection sentences of all objection types, the number of objection sentences of each objection type, the number of resolutions that were resolved in the objection sentences of each objection type, and the number of qualified resolutions that were resolved in the objection sentences of each objection type;

[0018] The handling willingness score and handling ability score of each objection type are calculated based on the statistical quantity, and the handling willingness score and handling ability score of the seat are generated based on the handling willingness score and handling ability score of the preset objection type.

[0019] Optionally, the handling willingness score and handling ability score of each objection type are calculated based on the statistical quantity, including:

[0020] The willingness to handle scores for each objection type are:

[0021]

[0022] The handling ability scores for each objection type are:

[0023]

[0024] Optionally, before inputting the objection sentences into the preset large model and identifying whether each objection sentence has been processed and whether the processing has met the requirements in combination with the preset prompt words, the method further includes:

[0025] The LoRA technology is used to fine-tune the preset large model. The fine-tuning process includes freezing other parameters in the preset large model except the target parameters, and only superimposing low-rank adaptation parameters on the specific layer where the target parameters are located to complete the fine-tuning of the preset large model.

[0026] Optionally, after generating the unstructured label of the agent according to the objection handling willingness score and the objection handling ability score, the method further includes:

[0027] generating a customer tag according to the customer demand information, wherein the customer tag includes an objection type related to the customer demand;

[0028] Matching the customer tag with the unstructured tag of the agent, calculating the matching degree between the customer tag and the unstructured tag of each agent, and determining the matching priority between the customer and each agent based on the matching degree;

[0029] Assign the customer to the agent with the highest matching priority.

[0030] Furthermore, in order to achieve the above-mentioned object, the present invention also provides a device for generating unstructured labels for seats, the device for generating unstructured labels for seats comprising:

[0031] A text embedding and encoding module obtains the voice text data of the agent, encodes each sentence in the voice text data into a sentence vector, and combines the sentence vector with the position encoding of the sentence to form a position embedded sentence vector;

[0032] An objection segment identification module, which processes the position embedding sentence vector through a deep learning model to identify and classify the objection type of each sentence;

[0033] The objection handling capability mining module evaluates the objection handling willingness and capability for each objection sentence belonging to the objection type, generates the objection handling willingness score and objection handling capability score of the agent, and generates an unstructured label for the agent based on the objection handling willingness score and the objection handling capability score.

[0034] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a seat unstructured label generation device, which includes a memory, a processor, and an agent unstructured label generation program stored in the memory and executable on the processor, and when the agent unstructured label generation program is executed by the processor, the steps of the agent unstructured label generation method as described above are implemented.

[0035] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer storage medium, on which is stored a seat unstructured label generation program, and when the seat unstructured label generation program is executed by a processor, the steps of the seat unstructured label generation method as described above are implemented.

[0036] The present invention relates to a method for generating unstructured labels for agents, which obtains the voice text data of the agents, encodes it into sentence vectors, and combines position encoding to form position-embedded sentence vectors; processes the position-embedded sentence vectors through a deep learning model, identifies and classifies the objection type of each sentence; evaluates the willingness and ability to handle sentences belonging to the objection type, and generates the agent's objection handling willingness score and processing ability score; finally, generates unstructured labels reflecting the comprehensive performance of the agents based on these scores. The present invention can automatically identify and evaluate the performance of agents in handling customer objections, and the generated unstructured labels can comprehensively and accurately reflect the comprehensive capabilities of the agents, improve the management efficiency of the customer service system, optimize customer resource allocation, and enhance the accuracy and satisfaction of customer service. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A schematic diagram of the structure of the device hardware operating environment involved in the embodiment of the device for generating unstructured labels for seats of the present invention;

[0038] Figure 2 This is a flow chart of a first embodiment of a method for generating unstructured labels for seats according to the present invention;

[0039] Figure 3 The figure is a functional module diagram of a preferred embodiment of the device for generating unstructured labels for seats of the present invention.

[0040] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0041] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0042] It should be noted that with the continuous development of information technology and the popularization of the Internet, more and more companies are beginning to use call centers and customer service systems to manage customer relationships and provide service support. In these systems, the work quality of agents (customer service personnel) directly affects customer satisfaction and the service efficiency of the company. However, due to the diversity and complexity of customer needs, there are significant differences in the performance of agents when handling different customer objections. How to effectively evaluate the work performance of agents and optimize customer resource allocation has become a problem that companies need to solve urgently.

[0043] Traditional customer service systems usually rely on manual management of agents and customer resources, mainly through manual evaluation and experience judgment to determine customer allocation and agent training. However, this method is often inefficient, unable to accurately reflect the performance of agents in specific business scenarios, and difficult to cope with complex changes in customer needs. At the same time, with the rapid development of artificial intelligence technology, more and more companies hope to use natural language processing (NLP) and deep learning technology to improve the intelligence level of customer service systems, so as to automatically and accurately process the interaction data between customers and agents.

[0044] In recent years, natural language processing technology, especially text embedding and semantic analysis methods based on deep learning, has made significant progress. These technologies can extract useful information from a large amount of customer-agent conversation data and generate labels that reflect agent performance, thereby supporting the intelligent allocation of customer resources.

[0045] At present, although there are some intelligent customer service systems based on NLP, most of them are still limited to simple text analysis and keyword matching, unable to deeply understand the specific content of customer objections, and difficult to achieve targeted customer resource allocation. In addition, how to quickly and accurately evaluate the performance of agents in a large-scale call center environment in real time and use the evaluation results to optimize customer allocation and agent training is still an urgent problem to be solved. Therefore, the development of an intelligent agent label generation and customer allocation system based on natural language processing and machine learning technology has important application value and market demand.

[0046] Based on the above defects, the present invention provides a seat unstructured label generation device, referring to Figure 1 , Figure 1This is a schematic diagram of the structure of the device hardware operating environment involved in the embodiment of the device for generating unstructured labels for seats of the present invention.

[0047] like Figure 1 As shown, the seat unstructured label generation device may include: a processor 1001, such as a CPU, a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the user interface 1003 may optionally include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a stable memory (non-volatile memory), such as a disk memory. The memory 1005 may optionally be a storage device independent of the aforementioned processor 1001.

[0048] Those skilled in the art will understand that Figure 1 The hardware structure of the seat unstructured label generating device shown in the figure does not constitute a limitation on the seat unstructured label generating device, and may include more or less components than shown in the figure, or combine certain components, or arrange the components differently.

[0049] like Figure 1 As shown, the memory 1005 as a storage medium may include an operating system, a network communication module, a user interface module, and an agent unstructured label generation program. The operating system is a program for managing and controlling the agent unstructured label generation device and software resources, and supports the operation of the network communication module, the user interface module, the agent unstructured label generation program, and other programs or software; the network communication module is used to manage and control the network interface 1004; and the user interface module is used to manage and control the user interface 1003.

[0050] exist Figure 1 In the hardware structure of the seat unstructured label generation device shown, the network interface 1004 is mainly used to connect to the backend server and perform data communication with the backend server; the user interface 1003 is mainly used to connect to the client and perform data communication with the client; the processor 1001 can call the seat unstructured label generation program stored in the memory 1005 and perform the following operations:

[0051] Acquire the voice text data of the seat, encode each sentence in the voice text data into a sentence vector, and combine the sentence vector with the position encoding of the sentence to form a position embedded sentence vector;

[0052] Processing the position embedding sentence vectors through a deep learning model to identify and classify the objection type of each sentence;

[0053] The willingness and ability to handle each objection sentence belonging to the objection type are evaluated to generate an objection handling willingness score and an objection handling ability score of the agent, and an unstructured label of the agent is generated according to the objection handling willingness score and the objection handling ability score.

