A method and system for handling customer consultation issues based on machine learning

Abstract

The present invention relates to the field of machine learning and discloses a method and system for processing customer consultation questions based on machine learning. The method comprises: obtaining a customer's consultation question; identifying the consultation question using a neural network semantic matching platform to obtain a question vocabulary for the consultation question; searching a consultation database based on the question vocabulary: when the search is successful, obtaining an answer text for the consultation question; when the search fails, generating an answer text for the consultation question in real time; obtaining a customer's evaluation of the answer text; obtaining an answer vocabulary for the answer text based on the neural network semantic matching platform and the evaluation; retraining a language processing model based on the answer vocabulary and the question vocabulary to obtain an updated language processing model, and transmitting the updated model back to the consultation database; obtaining the user's real-time question, and returning to step S2 to obtain a real-time answer to the real-time question. The present invention can improve the efficiency of real-time processing of consultation questions and the quality of answers.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a method and system for handling customer consultation issues based on machine learning. Background Art

[0002] In machine learning-based customer inquiry processing, slow real-time response speed and low-quality responses are particularly prominent. When handling customer inquiries, the system's real-time responsiveness is significantly compromised due to a fragmented interactive experience and high real-time data processing volume, resulting in frequent response delays. Furthermore, due to factors such as limited training data, high costs for adapting domain knowledge, and a lack of feedback, the system is prone to providing ambiguous or incorrect answers to low-frequency, complex questions. The generated answers generally lack empathy and are of questionable quality.

[0003] Traditional machine learning methods cannot directly support real-time data analysis and integration, and have low data processing efficiency. Multi-round dialogue context maintenance methods have complex feature engineering problems, and dialogue strategy decision-making methods have rigid answers. These traditional methods not only significantly reduce the real-time processing speed of consulting questions, but also seriously affect the quality of answers, making it difficult to effectively solve customer consulting problems and greatly reducing the customer experience. Summary of the Invention

[0004] The present invention provides a method and system for handling customer consultation questions based on machine learning, the main purpose of which is to address the problems of low efficiency in real-time answering to consultation questions and low answer quality.

[0005] To achieve the above objectives, the present invention provides a method for handling customer consultation issues based on machine learning, comprising:

[0006] 1. A method for handling customer consultation issues based on machine learning, characterized in that the method comprises:

[0007] S1. Obtain customer consultation questions;

[0008] S2. Identify the consulting question using a neural network semantic matching platform to obtain a vocabulary list of the consulting question;

[0009] S3. Search the consultation database according to the question vocabulary:

[0010] S31. When the search is successful, obtain the answer text of the consultation question;

[0011] S32. When the search fails, generating an answer text to the consultation question in real time based on the question vocabulary and the language processing model;

[0012] S4. Obtaining the customer's evaluation of the answer text;

[0013] S5. Based on the neural network semantic matching platform and the evaluation, the answer text is disassembled to obtain an answer vocabulary list for the answer text;

[0014] S6. Retrain the language processing model based on the answer vocabulary and the question vocabulary to obtain an updated language processing model, and transmit the updated language processing model, the question vocabulary, and the answer vocabulary back to the consultation database;

[0015] S7. Obtain the user's real-time question and return to step S2 to obtain a real-time answer to the real-time question.

[0016] In a preferred embodiment, the use of a neural network semantic matching platform to identify the consulting question and obtain a question vocabulary for the consulting question includes:

[0017] Decomposing the consulting question to obtain the question text;

[0018] Performing text vectorization processing on the question text to obtain a question vector of the question text;

[0019] The question vectors are semantically arranged based on a neural network semantic matching platform to obtain a question vocabulary list for the consulting question.

[0020] In a preferred embodiment, when the retrieval fails, generating the answer text of the consulting question in real time based on the question vocabulary and the language processing model includes:

[0021] Answer the question vocabulary based on the language processing model to obtain an answer vocabulary vector for the question vocabulary;

[0022] The answer word vector is processed based on an autoregressive algorithm to obtain the answer text of the consultation question, wherein the autoregressive algorithm formula is:

[0023]

[0024] Where, For the Step generated words, is the autoregressive function, For the generation step, To generate vocabulary, For vocabulary.

[0025] In a preferred embodiment, the processing of the answer vocabulary vector based on the autoregressive algorithm to obtain the answer text of the consultation question includes:

[0026] The optimal answer text is selected based on a re-ranking algorithm, wherein the re-ranking algorithm is:

[0027]

[0028] Where, is the optimal answer text, is the generation probability score of the candidate answer, For language fluency, For the relevance of the problem, and is the probability coefficient, For the generation step, To generate vocabulary, For the vocabulary list;

[0029] The optimal answer text is used as the answer text of the consultation question.

