CN117332066B - Intelligent agent text processing method based on large model - Google Patents
Intelligent agent text processing method based on large model Download PDFInfo
- Publication number
- CN117332066B CN117332066B CN202311402059.9A CN202311402059A CN117332066B CN 117332066 B CN117332066 B CN 117332066B CN 202311402059 A CN202311402059 A CN 202311402059A CN 117332066 B CN117332066 B CN 117332066B
- Authority
- CN
- China
- Prior art keywords
- customer service
- task
- professional
- processing
- field
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000003672 processing method Methods 0.000 title claims abstract description 16
- 238000012545 processing Methods 0.000 claims abstract description 89
- 238000011156 evaluation Methods 0.000 claims abstract description 37
- 238000012549 training Methods 0.000 claims abstract description 28
- 238000000034 method Methods 0.000 claims abstract description 27
- 238000004422 calculation algorithm Methods 0.000 claims description 39
- 230000008451 emotion Effects 0.000 claims description 37
- 239000003795 chemical substances by application Substances 0.000 claims description 24
- 239000013598 vector Substances 0.000 claims description 17
- 238000013441 quality evaluation Methods 0.000 claims description 7
- 238000012163 sequencing technique Methods 0.000 claims description 7
- 238000007635 classification algorithm Methods 0.000 claims description 6
- 238000002372 labelling Methods 0.000 claims description 6
- 230000006399 behavior Effects 0.000 claims description 5
- 230000005587 bubbling Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 239000003086 colorant Substances 0.000 claims description 3
- 238000004891 communication Methods 0.000 claims description 3
- 238000011835 investigation Methods 0.000 claims description 3
- 230000008520 organization Effects 0.000 claims description 3
- 230000004044 response Effects 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 238000000354 decomposition reaction Methods 0.000 claims description 2
- 230000005012 migration Effects 0.000 description 5
- 238000013508 migration Methods 0.000 description 5
- 230000004048 modification Effects 0.000 description 5
- 238000012986 modification Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 230000007935 neutral effect Effects 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000013145 classification model Methods 0.000 description 3
- 238000003058 natural language processing Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- RWSOTUBLDIXVET-UHFFFAOYSA-N Dihydrogen sulfide Chemical compound S RWSOTUBLDIXVET-UHFFFAOYSA-N 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010219 correlation analysis Methods 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000008713 feedback mechanism Effects 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000003064 k means clustering Methods 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000009827 uniform distribution Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/332—Query formulation
- G06F16/3329—Natural language query formulation or dialogue systems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/3331—Query processing
- G06F16/334—Query execution
- G06F16/3344—Query execution using natural language analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/35—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
- G06F18/24155—Bayesian classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
- G06N3/0455—Auto-encoder networks; Encoder-decoder networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/01—Customer relationship services
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Databases & Information Systems (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Software Systems (AREA)
- Business, Economics & Management (AREA)
- Evolutionary Biology (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Accounting & Taxation (AREA)
- Development Economics (AREA)
- Economics (AREA)
- Finance (AREA)
- Marketing (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Probability & Statistics with Applications (AREA)
- Human Computer Interaction (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention relates to an intelligent agent text processing method based on a large model, which belongs to the technical field of text processing and comprises the following steps: decomposing questions asked by the user; acquiring a problem complexity evaluation value and a problem attribute, and classifying task types; determining the professional field of the problem according to the type of the problem; matching the professional field of the problem with the professional field of the intelligent customer service of the seat; task allocation is carried out according to the field matching degree evaluation value and the customer service busyness; judging whether the capability of the customer service meets the task requirement according to the historical performance, training records and user evaluation data of the customer service; decomposing the task into a plurality of sub-problems according to the complexity of the problem; and determining the priority of the task for processing according to the emergency degree and importance of the sub-problem, so as to obtain a final solution. The customer service system and the customer service method have the advantages that customer service suitable for processing the customer problem is selected by analyzing the customer problem and customer service conditions, so that customer service efficiency and satisfaction are improved.
