Intelligent customer service system based on large language model and high-precision document analysis

Abstract

The invention discloses an intelligent customer service system based on a large language model and high-precision document analysis, relates to the technical field of intelligent semantic processing, and solves the problems that in the prior art, answering information is obtained according to a question and answer database, contexts are easily lost after multiple rounds of dialogues, and repeated questions or unanswered questions are caused. And the use experience of the user is influenced. The method comprises the steps of obtaining file information uploaded by a user; extracting key information in the file information by using a target detection algorithm; identifying intention information of the user according to the key information; supplementing the intention information to obtain question information of the user; evaluating the risk of the question information by using a bidirectional LSTM model to obtain a risk score; generating a sensitive prompt signal according to the risk score; corresponding content is matched based on the questioning information of the user to generate interpretation information; the user demand can be accurately analyzed according to the context relation, the accuracy of explaining information is improved, and the user experience is improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent semantic processing, relates to intelligent customer service technology for document parsing, and specifically is an intelligent customer service system based on large language models and high-precision document parsing. Background Art

[0002] Intelligent customer service is an automated service system built based on technologies such as artificial intelligence, natural language processing, and machine learning, which can simulate human conversations; through text, voice, or multi-modal interaction methods, it can respond to user inquiries, solve problems or guide business processes; its core purpose is to replace or assist human customer service to improve service efficiency and user experience; intelligent customer service can automatically handle high-frequency and repetitive problems, reduce the need for human customer service, and lower labor costs; and it can serve multiple customers simultaneously, with a processing speed far exceeding that of humans, significantly improving efficiency during peak periods; it can respond to customer requests without interruption, covering non-working hours, and avoiding customer loss due to waiting; intelligent customer service can avoid fluctuations in service quality caused by emotions and experience differences of human customer service, ensuring that each reply is professional and consistent.

[0003] The prior art (a patent application with publication number CN110570215A) discloses an intelligent customer service system, which includes: a knowledge base management module, a robot chat module, and a human customer service chat module; the knowledge base management module stores a question-and-answer database related to the business field; the robot chat module is used to receive the conversation above data, match the alternative conversation below data from the question-and-answer database; obtain the relevance between the alternative conversation below data and the conversation above data; in response to the relevance exceeding a preset threshold, determine the alternative conversation below data as the conversation below data; the human customer service module is used to receive the human customer service chat request generated when the human customer service confirms that the relevance is lower than the preset threshold; and receive the conversation below data from the human customer service; in the prior art, the conversation below data is matched from the question-and-answer database according to the relevance through the conversation above data; however, in the prior art, when obtaining the answer information according to the question-and-answer database, it is easy to lose the context after multiple rounds of conversations, resulting in repeated questions or irrelevant answers, affecting the user experience.

[0004] The present invention provides an intelligent customer service system based on large language models and high-precision document parsing to solve the above technical problems. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; for this purpose, the present invention proposes an intelligent customer service system based on a large language model and high-precision document parsing, which is used to solve the technical problem that in the prior art, when obtaining answer information according to a question-and-answer database, it is easy to lose the context after multiple rounds of conversations, resulting in repeated questions or irrelevant answers, affecting the user experience.

[0006] To achieve the above object, a first aspect of the present invention provides an intelligent customer service system based on a large language model and high-precision document parsing, including: a document parsing module, an information collection module connected thereto, and a terminal processing module;

[0007] The information collection module: is used to obtain the file information uploaded by the user; and extract the key information in the file information by using an object detection algorithm;

[0008] The document parsing module: is used to identify the intention information of the user according to the key information; supplement the intention information to obtain the user's question information; use a bidirectional LSTM model to evaluate the risk of the question information to obtain a risk score; and generate a sensitive prompt signal according to the risk score;

[0009] The terminal processing module: is used to match the corresponding content based on the user's question information to generate an explanation information.