[0054] Furthermore, the voice and text data of the agent are obtained, including:

[0055] Acquire the call voice data of the seat that meets the preset conditions, and convert the call voice data of the seat into the voice text data of the seat through automatic speech recognition technology.

[0056] Furthermore, the step of processing the position embedding sentence vector by a deep learning model to identify and classify the objection type of each sentence includes:

[0057] The position embedding sentence vector is input into the Transformer encoder to obtain sentence vector information, and then the sentence vector information is predicted through the fully connected layer and the softmax network layer to identify and classify the objection type of each sentence.

[0058] Furthermore, the willingness and ability to handle each objection sentence belonging to the objection type are evaluated to generate the objection handling willingness score and objection handling ability score of the seat, including:

[0059] Input the objection sentences into the preset large model, and identify whether each objection sentence has been processed and whether it has been processed in accordance with the requirements in combination with the preset prompt words;

[0060] Count the total number of objections for all objection sentences of all objection types, the number of objection sentences of each objection type, the number of resolutions that were resolved in the objection sentences of each objection type, and the number of qualified resolutions that were resolved in the objection sentences of each objection type;

[0061] The handling willingness score and handling ability score of each objection type are calculated based on the statistical quantity, and the handling willingness score and handling ability score of the seat are generated based on the handling willingness score and handling ability score of the preset objection type.

[0062] Furthermore, the handling willingness score and handling ability score of each objection type are calculated based on the statistical quantity, including:

[0063] The willingness to handle scores for each objection type are:

[0064]

[0065] The handling ability scores for each objection type are:

[0066]

[0067] Furthermore, before inputting the objection sentence into the preset large model and identifying whether each objection sentence has been processed and whether the processing has met the requirements in combination with the preset prompt words, it also includes:

[0068] The LoRA technology is used to fine-tune the preset large model. The fine-tuning process includes freezing other parameters in the preset large model except the target parameters, and only superimposing low-rank adaptation parameters on the specific layer where the target parameters are located to complete the fine-tuning of the preset large model.

[0069] Furthermore, after generating the unstructured label of the agent according to the objection handling willingness score and the objection handling ability score, the method further includes:

[0070] generating a customer tag according to the customer demand information, wherein the customer tag includes an objection type related to the customer demand;

[0071] Matching the customer tag with the unstructured tag of the agent, calculating the matching degree between the customer tag and the unstructured tag of each agent, and determining the matching priority between the customer and each agent based on the matching degree;

[0072] Assign the customer to the agent with the highest matching priority.

[0073] The specific implementation of the device for generating unstructured seat labels of the present invention is substantially the same as the embodiments of the method for generating unstructured seat labels described below, and will not be described in detail herein.

[0074] The present invention also provides a method for generating unstructured seat tags based on the unstructured seat tag generating device.

[0075] Reference Figure 2 , Figure 2 The figure is a flowchart of the first embodiment of the method for generating unstructured labels for seats according to the present invention.

[0076] The embodiment of the present invention provides an embodiment of the method for generating unstructured labels for agents. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in an order different from that shown here.

[0077] In various embodiments of the seat unstructured label generation method, the execution subject is a seat unstructured label generation device.

[0078] The seat unstructured label generation method is applied to a seat unstructured label generation device, and the seat unstructured label generation method includes:

[0079] Step S10, obtaining the voice text data of the seat, encoding each sentence in the voice text data into a sentence vector, and combining the sentence vector with the position encoding of the sentence to form a position embedded sentence vector;

[0080] With the continuous development of information technology and the popularization of the Internet, more and more companies are beginning to use call centers and customer service systems to manage customer relationships and provide service support. In these systems, the work quality of agents (customer service personnel) directly affects customer satisfaction and the company's service efficiency. However, due to the diversity and complexity of customer needs, there are significant differences in the performance of agents when handling different customer objections. How to effectively evaluate the work performance of agents and optimize customer resource allocation has become a problem that companies need to solve urgently.

[0081] Traditional customer service systems usually rely on manual management of agents and customer resources, mainly through manual evaluation and experience judgment to determine customer allocation and agent training. However, this method is often inefficient, unable to accurately reflect the performance of agents in specific business scenarios, and difficult to cope with complex changes in customer needs. At the same time, with the rapid development of artificial intelligence technology, more and more companies hope to use natural language processing (NLP) and deep learning technology to improve the intelligence level of customer service systems, so as to automatically and accurately process the interaction data between customers and agents.

[0082] In recent years, natural language processing technology, especially text embedding and semantic analysis methods based on deep learning, has made significant progress. These technologies can extract useful information from a large amount of customer-agent conversation data and generate labels that reflect agent performance, thereby supporting the intelligent allocation of customer resources.

[0083] At present, although there are some intelligent customer service systems based on NLP, most of them are still limited to simple text analysis and keyword matching, unable to deeply understand the specific content of customer objections, and difficult to achieve targeted customer resource allocation. In addition, how to quickly and accurately evaluate the performance of agents in a large-scale call center environment in real time and use the evaluation results to optimize customer allocation and agent training is still an urgent problem to be solved. Therefore, the development of an intelligent agent label generation and customer allocation system based on natural language processing and machine learning technology has important application value and market demand.

[0084] In this embodiment, voice recognition technology (ASR, Automatic Speech Recognition) can be used to convert voice data into text data in real time while the agent is talking to the customer. This method is suitable for scenarios that require immediate processing and analysis. It is also possible to obtain historical call recording files from the call center's database or storage system, and then use voice recognition technology to convert these recording files into text data. It is also possible to obtain transcribed text data directly from the call center system. These data may be previous call records that have been transcribed by the voice recognition system and stored in the text database. Similarly, the audio data of the agent's call can be uploaded to the cloud through a third-party voice-to-text service provider, and the third-party platform will perform voice-to-text processing and return text data.

[0085] Convert each sentence in the text data into a vector (i.e., a digital representation), usually using a pre-trained word embedding model (such as BERT, GPT, etc.). These vectors represent the semantic information of the sentence, making it easier for the model to understand and process it.

[0086] The vector representation of each sentence is combined with the position encoding information to form a position embedded sentence vector. Position encoding helps the model understand the relative position of the sentence in the text, which is especially important for processing contextual information and sequence data.

[0087] Specific implementations may be:

[0088] Implementation method 1: Acquiring voice data: When an agent is talking to a customer, voice data is acquired through a real-time monitoring system of a call center.

[0089] Speech-to-text: Through the built-in ASR system, voice data is converted into text data in real time.

[0090] Sentence encoding: Input the transformed text data into a pre-trained deep learning model, such as BERT, to encode each sentence into a sentence vector.

[0091] Position encoding: Perform position encoding on the generated sentence vector and convert it into a position embedded sentence vector.

[0092] Implementation method 2: Acquiring voice data: extracting historical call recording files from the database of the call center.

[0093] Speech-to-text: Use a dedicated ASR tool to convert these recording files into text data in batches.

[0094] Sentence encoding: Encode the sentences in the text data through a deep learning model (such as GPT) to generate sentence vectors.

[0095] Position encoding: The position encoding algorithm is applied to combine the sentence vector with its position information in the text to form a position embedded sentence vector.