[0030] In a preferred embodiment, the evaluation includes:

[0031] Whether the answer solves the customer's problem, the degree of emotion contained in the answer, the clarity of the answer, the completeness of the answer, and the customer's satisfaction with the consultation process.

[0032] In a preferred embodiment, the answer text is disassembled based on the neural network semantic matching platform and the evaluation to obtain an answer vocabulary for the answer text, including:

[0033] The answer text is split based on the BERT algorithm in the neural network semantic matching platform and the evaluation to obtain an answer vector for the answer text, wherein the BERT algorithm is:

[0034]

[0035] Where, is the answer vector for the answer text, is a matrix arrangement of the characters in the answer text, is the transpose of the matrix permutation, is the vector set of answer vectors, is the vector dimension;

[0036] An association relationship between the question vector and the answer vector is established, and an answer vocabulary of the answer text is generated according to the association relationship.

[0037] In a preferred embodiment, the retraining of the language processing model based on the answer vocabulary and the question vocabulary to obtain an updated language processing model includes:

[0038] Performing vector extraction on the answer vocabulary and the consulting question vocabulary respectively to obtain a question vector and an answer vector of the consulting question;

[0039] The language processing model is trained based on the question vector and the answer vector to obtain an updated language processing model.

[0040] In a preferred embodiment, the training of the language processing model based on the question vector and the answer vector to obtain an updated language processing model includes:

[0041] The language processing model is weighted based on a self-attention mechanism algorithm, the question vector, and the answer vector, wherein the self-attention mechanism algorithm is:

[0042]

[0043] Where, is the updated weight of the language processing model, is the corresponding parameter index, is an input vector determined by the question vector and the answer vector, is the value of the input vector in the question vocabulary and the answer vocabulary, is the normalized exponential function, is the identity of the matrix transpose, is the length of the input vector;

[0044] Parameters of the language processing model are updated based on the updated weights to obtain an updated language processing model.

[0045] In a preferred embodiment, the step of obtaining the user's real-time question and returning to step S2 to obtain a real-time answer to the real-time question includes:

[0046] Obtaining a user's real-time question, searching for an answer to the real-time question in the consultation database, and when the search is successful, outputting the answer in the consultation database as a real-time answer to the real-time question;

[0047] When the retrieval fails, a real-time answer to the real-time question is obtained based on the updated language processing model.

[0048] In order to solve the above problems, the present invention also provides a customer consultation problem processing system based on machine learning, the system comprising:

[0049] Customer consultation module: used to obtain customer consultation questions;

[0050] A question vocabulary generation module is used to identify the consulting question using a neural network semantic matching platform to obtain a question vocabulary for the consulting question;

[0051] Answer text module: used to search the consultation database according to the question vocabulary, and when the search is successful, obtain the answer text of the consultation question; when the search fails, generate the answer text of the consultation question in real time based on the question vocabulary and language processing model;

[0052] Customer evaluation module: used to obtain the customer's evaluation of the answer text;

[0053] An answer vocabulary generation module is configured to disassemble the answer text based on the neural network semantic matching platform and the evaluation to obtain an answer vocabulary for the answer text;

[0054] Model retraining module: used for retraining the language processing model based on the answer vocabulary and the question vocabulary to obtain an updated language processing model, and returning the updated language processing model, the question vocabulary and the answer vocabulary to the consultation database;

[0055] Database update module: used to obtain the user's real-time questions and return to step S2 to obtain real-time answers to the real-time questions.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] 1. Utilize a neural network semantic matching platform and language processing models to conduct multiple deep learning exercises on the same question, breaking down deeper logical connections, effectively maintaining contextual associations, and improving semantic recognition accuracy in multilingual scenarios, thereby increasing customer satisfaction with answers.

[0058] 2. By storing customer consultation issues and solutions in a consultation database, when the database receives the same or similar consultation request again, it can rely on the neural network semantic matching platform to prioritize the relevant solutions from the database. This mechanism effectively reduces the real-time data processing load of repeated routine consultations, thereby shortening the average waiting time for customers;

[0059] 3. Based on customers' actual evaluation and feedback data on solution solutions, a continuously optimized machine learning model was established. By analyzing the feedback differences in different scenarios, the model can dynamically adjust the answer strategy and continuously improve the accuracy of generated answers and service adaptability. This intelligent improvement not only significantly improves service efficiency, but also realizes self-improvement of the database through data closure, which significantly improves the efficiency of real-time problem solving. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 A flowchart of a method for solving customer consultation problems based on machine learning provided by one embodiment of the present invention;

[0061] Figure 2 A functional module diagram of a customer consultation system based on machine learning provided by one embodiment of the present invention;

[0062] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0063] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0064] The embodiment of the present application provides a method and system for handling customer consultation issues based on machine learning. The execution subject of the method and system for handling customer consultation issues based on machine learning includes but is not limited to at least one of the electronic devices such as a server, a terminal, etc. that can be configured to execute the method provided by the embodiment of the present application. In other words, the method and system for handling customer consultation issues based on machine learning can be executed by software or hardware installed on a terminal device or a server device. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc. The server can be an independent server, or it can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0065] Reference Figure 1 FIG2 is a flow chart of a method and system for handling customer consultation issues based on machine learning according to an embodiment of the present invention. In this embodiment, the method and system for handling customer consultation issues based on machine learning include:

[0066] S1. Get the customer's consultation questions.

