Customer service intelligent routing method fusing knowledge graph

Abstract

The invention discloses a customer service intelligent routing method fused with a knowledge graph, and relates to the technical field of data processing, and the method comprises the following steps: S1, optimizing a text sequence, S2, constructing the knowledge graph, S3, determining a question subject, and S4, outputting a question answer. After the dialogue voice data of the user is obtained, errors in voice recognition are directly corrected, so that the text better conforms to the content actually expressed by the user, the workload of manual proofreading is reduced, the quality of question and answer pairs in a corpus is improved, meanwhile, the question text provided by the user is expanded in combination with keywords and adjacent words, and the user experience is improved. The method is ensured to have relatively strong adaptability to different question methods, even if the knowledge graph is incomplete, the Skip-gram model can also determine adjacent words corresponding to each keyword, the occurrence of invalid responses is reduced, and the question and answer pair most fitted with the question text to be detected is screened out and then verified again, so that the detection efficiency is improved. And it is ensured that the method can provide better service for the user.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and specifically to an intelligent routing method for customer service integrating a knowledge graph. Background Art

[0002] The intelligent routing method for customer service integrating a knowledge graph enhances semantic understanding through structured domain knowledge, realizes accurate user intention recognition, allocates customer service ports for users by combining the capabilities and real-time loads of service nodes, and continuously optimizes strategies using feedback data to improve service efficiency and satisfaction. In the invention patent with the application number 202310601666.1, "An intelligent customer service method and system, belonging to the technical field of data processing, the method includes: receiving a voice consultation from a customer; recognizing the voice consultation according to the acoustic features and language features of the voice consultation through a voice recognition model; searching for the most relevant reply data to the voice consultation through a nearest neighbor query algorithm, and the robot automatically replies to the customer according to the reply data; in the case where the robot has difficulty solving the voice consultation, transferring to a human customer through a peak period intelligent queuing algorithm; in the case where the human customer service has difficulty solving the voice consultation, creating a work order and performing privacy protection processing on the content of the work order; dispatching professional personnel to process the work order. In the present invention, the voice consultation is recognized according to the acoustic features and language features of the voice consultation through a voice recognition model, which greatly improves the accuracy of voice recognition compared with relying solely on acoustic features."

[0003] The above-mentioned prior art solves problems such as the low correspondence when the customer service robot selects reply corpora for consultation questions. However, when this method is used, since it directly outputs the text result of the original voice recognition and cannot automatically correct errors, the text for subsequent processing does not match the actual content expressed by the user, and it completely relies on manual proofreading, thus increasing the cost of corpus construction. Moreover, this method can only process the original questions input by users and cannot expand similar question forms through keywords and adjacent words, and may not be able to match with the question-and-answer pairs in the corpus in the future, with poor adaptability. At the same time, this method does not verify the adaptability of the question-and-answer pairs, affecting the user experience. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent routing method for customer service integrating a knowledge graph to solve the problems raised in the above background art.

[0005] To achieve the above purpose, the present invention provides the following technical solution: An intelligent routing method for customer service integrating a knowledge graph, including the following steps:

[0006] S1. Optimize the text sequence: After obtaining the user's dialogue voice data, extract the text sequence corresponding to the current voice data, read the target characters, and determine the associated characters of each target character. Use the Transformer model to calculate the error probability of each character in the text sequence, select the characters with high error probability as target characters, replace the target characters with the associated characters with the lowest error probability, and output the replaced text sequence;

[0007] S2. Construct a knowledge graph: Receive the text sequence to generate question-answer pairs, use the vocabulary with the highest word frequency in the question-answer pairs as keywords. After setting the window size, obtain the adjacent words corresponding to each keyword, and then transmit the feature vectors of the keywords and adjacent words to the Skip-gram model. Update the model parameters by predicting the offset value between the adjacent words and the actual adjacent words, and construct a knowledge graph using the adjacent words corresponding to each keyword;