[0096] Implementation method 3: Obtain voice data: Upload the call recording to a third-party speech recognition service (such as Google Cloud Speech-to-Text).

[0097] Speech-to-text: Use a third-party service to transcribe speech into text data and then download it to your local computer.

[0098] Sentence encoding: Import the downloaded text data into a word embedding model, such as Word2Vec, to encode the sentences into vectors.

[0099] Position encoding: Encode the position information of the sentence to obtain the position embedded sentence vector.

[0100] Example description:

[0101] Assume that a call center needs to process the call records of agent A. This can be achieved by following the steps below:

[0102] Step 1: Obtain voice text data: The system obtains audio data from the real-time call of agent A and converts it into the following text through ASR technology:

[0103] Customer: I am not satisfied with the delivery speed of the product.

[0104] Agent: We will handle it as soon as possible, please rest assured.

[0105] Step 2: Sentence vectorization: Use the pre-trained BERT model to transform the sentence "I am not satisfied with the delivery speed of the product" into a vector [0.23, -0.15, 0.78, ...].

[0106] Step 3: Position encoding: Take "I am not satisfied with the delivery speed of the product" as the first sentence of the conversation, add position encoding to it, and combine to generate the position embedding sentence vector [0.45, -0.20, 0.85, ...].

[0107] Step 4: Processing: The positional embedded sentence vector is input into the downstream deep learning model for identifying the objection type and further processing.

[0108] By obtaining the voice text data of the agent and converting it into a processable sentence vector, and then combining it with position encoding to form a position embedded sentence vector, the structured processing of text data is achieved. This method can effectively capture the semantic information and position relationship in the conversation, and provide accurate input data for subsequent deep learning model analysis. It can more accurately identify and classify customer objection types, and generate unstructured labels reflecting agent performance based on these data. It significantly improves the intelligence level of the customer service system in handling complex contexts and dynamic conversations, helps optimize customer resource allocation and agent management, and ultimately improves customer satisfaction and service efficiency.

[0109] Step S20, processing the position embedded sentence vector through a deep learning model to identify and classify the objection type of each sentence;

[0110] In this embodiment, the encoded sentence vectors containing semantic information and position information are input into a deep learning model for processing. A deep learning model, such as a Transformer or other neural network architecture, will perform semantic understanding and feature extraction on these vectors to capture important information in the sentence.

[0111] After being processed by the deep learning model, the system can identify whether each sentence belongs to a certain type of customer objection (such as price objection, quality objection, etc.) and classify it into the corresponding category. This process is usually completed through a classifier, which can give a probability prediction based on the Softmax layer or other output layers.

[0112] Specific implementations may be:

[0113] Implementation 1: Input the position embedding sentence vector into the pre-trained Transformer model. The model uses the self-attention mechanism to model the relationship between sentences and extract features related to objections. The objection classification result of each sentence is output through the last layer of the model (such as the Softmax layer). It is suitable for processing large-scale and complex customer conversation data and can capture long-distance semantic dependencies.

[0114] Implementation 2: The position embedding sentence vector is input into the bidirectional LSTM model (BiLSTM), which processes sentence sequence information in both the forward and backward directions. The hidden state of LSTM can capture the semantic dependency of sentences in context. The output of LSTM is classified using the fully connected layer and the Softmax layer to identify the objection type of the sentence. It is suitable for processing conversation data with time series characteristics, especially when there is a strong sequential dependency between sentences.

[0115] Implementation 3: Input the position embedding sentence vector into the convolutional neural network (CNN) and extract local features through the convolution kernel. Extract the key features of the sentence through multi-layer convolution and pooling operations. Classify the features through the Softmax layer to identify the type of objection. It is suitable for processing dialogue data with strong structure and obvious local features, and can quickly extract and classify features.

[0116] Example description:

[0117] Input model: Input the previously generated positional embedding sentence vector into the pre-trained Transformer model. For example, after vectorization and positional encoding, the sentence "I think your product prices are too high" generates a vector of [0.25, -0.03, 0.67, ...], which is passed as input to the Transformer model.

[0118] Feature extraction and modeling: The Transformer model processes the input vector and uses the self-attention mechanism to capture the features related to "price" in the sentence. The multi-layer structure of the model enables it to understand the context and semantic relationships of the sentence, especially the expression of objection such as "the price is too high".

[0119] Classification output: After multiple layers of processing, the classifier in the last layer of the model (such as the Softmax layer) will generate a probability distribution to determine which objection type the sentence belongs to. For example, the model output may be [0.05 (no objection), 0.10 (quality objection), 0.80 (price objection), ...], where the highest probability corresponds to the "price objection" type.

[0120] Objection handling: Based on the classification results, the system attaches a "price objection" label to the sentence, preparing it for further processing willingness and processing ability evaluation. These classification results will be input into subsequent modules to generate unstructured labels that reflect the agent's processing performance.

[0121] By using a deep learning model to process position-embedded sentence vectors, we can accurately identify and classify the types of objections in customer conversations. Whether using the Transformer, LSTM, or CNN model, we can accurately determine the type of customer objection by deeply analyzing the semantics and positional relationships in the conversation content. This not only improves the efficiency of objection handling, but also provides accurate data support for subsequent customer service optimization and resource allocation, ultimately improving customer satisfaction and the company's service quality.

[0122] Step S30, evaluating the willingness and ability to handle each objection sentence belonging to the objection type, generating an objection handling willingness score and an objection handling ability score for the agent, and generating an unstructured label for the agent based on the objection handling willingness score and the objection handling ability score.

[0123] In this embodiment, the identified objection type sentences are further analyzed to evaluate the agent's performance in dealing with these objections. The handling willingness evaluation focuses on whether the agent responds to the customer's objection in a timely and positive manner, while the handling ability evaluation focuses on the effectiveness of the agent in handling the objection.

[0124] Each objection type is evaluated for willingness and ability, and a corresponding score is generated. These scores quantify the agent's performance in handling customer objections. The higher the score, the more proactive and effective the agent's performance.

[0125] One or more unstructured labels reflecting the comprehensive performance of the agent are generated based on the above scores. These labels can describe the characteristics and capabilities of the agent in handling specific types of objections, such as "good at handling price objections" and "fast response speed".

[0126] Specific implementations may be:

[0127] Implementation method 1: For each objection sentence, check the agent's response content and time. Set predefined rules (such as response time within 5 seconds means high willingness to handle) to evaluate handling willingness. Evaluate handling ability based on customer feedback on the agent's response or response results (such as whether the problem is solved). Convert these rules into scores to generate handling willingness scores and handling ability scores. Suitable for relatively standardized customer service processes, easy to implement and apply quickly.

[0128] Implementation method 2: Use a pre-trained deep learning model (such as BERT or Transformer) to analyze the agent's response text and automatically extract features related to willingness and ability to handle. The model is trained based on a large amount of historical data and can identify willingness and ability to handle in complex contexts. Generate a willingness and ability score for each objection sentence, and use these scores to further generate an overall score. Suitable for large-scale data processing and complex customer service scenarios, it can capture subtle semantic differences and performance features.

[0129] Implementation method 3: Evaluate the agent's handling effect based on the customer's feedback after the conversation (such as satisfaction survey). Combine the customer's subjective feedback with objective data (such as response time and resolution rate) to generate a handling willingness score and a handling ability score. The system generates corresponding unstructured tags based on the comprehensive score. It is suitable for service scenarios centered on customer experience and can directly reflect the customer's perception of the agent's performance.