[0010] Preferably, the extraction of the key information in the file information by using the object detection algorithm includes:

[0011] Retrieve the file information; wherein, the file information includes: table information, text information, and image information; extract the table information in the file information; use the YOLO algorithm to detect the coordinate position of the table information, and according to the coordinate position, use image processing technology to segment the rows and columns of the table information and convert it into two-dimensional array information;

[0012] Extract the text information in the file information; use word segmentation technology to divide the text information into several text words, remove the stop words in the several text words, and then use a word embedding vector model to convert the several text words into text vector information;

[0013] Extract the image information in the text information; mark the image information as I; perform convolution calculation through the formula F(I) = σ × (W0 * I + b0) to identify the image feature information of the image information; wherein, W0 represents the convolution kernel, b0 represents the bias, σ represents the activation function, and * represents the convolution operation; integrate the two-dimensional array information, text vector information, and image feature information into key information.

[0014] It should be noted that the YOLO algorithm is a method in the field of object detection. Its core concept is to transform the object detection task into a regression problem, and directly predict the object position and type in the image through a single forward propagation. In the present invention, the coordinate position of the table information is detected by the YOLO algorithm, and the table information can be transformed into two-dimensional array information.

[0015] The present invention processes the table information, text information, and image information in the file information respectively, and performs corresponding processing for different file types, which can screen the key information in the file information and lay a data foundation for subsequent analysis of the file information.

[0016] Preferably, the recognition of the user's intention information based on the key information includes:

[0017] Obtain the user's input information and retrieve the text vector information in the key information; transform the user's input information into input vector information;

[0018] Mark the text vector information and the input vector information as Ec and Eu respectively; integrate the text vector and the input vector into the context vector V through the function V = f(Eu, Ec); match the context vector with the text content to obtain the user's intention information; where f(*) is the integration function.

[0019] The present invention integrates the input information with the text vector information according to the user's input information to obtain the context vector, analyzes the user's intention based on the context vector to obtain the user's intention information, can analyze the customer's needs, and provides a data foundation for subsequent answering the user's needs.

[0020] Preferably, the supplementation of the intention information to obtain the user's question information includes:

[0021] Retrieve the user's intention information; divide the user's intention information into several fields and determine whether the several fields contain all types of fields; if so, analyze the grammar of the intention information; if not, analyze the context weight of the intention information;

[0022] Divide the intention information to obtain several elements; allocate the context weight to the intention information according to the several elements, supplement the user's intention information according to the context weight to obtain the supplementary content; send the supplementary content to the user for selection; supplement the intention information according to the user's selection to obtain the question information.

[0023] The present invention analyzes the integrity of the user's intent information. When the user's intent information fails to clearly understand the user's needs, the user's intent information is supplemented to obtain the user's question information. It can clearly analyze the user's intent information, timely understand the user's actual needs, and is beneficial to improving the accuracy of the intelligent customer service analysis results.

[0024] Preferably, the method for obtaining the weight of the context includes:

[0025] The attention mechanism is used to allocate context weights to the intent information. The calculation formula for context weight allocation is:

[0026] where, e i = W T h i + b; e i represents the score of the i-th element; e j represents the score of the j-th element; W represents the weight matrix; b is the bias parameter; h_i represents the hidden state of the i-th element.

[0027] Preferably, the method for evaluating the risk of the question information by using the bidirectional LSTM model to obtain a risk score includes:

[0028] Retrieve the user's question information and input the question information into the bidirectional LSTM model; calculate the risk score of the question information according to the self-attention mechanism;

[0029] The calculation formula for the risk score is: F_X = θ × (W a [h f ; h b + b a ); where, h f is the forward hidden state, h b is the backward hidden state; W a represents the weight parameter; b a represents the bias parameter.

[0030] The present invention evaluates the risk of the user's question information according to the user's question information. Since the user may input some sensitive words during the process of information input, resulting in risks in the question information. Therefore, evaluating the risk of the user's question information can improve the security of the intelligent customer and is beneficial to avoiding potential security hazards caused by sensitive words for the intelligent customer service.