[0008] S3. Determine the question subject: After extracting the voice question sent by the user, use the Conformer model to analyze the corresponding text sequence, extract the keywords in the sequence. If the keywords exist in the knowledge graph, directly obtain the adjacent words; otherwise, predict the adjacent words through the Skip-gram model. Combine the keywords and adjacent words to generate a new text sequence, which is used as the question subject to be detected;

[0009] S4. Output the question answer: Extract the question subject to be detected. After determining the attribute features of the question to be detected and the attribute features of all question-answer pairs in the corpus, select the question-answer pair with the highest similarity in the corpus as the candidate question-answer pair, and verify it through the random forest model. If it is correct, output the answer of the candidate question-answer pair; if it is wrong, re-screen. If the user requests to transfer to the artificial port, assign it to the corresponding customer service port according to the keywords.

[0010] Preferably, S1 includes the following steps:

[0011] S101. After obtaining the user's dialogue voice data, divide the voice data into multiple data frames. After converting the time-domain signal of each data frame into a frequency-domain signal through Fourier transform, extract the voice features corresponding to each frame according to the time-domain signal and frequency-domain signal. After sorting the voice features along the time axis, transmit them to the Conformer model for analysis and output the text sequence;

[0012] S102. Extract multiple different characters as target characters. After reading multiple characters similar in pronunciation to each target character, use them as the associated characters of the current target character and store them in the candidate set of the target character.

[0013] Preferably, S1 further includes the following steps:

[0014] S103: After extracting the text sequence, perform a word segmentation operation on the text sequence, transmit the word segmentation result to the Transformer model for analysis, and obtain the error probability of each character. If the error probability is higher than a first threshold, the current character is used as the target character; otherwise, the current character is not used as the target character;

[0015] S104: Replace the original target characters one by one with the associated characters in the target character candidate set, calculate the error probability of the replaced characters, select the associated character with the lowest error probability as the optimal character at the current position, and output the replaced text sequence.

[0016] Preferably, the step S2 specifically includes the following steps:

[0017] S201: After receiving all text sequences, generate multiple question-answer pairs based on the text sequences, store them in a corpus, count the frequency of occurrence of each word in each question-answer pair in the corpus, and use words with a frequency higher than a preset value as keywords for the current question-answer pair;

[0018] S202. Set the window size, select the keywords of the question-answer pair, determine the adjacent words corresponding to each keyword according to the window size, count the number of all keywords, determine the vector dimension according to the number of keywords, and set the corresponding feature vector for each keyword according to the vector dimension.

[0019] Preferably, the step S2 further includes the following steps:

[0020] S203: Sample data is constructed using the feature vectors of the keywords and adjacent words. The sample data is transmitted to the Skip-gram model for analysis, and a conditional constraint formula is constructed. The predicted adjacent words of each keyword are calculated according to the conditional constraint formula. The model parameters are updated by the offset value between the predicted adjacent words and the actual adjacent words, thereby completing the training of the Skip-gram model. The conditional constraint formula is specifically:

[0021]

[0022] Among them, n represents the total number of words, k represents the position range of adjacent words corresponding to the keyword, and u s Indicates keywords, u r+s represents adjacent words, f(·) represents the conditional function value, and s, r represent parameters;

[0023] S204: After counting all keywords and their corresponding adjacent words, use the adjacent words corresponding to the keywords to establish a relationship between the current keyword and other keywords, and construct a knowledge graph based on the relationship between each keyword.

[0024] Preferably, the step S3 specifically includes the following steps:

[0025] S301. After extracting the voice question sent by the user, analyze the corresponding voice features according to the voice question, and transmit them to the Conformer model for analysis to obtain a text sequence. After performing word segmentation on the voice question using a word segmentation tool, screen out the keywords in the voice question and search in the knowledge graph according to the keywords.

[0026] S302. If the keyword exists in the knowledge graph, output the adjacent words corresponding to the keyword according to the knowledge graph. Otherwise, transmit the keyword to the Skip-gram model for analysis to obtain the predicted adjacent words of the keyword.

[0027] S303. Read the adjacent words corresponding to each keyword, form a new text sequence according to the keyword and the adjacent words, and use it as the problem body to be detected.

