AI customer service robot automatic reply system based on knowledge base

Abstract

The invention discloses an AI customer service robot automatic reply system based on a knowledge base, and relates to the technical field of intelligent customer service. Personal information of a user and a new question text sent by a user side are obtained, and user portrait analysis is carried out based on the personal information of the user; the emotion fluctuation state of the user is analyzed according to the historical dialogue record of the user, client classification is carried out, client classification results comprise emotion stable clients, pacifying demand clients and efficiency demand clients, and question feature words and emotion feature words are extracted based on the question text; querying a corresponding answer text in the knowledge base management unit according to the question feature words, judging the emotion of the user according to a customer classification result and the emotion feature words, correspondingly generating an emotion safety response text, and training based on a historical training text to obtain a text generation language model; and integrating the answer text and the emotion demand word into an answer text through a text generation language model, and outputting the answer text to a user side.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent customer service, and particularly to an AI customer service robot automatic reply system based on a knowledge base. Background Art

[0002] A customer service dialogue system is a dialogue system in a specific field. Currently, this field is a relatively cutting-edge research content. The main goal is to achieve automatic content reply of customer service while improving the efficiency of solving customer problems.

[0003] However, the current customer service automatic reply system often adopts a retrieval-based reply scheme. Although this scheme ensures reply accuracy, the retrieval-based reply scheme can often only reply to fixed questions and does not have generalization ability. Moreover, the retrieval-based reply often does not consider context information during the extraction and reply process of question phrases, resulting in low reply accuracy, difficulty in considering the user's emotions, inability to provide emotional value, and poor user experience.

[0004] In view of the above technical defects, a solution is proposed. Summary of the Invention

[0005] The purpose of the present invention is to integrate the answer text and emotional demand words into an answer text through a text generation language model and output it to the user side, so as to provide emotional comfort in combination with the user's emotions and ensure the satisfaction of the dialogue.

[0006] To achieve the above purpose, the present invention adopts the following technical solution: an AI customer service robot automatic reply system based on a knowledge base, including an information acquisition unit, a user portrait analysis unit, an emotion analysis unit, a reply generation unit, and a knowledge base management unit;

[0007] The knowledge base management unit includes a user data management module and a text storage module. The user data management module is used to store the personal information of users and update the user's personal information according to the user's dialogue records; the text storage module is used to preprocess the comprehensive data to obtain standard data, store the historical question text, historical answer text, and relevant search data in the standard data, and also store emotion coping texts for different user types;

[0008] The information acquisition unit is used to obtain the personal information of users from the knowledge base management unit and obtain the new question text sent from the user side, and send the personal information of users to the user portrait analysis unit and send the new question text to the emotion analysis unit;

[0009] The user portrait analysis unit analyzes the user portrait based on the user's personal information and the user's historical conversation records, and classifies customers by analyzing the user's emotional fluctuation state according to the historical conversation records. The customer classification results include emotionally stable customers, customers with soothing needs, and customers with efficiency needs;

[0010] The emotion analysis unit is used to obtain and process the question text, extract question feature words and emotion feature words based on the question text, query the corresponding answer text in the knowledge base management unit according to the question feature words and send it to the reply generation unit, judge the user's emotion according to the customer classification result and the emotion feature words, and generate corresponding emotion response texts and send them to the reply generation unit;

[0011] The reply generation unit is used to obtain and process the answer text and emotion demand words, integrate the historical training text based on the user conversation records in the knowledge base management unit, train the text generation language model based on the historical training text, and integrate the answer text and emotion demand words into the answer text through the text generation language model and output it to the user terminal.