[0067] Specifically, the methods of obtaining customer consultation questions include online platform collection, manual collection, etc.

[0068] Specifically, the consulting questions include: "How to improve the generalization ability and explainability of AI models?", "How to select and optimize the technology stack?" and "How to design a subscription system and results-sharing model?" etc.

[0069] S2. Using a neural network semantic matching platform to identify the consulting question, and obtain a question vocabulary list of the consulting question.

[0070] In an embodiment of the present invention, the use of a neural network semantic matching platform to identify the consulting question and obtain a question vocabulary for the consulting question includes:

[0071] Decomposing the consulting question to obtain the question text;

[0072] Performing text vectorization processing on the question text to obtain a question vector of the question text;

[0073] The question vectors are semantically arranged based on a neural network semantic matching platform to obtain a question vocabulary list for the consulting question.

[0074] Specifically, the consultation questions are segmented and tagged with parts of speech, the content after segmentation is filtered according to the stop word list, and the meaningful words after the text segmentation are retained according to the parts of speech.

[0075] For example, the content of the consultation question is: "How does the smart home system developed by a technology company ensure user data security and prevent privacy leakage?" The question is segmented, and the result is: "Technology|Company|Developed|Smart Home|System|,|How|To|Ensure|User Data|Security|,|Prevent|Privacy|Leakage|?", and the corresponding part-of-speech tagging is: "Technology (noun)|Company (noun)|Developed (verb)|of (particle)|Smart Home (noun)|System (noun)|, (punctuation)|How (pronoun)|Ensure (verb)|User Data (noun)|Security (noun)|, (punctuation)|Prevent (verb)|Privacy (noun)|Leakage (verb)|? (punctuation)"; according to the part-of-speech, the meaningful words after the text segmentation are retained, and the results are: smart home, user data, security, privacy, leakage.

[0076] Specifically, the question text is vectorized based on the Word2Vec algorithm, wherein the calculation formula of the Word2Vec algorithm is:

[0077]

[0078] Where, is the vocabulary similarity of the question text, The first words, The first words, is the cosine value between two words.

[0079] Furthermore, in the Word2Vec algorithm, the higher the word similarity, the higher the degree of connection between the two words. The more connected the words are, the more likely they are to be meaningful words.

[0080] Specifically, the neural network semantic matching platform uses the TextRank algorithm for semantic ranking and constructs the required question vocabulary by calculating the importance of words. The TextRank algorithm is:

[0081]

[0082] Where, For the importance of vocabulary, is the position coefficient , is the position coefficient, is the position coefficient The impact value of is the influence value of the coefficient at any position, is the position coefficient for any position pair Find the derivative.

[0083] Furthermore, The higher the value, the more important the word is in the sentence. The more important the word is, the more likely it is to become an important part of the sentence.

[0084] Furthermore, Indicates the position coefficient The impact value of The calculation formula is:

[0085]

[0086] In the formula, ε is a parameter greater than 1.

[0087] Overall, using a neural network semantic matching platform to identify consulting questions and generate a question vocabulary is a significant initiative. Leveraging its powerful natural language processing capabilities, the platform can deeply analyze the semantic structure of consulting questions, accurately extract key information from complex sentences, and break down consulting questions into meaningful lexical units, which are then combined to generate a question vocabulary.

[0088] In general, the question vocabulary constructed in this way not only fully covers core elements such as technical fields, involved objects, and key behaviors, but also effectively filters out redundant and irrelevant information. In subsequent consultation database searches, these precise terms can quickly locate matching records and solutions for the consultation questions, significantly improving search accuracy and efficiency and laying a solid foundation for efficiently resolving user inquiries.

[0089] S3. Search the consultation database according to the question vocabulary.

[0090] S31. When the retrieval is successful, the answer text of the consultation question is obtained.

[0091] Specifically, when processing the search results, if the search results returned by the database contain duplicate records, or there are records that do not fully match the actual needs, the results need to be screened to remove duplicate records; at the same time, according to actual business needs, further filter out records that do not meet the requirements.