[0031] Preferably, the method for generating a sensitive prompt signal according to the risk score includes:

[0032] Retrieve the risk score; compare the risk score with the risk threshold; when the risk score is greater than the risk threshold, a sensitive reminder signal is generated;

[0033] Otherwise, divide the question information corresponding to the risk score into several elements; integrate the several elements into a sensitive input sequence; use a sensitive detection model of deep machine learning to analyze the sensitive input sequence to generate a sensitive reminder signal; return a standard prompt message according to the sensitive reminder signal.

[0034] The present invention analyzes the risk degree of sensitive words in the user's question information according to the risk score. When the risk score is greater than the risk threshold, a sensitive reminder signal is generated; the user is reminded to re-enter the question information, and the sensitive detection model is used to analyze the question information, which can further analyze the sensitive words existing in the question information, is beneficial to reducing the risk of the intelligent customer service being blocked, and improves the security of the intelligent customer service.

[0035] Preferably, the generating the explanatory information by matching the corresponding content based on the user's question information includes:

[0036] Retrieve the user's question information, and use a word vector model to convert the question information into a semantic vector; use a cross-modal alignment loss optimization formula to match the semantic vector with the multi-modal data in the knowledge base to obtain an explanatory content; sort and integrate the explanatory content according to the set type priority to obtain the explanatory information, and send the explanatory information to the user; wherein, the multi-modal data includes: text, pictures and videos.

[0037] The expression of the cross-modal alignment loss optimization formula is: wherein, L represents the loss function of cross-modal alignment; f T and f I are respectively the feature mapping functions of the text and image modalities; T k represents the k-th text sample; I k represents the k-th image sample; ∥*∥ represents the norm operation.

[0038] The present invention converts the question information into a semantic vector, matches the semantic vector with the multi-modal data in the knowledge base, and sorts and integrates the multi-modal data according to the set priority to obtain the explanatory information; can target the user's needs and provide comprehensive explanatory data through; is beneficial to fully answering the user's needs and improving the user experience.

[0039] Preferably, the updating method of the multi-modal data in the knowledge base includes:

[0040] Obtain the multi-modal data in the knowledge base; perform clustering analysis on the multi-modal data by using a clustering analysis algorithm; when there is multi-modal data in the knowledge base that does not match the question information, generate a knowledge point supplement signal and send the knowledge point supplement signal to the management personnel.

[0041] Obtain the number of times of obtaining answers to several multimodal data; when the number of times the multimodal data appears is greater than the number threshold, filter out the optimal data from the multimodal data in the same cluster as the answer content; otherwise, delete according to the importance of the multimodal data.

[0042] The present invention updates the multimodal data in the knowledge base, filters out the optimal data from several interpreted multimodal data to ensure the consistency of the interpreted content, and selectively deletes the answer content with low frequency occurrences, which can reduce the load on the memory storing the knowledge base and is beneficial to improving the consistency and practicality of the data stored in the knowledge base.

[0043] Preferably, the system further includes a prediction adjustment module, including:

[0044] Obtain the user's usage time period, generate a usage heat map according to the user's usage time period; adjust the on-duty personnel based on the usage heat map;

[0045] Obtain feedback data; wherein, the feedback data includes: user satisfaction and problem-solving rate; use a regression model to predict the feedback data to obtain prediction data; generate an optimization signal according to the prediction data and send the optimization signal to the corresponding administrator;

[0046] The expression of the regression model is: wherein, X d represents the d-th feature variable; γ d represents the regression coefficient corresponding to the d-th feature variable; Y represents the predicted value, ∈ is the error term; β represents the benchmark predicted value.

[0047] The present invention generates a usage heat map according to the user's usage time period, adjusts the on-duty personnel according to the usage heat map, and when the number of questioners is large, increases the on-duty personnel, which can assist personnel when the system is busy and avoid longer user waiting times; predicts future feedback data according to the feedback data, and generates an optimization signal according to the prediction result, which can timely adjust the intelligent customer service and is beneficial to improving the usage performance of the intelligent customer service.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] 1. The present invention processes the table information, text information, and image information in the document information respectively, and performs corresponding processing for different document types, which can screen the key information in the document information and lay a data foundation for subsequent analysis of the document information; according to the user's input information, the input information is integrated with the text vector information to obtain a context vector, and the user's intention is analyzed according to the context vector to obtain the user's intention information, which can analyze the customer's needs and lay a data foundation for subsequent answering the user's needs; analyze the integrity of the user's intention information. When the user's intention information cannot clearly understand the user's needs, the user's intention information is supplemented to obtain the user's question information; it can clearly analyze the user's intention information and timely understand the user's actual needs, which is beneficial to improving the accuracy of the intelligent customer service analysis results.