[0012] The AI customer service robot automatic reply system based on the knowledge base according to claim 1, wherein the specific process of customer classification is as follows:

[0013] S101. Obtain the user's personal information, where the user's personal information includes gender, age, education level, and occupation, and characterize the user according to the personal information;

[0014] S102. Obtain the user's historical conversation records, extract the emotion feature words in the question text sent by the user according to the historical conversation records, classify each question text of the user according to the emotion feature words, and generate corresponding emotion labels;

[0015] S103. Record the frequency of occurrence of each emotion label, and arrange them in descending order according to the frequency to obtain an emotion feature arrangement table. Use the emotion label in the first place on the emotion feature arrangement table as the main emotion feature of the user portrait, and assign a value of R to different emotion labels. The greater the emotional volatility of the emotion label, the greater the assigned value. Calculate the user's emotion feature value Yi according to the following formula: where i = 1, 2, 3,..., n, and n is the number of emotion labels, is the average value of the emotion assignment. The emotion feature value is used to reflect the emotion trend represented by the user's emotion label. The larger the emotion feature value, the greater the emotional volatility of the user. On the contrary, the smaller the emotion feature value, the smaller the emotional volatility of the user;

[0016] The emotion tags include anger, disgust, anxiety, surprise, confusion, calmness, and pleasure, with corresponding emotion assignments from large to small;

[0017] S104. Obtain the preset emotion feature judgment range (Ymin, Ymax). If the user's emotion feature value Yi is less than or equal to Ymin, the customer classification result is an efficiency - demand type customer;

[0018] If the user's emotion feature value Yi is less than Ymax and greater than Ymin, the customer classification result is an emotion - stable type customer;

[0019] If the user's emotion feature value Yi is greater than or equal to Ymax, the customer classification result is a soothing - demand type customer.

[0020] Further, the specific process of querying the corresponding answer text is as follows:

[0021] S201. Perform normalization processing on the question text through a language processing tool, detect and correct typos and grammar errors in the question text to obtain a standard question text;

[0022] S202. Identify the question feature words and emotion feature words in the standard question text, and integrate the question feature words and emotion feature words into a question set and an emotion set respectively;

[0023] S203. Obtain the question set, query the comparison question text containing the question feature words from the knowledge base management unit, and obtain the target answer text corresponding to the comparison question text. Convert the comparison question text into a standard question vector through an Embedding model, and convert the target answer text into a matching question vector;

[0024] S204. Calculate the vector similarity between the standard question vector and the matching question vector, sort the matching question texts from large to small according to the vector similarity, and output the target answer text with the largest vector similarity as the answer text.

[0025] Further, the specific process of generating an emotion response text is as follows:

[0026] S301. Determine the user type according to the customer classification result, and query the corresponding emotion - coping text from the knowledge base management unit according to the user type. The emotion - coping text has multiple groups of standard coping texts for the above - mentioned emotion tags, and the standard coping texts include emotion soothing texts, standard response texts, and apology texts;

[0027] S302. Obtain the emotion feature words in the question text, judge the user's emotion according to the proportion of the emotion tags corresponding to the emotion feature words: If the user's emotion tends to be a soothing demand, select an emotion soothing text from the standard coping texts as the emotion response text;

[0028] If the user's emotion tends to be stable, select the standard response text from the standard response texts as the emotion response text;

[0029] If the user's emotion tends to be emotional venting, select the apology text from the standard response texts as the emotion response text.

[0030] Furthermore, the specific process of obtaining the text generation language model is as follows:

[0031] S401. Obtain the user's conversation record, obtain the question text and answer text sent by the user according to the time series in the user's conversation record, extract the emotion feature words from each question text, and extract the answer feature words from the answer text;

[0032] S402. Integrate the emotion feature words and answer feature words as the words to be analyzed, obtain the embedding vectors of the target elements in the words to be analyzed and each word element in the preset word list, where the target element is a preset number of word elements before the word to be analyzed in the text where the word to be analyzed is located, the word to be analyzed is any word element in the preset word list, and the word element is a word and punctuation mark with a word meaning and a tone;

[0033] S403. Traverse each word element in the preset word list, respectively determine the similarity between the embedding vector of each word element and the embedding vector of the word to be analyzed, construct the label vector of the word to be analyzed according to each similarity to obtain the training data set, each component of the label vector is a word meaning identifier and an emotion identifier, the word meaning representation is used to represent that the similarity corresponding to the component satisfies the preset word meaning expression condition, and the emotion identifier is used to represent that the similarity corresponding to the component satisfies the preset emotion expression condition;

[0034] S404. Train the preset network based on the training data set, and exit the training when the training exit condition is met to obtain the trained text generation language model.