[0092] For example, to retrieve records containing "smart home" and "user data" at the same time, the query statement can be "SELECT * FROM consulting_database WHERE question_content LIKE '% smart home%' AND question_content LIKE'% user data%'".

[0093] Furthermore, the search results are sorted according to certain rules, and the processed search results are presented to the user in a suitable form or for use in subsequent processing procedures; the consulting questions and corresponding solutions can be displayed in a list form, or the results can be displayed in the form of charts, reports, etc. according to specific application scenarios for users to view and analyze.

[0094] In general, searching the consultation database based on the problem vocabulary can quickly locate records related to the current problem vocabulary from massive data; at the same time, by accurately matching vocabulary units, historically similar consultation cases and their solutions can be screened out, providing valuable data support for subsequent analysis; and combining language processing models with training is the key link in achieving deep data mining.

[0095] S32. When the retrieval fails, the answer text of the consulting question is generated in real time based on the question vocabulary and language processing model.

[0096] In an embodiment of the present invention, when the retrieval fails, generating the answer text of the consulting question in real time based on the question vocabulary and the language processing model includes:

[0097] Answer the question vocabulary based on the language processing model to obtain an answer vocabulary vector for the question vocabulary;

[0098] The answer word vector is processed based on an autoregressive algorithm to obtain the answer text of the consultation question, wherein the autoregressive algorithm formula is:

[0099]

[0100] Where, For the Step generated words, is an autoregressive function, is a generation step, is a generated word, is a vocabulary.

[0101] In an embodiment of the present invention, processing the answer vocabulary vector based on the autoregressive algorithm to obtain the answer text of the consultation question includes:

[0102] Selecting the optimal answer text based on a re-ranking algorithm, where the re-ranking algorithm is:

[0103]

[0104] In the formula, is the optimal answer text, is the generation probability score of the candidate answer, is the language fluency, is the relevance of the question, and is the probability coefficient, is a generation step, is a generated word, is a vocabulary.

[0105] Taking the optimal answer text as the answer text of the consultation question.

[0106] Specifically, when performing data preprocessing and vocabulary list collation, first check the given question vocabulary, and剔除 duplicate words and misspelled words that may exist; secondly, perform standardization operations such as unifying case and removing special symbols on the vocabulary; finally, check the data of the consultation questions and related solutions stored in the consultation database to ensure the integrity and accuracy of the data.

[0107] For example, if both "智家居" and "智能家居" appear in the vocabulary, they need to be unified into the correct "智能家居"; if "user-data" appears in the vocabulary, it is unified into "user_data". Specifically, when performing the retrieval operation, for each word in the vocabulary, construct a corresponding query statement; if there are multiple words in the vocabulary, the queries corresponding to these words need to be reasonably combined. Logical operators (such as AND, OR) can be used to connect multiple query conditions; then, submit the constructed query statement to the database management system, and the database performs the query operation to find matching records in the index and data storage areas.

[0108] Specifically, based on a simple combination model of GCN and BERT, the preprocessed vocabulary is combined, and the answer text of the consultation question is obtained by learning the local and global features of the text. The calculation formula of the simple combination model of GCN and BERT is:

[0109]

[0110] Where, is the final vector representation, The text content processed by the GCN algorithm, This is the text content processed by the BERT algorithm.

[0111] Specifically, the obtained answer text will be sorted according to the re-sorting algorithm, and the processed answer text will be presented to the user in a suitable form or for use in subsequent processing processes. The consulting questions and corresponding solutions can be displayed in a list form, or the results can be displayed in the form of charts, reports, etc. according to the specific application scenarios for users to view and analyze.

[0112] In general, convolutional neural networks possess powerful feature extraction capabilities. They can deeply learn the relationships between words in the question vocabulary and related data retrieved from the database, capturing semantic patterns and underlying regularities. During the ongoing training process, the model continuously optimizes its parameters, gradually improving its ability to understand and answer consulting questions, ultimately generating high-quality answer text. This answer text not only integrates existing knowledge in the database but also leverages the model's intelligent analytical capabilities to better meet the actual needs of the current consultation question, thereby providing users with accurate and effective answers.

[0113] S4. Obtaining the customer's evaluation of the answer text;

[0114] In an embodiment of the present invention, the evaluation includes: whether the answer solves the customer's problem, the degree of emotion contained in the answer, the clarity of the answer, the completeness of the answer, and the customer's satisfaction with the consultation process.

[0115] Specifically, customers' evaluations of answers typically cover multiple aspects, starting with whether the answer addresses the customer's problem. This includes whether the answer accurately addresses the inquiry question and contains no errors or misleading information. For example, for the question "How do smart home systems ensure user data security?", the answer should accurately describe security measures such as encryption and access control.

