Customer service method based on seat text processing analysis

Abstract

The invention relates to a customer service method based on seat text processing analysis. A pre-trained pre-analysis model analyzes the business field of customer demands and provides the reply difficulty of the customer demands; when the difficulty does not exceed a difficulty threshold value, the pre-analysis model adds the business field to a cue word of a pre-trained first language model; matching reply or configuration operations related to customer demands in cue words from contents of corresponding business fields in a customer service vector library through correlation matching, wherein the first language model is limited by the system cue words to organize reply or configuration operations with correlation ranked in front as reply results; submitting a reply result obtained by the customer demand and the first language model to the second language model to verify whether the reply result is valid, if so, replying to the customer, otherwise, providing an invalid reason by the second language model; and when the difficulty exceeds a difficulty threshold value, the customer demand is switched to a manual service module according to the business field, and idle customer service in the corresponding field is configured for service.

Description

Technical Field

[0001] The present invention relates to the technical field of large language model application, and in particular to a customer service method based on agent text processing and analysis. Background Art

[0002] In order to provide customer service conveniently and efficiently, many industry enterprises have opened official hotlines. Under the premise of knowing the customer, the agent records the customer interaction content in text, and accumulates a large amount of customer interaction service data. With the development of large language model technology, large language models have the ability to analyze agent texts, learn agent text content, and interact with customers based on the learned agent text content. However, on the one hand, limited by the current context understanding ability of large language models, although large language models can process longer texts, there are still challenges in understanding and generating extremely long texts, that is, large language models may lose context information when processing long texts, and their performance in customer service interactions is. When the response content to customer needs is long, some obvious incomplete responses often appear; on the other hand, in order to meet the requirements of the generalization ability of large language models, it is inevitable that certain erroneous feedback and hallucinations will occur. How to ensure that the large language model provides services in the domain where it can accurately answer and ensure the service quality of the large language model has become a problem. Summary of the invention

[0003] In order to solve the above technical problem or at least partially solve the above technical problem, the present invention provides a customer service method based on agent text processing and analysis.

[0004] In a first aspect, the present invention provides a customer service method based on agent text processing and analysis, comprising: constructing a customer service system including a pre-analysis model, a first language model, a second language model, a customer service vector library, a manual service module, a business configuration interface, and a business module;

[0005] The customer service provides customer needs to the customer service system;

[0006] The pre-trained pre-analysis model analyzes the business field of customer needs and provides the difficulty of answering the customer needs;

[0007] When the difficulty does not exceed the difficulty threshold, the pre-analysis model adds the business domain to the prompt words of the pre-trained first language model; matches the customer demand-related replies or configuration operations in the prompt words from the content of the corresponding business domain in the customer service vector library through relevance matching, and the first language model is restricted by the system prompt words to organize the replies or configuration operations with the highest relevance as the reply result;

[0008] Submit the customer's needs and the answer obtained by the first language model to the second language model to verify whether they are valid. If they are valid, reply to the customer and execute them in the business module through the business configuration interface. Otherwise, the second language model provides an invalid reason.

[0009] When the difficulty exceeds the threshold, the customer demand will be transferred to the manual service module according to the business field and the idle customer service in the corresponding field will be configured to provide service.

[0010] Furthermore, the customer service system includes an authentication module. After the customer's identity is authenticated by the authentication module, the customer is allowed to use the customer service system; an association relationship between the customer identity and the business field to which the customer's current business, which is bound to the customer identity, belongs is pre-built. After the authentication module determines the customer's identity, the business field that the customer may be involved in is determined based on the customer identity, and the business field that the customer may be involved in is added to the prompt words of the first language model.

[0011] Furthermore, the customer service vector library includes vectorized fragments of customer service reference files, and the construction method includes:

[0012] Pre-analyze agent texts, set corresponding responses or business configuration operations based on customer needs in agent texts, and obtain customer service reference documents;

[0013] Add business domain labels to the requirements of different business domains and their corresponding responses or configuration operations in the customer service reference document;

[0014] The content of the customer service reference file is formatted and segmented, and the content of the segmented customer service reference file is vectorized using the Embedding model to obtain vectorized segments that can be recognized by the language model; and the vectorized segments are sequentially assigned to corresponding business areas.