[0130] Example description:

[0131] Assume that in a call center, the call record between agent A and customer B has been processed through the previous steps, and the system has identified the following types of objection sentences:

[0132] Customer Sentence: I think your product is too expensive.

[0133] The agent responded: The quality of our products is indeed very good, and we provide after-sales guarantee, and the price is reasonable.

[0134] Step 1: Assessment of willingness to process

[0135] Time response analysis: The system first analyzes the speed of the agent's response. Suppose agent A responds within 3 seconds after the customer raises an objection. The system compares this response time with the predefined handling willingness standard. For example, a response time of less than 5 seconds can be considered to have a high handling willingness.

[0136] Semantic analysis: The system further analyzes the agent's response and uses a deep learning model to identify whether Agent A mentioned a solution or explanation for the objection in his response. In this case, the agent mentioned product quality and after-sales guarantee, and the system regarded this response as a positive response.

[0137] Generate a willingness to handle score: Based on the above analysis, the system generates a willingness to handle score for Agent A's response. For example, the system sets the willingness to handle score to 0.85 (out of a maximum score of 1) based on the response speed and content quality.

[0138] Step 2: Processing Capacity Assessment

[0139] Customer feedback analysis: The system then analyzes customer B's feedback during the conversation to determine whether the customer accepts or approves the agent's response. If the customer then says "I still think it's a bit expensive, but I'll consider it," this indicates that the customer partially accepts the agent's explanation. The system compares this feedback with the standard of handling ability and believes that agent A is more capable in handling this objection.

[0140] Solution quality analysis: The system can also further evaluate whether the solution provided by the agent meets the company's standards or is innovative. Assume that the solution provided by agent A is consistent with the company's recommendation, and the processing ability score is improved.

[0141] Generate a processing capability score: Based on customer feedback and solution quality, the system generates a score for agent A's processing capability. For example, the processing capability score is 0.75.

[0142] Step 3: Generate unstructured tags

[0143] Label generation: Based on the generated willingness score (0.85) and ability score (0.75), the system classifies Agent A's performance as "high willingness, strong ability". The system generates an unstructured label, such as "good at handling price objections, active in handling but average customer acceptance".

[0144] Tag application: These tags will be recorded in Agent A’s profile for future customer resource allocation or performance evaluation. For example, the next time a similar price objection is encountered, the system may prioritize assigning such customers to Agent A.

[0145] By evaluating the willingness and ability to handle the identified objection type sentences, the agent's objection handling willingness score and handling ability score are generated, and unstructured tags are generated based on these scores. It can quantify and fully reflect the agent's performance in handling customer objections, which not only improves the transparency and management efficiency of the customer service process, but also optimizes customer resource allocation based on these tags, improves customer satisfaction and agent performance. Through this automated evaluation and label generation mechanism, the deviation of manual intervention is reduced, the system's intelligence and decision-making support capabilities are enhanced, and it helps companies better manage customer service teams and improve overall service quality.

[0146] The present invention relates to a method for generating unstructured labels for agents, which obtains the voice text data of the agents, encodes it into sentence vectors, and combines position encoding to form position-embedded sentence vectors; processes the position-embedded sentence vectors through a deep learning model, identifies and classifies the objection type of each sentence; evaluates the willingness and ability to handle sentences belonging to the objection type, and generates the agent's objection handling willingness score and processing ability score; finally, generates unstructured labels reflecting the comprehensive performance of the agents based on these scores. The present invention can automatically identify and evaluate the performance of agents in handling customer objections, and the generated unstructured labels can comprehensively and accurately reflect the comprehensive capabilities of the agents, improve the management efficiency of the customer service system, optimize customer resource allocation, and enhance the accuracy and satisfaction of customer service.

[0147] Furthermore, a second embodiment of the method for generating unstructured labels for agents of the present invention is proposed. In the above step S10, the voice text data of the agent is obtained, including:

[0148] Step a1, acquiring the call voice data of the agent that meets the preset conditions, and converting the call voice data of the agent into the voice text data of the agent through automatic speech recognition technology.

[0149] In this embodiment, call data of agents can be collected from preset specific areas and time ranges. The preset area may include a geographic area, a service area, etc., and the preset time includes the specific time period or duration of the call. This screening method ensures that the collected call data is relevant to specific analysis needs. By presetting conditions for area and time, the system can focus on analyzing agent performance in specific situations, such as customer service performance during a peak period or in a specific area, so as to achieve targeted optimization.

[0150] Automatic Speech Recognition (ASR) technology is used to convert the voice content of the call into processable text data. ASR technology can recognize and transcribe the voice content, so that the system can perform subsequent analysis and processing on the text data.

[0151] Specific implementations may be:

[0152] Implementation method 1: According to the preset geographical area (such as a province or city) and working hours (such as 9:00 to 18:00), the qualified voice data is screened from the call center's call records. The screened voice data is input into the ASR system and converted into text data. It is suitable for analyzing customer service performance in a specific area or time period, such as customer service quality in a specific market area.

[0153] Implementation method 2: According to the preset service type (such as technical support, after-sales service) and time conditions (such as night service period), the call data of the corresponding seat is screened. The ASR technology is used to convert these call data into text data for subsequent analysis. It is suitable for evaluating customer interactions in different types of services and different time periods, and optimizing specific service processes.

[0154] Implementation method 3: Based on a preset specific event (such as a product launch or promotion period), filter out the call records during the event. Use ASR technology to convert voice data into text data to analyze the impact of the event on customer service. Suitable for analyzing the impact of specific events on customer interaction, such as customer reactions and problem feedback during promotions.

[0155] Example description:

[0156] Suppose a company wants to analyze the customer service quality of agents in Beijing between 9:00 am and 12:00 pm every day. The goal is to obtain call data in this area and time period and convert it into text data for subsequent analysis.

[0157] Step 1: Data screening

[0158] Preset regional conditions: The system sets the geographical area to Beijing and filters out all agent call data from this area. This may involve filtering based on information such as IP address and agent location.

[0159] Preset time condition: The system sets the time period from 9 am to 12 pm and filters out the call records within this time period. For example, a company can configure a time filter in the system to only extract call data made within this time range.

[0160] Filter results: The system filters out all call recordings within a specific area and time period. For example, if 50 agents have made calls with customers within this time period, these call recordings will be extracted.

[0161] Step 2: Speech to Text

[0162] Transcription using ASR technology: The system inputs the 50 selected call recording files into the automatic speech recognition (ASR) system. The system processes these audio files one by one, recognizes the speech content and converts it into text. For example:

[0163] Original audio: Customer said: "I would like to inquire about your product prices."

[0164] Transformed text: The ASR system transforms it into text: "I would like to inquire about your product prices."

[0165] Text data generation: The ASR system generates text data of all recordings for subsequent processing. Assuming that each call lasts an average of 5 minutes, the generated text may include multiple conversation records between customers and agents.

[0166] Step 3: Post-processing

[0167] Application of text data: The generated text data can be used for a variety of analytical purposes. For example, companies can perform sentiment analysis, keyword extraction, or objection type identification on these text data to further evaluate the performance of agents and customer needs.

[0168] Example analysis:

[0169] During the analysis, the system may find that customers in Beijing are more concerned about product prices in the morning. Based on this information, the company can adjust the training plan for agents or optimize the price communication strategy accordingly.