[0050] 2. The present invention evaluates the risk of the user's question information according to the user's question information; since the user may input some sensitive words during the information input process, resulting in risks in the question information; therefore, evaluating the risk of the user's question information can improve the security of the intelligent customer and is beneficial to avoiding security risks of the intelligent customer service caused by sensitive words; analyze the risk degree of the sensitive words in the user's question information according to the risk score. When the risk score is greater than the risk threshold, a sensitive reminder signal is generated; remind the user to re-enter the question information, and use the sensitive detection model to analyze the question information, which can further analyze the sensitive words existing in the question information and is beneficial to reducing the risk of the intelligent customer service being blocked and improving the security of the intelligent customer service; convert the question information into a semantic vector, match the multi-modal data in the semantic vector knowledge base, and sort and integrate the multi-modal data according to the set priority to obtain the explanation information; it can provide comprehensive explanation data for the user's needs; it is beneficial to fully answer the user's needs and improve the user experience; update the multi-modal data in the knowledge base, screen the optimal data from several explanatory multi-modal data to ensure the consistency of the explanation content, and selectively delete the frequently-occurring answer content, which can reduce the load of the memory storing the knowledge base and is beneficial to improving the consistency and practicality of the data stored in the knowledge base; generate a usage heat map according to the user's usage time period, and adjust the duty personnel according to the usage heat map. When the number of questioners is large, the number of duty personnel is increased, which can assist personnel when the system is busy and avoid longer user waiting times; predict the future feedback data according to the feedback data, and generate an optimization signal according to the prediction result, which can timely adjust the intelligent customer service and is beneficial to improving the usage performance of the intelligent customer. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0052] Figure 1 Schematic diagram of the overall steps of the present invention;

[0053] Figure 2 Schematic diagram of the steps for extracting question information of the present invention;

[0054] Figure 3 Schematic diagram of the steps for explaining information and risk assessment of the present invention. Detailed implementation manners

[0055] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0056] Please refer to Figure 1 , the first aspect embodiment of the present invention provides an intelligent customer service system based on a large language model and high-precision document parsing, including: a document parsing module, an information collection module connected thereto, and a terminal processing module;

[0057] Information collection module: used to obtain the file information uploaded by the user; extract the key information in the file information using the target detection algorithm;

[0058] Document parsing module: used to identify the intention information of the user according to the key information; supplement the intention information to obtain the question information of the user; use the bidirectional LSTM model to evaluate the risk of the question information to obtain a risk score; generate a sensitive prompt signal according to the risk score;

[0059] Terminal processing module: used to match the corresponding content based on the question information of the user to generate explanatory information.

[0060] Please refer to Figure 2, obtain the file information uploaded by the user; among them, the file information includes: table information, text information, and image information; extract the table information in the file information; use the YOLO algorithm to detect the coordinate positions of the table information, and according to the coordinate positions, use image processing technology to segment the rows and columns of the table information and convert it into two-dimensional array information; extract the text information in the file information; use the word segmentation technology to divide the text information into several text words, and after removing the stop words in the several text words, use the word embedding vector model to convert the several text words into text vector information; extract the image information in the text information; mark the image information as I; perform convolution calculation through the formula F(I)=σ×(W0*I + b0) to identify the image feature information of the image information; where, W0 represents the convolution kernel, b0 represents the bias, σ represents the activation function, and * represents the convolution operation; integrate the two-dimensional array information, text vector information, and image feature information into key information.