[0015] Furthermore, the process of training the pre-analysis model to analyze the business field of customer needs and provide difficulty in answering customer needs includes: creating the pre-analysis model based on a Boosting algorithm;

[0016] Set response difficulty labels for each customer need in the customer service reference file based on factors that affect the response difficulty of customer needs;

[0017] After configuring the response difficulty label, the customer demand containing the response difficulty label and the business field label is used as the training data of the pre-analysis model; the sum of the first cross entropy between the business field label of the customer demand and the business field predicted by the pre-analysis model and the second cross entropy between the response difficulty label of the customer demand and the difficulty predicted by the pre-analysis model is used as the first loss function for training the pre-analysis model, and the pre-analysis model is trained with the minimum value of the first loss function as the goal until the first loss function is less than the set first threshold, so that the pre-analysis model can output the business field and response difficulty of the customer demand.

[0018] Furthermore, factors affecting the difficulty of answering customer needs include: the length of the answer or configuration operation that meets the customer needs and the uncertainty of the first language model predicting the correct answer or configuration operation for the customer needs; then, the answer difficulty label is calculated as follows:

[0019]

[0020] length() indicates the length of the reply or configuration operation to meet any customer demand; I is the total number of customer demands in the customer service reference file, A i , A j Any customer requirement R in the customer service reference file i , R j Reply or configuration operation, α 1 is the contribution coefficient of normalized length to difficulty;

[0021] To meet customer needs j A reply or configuration operation j The normalized length of

[0022] P(A j |R j ) is the first language model prediction to meet customer needs R j A reply or configuration operation j The probability of; H() is the entropy function, H(P(A j |R j ))=-P(A j |R j )logP(A j |R j ), indicating that the first language model predicts that it meets customer demand R j A reply or configuration operation j The entropy of the probability; α 2 is the contribution coefficient of normalized entropy to difficulty;

[0023] Predicting customer needs for the first language modelj A reply or configuration operation j The normalized entropy of the probability of w ,w∈W is the first language model prediction and customer demand R j Related reply or configuration operations, W is an index set of reply or configuration operations determined by relevance, and the total number of indexes in the index set is N.

[0024] Furthermore, the customer service vector library is used in conjunction with the first language model to construct a search enhancement architecture, including: using system prompt words to constrain the first language model so that the first language model matches the customer needs based on the content of the corresponding business field in the customer service vector library or the configuration operation;

[0025] Obtain customer needs as prompt words, or obtain customer needs and customer demand business fields as prompt words, and vectorize customer needs through the Embedding model to obtain a query vector;

[0026] By using correlation matching, content related to the query vector is searched from the customer service vector library as a response vector, or content related to the query vector is searched from a reference domain of a corresponding business field in the customer service vector library as a response vector;

[0027] The query vector and the response vector are combined into a feedback prompt word. Within the domain defined by the feedback prompt word, the first language model organizes the replies or configuration operations with the highest relevance as the answer result.

[0028] Furthermore, customer needs are randomly generated and input into the retrieval enhancement architecture to obtain a service solution output by the retrieval enhancement architecture; the effectiveness of the output of the retrieval enhancement architecture is manually evaluated, and invalid reasons are given when it is not feasible; customer needs, service solutions output by the retrieval enhancement architecture, and corresponding evaluation contents are used as training data to fine-tune the second language model; during the fine-tuning process, the cross entropy between the evaluation results and invalid reasons output by the second language model and the actual evaluation results and invalid reasons is used as a second loss function, and the parameters of the second language model are fine-tuned based on the second loss function value until the value of the second loss function is less than a second threshold, so that the second language model has the ability to evaluate service solutions.

[0029] Furthermore, when the reply result involves business configuration operations, the corresponding business configuration interface is called according to the business configuration operation content in the reply result to form a service plan, and the second language model verifies whether the service plan is valid; if valid, the service plan is fed back to the customer, and after the user authorizes and agrees to the service plan, the service plan is executed through the business module.

[0030] In a second aspect, the present invention provides a customer service device based on agent text processing and analysis, comprising: at least one processing unit, the processing unit is connected to a storage unit via a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the customer service method based on agent text processing and analysis is implemented.