[0035] Furthermore, it further includes a dialogue evaluation unit, which is used to obtain the current user's conversation record, sort the question texts sent by the user according to the time series in the user's conversation record, obtain the emotion feature words in the question texts one by one, calculate the user's emotion satisfaction according to the emotion labels and occurrence frequencies corresponding to the emotion feature words, judge the user's satisfaction with the automatic reply process according to the preset satisfaction evaluation range, and send the satisfaction result to the knowledge base management unit.

[0036] Furthermore, the specific process of judging the user's satisfaction with the automatic reply process is as follows:

[0037] S501. Obtain the emotional feature words in the question text one by one, obtain the corresponding emotional labels according to the emotional feature words, similarly assign the value of R to different emotional labels, and record the frequency M of each emotional label appearance;

[0038] S502. Calculate the user's emotional satisfaction degree Ui according to the following formula: where m is the number of emotional labels, and α is a preset proportional coefficient. The user's emotional satisfaction degree is used to reflect the degree of emotional fluctuation of the customer during the Q&A process. The greater the user's emotional satisfaction degree, the greater the degree of emotional fluctuation of the customer during the Q&A process. On the contrary, the smaller the user's emotional satisfaction degree, the smaller the degree of emotional fluctuation of the customer during the Q&A process;

[0039] S503. Obtain the preset satisfaction evaluation range (Umin, Umax). If the user's emotional satisfaction degree Ui is less than or equal to Umin, it means that the customer satisfaction is high;

[0040] If the user's emotional satisfaction degree Ui is less than Umax and greater than Umin, it means that the customer satisfaction is medium;

[0041] If the user's emotional satisfaction degree Ui is less than or equal to Umax, it means that the customer satisfaction is low.

[0042] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0043] The AI customer service robot automatic reply system based on the knowledge base obtains the personal information of the user and the new question text sent from the user side, analyzes the user portrait based on the personal information of the user, analyzes the emotional fluctuation state of the user according to the user's historical conversation records for customer classification. The customer classification results include emotionally stable customers, customers in need of appeasement, and customers in need of efficiency. Extract question feature words and emotional feature words based on the question text, query the corresponding answer text in the knowledge base management unit according to the question feature words, judge the user's emotion according to the customer classification result and the emotional feature words, and generate corresponding emotional response texts accordingly. Train the text generation language model based on the historical training text, and integrate the answer text and the emotional demand words into the answer text through the text generation language model and output it to the user side, which can provide emotional appeasement in combination with the user's emotion and ensure the satisfaction of the conversation. Brief Description of the Drawings

[0044] Figure 1 Shows the overall external structure schematic diagram of the present invention. Detailed Embodiment

[0045] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0046] Embodiment 1:

[0047] As Figure 1 shown, the AI customer service robot automatic reply system based on the knowledge base includes an information acquisition unit, a user portrait analysis unit, an emotion analysis unit, a reply generation unit, and a knowledge base management unit;

[0048] The knowledge base management unit includes a user data management module and a text storage module. The user data management module is used to store the personal information of users and update the user's personal information according to the user's conversation records; the text storage module is used to preprocess the comprehensive data to obtain standard data, store the historical question text, historical answer text, and relevant search data in the standard data, and also store emotion coping texts for different user types;

[0049] The information acquisition unit is used to obtain the personal information of users from the knowledge base management unit and obtain the new question text sent from the user terminal, and send the personal information of users to the user portrait analysis unit and the new question text to the emotion analysis unit;