[0028] Preferably, the S4 specifically includes the following steps:

[0029] S401. Extract the problem body to be detected, and transmit the problem body to be detected and all the question-and-answer pairs in the corpus to the BERT model for analysis to obtain the attribute features of the problem to be detected and the attribute features of all the question-and-answer pairs in the corpus respectively.

[0030] S402. Use a similarity analysis algorithm to calculate the attribute features of the problem to be detected and the attribute features of all the question-and-answer pairs in the corpus, so as to determine the similarity between the problem body to be detected and all the question-and-answer pairs in the corpus, and select the question-and-answer pair with the highest similarity as the candidate question-and-answer pair for the current problem body to be detected.

[0031] Preferably, the S4 specifically further includes the following steps:

[0032] S403. Transmit the attribute features of the problem to be detected and the candidate question-and-answer pair to the random forest model for analysis to obtain the corresponding prediction result. If the result is correct, use the answer in the candidate question-and-answer pair as the answer to the problem to be detected. If the result is wrong, re-screen the candidate question-and-answer pair.

[0033] S404. After determining the answer to the problem to be detected, use the NaturalReader software to convert the answer to the problem to be detected from a text sequence into voice data, and transmit it to the user port. If the user port returns a transfer to artificial instruction, determine the consultation field according to the keyword of the problem to be detected, and select the corresponding artificial customer service port according to the consultation field.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] After obtaining the user's dialogue voice data, the present invention directly corrects the errors in speech recognition, making the text more in line with the actual content expressed by the user, reducing the workload of manual proofreading, improving the quality of the question-and-answer pairs in the corpus. At the same time, it combines keywords with adjacent words to expand the question text provided by the user, ensuring that the method has strong adaptability to different question forms. Even if the knowledge graph is imperfect, the Skip-gram model can determine the adjacent words corresponding to each keyword, reducing the occurrence of invalid responses. And after screening out the question-and-answer pair that best fits the text of the question to be detected, it will be verified again, ensuring that the method can provide better services for users. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flowchart of the overall method provided by an embodiment of the present invention;

[0037] Figure 2 It is a flowchart of the method for optimizing the text sequence provided by an embodiment of the present invention;

[0038] Figure 3 It is a flowchart of the method for constructing a knowledge graph provided by an embodiment of the present invention;

[0039] Figure 4 It is a flowchart of the method for determining the question subject provided by an embodiment of the present invention;

[0040] Figure 5 It is a flowchart of the method for outputting the question answer provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] 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. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0042] Please refer to Figures 1-5 , the present invention provides a technical solution: a customer service intelligent routing method integrating a knowledge graph, including the following steps:

[0043] S1. Optimize the text sequence: After obtaining the user's dialogue voice data, extract the text sequence corresponding to the current voice data, read the target characters, and determine the associated characters of each target character. Use the Transformer model to calculate the error probability of each character in the text sequence, screen out the characters with high error probability as target characters, replace the target characters with the associated characters with the lowest error probability, and output the replaced text sequence;

[0044] S2. Construct a knowledge graph: Receive a text sequence to generate question-answer pairs, use the vocabulary with the highest word frequency in the question-answer pairs as keywords. After setting the window size, obtain the adjacent words corresponding to each keyword, and then transmit the feature vectors of the keywords and adjacent words to the Skip-gram model. Update the model parameters by predicting the offset value between the adjacent word and the actual adjacent word, and construct a knowledge graph using the adjacent words corresponding to each keyword;

[0045] S3. Determine the question subject: After extracting the voice question sent by the user, use the Conformer model to analyze the corresponding text sequence, and extract the keywords in the sequence. If the keyword exists in the knowledge graph, directly obtain the adjacent words; otherwise, predict the adjacent words through the Skip-gram model, combine the keyword and the adjacent words to generate a new text sequence, and use it as the question subject to be detected;

[0046] S4. Output the question answer: Extract the question subject to be detected, determine the attribute features of the question to be detected and the attribute features of all question-answer pairs in the corpus, select the question-answer pair with the highest similarity in the corpus as the candidate question-answer pair, and verify it through the random forest model. If it is correct, output the answer of the candidate question-answer pair; if it is wrong, re-screen. If the user requests to transfer to the manual port, allocate it to the corresponding customer service port according to the keyword.