[0031] In a third aspect, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the customer service method based on agent text processing and analysis as described above.

[0032] The above technical solution provided by the embodiment of the present invention has the following advantages compared with the prior art:

[0033] This application uses a pre-analysis model to analyze the business field to which the customer demand belongs and the difficulty of responding. After understanding the business field to which the customer demand belongs, the business field is added to the prompt words of the first language model. The search enhancement architecture based on the business field can match the corresponding reference domain. The search enhancement architecture selects content from the reference domain to respond to customer needs, which greatly reduces the search scope of the search enhancement architecture and reduces the number of tokens analyzed by the first language model. In addition, since the current language model often has poor response effects on high-difficulty demands, this application designs the difficulty of responding to customer needs, allocates customer needs according to difficulty, allocates customer needs that can be accurately processed by the first language model to the first language model, and transfers customer needs with low processing accuracy by the first language model to manual customer service processing. While ensuring the overall customer service effect, minimize the manual workload. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0036] Figure 1 A flowchart of a customer service method based on agent text processing and analysis provided by an embodiment of the present invention;

[0037] Figure 2 An architectural diagram of a customer service system provided by an embodiment of the present invention;

[0038] Figure 3A flowchart of a process for building a customer service vector library provided by an embodiment of the present invention;

[0039] Figure 4 A flowchart of using a customer service vector library in conjunction with a first language model to construct a search enhancement architecture provided by an embodiment of the present invention;

[0040] Figure 5 A flowchart of training a second language model provided by an embodiment of the present invention;

[0041] Figure 6 A schematic diagram of a customer service device based on agent text processing and analysis provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] It should be noted that, in this article, the terms "include", "comprises" or any other variations 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, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0044] Example 1

[0045] The technology of the present invention implements a customer service method based on agent text processing and analysis, including:

[0046] like Figure 2 As shown, a customer service system is constructed, which includes an authentication module, a pre-analysis model, a first language model, a second language model, a customer service vector library, a manual service module, a business configuration interface and a business module.

[0047] The customer service vector library is the information reference for the first language model to respond to customer needs. The content of the customer service vector library needs to be recognizable by the first language model, and the content of the customer service vector library supports matching through correlation analysis. To meet the above requirements, Figure 3 As shown in the figure, the process of building a customer service vector library includes:

[0048] Pre-analyze the agent text, summarize and remove duplicate customer needs in the agent text. Set corresponding replies or configuration operations to meet the customer needs in the agent text, and obtain a customer service reference file; wherein the customer service reference file contains customer needs and corresponding replies or configuration operations to meet the needs. Add business domain labels to the needs of different business fields and their corresponding replies or configuration operations to meet the needs in the customer service reference file. Format and segment the content of the customer service reference file, and use the Embedding model to vectorize the content of the segmented customer service reference file to obtain vectorized fragments that can be recognized by the language model; and assign the vectorized fragments to the corresponding business fields in order. For large-scale business scenarios and large-scale needs, the scale of customer service reference files is often very large. This application classifies the content in the customer service reference file by business field to obtain several reference domains with relatively small ranges.

[0049] like Figure 4 As shown in the figure, the customer service vector library is used in conjunction with the first language model to build a retrieval enhancement architecture, including:

[0050] The system prompt words are used to constrain the first language model so that the first language model matches replies or configuration operations according to customer needs based on the content of the corresponding business field in the customer service vector library, and selects the replies or configuration operations with the highest relevance as the reply results.

[0051] Customer needs are obtained as prompt words, or customer needs and customer demand business fields are obtained as prompt words, and customer needs are vectorized through the Embedding model to obtain a query vector.

[0052] By using correlation matching, content related to the query vector is searched from the customer service vector library as a response vector, or content related to the query vector is searched from a reference domain of a corresponding business field in the customer service vector library as a response vector;

[0053] The query vector and the response vector are combined into a feedback prompt word. Within the domain defined by the feedback prompt word, the first language model organizes the replies or configuration operations with the highest relevance as the answer result.