Description
Technical Field
The invention belongs to the technical field of text processing, and particularly relates to an intelligent agent text processing method based on a large model.
Background
With the popularization of the internet and the improvement of the living standard of people, the demand of customer service is increasing. Traditional customer service solutions often rely on manual intervention, are inefficient and are susceptible to personal abilities, experience, and emotion of the customer service personnel. The traditional seat customer service has large customer service working pressure, and tasks cannot be reasonably distributed according to the customer service working strength; the malicious questions are difficult to find in time, and the service time is wasted; the customer service capacity and the professional field are difficult to judge, and the problem can not be solved pertinently according to the customer service field; the questioning cannot analyze the questioning difficulty and customer service cannot be distributed according to the urgency of the questions; meanwhile, in the face of customer questioning, manual customer service is often distributed, the customer service manually judges the direction of the customer questioning, and then customer service personnel in corresponding areas are contacted, so that the efficiency is low.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides an intelligent seat text processing method based on a large model, which improves the customer service efficiency and satisfaction through analyzing the user problems and selecting customer service suitable for processing the user problems under the customer service conditions.
The aim of the invention can be achieved by the following technical scheme:
An intelligent agent text processing method based on a large model comprises the following steps:
S1, decomposing questions asked by a user by using a large language model ChatGLM, judging attributes of the questions of the user, analyzing emotion colors of the questions of the user, identifying intentions and information requirements of the questions of the user, and labeling keywords in the questions;
S2, acquiring a problem type and a complexity evaluation value according to the problem attribute, and integrating the problem type, the emergency degree, the professional vocabulary and the professional problem attribute to determine the complexity of the problem; the problem types comprise professional after-sale problems, technical support problems and complex relation problems;
s3, acquiring domain information of a user question according to the type of the question, judging whether the question is a question attack and a malicious guide of a malicious client, and if the question is not a malicious error guide, determining the professional domain of the question by comparing the question with a professional domain knowledge base;
S4, acquiring the professional field of the intelligent customer service of the seat, and comparing the professional field of the problem with the professional field of the intelligent customer service of the seat to obtain a field matching degree evaluation value;
S5, distributing tasks to proper customer service according to the field matching degree evaluation value and the customer service busyness;
s6, after tasks are distributed, judging whether the capability of the customer service meets the task requirements according to the historical performance of the customer service, training records and user evaluation data;
s7, decomposing the task according to the complexity of the problem to obtain a plurality of sub-problems which are more suitable for the customer service knowledge field;
and S8, determining the priority of the task for processing according to the emergency degree and the importance of the sub-problem, and obtaining a final solution.
Further, in the step S1, the attribute of the problem includes help, advice, and complaints; wherein, analyzing the emotion color of the user question adopts an LSTM model.
Further, in the step S2, the determining of the problem type includes:
if the questions include the product model, purchase date and warranty status attributes, judging that the questions are professional after-sale questions; if the problem comprises the operating system, the software version, the equipment model and the network environment attribute, judging the problem as a technical support problem; if the problem includes personnel identity, job responsibilities and team organization structure attributes, then the problem is determined to be a complex relationship problem.
Further, in the step S2, the determining of the complexity evaluation value specifically includes the following steps:
Judging whether a large number of same problems exist in a short time according to the problem description and the history record, if so, recording the number and the frequency of the same problems, judging that the number of questions is higher than a preset number value to be an urgent problem, and if so, recording the urgent degree of the problem and the required solving time;
determining whether a large number of professional terms and domain-specific knowledge are involved in the problem through the problem description and related documents, judging whether the problem belongs to the professional problem according to the professional vocabulary, the problem description and related knowledge involved in the problem, and recording the professionality of the problem and the participation degree of required professional technicians if the problem belongs to the professional problem.