[0170] By screening and processing data in specific situations, this embodiment can provide targeted analysis results, helping enterprises to more accurately evaluate and optimize customer service performance. It not only improves the efficiency of data processing, but also adjusts the scope and conditions of data collection according to specific needs, with strong flexibility and adaptability to a variety of application scenarios. This approach also enables enterprises to better respond to service challenges in different regions and time periods, ultimately improving overall customer satisfaction and operational efficiency.

[0171] Furthermore, a third embodiment of the method for generating unstructured seat tags of the present invention is proposed, wherein the above step S20 includes:

[0172] Step b1, input the position embedding sentence vector into the Transformer encoder to obtain sentence vector information, and then predict the sentence vector information through the fully connected layer and the softmax network layer to identify and classify the objection type of each sentence.

[0173] In this example, the position embedding sentence vector is input into the Transformer encoder, which is a deep learning model based on the self-attention mechanism that can process input sequences and capture the contextual relationship between sentences. Position embedding enables the model to understand the relative position of sentences in the text.

[0174] After being processed by the Transformer encoder, the input sentence vectors are converted into higher-level semantic representations that capture the key semantic features in the sentence.

[0175] The processed sentence vector information is input into the fully connected layer for feature mapping and dimensionality reduction, and then the probability distribution is output through the Softmax network layer to predict the objection type of each sentence.

[0176] Specific implementations may be:

[0177] Implementation 1: The position embedding sentence vector is input into a standard Transformer encoder, which usually contains multiple self-attention heads and a feed-forward neural network. The processed sentence vector information is feature mapped through a fully connected layer, and finally the objection type is classified through a Softmax layer. It is suitable for scenarios where long sequences of text need to be processed and the context dependencies are complex.

[0178] Implementation 2: Input to a lightweight Transformer model (such as DistilBERT), which retains the key features of the Transformer while reducing the amount of computation. The processed sentence vector information is passed through the fully connected layer and the Softmax layer to classify the objection type. It is suitable for scenarios with limited computing resources, especially mobile terminals or systems with high real-time processing requirements.

[0179] Implementation method 3: Input to a Transformer model (such as BERT) that has been pre-trained on a large-scale text corpus and fine-tune it on a specific task. The fine-tuned processing results are classified through the fully connected layer and the Softmax layer. It is suitable for scenarios that require high-precision classification results, and fine-tuning makes the model more suitable for specific tasks.

[0180] Example description:

[0181] Suppose you need to process a conversation between a customer and an agent to identify whether there is an objection and classify the objection type. Here are the specific steps:

[0182] Sentence vectorization and positional encoding: The sentence “I think your product prices are too high” has been transformed into positional embedding sentence vectors such as [0.34, -0.12, 0.44, ...].

[0183] Input Transformer encoder: The system embeds this position into the sentence vector and inputs it into the pre-trained Transformer model. The model processes the input through a multi-layer self-attention mechanism to capture the semantic features in the sentence and its relationship with the context. For example, after processing, the Transformer model outputs a higher-dimensional vector representation, such as [0.52, 0.14, 0.67, ...], which contains semantic information about the relationship between "price" and "too high".

[0184] Fully connected layer and Softmax layer classification: The output vector is input to the fully connected layer for linear transformation and dimensionality reduction for classification. Finally, the Softmax layer outputs the probability distribution of each objection type, for example, [0.02 (no objection), 0.03 (quality objection), 0.90 (price objection), ...], and the system determines that the sentence belongs to the "price objection" type.

[0185] Application of classification results: The system attaches a "price objection" label to the sentence, which can be used for subsequent agent performance evaluation or customer service strategy optimization.

[0186] This embodiment achieves accurate analysis of customer-agent conversations by inputting position-embedded sentence vectors into the Transformer encoder and combining the fully connected layer and the Softmax network layer to identify and classify objection types. In this way, the customer's specific objections can be efficiently identified and classified according to semantic information, significantly improving the intelligence level of the customer service system and its ability to handle complex semantics. This not only improves the response speed of the customer service system, but also provides accurate data support for agent management and customer resource allocation, thereby optimizing service processes and improving customer satisfaction.

[0187] Furthermore, a fourth embodiment of the method for generating unstructured labels for agents of the present invention is proposed. In the above step S30, each objection sentence belonging to the objection type is evaluated for its willingness and ability to handle, and an objection handling willingness score and an objection handling ability score of the agent are generated, including:

[0188] Step c1, input the objection sentence into the preset large model, and identify whether each objection sentence has been processed and whether it has been processed in accordance with the requirements in combination with the preset prompt words;

[0189] Step c2, counting the total number of objections of all objection sentences of all objection types, the number of objection sentences of each objection type, the number of processed objection sentences of each objection type, and the number of qualified processed objection sentences of each objection type that meet the requirements;

[0190] Step c3, calculating the processing willingness score and processing ability score of each objection type based on the statistical quantity, and generating the processing willingness score and processing ability score of the seat based on the preset processing willingness score and processing ability score of the objection type.

[0191] In this embodiment, the identified objection sentence is input into a preset large language model (such as BERT, GPT, Phoenix large model, etc.), which combines the preset prompt words to determine whether the sentence has been responded to and whether the response meets the specified standards. The prompt words can be some specific keywords or phrases to help the model understand how to judge the handling situation.

[0192] All identified objection sentences are counted, including the total number of sentences of each objection type, the number of sentences actually processed, and the number of sentences that meet the requirements. These data provide a basis for calculating the agent's willingness and ability to handle.

[0193] Based on the statistical number of processed and qualified processed cases, the processing willingness score and processing ability score of each objection type are calculated, and finally the overall seat score is generated. The processing willingness score is usually based on the timeliness and enthusiasm of the processing, while the processing ability score is based on the effectiveness of the processing results.

[0194] Specific implementations may be:

[0195] Implementation 1:

[0196] Step 1: Enter your objection sentence

[0197] Use a standard preset large model (such as GPT-3) combined with fixed prompt words to evaluate the willingness and ability to handle the identified objection sentences. For example, input the sentence "I think the price of this product is unreasonable and too expensive" into the model, and use prompt words such as "handle" and "respond" to assist the model in making judgments.

[0198] Step 2: Determine the processing result

[0199] The model analyzes the agent's response to determine whether the customer's objection has been effectively handled. If the response is reasonable and meets the requirements, it is marked as "qualified handling".

[0200] Step 3: Statistics and score calculation

[0201] The system counts the total number of objection sentences, the number of processed sentences, and the number of qualified processed sentences. The system calculates the processing willingness score and processing ability score for each objection type. For example, processing willingness score = processing number / total number of objection sentences, processing ability score = qualified processing number / processing number.

[0202] Step 4: Generate agent scores and labels

[0203] The overall processing willingness score and processing ability score are generated based on the statistical results. The system generates unstructured labels, such as "actively responding to customer objections and strong processing ability."

[0204] Implementation 2:

[0205] Step 1: Enter your objection sentence

[0206] The system uses a preset large model and customized prompt words based on different business needs to analyze the objection sentences. For example, when handling technical support objections, the prompt words can be "technical problem" or "solution".

[0207] Step 2: Determine the processing result

[0208] Combined with custom prompt words, the model determines the effectiveness of the agent's response, identifies whether it meets the processing standards, and marks it as "qualified processing" or "unqualified processing."