[0054] Specifically, the pre-analysis model is trained to analyze the customer demand business area and provide answers to customer demand difficulties, including:

[0055] The pre-analysis model is created based on a Boosting algorithm; the Boosting algorithm combines multiple decision tree-based weak classifiers into a strong classifier.

[0056] Set response difficulty labels for each customer need in the customer service reference file based on factors that affect the response difficulty of customer needs;

[0057] In this application, the difficulty of answering the defined customer needs depends on the following factors:

[0058] 1. The length of the reply or configuration operation that meets customer needs. On the one hand, a customer may use complex language expressions to describe a simple demand. Although the customer's description is complex, the reply or configuration operation to meet the customer's needs is relatively simple. Due to the professionalism of customer service, the responses to customer needs in the customer service reference documents are relatively concise. The length of the concise reply or configuration operation that meets customer needs can well reflect the difficulty of the language model in responding to customer needs. On the other hand, the longer the reply or configuration operation, the weaker the correlation between the contexts as the text length increases. The language model is limited by its ability to understand the context, and its ability to respond to long texts with weak correlation is poor. That is, many language models often have incomplete responses when dealing with long text problems.

[0059] 2. The uncertainty of the first language model predicting the correct response or configuration operation based on customer needs. The uncertainty is related to the probability of the language model predicting the correct response or configuration operation based on customer needs. When the probability of the correct response or configuration operation is close to the probability of the wrong response or configuration operation, and the optimal one cannot be determined, the uncertainty is large and the certainty is small; conversely, when the probability of the correct response or configuration operation is greater than the probability of the wrong response or configuration operation, the certainty of the correct response or configuration operation is large. This application constructs system prompt words based on customer needs and their corresponding domain labels in the customer service reference file and provides them to the pre-trained first language model. The first language model predicts each corresponding reply and gives the probability of each reply; the entropy of the probability of the correct reply or configuration operation in the reply is used as an uncertainty indicator.

[0060] Then, the answer difficulty label is calculated as follows:

[0061] in,

[0062] length() indicates the length of the reply or configuration operation to meet any customer demand; I is the total number of customer demands in the customer service reference file, A i , A j Any customer requirement R in the customer service reference file i , R j Reply or configuration operation, α 1 is the contribution coefficient of normalized length to difficulty;

[0063] To meet customer needs j A reply or configuration operation j The normalized length of

[0064] P(A j |R j ) is the first language model prediction to meet customer needs R j A reply or configuration operation j The probability of; H() is the entropy function, H(P(A j |R j ))=-P(A j |R j )logP(A j |R j ), indicating that the first language model predicts that it meets customer demand R j A reply or configuration operation j The entropy of the probability; α 2 is the contribution coefficient of normalized entropy to difficulty;

[0065] Predicting customer needs for the first language model j A reply or configuration operation j The normalized entropy of the probability of w ,w∈W is the first language model prediction and customer demand R j Related reply or configuration operations, W is an index set of reply or configuration operations determined by relevance, and the total number of indexes in the index set is N.

[0066] After configuring the answer difficulty label, the customer demand containing the answer difficulty label and the business field label is used as the training data of the pre-analysis model; the logarithm of the error rate between the business field label of the customer demand and the business field predicted by the pre-analysis model and the weighted sum of the mean square error between the answer difficulty label of the customer demand and the difficulty predicted by the pre-analysis model are used as the first loss function for training the pre-analysis model, and the pre-analysis model is trained with the minimum value of the first loss function as the goal until the first loss function is less than the set first threshold, so that the pre-analysis model can output the business field and difficulty of the customer demand.

[0067] like Figure 5 As shown, the training process of the second language model includes:

[0068] Generate customer needs randomly and input them into the retrieval enhancement architecture to obtain the service solutions output by the retrieval enhancement architecture, i.e., responses or configuration operations for customer needs. Humanly evaluate the effectiveness of the output of the retrieval enhancement architecture and give invalidation reasons when it is not feasible. Fine-tune the second language model by combining customer needs, service solutions output by the retrieval enhancement architecture, and corresponding evaluation contents into training data; during the fine-tuning process, use the cross entropy of the evaluation results and invalidation reasons output by the second language model with the actual evaluation results and invalidation reasons as the second loss function, and fine-tune the parameters of the second language model based on the second loss function value until the value of the second loss function is less than the second threshold, so that the second language model has the ability to evaluate service solutions.