Further, the step S3 specifically includes the following steps:
Using TF-IDF text classification algorithm to obtain specific class of problem and information of user questioning field; detecting the sensitive words by using a DFA algorithm, judging whether malicious information or guiding keywords are contained in the problem content, and marking the problem content as malicious behaviors if the malicious information or the guiding keywords are detected; information retrieval using inverted indexes, acquiring professional domain content related to the problem from a domain knowledge base; calculating the matching degree of the problem and the knowledge base content by adopting cosine similarity; using a TF-IDF text classification algorithm to match a vector machine, marking the problems and comparing the problems with data labels in a domain knowledge base; determining whether the problem is in a certain professional field by using a threshold judgment method according to the data tag comparison result in the field knowledge base; comparing the questions with the knowledge base content, screening out the most matched answers, and returning the most matched answers to the user;
Specifically, judging whether the problem content contains emotion of the problem by adopting a naive Bayesian combined with an Adaboost algorithm, carrying out emotion classification on the problem to be analyzed by using a trained naive Bayesian classifier and an Adaboost classifier, and obtaining the emotion type of the problem and a judging result of whether the problem contains malicious guidance by comparing the emotion type of the problem and the characteristic matching degree of related properties of the malicious guidance;
Specifically, a BERT model is adopted in the professional field of the determined problem, a problem text is input into the trained model, a BERT classification result is obtained, semantic matching is carried out on the new problem text and each item in the knowledge base, the problem text and the knowledge base item are input into the BERT model, an BERT output representation is obtained, then cosine similarity is used for calculating the similarity between the problem text and the knowledge base item, and the item with the highest similarity is selected as a matching result.
Further, in the step S4, the step of obtaining the intelligent customer service professional field of the seat includes:
The method comprises the steps of obtaining the number of sales problems, after-sales problems and technical support problems processed by a seat customer service by recording the processing records of the seat customer service on customer problems; comprehensively evaluating customer satisfaction according to product or service quality evaluation, delivery time rate, after-sales service quality evaluation, customer feedback and complaint handling condition attributes; judging the processing capacity of the seat customer service through the performance of the seat customer service in the aspects of processing sales problems, after-sales problems and technical support, including complex problem processing capacity, response speed and communication skills; judging the professional field of the customer satisfaction degree and the processing capacity according to the number of customer service processing;
The step of obtaining the field matching degree evaluation value comprises the following steps:
Obtaining keywords and semantics in the problem description, and primarily judging the professional field related to the problem by analyzing the keywords and semantics in the problem description; constructing a knowledge base of intelligent customer service of the seat, which contains related knowledge and solutions in each professional field; matching the problems with data in a knowledge base by using a vector space model matching algorithm to obtain a matching result, and distributing the problems to corresponding professional fields according to the matching result; establishing a domain professional agent customer service library according to the related information of each professional domain and the professional agent customer service, and matching the problems with the professional domain of the agent customer service by using a vector space model matching algorithm; and obtaining a domain matching degree evaluation value by using the similarity between the professional domain of the cosine similarity calculation problem and the professional domain of the intelligent customer service of the seat.
Further, in the step S5, the task allocation includes the following steps:
Judging the current workload and busy degree of the customer service by examining the workload, the processing task number and the processing time of the customer service; the capacity and efficiency of customer service processing tasks are evaluated through the working experience of the customer service and the speed and quality of solving the problems; determining whether the customer service can be used for processing tasks or not according to the working time, vacation and off-duty of the customer service; the weighted average method is adopted, the workload of customer service, the number of tasks and the time of processing are synthesized, the working experience of the customer service, the speed and the quality of solving the problem are synthesized, and the availability score of the customer service is obtained;
Judging the emergency degree of the task according to the deadline, importance and influence on the client of the task, and sequencing the priority of task processing according to the emergency degree of the task; and distributing the tasks to customer services with field matching degree higher than preset matching degree and availability degree higher than preset grading according to the priority ordering sequence of task processing.