[0061] It should be noted that when using the YOLO algorithm to detect the coordinate positions of the table information, the loss function of the detection target is defined as P = P conf +λ1P coord +λ2P cls , where, P conf is the confidence loss, P coord is the position loss, P cls is the class loss, and λ1 and λ2 are weight parameters.

[0062] It should be noted that the convolution kernel usually uses odd-sized square convolution kernels such as 3×3 and 5×5 to clarify the center point and reduce the number of parameters; the bias is used to adjust the output of the activation function to help the model better fit the data; the activation functions include ReLU, Sigmoid, Softmax; ReLU is preferred, and the output layer selects Sigmoid or Softmax according to the task.

[0063] Obtain the input information of the user, and retrieve the text vector information in the key information; convert the input information of the user into input vector information; mark the text vector information and the input vector information as Ec and Eu respectively; integrate the text vector and the input vector into the context vector V through the function V = f(Eu, Ec); match the context vector with the text content to obtain the intention information of the user; where, f(*) is the integration function;

[0064] Divide the user's intent information into several fields, and determine whether the several fields contain all types of fields; if so, analyze the grammar of the intent information; if not, analyze the context weight of the intent information; divide the intent information to obtain several elements; allocate context weights to the intent information according to the several elements, and supplement the user's intent information according to the context weight to obtain supplementary content; send the supplementary content to the user for selection; supplement the intent information according to the user's selection to obtain a question message.

[0065] It should be noted that the methods for obtaining the context weights include:

[0066] Use the attention mechanism to allocate context weights to the intent information. The calculation formula for context weight allocation is:

[0067] Among them, e i = W T h i + b; e i represents the score of the i-th element; e j represents the score of the j-th element; W represents the weight matrix; b is the bias parameter; h_i represents the hidden state of the i-th element.

[0068] It should be noted that the weight matrix maps the input data to the query, key, and value vector spaces to capture different semantic roles; the bias parameter is used to help the model better fit the data; the hidden state is the internal state calculated by the RNN when processing the input at the i-th time, and extracts the key information of the input data through a non-linear transformation (such as an activation function) to form a compact representation of the input.

[0069] Please refer to Figure 3 , retrieve the user's question message, and input the question message into a bidirectional LSTM model; calculate the risk score of the question message according to the self-attention mechanism; the calculation formula for the risk score is: F_X = θ × (W a [h f ; h b + b a ); among them, h f is the forward hidden state, h b is the backward hidden state; W a represents the weight parameter; b a represents the bias parameter.

[0070] It should be noted that the weight parameter maps the concatenated vector of the forward hidden state hf and the backward hidden state hb to the target space to capture context information; the bias parameter is used to better fit the data; the forward hidden state is the hidden state passed from front to back when the recurrent neural network processes sequential data, capturing the historical information in the sequence and used for predicting the future or understanding the context; the backward hidden state is the hidden state passed from back to front when the recurrent neural network processes sequential data, capturing the future information in the sequence and used for understanding the context or performing bidirectional modeling.

[0071] Compare the risk score with the risk threshold; when the risk score is greater than the risk threshold, generate a sensitive reminder signal; otherwise, divide the question information corresponding to the risk score into several elements; integrate the several elements into a sensitive input sequence; use the sensitive detection model of deep machine learning to analyze the sensitive input sequence and generate a sensitive reminder signal; return the standard prompt according to the sensitive reminder signal;

[0072] Retrieve the user's question information, and use the word vector model to convert the question information into a semantic vector; use the cross-modal alignment loss optimization formula to match the semantic vector with the multi-modal data in the knowledge base to obtain the explanation content; sort and integrate the explanation content according to the set type priority to obtain the explanation information, and send the explanation information to the user; the expression of the cross-modal alignment loss optimization formula is: where L represents the loss function of cross-modal alignment; f T and f I are the feature mapping functions of the text and image modalities respectively; T k represents the k-th text sample; I k represents the k-th image sample; ∥*∥ represents the norm operation.