[0209] Step 3: Statistics and score calculation

[0210] The total number of sentences, number of processed and qualified processed for each objection type are counted, and the corresponding scores are calculated. The processing willingness score and processing ability score are calculated based on statistical data to provide a quantitative assessment of the agent's performance.

[0211] Step 4: Generate agent scores and labels

[0212] The system generates a comprehensive willingness-to-handle score and a capability-to-handle score, and generates unstructured labels, such as “excellent at handling technical support-related objections.”

[0213] Implementation 3:

[0214] Step 1: Enter your objection sentence

[0215] The system uses a large model combined with machine learning technology to input objection sentences and dynamically generate prompt words based on real-time data. For example, the system automatically generates prompt words related to current customer needs based on the current customer service environment.

[0216] Step 2: Determine the processing result

[0217] Dynamic prompt words help the model judge the agent's handling status more accurately. The system automatically identifies the handling status of each sentence and marks it as "qualified" or "unqualified".

[0218] Step 3: Statistics and score calculation

[0219] The processing of objection sentences is counted in real time, and scores are calculated based on the processing results. The system updates the processing willingness score and processing ability score based on real-time data to ensure the timeliness and accuracy of the evaluation results.

[0220] Step 4: Generate agent scores and labels

[0221] The system generates a composite score and generates unstructured labels based on the latest data, such as "quick response, accurate processing."

[0222] Example description:

[0223] Suppose a customer service system is processing a call record and the system recognizes the following objection sentence:

[0224] Customer objection: I think the price of this product is unreasonable and too expensive.

[0225] Agent response: The pricing of our products is based on market research and quality assurance. You can consider it again.

[0226] Step 1: Pre-set large model evaluation

[0227] Input objection sentence: The system inputs the customer's objection sentence "I think the price of this product is unreasonable and too expensive" into the preset big model, and combines it with prompt words such as "handle", "respond", and "solution". The big model analyzes the agent's response and determines whether the customer's objection has been handled and whether the handling meets the requirements. For example, the model may identify "pricing is based on market research and quality assurance" as the core part of the response and judge its rationality.

[0228] Determine the processing result: The system determines whether the response meets the requirements. If the model concludes that the agent's response is effective and has addressed the customer's concerns, it is marked as "qualified processing".

[0229] Step 2: Statistics and score calculation

[0230] Statistical results: The system counts the number of objection sentences (1), the number of processed sentences (1), and the number of qualified processed sentences (1) in this call. Assume that in the past 10 calls, the agent responded to 15 objection sentences, of which 10 sentences were effectively processed.

[0231] Calculate the score: The system calculates the processing willingness score, for example: processing willingness score = processing number / total number of objection sentences = 1 / 1 = 1.0. Calculate the processing ability score, for example: processing ability score = qualified processing number / processing number = 1 / 1 = 1.0.

[0232] Step 3: Generate agent scores and labels

[0233] Generate a comprehensive score: The system generates a comprehensive score for the agent based on the willingness score and the ability score. For example, a combination of a willingness score of 1.0 and an ability score of 1.0 reflects that the agent performed very well in this call.

[0234] Generate unstructured tags: Based on the comprehensive score, the system generates an unstructured tag, such as "actively respond to customer objections and have strong handling capabilities."

[0235] Through the above steps, this embodiment can automatically evaluate the agent's willingness and ability to handle objection sentences, and generate quantitative scores and unstructured labels. This method significantly improves the automation of the customer service system, reduces manual intervention, and improves the accuracy and consistency of the evaluation. Based on these quantitative indicators and labels, enterprises can better manage agent performance, optimize customer resource allocation, and improve overall customer satisfaction and service efficiency.

[0236] Furthermore, a fifth embodiment of the method for generating unstructured labels for agents of the present invention is proposed. In the above step c3, the processing willingness score and processing ability score of each objection type are calculated according to the statistical quantity, including:

[0237] Step d1: The willingness score for handling each objection type is:

[0238]

[0239] Step d2: The handling capability score for each objection type is:

[0240]

[0241] In this embodiment, the formula for calculating the processing willingness score takes into account two key factors: the ratio of the number of objection sentences of the objection type to the total number of objections: this part reflects the proportion of this type of objection in all objections. The ratio of the number of processing to the number of objection sentences: this part reflects how many of this type of objection are effectively processed. Formula:

[0242]

[0243] The formula for the processing ability score takes into account the following factors: The ratio of the number of objection sentences of this objection type to the total number of objections: This also reflects the relative importance of this type of objection among all objections. The ratio of the number of qualified handlings to the number of objection sentences: This partly reflects how many of the objections that have been handled were handled in accordance with the standards. Formula:

[0244]

[0245] It should be noted that the reason for adding 0.1 to the formula is to prevent the denominator from being zero, thereby avoiding errors in calculation. In addition, adding a small positive number (such as 0.1) can improve the stability and robustness of the calculation in certain scenarios.

[0246] Specific implementations may be:

[0247] Implementation 1:

[0248] Step 1: Data Collection

[0249] The system collects objection type data for each agent, including the total number of objections, the number of sentences for each objection type, the number of objections handled, and the number of qualified handlings.

[0250] Step 2: Formula Application

[0251] Use the formula to calculate the willingness and ability to handle scores for each type of objection.

[0252] Step 3: Synthesis of results

[0253] The scores of all objection types are aggregated to generate the agent's overall handling willingness score and handling ability score.

[0254] Implementation 2:

[0255] Step 1: Dynamic data update

[0256] According to actual business needs, the system updates the number of objections and processing data in real time during the objection processing process.

[0257] Step 2: Formula Application

[0258] The score is calculated according to the formula, but weight adjustment is introduced in dynamic business scenarios to ensure that the score calculation can reflect the changing business environment in a timely manner.

[0259] Step 3: Apply the results

[0260] Generate agent performance scores in real time and consider them when allocating customer resources.

[0261] Example description:

[0262] Suppose in a customer service system, an agent handles the following objections:

[0263] Total objections: 100

[0264] Price objection: Number of objection sentences: 20; Number of processing: 15; Number of qualified processing: 10.

[0265] Quality objections: Number of objection sentences: 30; Number of processed: 20; Number of qualified processed: 18.

[0266] Step 1: Calculate the treatment willingness score

[0267]

[0268] Step 2: Calculate the processing power score

[0269]

[0270] Step 3: Generate the total agent score

[0271] The scores of each objection type are aggregated to generate the agent's overall willingness and ability scores. These scores can be further used to generate unstructured labels, such as "average performance in handling price objections, but strong performance in handling quality objections."

[0272] This embodiment uses a clear formula to calculate the scores of processing willingness and processing ability, which helps enterprises to quantify the performance of agents and more accurately identify the strengths and weaknesses of each agent. The statistics and calculation of scores not only improve the objectivity and consistency of the evaluation, but also provide important data support for subsequent customer resource allocation and agent training.

[0273] Furthermore, a sixth embodiment of the method for generating unstructured labels for agents of the present invention is proposed. In the above step c1, before inputting the objection sentences into the preset large model and identifying whether each objection sentence has been processed and whether the processing has met the requirements in combination with the preset prompt words, the following is further included:

[0274] Step e1, using LoRA technology to fine-tune the preset large model. The fine-tuning process includes freezing other parameters in the preset large model except the target parameters, and only superimposing low-rank adaptation parameters on the specific layer where the target parameters are located to complete the fine-tuning of the preset large model.