Further, in the step S7, the task decomposition includes the following steps:
When a plurality of possible answers or solutions exist for the problem proposed by the customer, the problem can be solved by applying knowledge in a plurality of fields, and fuzzy information or incomplete conditions exist in the problem, more information needs to be acquired through further inquiry or investigation, and the complex task is judged; acquiring professional fields contained in a complex task, wherein the professional fields comprise product knowledge, technical knowledge and industry knowledge fields; according to the field contained by the task, decomposing the complex task, and according to the decomposed field, generating a plurality of sub-problems more suitable for the customer service knowledge field.
Further, the step S8 includes the steps of:
Task weights are distributed according to the emergency degree and the importance, and a comprehensive score is generated for the tasks corresponding to the sub-problems according to a simple weighting method; determining a preliminary priority of each task by using an bubbling sequencing algorithm according to the comprehensive scores; determining the final task priority by using a priority queue method according to the preliminary priorities of all the tasks; selecting a corresponding branch limit algorithm solving strategy according to the final task priority; performing specific processing and operation on the task by using a selected algorithm; verifying whether the task is completed according to the processing result; if the task is not completed, recalculating the comprehensive score and determining the priority; and obtaining a final solution according to the processing results of all the tasks.
The beneficial effects of the invention are as follows:
The invention combines TF-IDF scoring algorithm to extract key words, and determines matching degree of each item in question and knowledge base by cosine similarity method, and uses NLP sensitive word detection technology based on language big model to judge whether the question content contains malicious information or guiding key words. And determining the priority of the task for processing according to the emergency degree and importance of the sub-problems, and corresponding the problems to the most appropriate customer service in the knowledge field.
Drawings
The present invention is further described below with reference to the accompanying drawings for the convenience of understanding by those skilled in the art.
FIG. 1 is a schematic diagram of the steps of the text processing method of the present invention.
Detailed Description
In order to further describe the technical means and effects adopted by the invention for achieving the preset aim, the following detailed description is given below of the specific implementation, structure, characteristics and effects according to the invention with reference to the attached drawings and the preferred embodiment.
Referring to fig. 1, an intelligent agent text processing method based on a large model includes the following steps:
s1, decomposing a question asked by a user by using a large language model ChatGLM to obtain the aim of asking by the user.
Inputting a question asked by a user into a large language model ChatGLM, decomposing the question asked by the user, acquiring the aim of the question asked by the user, and judging the attribute of the question, including help, suggestion and complaint; inputting the user questions into an LSTM neural network for emotion analysis, and analyzing emotion colors of the client questions; identifying the intention of a client to ask questions and determining the purpose of the questions; identifying information requirements, identifying information requirements required by customer questioning; identifying the problem type, and determining the problem type of the client question; determining key words and phrases of a client question; and adjusting by combining with specific scenes to improve the classification accuracy and classification effect.
For example, a telecommunications agent system needs to analyze and classify questions asked by a user. The telecom online customer service receives a user question: i want to know how to handle the phone card. The problem is resolved using a large language model ChatGLM. First, it is determined what the user asks for, i.e., the nature of the question is judged. User questions can be categorized into help, advice, and complaints by training a large language model ChatGLM. The model recognizes that the user's question is for the purpose of assistance. Next, emotion analysis may be performed using the LSTM neural network to analyze the emotion color of the user question. And taking the user question as input, and obtaining the emotion score through the LSTM model. Assume that the emotion score for the user question is 8, indicating that the user question has a positive emotion. Then, the intention of the user to ask is identified to determine the purpose of the question, and the intention of the user to ask is identified to transact a telephone card. And determining which information is required for the user to ask a question by using an information requirement identification method, and identifying that the information requirement for the user to ask the question is a handling flow. Meanwhile, a problem type recognition method is used for determining the problem type of the user question and recognizing how the problem type of the user question is. And finally, carrying out slot annotation on the user question to determine key words and phrases. Keywords and phrases in user questions, including business and phone cards, are identified using natural language processing techniques.