[0050] The user portrait analysis unit analyzes the user portrait based on the personal information of users and the user's historical conversation records, and analyzes the emotional fluctuation state of users according to the user's historical conversation records for customer classification. The customer classification results include emotionally stable customers, customers with soothing needs, and customers with efficiency needs;

[0051] The specific process of customer classification is as follows:

[0052] S101. Obtain the personal information of users. The personal information of users includes gender, age, education level, and occupation. Characterize the portrait of users according to the personal information of users;

[0053] S102. Obtain the historical conversation records of users. Extract the emotional feature words in the question text sent by users according to the historical conversation records, classify each question text of users according to the emotional feature words, and generate corresponding emotional labels;

[0054] S103. Record the frequency of each emotion label, arrange them in descending order according to the frequency to obtain an emotion feature arrangement table. Use the emotion label in the first place on the emotion feature arrangement table as the main emotion feature of the user portrait, and assign a value of R to different emotion labels. The greater the emotional volatility of the emotion label, the greater the assigned value. Calculate the user's emotion feature value Yi according to the following formula: where i = 1, 2, 3, …, n, and n is the number of emotion labels. is the average value of the emotion assignment. The emotion feature value is used to reflect the emotion trend represented by the user's emotion label. The larger the emotion feature value, the greater the emotional volatility of the user. On the contrary, the smaller the emotion feature value, the smaller the emotional volatility of the user.

[0055] The emotion labels include anger, disgust, anxiety, surprise, doubt, calmness, and pleasure, and the corresponding emotion assignments are from large to small.

[0056] S104. Obtain the preset emotion feature judgment range (Ymin, Ymax). If the user's emotion feature value Yi is less than or equal to Ymin, the customer classification result is an efficiency-demand type customer;

[0057] If the user's emotion feature value Yi is less than Ymax and greater than Ymin, the customer classification result is an emotion-stable type customer;

[0058] If the user's emotion feature value Yi is greater than or equal to Ymax, the customer classification result is a soothing-demand type customer.

[0059] The emotion analysis unit is used to obtain and process the problem text, extract problem feature words and emotion feature words based on the problem text, query the corresponding answer text in the knowledge base management unit and send it to the reply generation unit, judge the user's emotion according to the customer classification result and emotion feature words, and generate corresponding emotion response text and send it to the reply generation unit;

[0060] The specific process of querying the corresponding answer text is as follows:

[0061] S201. Normalize the problem text through a language processing tool, detect and correct the typos and grammar errors in the problem text to obtain a standard problem text;

[0062] S202. Identify the problem feature words and emotion feature words in the standard problem text, and integrate the problem feature words and emotion feature words into a problem set and an emotion set respectively;

[0063] S203. Obtain a set of questions, query the comparison question text containing the question feature words from the knowledge base management unit, and obtain the target answer text corresponding to the comparison question text. Convert the comparison question text into a standard question vector through the Embedding model, and convert the target answer text into a matching question vector;

[0064] S204. Calculate the vector similarity between the standard question vector and the matching question vector, sort the matching question texts from largest to smallest according to the vector similarity, and output the target answer text with the largest vector similarity as the answer text.

[0065] The specific process of generating the emotional response text is as follows:

[0066] S301. Determine the user type according to the customer classification result, and query the corresponding emotional coping text from the knowledge base management unit according to the user type. There are multiple groups of standard coping texts for the above emotional labels, and the standard coping texts include emotional soothing texts, standard response texts, and apology texts;

[0067] S302. Obtain the emotional feature words in the question text, judge the user's emotion according to the emotional label corresponding to the emotional feature word and the proportion of the emotional label: If the user's emotion tends to the need for soothing, select the emotional soothing text from the standard coping texts as the emotional response text;

[0068] If the user's emotion tends to be stable, select the standard response text from the standard coping texts as the emotional response text;

[0069] If the user's emotion tends to emotional venting, select the apology text from the standard coping texts as the emotional response text.