[0073] Obtain the multi-modal data in the knowledge base; use the clustering analysis algorithm to perform clustering analysis on the multi-modal data; when there is multi-modal data in the knowledge base that does not match the question information, generate a knowledge point supplement signal and send the knowledge point supplement signal to the administrator; obtain the number of times of answering for several multi-modal data; when the number of times the multi-modal data appears is greater than the number threshold, select the optimal data from the multi-modal data in the same cluster as the answer content; otherwise, delete according to the importance of the multi-modal data.

[0074] It should be noted that: use the clustering analysis algorithm to perform clustering analysis on the multi-modal data; its objective function is: where C i represents the i-th cluster, and μ i is the cluster center point.

[0075] Obtain the user's usage time period, generate a usage heat map based on the user's usage time period; adjust the on-duty personnel based on the usage heat map; obtain feedback data; wherein, the feedback data includes: user satisfaction and problem-solving rate; use a regression model to predict the feedback data to obtain prediction data; generate an optimization signal according to the prediction data and send the optimization signal to the corresponding administrator; the expression of the regression model is: wherein, X d represents the d-th feature variable; γ d represents the regression coefficient corresponding to the d-th feature variable; Y represents the predicted value, ∈ is the error term; β represents the baseline predicted value.

[0076] Some of the data in the above formula are calculated by removing the dimension and taking its numerical value. The formula is obtained by software simulation of a large amount of collected data to get a formula closest to the actual situation; the preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained by simulation of a large amount of data.

[0077] The working principle of the present invention: The present invention obtains the file information uploaded by the user; uses a target detection algorithm to extract the key information in the file information; identifies the user's intention information according to the key information; supplements the intention information to obtain the user's question information; uses a bidirectional LSTM model to evaluate the risk of the question information to obtain a risk score; generates a sensitive prompt signal according to the risk score; matches the corresponding content based on the user's question information to generate an explanation information.

[0078] The above embodiments are only used to illustrate the technical method of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. An intelligent customer service system based on a large language model and high-precision document parsing, characterized in that, Including: A document parsing module, an information collection module connected thereto, and a terminal processing module; Information collection module: used to obtain the file information uploaded by the user; extract the key information in the file information using a target detection algorithm; Document parsing module: used to identify the user's intention information based on the key information; Supplement the intention information to obtain the user's question information; Use a bidirectional LSTM model to evaluate the risk of the question information to obtain a risk score; Generate a sensitive prompt signal according to the risk score; Terminal processing module: used to match the corresponding content based on the user's question information to generate an explanation information.

2. The intelligent customer service system based on large language models and high-precision document parsing according to claim 1, characterized in that, The extracting the key information in the file information using the target detection algorithm includes: Retrieve the file information; wherein, the file information includes: table information, text information and image information; extract the table information in the file information; use the YOLO algorithm to detect the coordinate position of the table information, and divide the rows and columns of the table information using image processing technology according to the coordinate position, and convert it into two-dimensional array information; Extract the text information in the file information; use the word segmentation technology to divide the text information into several text words, remove the stop words in the several text words, and then use the word embedding vector model to convert the several text words into text vector information; Extract the image information in the text information; mark the image information as I; perform convolution calculation through the formula F(I)=σ×(W0*I + b0) to identify the image feature information of the image information; where, W0 represents the convolution kernel, b0 represents the bias, σ represents the activation function, and * represents the convolution operation; integrate the two-dimensional array information, text vector information and image feature information into key information.

3. The intelligent customer service system based on large language models and high-precision document parsing according to claim 1, wherein The identifying the user's intention information based on the key information includes: Obtain the user's input information, and retrieve the text vector information in the key information; convert the user's input information into input vector information; Mark the text vector information and the input vector information as Ec and Eu respectively; integrate the text vector and the input vector into a context vector V through the function V = f(Eu, Ec); match the context vector with the text content to obtain the user's intention information; where, f(*) is an integration function.