S2, acquiring a specific category and a complexity evaluation value of the problem according to the problem attribute.
According to the question attributes, questions are classified into three categories: professional after-market problems, technical support problems, and complex relationship problems. If the problems include product model, purchase date and warranty status attributes, judging that the problems are professional after-sale problems; if the problem comprises the operating system, the software version, the equipment model and the network environment attribute, judging the problem as a technical support problem; if the problem includes personnel identity, job responsibilities and team organization structure attributes, then the problem is determined to be a complex relationship problem.
For professional after-sales problems, the after-sales problems can be divided into simple after-sales tasks, general after-sales tasks and complex after-sales tasks; for the technical support problem, the method can be divided into a simple technical support task, a general technical support task and a complex technical support task; for complex relationship problems, it can be classified into simple relationship problems, general relationship problems, and complex relationship problems.
And judging whether a large number of same problems exist in a short time according to the problem description and the history. If the number and the frequency of the same questions are recorded, the questions are higher than the preset number value, and the urgent questions are judged. If an emergency problem, the degree of emergency and the required time to resolve the problem are recorded. From the description of the problem and the related documents, it is determined whether a large number of terms of art and domain-specific knowledge are involved in the problem. Judging whether the problem belongs to the professional problem according to the professional vocabulary, the problem description and the related knowledge involved in the problem. If a professional question, the professionality of the question and the degree of involvement of the professional technician required are recorded. The complexity of the problem is determined by integrating the problem type, urgency, specialized vocabulary and specialized problem attributes.
For example, one problem is described as follows: when the mobile phone is restarted, the mobile phone still cannot be connected, and how to connect the mobile network is asked, is the base station problematic? From the description of the problem, it can be judged that this is a technical problem because the user encounters difficulty in network connection. From the history, if it is found that a plurality of users have encountered the same network connection problem in a short time, it can be judged that there are a large number of the same problems. 400 feedback about network connection are received in the last day, and the emergency problem is judged if the feedback exceeds a preset quantity value. Further analysis of the problem description, if the user is found to refer to a number of terms in describing the problem, including restarting the handset, base station, it can be determined that this is a professional problem, requiring the participation of a professional technician. The problem can be determined to be a technical problem by integrating the problem type, the degree of urgency, the specialized vocabulary and the specialized problem attributes, the degree of urgency is high, the specialized terms are involved and the specialized technician is required to participate, and thus the complexity of determining the problem is high.
S3, acquiring domain information of a user question according to the type of the question, judging whether the question is a question attack and a malicious guide of a malicious client, and if the question is not a malicious error guide, determining the professional domain of the question by comparing the question with a professional domain knowledge base.
And obtaining the specific class of the problem and the information of the user questioning field by using a TF-IDF text classification algorithm. And detecting the sensitive words by using a DFA algorithm, and judging whether malicious information or guiding keywords are contained in the problem content. If malicious information or guiding keywords are detected, marking as malicious behaviors. And acquiring professional domain content related to the problem from a domain knowledge base by using information retrieval of the inverted index. And calculating the matching degree of the problem and the knowledge base content by adopting cosine similarity. And labeling the problems by using a TF-IDF text classification algorithm in combination with a vector machine, and comparing the problems with data labels in a domain knowledge base. And determining whether the problem is in a certain professional field by using a threshold judgment method according to the data tag comparison result in the field knowledge base. And comparing the questions with the knowledge base content, screening out the best matched answers, and returning the best matched answers to the user.
And analyzing the problem emotion by using naive Bayes and an Adaboost algorithm, and judging whether the problem contains malicious guidance.