[0070] The reply generation unit is used to obtain the answer text and the emotional demand words and process them, integrate the historical training text based on the user conversation record in the knowledge base management unit, train the text generation language model based on the historical training text, and integrate the answer text and the emotional demand words into the answer text through the text generation language model and output it to the user terminal.

[0071] The specific process of obtaining the text generation language model is as follows:

[0072] S401. Obtain the user conversation record, obtain the question text and the answer text sent by the user according to the time series in the user conversation record, extract the emotional feature words from each question text, and extract the answer feature words from the answer text;

[0073] S402. Integrate the emotion feature words and the Q&A feature words as the words to be analyzed, and obtain the embedding vectors of the target elements in the words to be analyzed and each word element in the preset word list. The target elements are the preset number of word elements before the word to be analyzed in the text where the word to be analyzed is located. The word to be analyzed is any word element in the preset word list, and the word element is a word and punctuation mark with a word meaning and a tone.

[0074] S403. Traverse each word element in the preset word list, respectively determine the similarity between the embedding vector of each word element and the embedding vector of the word to be analyzed, and construct the label vector of the word to be analyzed according to each similarity to obtain the training data set. Each component of the label vector is a word meaning identifier and an emotion identifier. The word meaning representation is used to represent that the similarity corresponding to the component meets the preset word meaning expression condition, and the emotion identifier is used to represent that the similarity corresponding to the component meets the preset emotion expression condition.

[0075] S404. Train the preset network based on the training data set, and exit the training when the training exit condition is met to obtain the trained text generation language model.

[0076] It also includes a dialogue evaluation unit. The dialogue evaluation unit is used to obtain the current user dialogue record, sort the question texts sent by the user according to the time series in the user dialogue record, obtain the emotion feature words in the question texts one by one, calculate the user's emotion satisfaction according to the emotion labels and occurrence frequencies corresponding to the emotion feature words, judge the user's satisfaction with the automatic reply process according to the preset satisfaction evaluation range, and send the satisfaction result to the knowledge base management unit.

[0077] The specific process of judging the user's satisfaction with the automatic reply process is as follows:

[0078] S501. Obtain the emotion feature words in the question texts one by one, obtain the corresponding emotion labels according to the emotion feature words, and also assign the value R to different emotion labels and record the occurrence frequency M of each emotion label.

[0079] S502. Calculate the user's emotion satisfaction Ui according to the following formula: where m is the number of emotion labels, and α is a preset proportional coefficient. The user's emotion satisfaction is used to reflect the degree of emotional fluctuation of the customer during the Q&A process. The greater the user's emotion satisfaction, the greater the degree of emotional fluctuation of the customer during the Q&A process. On the contrary, the smaller the user's emotion satisfaction, the smaller the degree of emotional fluctuation of the customer during the Q&A process.

[0080] S503. Obtain the preset satisfaction evaluation range (Umin, Umax). If the user's emotion satisfaction Ui is less than or equal to Umin, it means that the customer satisfaction is high.

[0081] If the user's emotional satisfaction level Ui is less than Umax and greater than Umin, it indicates medium customer satisfaction;

[0082] If the user's emotional satisfaction level Ui is less than or equal to Umax, it indicates low customer satisfaction.

[0083] The present invention obtains the user's personal information and the new problem text sent from the user terminal, performs user portrait analysis based on the user's personal information, analyzes the emotional fluctuation state of the user according to the user's historical conversation records for customer classification. The customer classification results include emotionally stable customers, customers with soothing needs, and customers with efficiency needs. It extracts problem feature words and emotional feature words based on the problem text, queries the corresponding answer text in the knowledge base management unit according to the problem feature words, judges the user's emotion according to the customer classification results and emotional feature words, and correspondingly generates an emotional response text. It trains a text generation language model based on historical training texts, integrates the answer text and emotional need words through the text generation language model and outputs the answer text to the user terminal, which can provide emotional soothing in combination with the user's emotion to ensure the satisfaction of the conversation.