[0275] In this embodiment, LoRA (Low-Rank Adaptation) technology is an effective model fine-tuning method. During the fine-tuning process, low-rank adaptation parameters are only superimposed on specific layers of the large model, retaining the main structure and parameters of the large model. In this way, fine-tuning can adapt to the needs of specific tasks while retaining the original capabilities of the model.

[0276] During fine-tuning, most of the parameters in the large model are frozen, ensuring that only the target parameters of specific layers in the model are fine-tuned. These target parameters are the only part updated during fine-tuning and are usually the most representative part of the model. Low-rank adaptation parameters are superimposed on the specific layer where the target parameters are located. These low-rank parameters can significantly improve the performance of the model on specific tasks by adjusting a small number of parameters.

[0277] Specific implementations may be:

[0278] Implementation 1: LoRA fine-tuning for a single task

[0279] Step 1: Determine the target task:

[0280] The system identifies large model target tasks that require fine-tuning, such as customer service objection handling.

[0281] Step 2: Apply LoRA technology:

[0282] Freeze all parameters in the large model and select only the layers relevant to the target task for fine-tuning. Stack low-rank adaptation parameters on these layers and fine-tune them.

[0283] Step 3: Fine-tuning completed:

[0284] After the model is fine-tuned, the system saves the updated model parameters for subsequent objection sentence processing.

[0285] Implementation 2: Multi-task Joint Fine-tuning

[0286] Step 1: Identify multi-task goals:

[0287] The system needs to handle multiple related tasks simultaneously, such as handling objection classification and customer sentiment analysis.

[0288] Step 2: Apply LoRA technology:

[0289] In a multi-task setting, model layers relevant to all tasks are selected for fine-tuning. Low-rank adaptation parameters are superimposed on these layers to adapt to the needs of multiple tasks at the same time.

[0290] Step 3: Joint fine-tuning is completed:

[0291] After fine-tuning, the model is better able to handle multiple tasks and improve overall performance.

[0292] Implementation 3: Online real-time fine-tuning

[0293] Step 1: Real-time data collection:

[0294] Collect data from online customer service environments in real time and fine-tune goals based on the latest customer needs and market changes.

[0295] Step 2: Real-time LoRA fine-tuning:

[0296] Apply LoRA technology online to fine-tune large models in real time. The model is constantly updated to adapt to real-time changing task requirements.

[0297] Step 3: Fine-tune the application in real time:

[0298] The model fine-tuned in real time is put into use immediately, improving the system's adaptability to new data and tasks.

[0299] Example description:

[0300] Suppose a large customer service system needs to handle various types of customer objections. In order to improve the accuracy of the model, the system decides to fine-tune the preset large model. The following are the specific steps:

[0301] Step 1: Determine the target task

[0302] The system identified the types of objections that needed to be handled, including price objections, quality objections, etc. It was decided to fine-tune the large model for objection handling to improve the model's performance on these tasks.

[0303] Step 2: Apply LoRA Technology

[0304] The system selects certain layers in the preset large model, such as the middle layer that is most relevant to the semantic understanding of the sentence, for fine-tuning. On these layers, the system uses LoRA technology to superimpose low-rank adaptation parameters. For example, the parameters of a certain layer in the model were originally matrix W. During the LoRA fine-tuning process, the system superimposed a low-rank matrix A on W to adjust the model's ability to handle objectionable sentences.

[0305] Step 3: Fine-tuning completed

[0306] After several rounds of training, the system is fine-tuned and the model's performance on the objection handling task is verified. Suppose the fine-tuned model improves the accuracy of "price objection" from 85% to 92%.

[0307] Step 4: Post-processing

[0308] The fine-tuned model is applied in actual customer service scenarios. When the system recognizes a customer's objection sentence, the model processes the sentence with the preset prompt words to determine whether it has been processed and whether the processing meets the requirements.

[0309] This embodiment uses LoRA technology to fine-tune the preset large model. The system can significantly improve its performance on specific tasks while maintaining the original performance of the large model. The fine-tuning process is efficient and computationally low-cost, allowing the large model to quickly adapt to new business needs. In addition, the strategy of freezing most parameters and only superimposing low-rank adaptation parameters on the target layer ensures the stability of the fine-tuning process and the reliability of the results, thereby achieving better processing effects in practical applications.

[0310] Furthermore, a seventh embodiment of the method for generating unstructured labels for agents of the present invention is proposed. In the above step S30, after generating the unstructured labels for agents according to the objection handling willingness score and the objection handling ability score, the method further includes:

[0311] Step f1, generating a customer tag according to the customer demand information, wherein the customer tag includes an objection type related to the customer demand;

[0312] Step f2, matching the customer tag with the unstructured tag of the agent, calculating the matching degree between the customer tag and the unstructured tag of each agent, and determining the matching priority between the customer and each agent based on the matching degree;

[0313] Step f3: assign the customer to the agent with the highest matching priority.

[0314] In this embodiment, corresponding customer tags are generated based on the customer's needs and historical data. These tags may include the customer's main concerns, common objection types, or specific service needs. For example, if the customer is often sensitive to price, the customer tag may include "price objection".

[0315] The generated customer tags are matched with the unstructured tags of each agent, and the matching degree between the two is calculated. The matching degree is usually calculated based on factors such as the similarity and relevance of the tags.

[0316] Based on the calculated matching degree, the system assigns a priority to each customer and agent combination. A combination with a high matching priority means that the agent is more suitable for handling the customer's needs.

[0317] Ultimately, customers are assigned to the most suitable agents based on matching priorities to ensure that they receive the best service experience.

[0318] Specific implementations may be:

[0319] Implementation 1: Rule-based matching

[0320] Step 1: Generate customer tags:

[0321] The system generates customer tags based on the customer's historical data, current needs, etc.

[0322] Step 2: Tag matching:

[0323] The match between the customer’s tags and the agent’s unstructured tags is calculated using predefined rules such as similarity of tags.

[0324] Step 3: Determine matching priority:

[0325] Based on the match, the system assigns a priority to the combination of customer and each agent.

[0326] Step 4: Customer Assignment:

[0327] Assign the customer to the agent with the highest matching priority.

[0328] Implementation 2: Machine Learning-Based Matching

[0329] Step 1: Generate customer tags:

[0330] Use machine learning models to analyze customer data and generate customer labels.

[0331] Step 2: Tag matching:

[0332] The matching degree between customer labels and agent unstructured labels is calculated through a trained matching algorithm (such as a matching model based on a neural network).

[0333] Step 3: Determine matching priority:

[0334] According to the matching degree output by the matching model, the system generates a matching priority.

[0335] Step 4: Customer Assignment:

[0336] Automatically assign customers to the agent with the highest matching priority.

[0337] Implementation 3: Real-time dynamic matching

[0338] Step 1: Generate customer tags:

[0339] Analyze customers' current needs in real time and generate dynamic customer tags.

[0340] Step 2: Tag matching:

[0341] The matching degree between customer tags and agent unstructured tags is calculated in real time, and the matching algorithm is dynamically adjusted based on the latest data.

[0342] Step 3: Determine matching priority:

[0343] The matching priority between customers and agents is determined in real time, and the priority is updated in real time as the data changes.

[0344] Step 4: Customer Assignment:

[0345] Dynamically assign customers to the most suitable agent based on real-time matching priority.

[0346] Example description:

[0347] Assume that the customer service system of an online shopping platform is processing a large number of customer requests. The system assigns the most suitable seat to the customer through the following steps:

[0348] Step 1: Generate customer tags

[0349] Customer A's historical data shows that he is often sensitive to prices. Based on this information, the system generates customer labels such as "price objection" and "sensitive customer".