[0069] like Figure 1 As shown, the service process of the entire customer service system is as follows:

[0070] The customer service system includes an authentication module, and after the customer's identity is authenticated by the authentication module, the customer is allowed to use the customer service system. In one embodiment, an association relationship between the customer identity and the business field to which the customer's current business, which is bound to the customer identity, belongs is pre-built. After the authentication module determines the customer identity, the business field that the customer may be involved in is determined based on the customer identity, and the business field that the customer may be involved in is added to the prompt words of the first language model.

[0071] The customer service provides customer needs to the customer service system; the customer needs participate in the analysis process of the pre-analysis model, the first language model, and the second language model.

[0072] Specifically, the pre-analysis model is first pre-trained to analyze the business field of the customer's needs and provide the difficulty of answering the customer's needs.

[0073] When the difficulty does not exceed the difficulty threshold, the pre-analysis model adds the business domain to the prompt words of the pre-trained first language model; the first language model is restricted by the system prompt words, so that the first language model matches the replies or configuration operations related to the customer needs in the prompt words from the content of the corresponding business domain in the customer service vector library, and selects the reply or configuration operation with the highest relevance as the reply result.

[0074] The customer's needs and the response results obtained by the first language model are submitted to the second language model to verify whether they are valid. If they are valid, the customer will be replied, otherwise the second language model will provide reasons for invalidity.

[0075] When the difficulty exceeds the threshold, the customer demand will be transferred to the manual service module according to the business field and the idle customer service in the corresponding field will be configured to provide service.

[0076] This application uses a pre-analysis model to analyze the business field to which the customer demand belongs and the difficulty of responding. After understanding the business field to which the customer demand belongs, the business field is added to the prompt words of the first language model. The search enhancement architecture based on the business field can match the corresponding reference domain. The search enhancement architecture selects content from the reference domain to respond to customer needs, which greatly reduces the search scope of the search enhancement architecture and reduces the number of tokens analyzed by the first language model. In addition, since the current language model often has poor response effects on high-difficulty demands, this application designs the difficulty of responding to customer needs, allocates customer needs according to difficulty, allocates customer needs that can be accurately processed by the first language model to the first language model, and transfers customer needs with low processing accuracy by the first language model to manual customer service processing. While ensuring the overall customer service effect, minimize the manual workload.

[0077] Example 2

[0078] See also Figure 6 As shown, an embodiment of the present invention provides a customer service device based on agent text processing and analysis, comprising: at least one processing unit, the processing unit is connected to a storage unit via a bus unit, the storage unit is a computer-readable storage medium, and can be used to store software programs, computer executable programs, and modules, such as the software programs, computer executable programs, and modules corresponding to a customer service method based on agent text processing and analysis in an embodiment of the present invention. The processing unit implements the above-mentioned customer service method based on agent text processing and analysis by running the software programs, computer executable programs, and modules stored in the storage unit, including:

[0079] Build a customer service system consisting of a pre-analysis model, a first language model, a second language model, a customer service vector library, a manual service module, a business configuration interface, and a business module;

[0080] The customer service provides customer needs to the customer service system;

[0081] The pre-trained pre-analysis model analyzes the business field of customer needs and provides the difficulty of answering the customer needs;

[0082] When the difficulty does not exceed the difficulty threshold, the pre-analysis model adds the business domain to the prompt words of the pre-trained first language model; matches the customer demand-related replies or configuration operations in the prompt words from the content of the corresponding business domain in the customer service vector library through relevance matching, and the first language model is restricted by the system prompt words to organize the replies or configuration operations with the highest relevance as the reply result;

[0083] Submit the customer's needs and the answer obtained by the first language model to the second language model to verify whether they are valid. If they are valid, reply to the customer and execute them in the business module through the business configuration interface. Otherwise, the second language model provides an invalid reason.

[0084] When the difficulty exceeds the threshold, the customer demand will be transferred to the manual service module according to the business field and the idle customer service in the corresponding field will be configured to provide service.