[0084] The setting of the size of the interval is for the convenience of comparison. Regarding the size of the interval, it depends on the amount of sample data and the base quantity set by those skilled in the art for each group of sample data; as long as it does not affect the proportional relationship between the parameters and the quantified values.

[0085] The above formulas are all dimensionless and take their numerical calculations. The formula is a formula obtained by collecting a large amount of data for software simulation to approximate the real situation. The preset parameters in the formula are set by those skilled in the art according to the actual situation;

[0086] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its inventive concept, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. An AI customer service robot automatic reply system based on a knowledge base, characterized in that, It includes an information acquisition unit, a user portrait analysis unit, an emotion analysis unit, a response generation unit, and a knowledge base management unit; The knowledge base management unit includes a user data management module and a text storage module. The user data management module is used to store the personal information of users and update the personal information of users according to the user conversation records; the text storage module is used to preprocess the comprehensive data to obtain standard data, store the historical question text, historical answer text, and relevant search data in the standard data, and also store emotion coping texts for different user types; The information acquisition unit is used to obtain the personal information of users from the knowledge base management unit and obtain the new question text sent from the user terminal, and send the personal information of users to the user portrait analysis unit and the new question text to the emotion analysis unit; The user portrait analysis unit analyzes the user portrait based on the personal information of users and the historical conversation records of users, and analyzes the emotional fluctuation state of users based on the historical conversation records of users for customer classification. The customer classification results include emotionally stable customers, comfort - seeking customers, and efficiency - seeking customers; The emotion analysis unit is used to obtain and process the question text, extract question feature words and emotion feature words based on the question text, query the corresponding answer text in the knowledge base management unit according to the question feature words and send it to the response generation unit, judge the user's emotion according to the customer classification result and the emotion feature words, and generate corresponding emotion response texts and send them to the response generation unit; The response generation unit is used to obtain and process the answer text and emotion demand words, integrate the historical training text based on the user conversation records in the knowledge base management unit, train a text generation language model based on the historical training text, and integrate the answer text and emotion demand words into an answer text through the text generation language model and output it to the user terminal.

2. The AI customer service robot automatic reply system based on a knowledge base according to claim 1, wherein The specific process of customer classification is as follows: S101. Obtain the personal information of users. The personal information of users includes gender, age, education level, and occupation, and represent the portrait of users according to the personal information of users; S102. Obtain the historical conversation records of users, extract the emotion feature words in the question text sent by users according to the historical conversation records, classify each question text of users according to the emotion feature words, and generate corresponding emotion labels; S103. Record the frequency of each emotion label, arrange them in descending order according to the frequency to obtain an emotion feature arrangement table. Use the emotion label in the first place on the emotion feature arrangement table as the main emotion feature of the user portrait, and assign a value of R to different emotion labels. The greater the emotional volatility of the emotion label, the greater the assigned value. Calculate the user's emotion feature value Yi according to the following formula: where i = 1, 2, 3, …, n, and n is the number of emotion labels, is the average value of the emotion assignment, and the emotion feature value is used to reflect the emotion trend represented by the user's emotion label; S104. Obtain the preset emotion feature judgment range (Ymin, Ymax). If the emotion feature value Yi of users is less than or equal to Ymin, the customer classification result is efficiency - seeking customers; If the emotion feature value Yi of users is less than Ymax and greater than Ymin, the customer classification result is emotionally stable customers; If the emotion feature value Yi of users is greater than or equal to Ymax, the customer classification result is comfort - seeking customers.

3. The AI customer service robot automatic reply system based on a knowledge base according to claim 1, wherein The specific process of querying the corresponding answer text is as follows: S201. Normalize the question text through a language processing tool, detect and correct the spelling mistakes and grammar mistakes in the question text to obtain a standard question text; S202. Identify the problem feature words and emotion feature words in the standard problem text, and integrate the problem feature words and emotion feature words into a problem set and an emotion set respectively; S203. Obtain the problem set, query the comparison problem text containing the problem feature words from the knowledge base management unit, and obtain the target answer text corresponding to the comparison problem text. Convert the comparison problem text into a standard problem vector through the Embedding model, and convert the target answer text into a matching problem vector; S204. Calculate the vector similarity between the standard problem vector and the matching problem vector, sort the matching problem texts from largest to smallest according to the vector similarity, and output the target answer text with the largest vector similarity as the answer text.