[0047] S1 includes the following steps:

[0048] S101. After obtaining the user's dialogue voice data, divide the voice data into multiple data frames, convert the time-domain signal of each data frame into a frequency-domain signal through Fourier transform, extract the voice features corresponding to each frame according to the time-domain signal and the frequency-domain signal, sort the voice features along the time axis, and then transmit them to the Conformer model for analysis to output a text sequence;

[0049] S102. Extract multiple different characters as target characters, read multiple characters similar in pronunciation to each target character, and use them as the associated characters of the current target character, and store them in the candidate set of the target character;

[0050] S1 also includes the following steps:

[0051] S103. After extracting the text sequence, perform a word segmentation operation on the text sequence, transmit the segmented result to the Transformer model for analysis to obtain the error probability of each character. If the error probability is higher than the first threshold, use the current character as the target character; otherwise, do not use the current character as the target character.

[0052] S104. After replacing the original target characters with the associated characters in the target character candidate set one by one, calculate the error probability of the replaced characters, select the associated character with the lowest error probability as the optimal character at the current position, and output the replaced text sequence;

[0053] S2 specifically includes the following steps:

[0054] S201. After receiving all text sequences, generate multiple question-and-answer pairs according to the text sequences, store them in the corpus, and after counting the occurrence frequencies of each vocabulary in each question-and-answer pair in the corpus, use the vocabulary with an occurrence frequency higher than the preset value as the keyword of the current question-and-answer pair;

[0055] S202. Set the window size, select the keywords of the question-and-answer pair, determine the adjacent words corresponding to each keyword according to the window size, count the number of all keywords, determine the vector dimension according to the number of keywords, and set the corresponding feature vector for each keyword according to the vector dimension;

[0056] S2 also specifically includes the following steps:

[0057] S203. Use the feature vectors of the keywords and adjacent words to construct sample data, transmit the sample data to the Skip-gram model for analysis, construct a conditional constraint formula, calculate the predicted adjacent words of each keyword according to the conditional constraint formula, and update the model parameters through the offset value between the predicted adjacent words and the actual adjacent words, thereby completing the training of the Skip-gram model. The conditional constraint formula is specifically:

[0058]

[0059] where n represents the total number of vocabularies, k represents the adjacent word position range corresponding to the keyword, u s represents the keyword, u r+s represents the adjacent word, f(·) represents the conditional function value, and s, r represent parameters;

[0060] S204. After counting all keywords and their corresponding adjacent words, use the adjacent words corresponding to the keywords to establish a relationship between the current keyword and other keywords, and construct a knowledge graph according to the relationships between each keyword;

[0061] S3 specifically includes the following steps:

[0062] S301. After extracting the voice question sentence sent by the user, analyze the corresponding voice features according to the voice question sentence, and transmit them to the Conformer model for analysis to obtain a text sequence. After using a word segmentation tool to perform word segmentation on the voice question sentence, screen out the keywords in the voice question sentence and search according to the keywords in the knowledge graph;

[0063] S302. If the keyword exists in the knowledge graph, the adjacent words corresponding to the keyword are output according to the knowledge graph. Otherwise, the keyword is transferred to the Skip-gram model for analysis to obtain the predicted adjacent words of the keyword;

[0064] S303, reading the adjacent words corresponding to each keyword, forming a new text sequence based on the keyword and the adjacent words, and using it as the subject of the question to be detected;

[0065] S4 specifically includes the following steps:

[0066] S401: Extract the question subject to be tested, transfer the question subject to be tested and all question-answer pairs in the corpus to the BERT model for analysis, and obtain the attribute features of the question to be tested and the attribute features of all question-answer pairs in the corpus respectively;