[0116] Specifically, the completeness of an answer refers to whether the answer comprehensively covers all the key points involved in the question. For example, when answering the question "How does a smart home system ensure user data security?", if the answer only mentions encryption but does not address aspects such as data storage security, it will be considered incomplete.

[0117] Specifically, the clarity of the answer refers to whether the text is clear and easy to understand, whether the logic is coherent, and whether it can allow customers to easily understand the content, avoiding the situation where professional terms are piled up without explanation.

[0118] Specifically, the relevance of the answer to the consulting question is also critical. The answer must be closely centered around the consulting question and cannot be irrelevant. At the same time, the customer's satisfaction with the consulting process is also reflected in whether the waiting time from submitting the consultation to obtaining the answer is within an acceptable range.

[0119] In general, customers' evaluations of consulting questions can add rich and diverse training data to the database; when customers evaluate the answers, whether they point out that the answers are wrong, incomplete, or the feedback is not clear enough, these evaluation information becomes a valuable data resource.

[0120] For example, suppose a customer's feedback indicates that the answer to "smart home systems ensure user data security" does not mention "network transmission security." This evaluation will allow the database to identify the missing points in the current answer, and then add content related to network transmission security when subsequently improving the answer.

[0121] In general, through a large number of such customer reviews, the database can be continuously updated and expanded to cover various actual feedback on more different types of consulting questions; these newly added training data enable the database to have a more comprehensive understanding of customer needs and also provide direction for optimizing answers; after continuous use of these evaluation data for training, the quality of answers generated by the model will gradually improve and better meet customer requirements, whether in terms of accuracy, completeness or clarity, and can provide customers with better quality and more tailored answers.

[0122] S5. Based on the neural network semantic matching platform and the evaluation, the answer text is disassembled to obtain an answer vocabulary of the answer text.

[0123] In an embodiment of the present invention, the answer text is disassembled based on the neural network semantic matching platform and the evaluation to obtain an answer vocabulary for the answer text, including:

[0124] The answer text is split based on the BERT algorithm in the neural network semantic matching platform and the evaluation to obtain an answer vector for the answer text, wherein the BERT algorithm is:

[0125]

[0126] Where, is the answer vector for the answer text, is a matrix arrangement of the characters in the answer text, is the transpose of the matrix permutation, is the vector set of answer vectors, is the vector dimension;

[0127] An association relationship between the question vector and the answer vector is established, and an answer vocabulary of the answer text is generated according to the association relationship.

[0128] Specifically, the answer text is preprocessed, the unstructured text is cleaned, and the question type, answer validity label and semantic association strength are marked. The questions and answers are segmented, stemmed and stop words are filtered respectively.

[0129] Specifically, the preprocessed answer text is annotated, an answer semantic cluster is generated, all answer texts contained in the answer semantic cluster are extracted, high-frequency core keywords are extracted, and the intra-cluster discrimination weight of the keywords is calculated. The representativeness of the keywords to the cluster is evaluated, common words (such as "please" and "operation") are filtered, and terms with high recognition (such as "password reset" and "bill export") are retained to generate a weighted candidate keyword list.

[0130] Furthermore, based on the neural network semantic matching platform, the candidate keyword list is deeply analyzed, the answers are broken down into weighted keywords and phrases, and the keyword weights are dynamically adjusted through customer feedback data. For example, answers with high follow-up rates will trigger the demotion of related words, forming a dynamic semantic map that reflects the real needs of business scenarios.

[0131] Furthermore, the cluster labels are integrated with the candidate keyword list to construct a hierarchical answer vocabulary: the first-level nodes are semantic cluster labels, and the second-level nodes are keywords and their weights. Then, through manual review and business rule calibration, ambiguous words are removed, synonyms are merged, and domain knowledge is injected. The final vocabulary is stored in a graph database to support weight-based semantic retrieval and association reasoning.

[0132] In general, the core role of the answer vocabulary is to transform unstructured answer text into a quantifiable and interpretable combination of semantic units through semantic decoupling and structured reorganization, thereby significantly improving the accuracy and controllability of the intelligent question-answering system; at the same time, this structured expression not only supports the rapid matching of question intent and answer core elements, but also automatically calls high-weight vocabulary to construct compliant language when generating new answers. At the same time, by monitoring low-frequency word combinations to identify knowledge blind spots, it drives targeted supplementation of the knowledge base, and achieves a dual enhancement of answer quality and domain adaptability.