4. The AI customer service robot automatic reply system based on a knowledge base according to claim 1, characterized in that, The specific process of generating the emotion response text is as follows: S301. Obtain the user type according to the customer classification result, and query the corresponding emotion response text from the knowledge base management unit according to the user type. The emotion response text has multiple groups of standard response texts for the above emotion labels. The standard response texts include emotion soothing texts, standard response texts, and apology texts; S302. Obtain the emotion feature words in the problem text, and judge the user's emotion according to the proportion of the emotion labels corresponding to the emotion feature words: If the user's emotion tends to the need for soothing, select the emotion soothing text from the standard response texts as the emotion response text; If the user's emotion tends to be stable, select the standard response text from the standard response texts as the emotion response text; If the user's emotion tends to emotional venting, select the apology text from the standard response texts as the emotion response text.

5. The AI customer service robot automatic reply system based on a knowledge base according to claim 1, characterized in that, The specific process of obtaining the text generation language model is as follows: S401. Obtain the user's conversation record, obtain the problem text and answer text sent by the user according to the time series in the user's conversation record, extract the emotion feature words from each problem text, and extract the question answering feature words from the answer text; S402. Integrate the emotion feature words and the question answering feature words as the words to be analyzed, and obtain the embedding vectors of the target elements and each word element in the preset word list in the words to be analyzed, where the target element is a preset number of word elements before the word to be analyzed in the text where the word to be analyzed is located, the word to be analyzed is any word element in the preset word list, and the word element is a word and punctuation mark with word meaning and tone; S403. Traverse each word element in the preset word list, determine the similarity between the embedding vector of each word element and the embedding vector of the word to be analyzed respectively, and construct the label vector of the word to be analyzed according to each similarity to obtain the training data set. Each component of the label vector is a word meaning identifier and an emotion identifier. The word meaning representation is used to characterize that the similarity corresponding to the component meets the preset word meaning expression condition, and the emotion identifier is used to characterize that the similarity corresponding to the component meets the preset emotion expression condition; S404. Train the preset network based on the training data set, and exit the training when the training exit condition is met to obtain the trained text generation language model.

6. The AI customer service robot automatic reply system based on a knowledge base according to claim 1, characterized in that, It further includes a dialogue evaluation unit, which is used to obtain the current user dialogue record, sort the question texts sent by the user according to the time series in the user dialogue record, obtain the emotional feature words in the question texts one by one, calculate the user's emotional satisfaction according to the emotional labels and occurrence frequencies corresponding to the emotional feature words, judge the user's satisfaction with the automatic reply process according to the preset satisfaction evaluation range, and send the satisfaction result to the knowledge base management unit.

7. The AI customer service robot automatic reply system based on a knowledge base according to claim 6, characterized in that, The specific process of judging the user's satisfaction with the automatic reply process is as follows: S501. Obtain the emotional feature words in the question texts one by one, obtain the corresponding emotional labels according to the emotional feature words, also assign the value of R to different emotional labels, and record the occurrence frequency M of each emotional label; S502. Calculate the user's emotional satisfaction degree Ui according to the following formula: where m is the number of emotion labels, α is a preset proportionality coefficient, and the user's emotional satisfaction degree is used to reflect the degree of emotional fluctuation of the customer during the Q&A process; S503. Obtain the preset satisfaction evaluation range (Umin, Umax). If the user's emotional satisfaction Ui is less than or equal to Umin, it means that the customer satisfaction is high; If the user's emotional satisfaction Ui is less than Umax and greater than Umin, it means that the customer satisfaction is medium; If the user's emotional satisfaction Ui is less than or equal to Umax, it means that the customer satisfaction is low.

Images