[0067] S402: Use a similarity analysis algorithm to calculate the attribute features of the question to be detected and the attribute features of all question-answer pairs in the corpus, thereby determining the similarity between the question subject to be detected and all question-answer pairs in the corpus, and select the question-answer pair with the highest similarity as the candidate question-answer pair for the current question subject to be detected. The specific similarity analysis algorithm is:

[0068]

[0069] Among them, q represents the attribute characteristics of the problem to be detected, Q c Represents the attribute features of the c-th question pair in the corpus, Con(q,Q c ) represents the attribute feature q of the question to be detected and the attribute feature Q of the question-answer pair in the corpus c The similarity between them, l(·) represents the distance function, q i Represents the i-th feature data in the problem attribute feature q to be detected, q j Represents the attribute features Q of the question-answer pair in the corpus c The jth feature data in , m represents the number of feature data of the problem attribute to be detected, i, j, c represent coefficients;

[0070] S4 specifically further includes the following steps:

[0071] S403: The attribute features of the question to be tested and the attribute features of the candidate question-answer pairs are transmitted to the random forest model for analysis to obtain the corresponding prediction results. If the result is correct, the answer in the candidate question-answer pair is used as the answer to the question to be tested. If the result is incorrect, the candidate question-answer pairs are re-screened.

[0072] S404. After determining the answer to the question to be detected, use NaturalReader software to convert the answer to the question to be detected from a text sequence into voice data, and then transmit it to the user port. If the user port returns a transfer-to-artificial instruction, determine the consultation field according to the keywords of the question to be detected, and select the corresponding artificial customer service port according to the consultation field.

[0073] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0074] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A customer service intelligent routing method integrating a knowledge graph, characterized in that The method comprises the following steps: S1. Optimize text sequences: After acquiring user conversation voice data, extract the text sequence corresponding to the current voice data, read the target characters, and determine the associated characters of each target character. Use the Transformer model to calculate the error probability of each character in the text sequence, select characters with high error probabilities as target characters, replace the target characters with the associated characters with the lowest error probability, and output the replaced text sequence. S2. Build a knowledge graph: Receive a text sequence to generate question-answer pairs. Use the most frequent word in the question-answer pair as the keyword. After setting the window size, obtain the adjacent words corresponding to each keyword. Transfer the feature vectors of the keyword and adjacent words to the Skip-gram model. Update the model parameters by the offset value between the predicted adjacent words and the actual adjacent words. Use the adjacent words corresponding to each keyword to build a knowledge graph. S3. Determine the question subject: After extracting the user's voice question, use the Conformer model to analyze the corresponding text sequence and extract the keywords in the sequence. If the keyword exists in the knowledge graph, the adjacent words are directly obtained. Otherwise, the adjacent words are predicted using the Skip-gram model. A new text sequence is generated by combining the keywords and adjacent words, which is used as the question subject to be detected. S4. Output the answer to the question: Extract the question subject to be tested, determine the attribute features of the question to be tested and the attribute features of all question-answer pairs in the corpus, select the question-answer pair with the highest similarity in the corpus as the candidate question-answer pair, and verify it through the random forest model. If it is correct, output the answer to the candidate question-answer pair. If it is wrong, re-screen it. If the user requests to transfer to the manual port, it will be assigned to the corresponding customer service port according to the keyword.

2. The customer service intelligent routing method integrating a knowledge graph according to claim 1, characterized in that: Said S1 comprises the following steps: S101: After acquiring user conversation voice data, the voice data is divided into multiple data frames. After converting the time domain signal of each data frame into a frequency domain signal through Fourier transform, the voice features corresponding to each frame are extracted based on the time domain signal and the frequency domain signal. After sorting the voice features according to the time axis, the voice features are transferred to the Conformer model for analysis and the text sequence is output; S102: extract multiple different characters as target characters, read multiple characters with similar pronunciation to each target character, use them as associated characters of the current target character, and store them in a candidate set of target characters.