[0133] S6. Retrain the language processing model based on the answer vocabulary and the question vocabulary to obtain an updated language processing model, and transmit the updated language processing model, the question vocabulary, and the answer vocabulary back to the consultation database;

[0134] In an embodiment of the present invention, retraining the language processing model based on the answer vocabulary and the question vocabulary to obtain an updated language processing model includes:

[0135] Performing vector extraction on the answer vocabulary and the consulting question vocabulary respectively to obtain a question vector and an answer vector of the consulting question;

[0136] The language processing model is trained based on the question vector and the answer vector to obtain an updated language processing model.

[0137] In an embodiment of the present invention, the training of the language processing model based on the question vector and the answer vector to obtain an updated language processing model includes:

[0138] Based on the self-attention mechanism algorithm, the question vector, and the answer vector, a weight calculation is performed on the language processing model, and the language processing model is updated based on the weight score obtained by the calculation to obtain an updated language processing model, wherein the self-attention mechanism algorithm is:

[0139]

[0140] Where, is the score of the self-attention mechanism, is the corresponding parameter index, an input vector determined by the question vector and the answer vector, is the value of the input vector in the question vocabulary and the answer vocabulary, is the normalized exponential function, is the identity of the matrix transpose, is the length of the input vector.

[0141] Specifically, the preprocessed consultation question text and the associated answer text are input into the self-attention mechanism formula. The long-distance dependencies within the text and the cross-text semantic associations are captured in parallel through multi-head attention layers. The model calculates the cross-attention weights of each word in the question and the answer word, identifies the core matching features between the question and the answer, and generates a joint semantic representation vector, providing a context-aware embedding basis for weight parsing.

[0142] Specifically, a pre-built question-answer logic vocabulary (including weighted business keywords and association rules) is used to perform structured analysis on the question and answer texts, and the resulting analysis is divided into the question side and the answer side.

[0143] Furthermore, the question side includes: matching question words with vocabulary keywords, calculating the comprehensive score of word frequency-inverse document frequency and vocabulary weight, and screening the Top-N high-scoring words as the core semantic labels of the question; the answer side includes: parsing the answer text into a combination of vocabulary terms, adjusting the term contribution based on self-attention weights (such as the attention score of "data" in the context increases its final weight), and generating a set of answer semantic labels with dynamic weights.

[0144] Furthermore, through the collaborative analysis of the self-attention mechanism and the structured vocabulary, high-precision semantic alignment and dynamic optimization of consulting questions and answers are achieved. The self-attention mechanism deeply explores the contextual associations of the question text (such as the implicit causal relationship between "account abnormality" and "login failure"), and combines the question-answering logic vocabulary to strengthen the weight of keywords (such as the weight of "password reset" is increased to 0.96), generating a business-explanatory question index that accurately represents user intentions; at the same time, the answer index is decomposed into quantifiable semantic units through the vocabulary (such as "verify identity → 0.88" and "submit work order → 0.75"), and the unit weights are dynamically adjusted based on real-time feedback data to update the language processing model.

[0145] In general, this bidirectional indexing system not only supports millisecond-level semantic matching, but also continuously eliminates inefficient answers and strengthens high-quality solutions through a weight decay mechanism and conflict resolution rules, allowing the system to maintain high robustness when responding to changes in business scenarios, and ultimately achieve full-link intelligence of the knowledge base's "understanding-matching-evolution".

[0146] S7. Obtain the user's real-time question and return to step S2 to obtain a real-time answer to the real-time question.

[0147] In the embodiment of the present invention, the step of obtaining the user's real-time question and returning to step S2 to obtain a real-time answer to the real-time question includes:

[0148] Obtaining a user's real-time question, searching for an answer to the real-time question in the consultation database, and when the search is successful, outputting the answer in the consultation database as a real-time answer to the real-time question;

[0149] When the retrieval fails, a real-time answer to the real-time question is obtained based on the updated language processing model.

[0150] Specifically, a dual index of the answer text is constructed. One part is the content index: based on the semantic tag set after the vocabulary is disassembled, a weighted tag vector is generated, and the relationship between the tags is modeled through the self-attention mechanism (such as the correlation strength between "data" and "account security") to form a structured answer description; the other part is the feedback enhancement index: integrating historical customer feedback data (such as follow-up rate, solution rate), and dynamically calibrating the answer tag weights (such as reducing the "operation steps" tag weight of answers with high follow-up rate by 20%); finally, the enhanced answer index is associated with the question index through a bidirectional mapping table, and the confidence level and version number are recorded to establish a "question-answer mapping table" in the consultation database.

[0151] Specifically, in the "Question-Answer Mapping Table" of the consultation database, question index IDs are bidirectionally bound to corresponding answer index IDs, and status information such as the number of associations and the time of the most recent call is recorded. For iteratively optimized answers, a version control mechanism is implemented: when an answer is modified based on customer feedback, the system retains a record of historical versions and distinguishes between new and old versions using timestamps to ensure traceability and rollback capabilities.