[0085] Of course, the computer program stored in the storage unit of the device for implementing the customer service method based on agent text processing and analysis provided in an embodiment of the present invention is not limited to the method operations described above, and can also execute related operations in the customer service method based on agent text processing and analysis provided in any embodiment of the present invention.

[0086] Example 3

[0087] An embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. When the computer program is executed, the customer service method based on agent text processing and analysis is implemented, including:

[0088] Build a customer service system consisting of a pre-analysis model, a first language model, a second language model, a customer service vector library, a manual service module, a business configuration interface, and a business module;

[0089] The customer service provides customer needs to the customer service system;

[0090] The pre-trained pre-analysis model analyzes the business field of customer needs and provides the difficulty of answering the customer needs;

[0091] When the difficulty does not exceed the difficulty threshold, the pre-analysis model adds the business domain to the prompt words of the pre-trained first language model; matches the customer demand-related replies or configuration operations in the prompt words from the content of the corresponding business domain in the customer service vector library through relevance matching, and the first language model is restricted by the system prompt words to organize the replies or configuration operations with the highest relevance as the reply result;

[0092] Submit the customer's needs and the answer obtained by the first language model to the second language model to verify whether they are valid. If they are valid, reply to the customer and execute them in the business module through the business configuration interface. Otherwise, the second language model provides an invalid reason.

[0093] When the difficulty exceeds the threshold, the customer demand will be transferred to the manual service module according to the business field and the idle customer service in the corresponding field will be configured to provide service.

[0094] A computer-readable storage medium provided in an embodiment of the present invention stores a computer program which is not limited to the method operations described above, but can also execute related operations in a customer service method based on agent text processing and analysis provided in any embodiment of the present invention.

[0095] In the embodiments provided by the present invention, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, structures or units, which can be electrical, mechanical or other forms.

[0096] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0097] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0098] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A customer service method based on agent text processing and analysis, characterized in that: include: Constructing a customer service system comprising a pre-analysis model, a first language model, a second language model, a customer service vector library, a manual service module, a business configuration interface and a business module, wherein the customer service vector library comprises vectorized segments of a customer service reference file; The customer service provides customer needs to the customer service system; The pre-trained pre-analysis model analyzes the business field of customer needs and provides the difficulty of answering the customer needs; When the difficulty of the reply does not exceed the difficulty threshold, the pre-analysis model adds the business field to the prompt words of the pre-trained first language model; through relevance matching, the content of the corresponding business field in the customer service vector library is matched with the reply or configuration operation related to the customer demand in the prompt word, and the first language model is restricted by the system prompt word to organize the reply or configuration operation with the highest relevance as the reply result; Submit the customer's needs and the answer obtained by the first language model to the second language model to verify whether they are valid. If they are valid, reply to the customer and execute them in the business module through the business configuration interface. Otherwise, the second language model provides an invalid reason. When the difficulty exceeds the threshold, the customer demand will be transferred to the manual service module according to the business field and the idle customer service in the corresponding field will be configured to provide service.

2. The customer service method based on agent text processing and analysis according to claim 1, characterized in that: The customer service system includes an authentication module. After the customer's identity is authenticated by the authentication module, the customer is allowed to use the customer service system; an association relationship between the customer identity and the business field to which the customer's current business, which is bound to the customer identity, belongs is pre-built. After the authentication module determines the customer's identity, the business field that the customer may be involved in is determined based on the customer identity, and the business field that the customer may be involved in is added to the prompt words of the first language model.

3. The customer service method based on agent text processing and analysis according to claim 1, characterized in that: The method of constructing the customer service vector library includes: Pre-analyze agent texts, set corresponding responses or business configuration operations based on customer needs in agent texts, and obtain customer service reference documents; Add business domain labels to the requirements of different business domains and their corresponding responses or configuration operations in the customer service reference document; The content of the customer service reference file is formatted and segmented, and the content of the segmented customer service reference file is vectorized using the Embedding model to obtain vectorized segments that can be recognized by the language model; and the vectorized segments are sequentially assigned to corresponding business areas.