3. The customer service intelligent routing method integrating a knowledge graph according to claim 2, wherein: Said S1 further comprises the following steps: S103: After extracting the text sequence, perform a word segmentation operation on the text sequence, transmit the word segmentation result to the Transformer model for analysis, and obtain the error probability of each character. If the error probability is higher than a first threshold, the current character is used as the target character; otherwise, the current character is not used as the target character; S104: Replace the original target characters one by one with the associated characters in the target character candidate set, calculate the error probability of the replaced characters, select the associated character with the lowest error probability as the optimal character at the current position, and output the replaced text sequence.

4. The customer service intelligent routing method integrating a knowledge graph according to claim 1, characterized in that: The S2 specifically includes the following steps: S201. After receiving all text sequences, generate multiple question-and-answer pairs according to the text sequences, store them in the corpus, and after counting the occurrence frequencies of each word in each question-and-answer pair in the corpus, use the words with occurrence frequencies higher than the preset value as the keywords of the current question-and-answer pair; S202. Set the window size. After selecting the keywords of the question-and-answer pair, determine the adjacent words corresponding to each keyword according to the window size, count the number of all keywords, determine the vector dimension according to the number of keywords, and set the corresponding feature vector for each keyword according to the vector dimension.

5. The customer service intelligent routing method integrating a knowledge graph according to claim 4, characterized in that: The specific steps of S2 also include the following: S203. Use the feature vectors of the keywords and adjacent words to construct sample data, transmit the sample data to the Skip-gram model for analysis, construct a conditional constraint formula, calculate the predicted adjacent words of each keyword according to the conditional constraint formula, and update the model parameters through the offset value between the predicted adjacent words and the actual adjacent words, so as to complete the training of the Skip-gram model; S204. After counting all keywords and their corresponding adjacent words, use the adjacent words corresponding to the keywords to establish relationships between the current keyword and other keywords, and construct a knowledge graph according to the relationships between each keyword.

6. The customer service intelligent routing method integrating a knowledge graph according to claim 1, characterized in that: The specific steps of S3 include the following: S301. After extracting the voice question sent by the user, analyze the corresponding voice features according to the voice question and transmit them to the Conformer model for analysis to obtain a text sequence. After using a word segmentation tool to perform word segmentation on the voice question, screen out the keywords in the voice question and search according to the keywords in the knowledge graph; S302. If the keyword exists in the knowledge graph, output the adjacent words corresponding to the keyword according to the knowledge graph. Otherwise, transmit the keyword to the Skip-gram model for analysis to obtain the predicted adjacent words of the keyword; S303. Read the adjacent words corresponding to each keyword, form a new text sequence according to the keyword and the adjacent words, and use it as the problem subject to be detected.

7. The customer service intelligent routing method integrating a knowledge graph according to claim 1, characterized in that: The specific steps of S4 include the following: S401. Extract the problem subject to be detected, transmit the problem subject to be detected and all question-and-answer pairs in the corpus to the BERT model for analysis, and obtain the problem attribute features to be detected and the attribute features of all question-and-answer pairs in the corpus respectively; S402. Use a similarity analysis algorithm to calculate the problem attribute features to be detected and the attribute features of all question-and-answer pairs in the corpus, so as to determine the similarity between the problem subject to be detected and all question-and-answer pairs in the corpus, and select the question-and-answer pair with the highest similarity as the candidate question-and-answer pair for the current problem subject to be detected.

8. The customer service intelligent routing method integrating a knowledge graph according to claim 7, wherein: The specific steps of S4 also include the following: S403. Transmit the problem attribute features to be detected and the attribute features of the candidate question-and-answer pair to the random forest model for analysis to obtain the corresponding prediction result. If the result is correct, use the answer in the candidate question-and-answer pair as the answer to the problem to be detected. If the result is wrong, re-screen the candidate question-and-answer pair; S404. After determining the answer to the problem to be detected, use the NaturalReader software to convert the answer to the problem to be detected from a text sequence into voice data, and then transmit it to the user port. If the user port returns a transfer-to-artificial instruction, determine the consultation field according to the keywords of the problem to be detected, and select the corresponding artificial customer service port according to the consultation field.

Images