[0152] Specifically, each time an answer is called, customer feedback data (such as satisfaction ratings, follow-up questions, and session termination rates) is collected in real time and associated with the corresponding answer index. This feedback data is periodically analyzed to trigger a re-evaluation process for answers with low satisfaction. This process uses a reinforcement learning model to adjust the answer generation strategy, update the semantic vector representation, and generate a new answer index version. The optimized index overwrites the old version, and the weight coefficients in the mapping table are updated to prioritize the retrieval of high-quality answers.

[0153] In general, the core function of the consulting database is to significantly improve the accuracy and adaptability of intelligent question and answering through a dynamic optimization mechanism driven by customer feedback. At the same time, the feedback mechanism collects the interaction data between customers and answers in real time, and deeply binds it with the corresponding answer index; based on the feedback analysis model, it automatically identifies inefficient answers and triggers the answer reconstruction process; finally, through reinforcement learning combined with manual review, the answer expression logic is adjusted or missing information is supplemented to generate an optimized new version of the answer index, and the semantic vector and the weight coefficient of the model in the language pen are updated at the same time.

[0154] In general, this closed-loop mechanism of "feedback-evaluation-iteration" not only enables the answer library to dynamically adapt to changes in customer needs, but also trains differentiated scenario response strategies through historical feedback data, ultimately achieving continuous improvement in both answer resolution rate and customer satisfaction.

[0155] like Figure 2 , which is a functional module diagram of a customer consultation problem processing system based on machine learning provided by one embodiment of the present invention.

[0156] The machine learning-based customer consultation problem handling system 100 of the present invention can be installed in an electronic device. Depending on the functions implemented, the machine learning-based customer consultation problem handling system 100 can include a customer consultation module 101, a question vocabulary generation module 102, an answer text module 103, a customer usage evaluation module 104, an answer vocabulary generation module 105, a model retraining module 106, and a database update module 107. The module of the present invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by an electronic device processor and can perform fixed functions, and is stored in the memory of the electronic device.

[0157] In this embodiment, the functions of each module / unit are as follows:

[0158] Customer consultation module 101: used to obtain customer consultation questions;

[0159] Question vocabulary generation module 102: used to identify the consulting question using a neural network semantic matching platform to obtain a question vocabulary for the consulting question;

[0160] Answer text module 103: used to search the consultation database according to the question vocabulary, and obtain the answer text of the consultation question when the search is successful; when the search fails, generate the answer text of the consultation question in real time based on the question vocabulary and language processing model;

[0161] The customer evaluation module 104 is used to obtain the customer's evaluation of the answer text;

[0162] An answer vocabulary generation module 105 is configured to decompose the answer text based on the neural network semantic matching platform and the evaluation to obtain an answer vocabulary for the answer text;

[0163] Model retraining module 106: configured to retrain the language processing model based on the answer vocabulary and the question vocabulary to obtain an updated language processing model, and transmit the updated language processing model, the question vocabulary, and the answer vocabulary back to the consultation database;

[0164] Database updating module 107: used to obtain the user's real-time question and return to step S2 to obtain the real-time answer to the real-time question.

[0165] In the several embodiments provided by the present invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the module division is merely a logical function division, and other division methods may be used in actual implementation.

[0166] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network elements. Some or all of the modules may be selected to achieve the purpose of the solution of this embodiment according to actual needs.

[0167] In addition, the functional modules in various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or hardware plus software functional modules.

[0168] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0169] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to achieve optimal results.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for handling customer consultation problems based on machine learning, characterized in that: The method comprises: S1. Obtain customer consultation questions; S2. Deconstruct the text of the consultation question to obtain the question text of the consultation question; Performing text vectorization processing on the question text to obtain a question vector of the question text; Semantically arranging the question vectors based on a neural network semantic matching platform to obtain a question vocabulary for the consulting question; S3. Search the consultation database according to the question vocabulary: S31. When the search is successful, obtain the answer text of the consultation question; S32. When the search fails, generating an answer text to the consultation question in real time based on the question vocabulary and the language processing model; S4. Obtaining the customer's evaluation of the answer text; S5. Split the answer text based on the BERT algorithm in the neural network semantic matching platform and the evaluation to obtain an answer vector for the answer text, wherein the BERT algorithm is: Among them, A vv is the answer vector of the answer text, X dv is a matrix arrangement of the characters in the answer text, is the transpose of the matrix arrangement, R v×v is the vector set of answer vectors, v is the vector dimension; Establishing an association relationship between the question vector and the answer vector, and generating an answer vocabulary for the answer text according to the association relationship; S6. Perform vector extraction on the answer vocabulary and the question vocabulary respectively to obtain a question vector and an answer vector of the consulting question; Training the language processing model based on the question vector and the answer vector to obtain an updated language processing model; transmitting the updated language processing model, the question vocabulary, and the answer vocabulary back to the consultation database; S7. Obtain the user's real-time question and return to step S2 to obtain a real-time answer to the real-time question.