4. The customer service method based on agent text processing and analysis according to claim 1, characterized in that: The process of training the pre-analysis model to analyze the business field of customer needs and provide the difficulty of answering customer needs includes: creating the pre-analysis model based on a Boosting algorithm; Set response difficulty labels for each customer need in the customer service reference file based on factors that affect the response difficulty of customer needs; After configuring the response difficulty label, the customer demand containing the response difficulty label and the business field label is used as the training data of the pre-analysis model; the weighted sum of the logarithm of the error rate between the business field label of the customer demand and the business field predicted by the pre-analysis model and the mean square error between the response difficulty label of the customer demand and the difficulty predicted by the pre-analysis model is used as the first loss function for training the pre-analysis model, and the pre-analysis model is trained with the minimum value of the first loss function as the goal until the first loss function is less than the set first threshold, so that the pre-analysis model can output the business field and response difficulty of the customer demand.

5. The customer service method based on agent text processing and analysis according to claim 4 is characterized in that: Factors that affect the difficulty of answering customer needs include: the length of the answer or configuration operation that meets the customer needs and the uncertainty of the first language model predicting the correct answer or configuration operation for the customer needs; then, the answer difficulty label is calculated as follows: in, length() indicates the length of the reply or configuration operation to meet any customer demand; I is the total number of customer demands in the customer service reference file; A i , A j Any customer requirement R in the customer service reference file i , R j The reply or configuration operation; α1 is the contribution coefficient of normalized length to difficulty; To meet customer needs j A reply or configuration operation j The normalized length of P(A j |R j ) is the first language model prediction to meet customer needs R j A reply or configuration operation j The probability of; H() is the entropy function, H(P(A j |R j ))=-P(A j |R j )logP(A j |R j ), indicating that the first language model predicts that it meets customer demand R j A reply or configuration operation j The entropy of the probability; α2 is the contribution coefficient of normalized entropy to difficulty; Predicting customer needs for the first language model j A reply or configuration operation j The normalized entropy of the probability of w ,w∈W is the first language model prediction and customer demand R j Related reply or configuration operations, W is an index set of reply or configuration operations determined by relevance, and the total number of indexes in the index set is N.

6. The customer service method based on agent text processing and analysis according to claim 1, characterized in that: The customer service vector library is used in conjunction with the first language model to construct a search enhancement architecture, including: using system prompt words to constrain the first language model so that the first language model performs a reply or configuration operation match for the customer needs in the prompt words based on the content of the corresponding business field in the customer service vector library; Obtain customer needs as prompt words, or obtain customer needs and customer demand business fields as prompt words, and vectorize customer needs through the Embedding model to obtain a query vector; By using correlation matching, content related to the query vector is searched from the customer service vector library as a response vector, or content related to the query vector is searched from a reference domain of a corresponding business field in the customer service vector library as a response vector; The query vector and the response vector are combined into a feedback prompt word. Within the domain defined by the feedback prompt word, the first language model organizes the replies or configuration operations with the highest relevance as the answer result.

7. The customer service method based on agent text processing and analysis according to claim 6, characterized in that: Randomly generate customer needs and input them into the retrieval enhancement architecture to obtain a service solution output by the retrieval enhancement architecture; manually evaluate the effectiveness of the output of the retrieval enhancement architecture and give invalid reasons when it is not feasible; form training data with customer needs, service solutions output by the retrieval enhancement architecture and corresponding evaluation contents to fine-tune the second language model; during the fine-tuning process, use the cross entropy of the evaluation results and invalid reasons output by the second language model and the actual evaluation results and invalid reasons as the second loss function, and fine-tune the parameters of the second language model based on the second loss function value until the value of the second loss function is less than a second threshold, so that the second language model has the ability to evaluate service solutions.

8. The customer service method based on agent text processing and analysis according to claim 1, characterized in that: When the reply result involves business configuration operations, the corresponding business configuration interface is called according to the business configuration operation content in the reply result to form a service plan, and the second language model verifies whether the service plan is valid; if valid, the service plan is fed back to the customer, and after the user authorizes and agrees to the service plan, the service plan is executed through the business module.

9. A customer service device based on agent text processing and analysis, characterized in that: include: At least one processing unit, the processing unit is connected to a storage unit via a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the customer service method based on agent text processing and analysis as described in any one of claims 1-8 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the customer service method based on agent text processing and analysis as described in any one of claims 1 to 8 is implemented.

Images