[0350] Step 2: Match customer and agent tags

[0351] The system analyzes the unstructured tags of all currently available agents. For example, Agent B's tags are "good at handling price objections" and "fast response speed". The system calculates the match between Customer A and Agent B and obtains a high match (for example, 0.85).

[0352] Step 3: Determine matching priority

[0353] The system sorts the matching degree between customer A and all seats, and concludes that seat B has the highest matching priority with customer A.

[0354] Step 4: Customer Assignment

[0355] The system ultimately assigns customer A to agent B for processing, ensuring that customer A receives targeted services and improving customer satisfaction.

[0356] In this embodiment, by matching customer tags with unstructured tags of agents and allocating customers based on the matching degree, the system can effectively optimize the configuration of customer service resources. Matching degree calculation and priority allocation not only improve the efficiency of customer service, but also ensure that each customer can get the most appropriate service experience. Such a system can significantly improve customer satisfaction, while optimizing the workload of agents, and ultimately improve the operational efficiency of the entire customer service system.

[0357] The invention also provides a device for generating unstructured labels for seats.

[0358] Reference Figure 3 , Figure 3 This is a functional module diagram of the first embodiment of the seat unstructured label generation device of the present invention. The seat unstructured label generation device comprises:

[0359] A text embedding and encoding module obtains the voice text data of the agent, encodes each sentence in the voice text data into a sentence vector, and combines the sentence vector with the position encoding of the sentence to form a position embedded sentence vector;

[0360] An objection segment identification module, which processes the position embedding sentence vector through a deep learning model to identify and classify the objection type of each sentence;

[0361] The objection handling capability mining module evaluates the objection handling willingness and capability for each objection sentence belonging to the objection type, generates the objection handling willingness score and objection handling capability score of the agent, and generates an unstructured label for the agent based on the objection handling willingness score and the objection handling capability score.

[0362] In addition, an embodiment of the present invention further provides a computer storage medium.

[0363] The storage medium stores a seat unstructured label generation program, and when the seat unstructured label generation program is executed by the processor, the steps of the seat unstructured label generation method as described above are implemented.

[0364] The specific implementation of the storage medium of the present invention is basically the same as the embodiments of the above-mentioned method for generating unstructured seat tags, and will not be described in detail here.

[0365] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims. All equivalent structures or equivalent process changes made using the contents of the specification and drawings of the present invention, or directly or indirectly used in other related technical fields, are protected by the present invention.

[0366] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to the multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or the flowchart, and the combination of boxes in the block diagram and / or the flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0367] In addition, each functional module in each embodiment of the present application can be integrated together to form an independent part, or each module can exist separately, or two or more modules can be integrated to form an independent part. If the function is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, RandomAccess Memory), a disk or an optical disk.

[0368] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0369] The above are only various implementations of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0370] It should be noted that if software tools or components other than those of our company appear in the embodiments of the present application, they are only used for illustrative purposes and do not represent actual use.

Claims

1. A method for generating unstructured labels for seats, characterized in that: The following steps are involved: Acquire the voice text data of the seat, encode each sentence in the voice text data into a sentence vector, and combine the sentence vector with the position encoding of the sentence to form a position embedded sentence vector; Processing the position embedding sentence vectors through a deep learning model to identify and classify the objection type of each sentence; The willingness and ability to handle each objection sentence belonging to the objection type are evaluated to generate an objection handling willingness score and an objection handling ability score of the agent, and an unstructured label of the agent is generated according to the objection handling willingness score and the objection handling ability score.

2. The method for generating unstructured seat tags according to claim 1, characterized in that: Get the agent's voice and text data, including: Acquire the call voice data of the seat that meets the preset conditions, and convert the call voice data of the seat into the voice text data of the seat through automatic speech recognition technology.

3. The method for generating unstructured seat tags according to claim 1, characterized in that: The step of processing the position embedding sentence vector by a deep learning model to identify and classify the objection type of each sentence includes: The position embedding sentence vector is input into the Transformer encoder to obtain sentence vector information, and then the sentence vector information is predicted through the fully connected layer and the softmax network layer to identify and classify the objection type of each sentence.

4. The method for generating unstructured seat tags according to claim 1, characterized in that: Evaluate the willingness and ability to handle each objection sentence belonging to the objection type, and generate the agent's objection handling willingness score and objection handling ability score, including: Input the objection sentences into the preset large model, and identify whether each objection sentence has been processed and whether it has been processed in accordance with the requirements in combination with the preset prompt words; Count the total number of objections for all objection sentences of all objection types, the number of objection sentences of each objection type, the number of resolutions that were resolved in the objection sentences of each objection type, and the number of qualified resolutions that were resolved in the objection sentences of each objection type; The handling willingness score and handling ability score of each objection type are calculated based on the statistical quantity, and the handling willingness score and handling ability score of the seat are generated based on the handling willingness score and handling ability score of the preset objection type.

5. The method for generating unstructured seat tags according to claim 4, characterized in that: The processing willingness score and processing ability score of each objection type are calculated based on the statistical quantity, including: The willingness to handle scores for each objection type are: The handling ability scores for each objection type are:

6. The method for generating unstructured seat tags according to claim 4, characterized in that: Before inputting the objection sentence into the preset large model and identifying whether each objection sentence has been processed and whether it has been processed in accordance with the requirements in combination with the preset prompt words, it also includes: The LoRA technology is used to fine-tune the preset large model. The fine-tuning process includes freezing other parameters in the preset large model except the target parameters, and only superimposing low-rank adaptation parameters on the specific layer where the target parameters are located to complete the fine-tuning of the preset large model.

7. The method for generating unstructured seat tags as claimed in claim 1, characterized in that: After generating the unstructured label of the agent according to the objection handling willingness score and the objection handling ability score, the method further includes: generating a customer tag according to the customer demand information, wherein the customer tag includes an objection type related to the customer demand; Matching the customer tag with the unstructured tag of the agent, calculating the matching degree between the customer tag and the unstructured tag of each agent, and determining the matching priority between the customer and each agent based on the matching degree; Assign the customer to the agent with the highest matching priority.

8. A device for generating unstructured labels for seats, characterized in that: The seat unstructured label generating device comprises: A text embedding and encoding module, used to obtain the voice text data of the agent, encode each sentence in the voice text data into a sentence vector, and combine the sentence vector with the position encoding of the sentence to form a position embedded sentence vector; An objection segment identification module, used to process the position embedding sentence vector through a deep learning model to identify and classify the objection type of each sentence; The objection handling capability mining module is used to evaluate the objection handling willingness and handling capability of each objection sentence belonging to the objection type, generate the objection handling willingness score and objection handling capability score of the agent, and generate an unstructured label of the agent based on the objection handling willingness score and the objection handling capability score.

9. A device for generating unstructured labels for seats, characterized in that: The seat unstructured label generation device includes a memory, a processor, and an agent unstructured label generation program stored in the memory and executable on the processor. When the seat unstructured label generation program is executed by the processor, the steps of the seat unstructured label generation method according to any one of claims 1 to 7 are implemented.

10. A computer storage medium, characterized in that: The storage medium stores a seat unstructured label generation program, and when the seat unstructured label generation program is executed by the processor, the steps of the seat unstructured label generation method according to any one of claims 1 to 7 are implemented.

Images