4. The intelligent customer service system based on a large language model and high-precision document parsing according to claim 1, characterized in that The supplementing the intention information to obtain the user's question information includes: Retrieve the user's intention information; divide the user's intention information into several fields, and determine whether the several fields contain all types of fields; if so, analyze the grammar of the intention information; if not, analyze the context weight of the intention information; Divide the intention information to obtain several elements; assign context weights to the intention information according to the several elements, supplement the user's intention information according to the context weights to obtain supplementary content; send the supplementary content to the user for selection; supplement the intention information according to the user's selection to obtain question information.

5. The intelligent customer service system based on a large language model and high-precision document parsing according to claim 4, wherein The way to obtain the weight of the context includes: The attention mechanism is used to assign context weights to the intent information, and the calculation formula for context weight assignment is as follows: where, e i = W T h i + b; e i represents the score of the i-th element; e j represents the score of the j-th element; W represents the weight matrix; b is the bias parameter; hi represents the hidden state of the i-th element.

6. The intelligent customer service system based on the large language model and high-precision document parsing according to claim 1, wherein, The using the bidirectional LSTM model to evaluate the risk of the question information to obtain a risk score includes: Retrieve the user's question information, and input the question information into the bidirectional LSTM model; calculate the risk score of the question information according to the self-attention mechanism; The calculation formula for the risk score is: FX = θ × (W a [h f ; h b +b a ); where h f is the forward hidden state, and h b is the backward hidden state; W a represents the weight parameter; b a represents the bias parameter.

7. The intelligent customer service system based on the large language model and high-precision document parsing according to claim 1, characterized in that, Generating a sensitive prompt signal based on the risk score includes: Retrieving the risk score; comparing the risk score with the risk threshold; when the risk score is greater than the risk threshold, generating a sensitive reminder signal; Otherwise, dividing the question information corresponding to the risk score into several elements; integrating the several elements into a sensitive input sequence; analyzing the sensitive input sequence using a sensitive detection model of deep machine learning to generate a sensitive reminder signal; and returning a standard prompt message according to the sensitive reminder signal.

8. The intelligent customer service system based on the large language model and high-precision document parsing according to claim 1, characterized in that Generating an explanatory message by matching corresponding content based on the user's question information includes: Retrieving the user's question information, converting the question information into a semantic vector using a word vector model; matching the semantic vector with multi-modal data in the knowledge base using a cross-modal alignment loss optimization formula to obtain explanatory content; sorting and integrating the explanatory content according to the set type priority to obtain an explanatory message, and sending the explanatory message to the user; wherein the multi-modal data includes: text, pictures, and videos; The expression of the cross-modal alignment loss optimization formula is as follows: where L represents the loss function of cross-modal alignment; f T and f I are the feature mapping functions of the text and image modalities respectively; T k represents the k-th text sample; I k represents the k-th image sample; ∥*∥ represents the norm operation.

9. The intelligent customer service system based on the large language model and high-precision document parsing according to claim 8, wherein, The update method of the multi-modal data in the knowledge base includes: Obtaining the multi-modal data in the knowledge base; performing clustering analysis on the multi-modal data using a clustering analysis algorithm; when there is multi-modal data in the knowledge base that does not match the question information, generating a knowledge point supplement signal and sending the knowledge point supplement signal to the administrator; Obtaining the number of times several multi-modal data are answered; when the number of times the multi-modal data appears is greater than the number threshold, screening out the optimal data from the multi-modal data in the same cluster as the answer content; otherwise, deleting according to the importance of the multi-modal data.

10. The intelligent customer service system based on large language models and high-precision document parsing according to claim 1, wherein, The system further includes a prediction adjustment module, including: Obtaining the user's usage time period, generating a usage heat map according to the user's usage time period; adjusting the on-duty personnel based on the usage heat map; Obtaining feedback data; wherein the feedback data includes: user satisfaction and problem-solving rate; predicting the feedback data using a regression model to obtain prediction data; generating an optimization signal according to the prediction data and sending the optimization signal to the corresponding administrator; The expression of the regression model is as follows: where X d represents the d-th feature variable; γ d represents the regression coefficient corresponding to the d-th feature variable; Y represents the predicted value, ∈ is the error term; β represents the benchmark predicted value.

Images