2. A method for handling customer consultation issues based on machine learning according to claim 1, characterized in that: When the search fails, generating the answer text of the consultation question in real time based on the question vocabulary and language processing model, including: Answer the question vocabulary based on the language processing model to obtain an answer vocabulary vector for the question vocabulary; The answer word vector is processed based on an autoregressive algorithm to obtain the answer text of the consultation question, wherein the autoregressive algorithm formula is: and t =Decoder(and t ,H) Among them, y t is the word generated in the tth step, Decoder is the autoregressive function, t is the generation step, y is the generated vocabulary, and H is the vocabulary.

3. The method for handling customer consultation problems based on machine learning according to claim 2 is characterized in that: The processing of the answer vocabulary vector based on the autoregressive algorithm to obtain the answer text of the consultation question includes: The optimal answer text is selected based on a re-ranking algorithm, wherein the re-ranking algorithm is: FinalScore(and 1:t )=Score(y 1:t )+λ·Fluency(y 1:t )+β·Relevance(and 1:t ,H) Among them, FinalScore(y 1:t ) is the optimal answer text, Score(y 1:t ) is the probability score of the candidate answer, Fluency(y 1:t ) is language fluency, Relevance(y 1:t ,H) is the relevance of the question, λ and β are probability coefficients, t is the generation step, y is the generated vocabulary, and H is the vocabulary; The optimal answer text is used as the answer text of the consultation question.

4. The method for handling customer consultation problems based on machine learning according to claim 1 is characterized in that: The evaluation includes: Whether the answer solves the customer's problem, the degree of emotion contained in the answer, the clarity of the answer, the completeness of the answer, and the customer's satisfaction with the consultation process.

5. The method for handling customer consultation problems based on machine learning according to claim 1, characterized in that: The training of the language processing model based on the question vector and the answer vector to obtain an updated language processing model includes: The language processing model is weighted based on a self-attention mechanism algorithm, the question vector, and the answer vector, wherein the self-attention mechanism algorithm is: Where Attention(Q,K,V) is the updated weight of the language processing model, Q is the corresponding parameter index, K is the input vector determined by the question vector and the answer vector, V is the value of the input vector in the question vocabulary and the answer vocabulary, softmax is the normalized exponential function, T is the symbol of the matrix transpose, and d k is the length of the input vector; Parameters of the language processing model are updated based on the updated weights to obtain an updated language processing model.

6. The method for handling customer consultation questions based on machine learning according to any one of claims 1 to 5, wherein obtaining the user's real-time question and returning to step S2 to obtain a real-time answer to the real-time question comprises: Obtaining a user's real-time question, searching for an answer to the real-time question in the consultation database, and when the search is successful, outputting the answer in the consultation database as a real-time answer to the real-time question; When the retrieval fails, a real-time answer to the real-time question is obtained based on the updated language processing model.

7. A customer consultation problem processing system based on machine learning, characterized in that: The system comprises: Customer consultation module: used to obtain customer consultation questions; Question vocabulary generation module: used for text decomposing the consulting question to obtain the question text of the consulting question; Performing text vectorization processing on the question text to obtain a question vector of the question text; Semantically arranging the question vectors based on a neural network semantic matching platform to obtain a question vocabulary for the consulting question; Answer text module: used to search the consultation database according to the question vocabulary, and when the search is successful, obtain the answer text of the consultation question; when the search fails, generate the answer text of the consultation question in real time based on the question vocabulary and language processing model; Customer evaluation module: used to obtain the customer's evaluation of the answer text; An answer vocabulary generation module is configured to split the answer text based on the BERT algorithm in the neural network semantic matching platform and the evaluation to obtain an answer vector for the answer text, wherein the BERT algorithm is: Among them, A vv is the answer vector of the answer text, X dv is a matrix arrangement of the characters in the answer text, is the transpose of the matrix arrangement, R v×v is the vector set of answer vectors, v is the vector dimension; Establishing an association relationship between the question vector and the answer vector, and generating an answer vocabulary for the answer text according to the association relationship; Model retraining module: used for performing vector extraction on the answer vocabulary and the question vocabulary respectively to obtain the question vector and the answer vector of the consulting question; Training the language processing model based on the question vector and the answer vector to obtain an updated language processing model; transmitting the updated language processing model, the question vocabulary, and the answer vocabulary back to the consultation database; Database update module: used to obtain real-time questions from users and return them to the question vocabulary generation module to obtain real-time answers to the real-time questions.

Images