And collecting a certain amount of tagged questions and corresponding emotion categories, and related attribute data of malicious guidance according to the requirements. The TF-IDF is used to convert the question text into features of a vector representation. Training the extracted feature vector and the corresponding emotion category through a naive Bayes algorithm to obtain a naive Bayes classifier. And initializing sample weights by using an Adaboost algorithm, performing multiple rounds of training, adjusting the sample weights according to the classification result of the previous round, training a plurality of weak classifiers, and combining the weak classifiers according to the sample weights to obtain a strong classifier.
And selecting the characteristics according to the text characteristics and the malicious guidance related attributes in the problems, analyzing the relevance between the text characteristics and the malicious guidance related attributes, and selecting the characteristics related to the malicious guidance. And performing emotion classification on the problem to be analyzed by using a trained naive Bayesian classifier and an Adaboost classifier, and judging whether the problem contains malicious guidance or not. And comparing the emotion type of the problem with the characteristic matching degree of the related attribute of the malicious guidance to obtain the emotion type of the problem and a judging result of whether the malicious guidance is contained or not.
For example, it is necessary to construct an emotion classifier for determining emotion categories of a question including positive, neutral, negative, and whether the question contains malicious guidance. 100 tagged problem samples were collected and labeled for their emotion classification and malicious guidance attribute. First, the TF-IDF algorithm is used to convert the question text into features of the vector representation. 1000 keywords are selected as feature words, and TF-IDF values of each feature word are calculated in each question text to obtain a 100-dimensional feature vector representation. Next, the extracted feature vectors and corresponding emotion categories are trained using a naive bayes algorithm. The emotion categories of the training set are distributed as follows, positive category 30 samples, neutral category 40 samples, and negative category 30 samples. The conditional probability of each feature in each emotion category can be calculated according to the principle of a naive bayes algorithm. For a certain feature word, the number of occurrences is 25 in the positive category, 10 in the neutral category, and 5 in the negative category. Then during training, the conditional probabilities of the feature vocabulary under the positive, neutral and negative categories can be calculated to be 62.5%, 25% and 12.5% respectively. And then, carrying out multiple rounds of training on the feature vector and the emotion category by using an Adaboost algorithm, adjusting the sample weight according to the classification result of the previous round, and training a plurality of weak classifiers. And 5 rounds of training are carried out, each round of training is carried out to obtain a weak classifier, and a strong classifier is obtained by combining the weak classifiers. Next, feature selection is performed to analyze the relevance between text features in the question and the malicious guidance related attributes. It was found that among all the problems, the problem that contains feature words is often of malicious guidance nature. Feature vocabulary related to malicious guidance is selected as an important feature. Finally, using a trained naive Bayes classifier and an Adaboost classifier to carry out emotion classification on the problem to be analyzed. For one problem to be analyzed: the product can not be used at all, and the emotion type can be obtained to be negative through feature extraction and judgment by a classifier. The emotion type of the problem and the judging result of whether the malicious guidance is contained or not can be obtained by comparing the emotion type of the problem with the characteristic matching degree of the related attribute of the malicious guidance. If a problem is judged to be of a positive category but contains both malicious guidance-related feature words, the problem may be suspected.
The problems are classified by using a pretrained model of BERT, and the professional field of the problem is determined.
According to the characteristics of the service field, a training data set and a knowledge base data set of problem classification are prepared. Labeling the training data set, and associating each problem with a corresponding classification label. The format of the knowledge base data set is determined, including common questions and corresponding answers. And training a problem classification model by using a pre-training model of the BERT, obtaining an output representation of the BERT by inputting the text of the problem into the BERT model, classifying the output representation of the BERT by inputting the output representation into a full-connection layer, and performing model training and parameter optimization by using a cross entropy loss function. Inputting new problem text into the trained problem classification model, obtaining the output of BERT, and inputting the output into the trained classifier. Judging the classification of the problem and outputting a classification result. And carrying out semantic matching on the new problem text and each item in the knowledge base. The output representation of the BERT is obtained by inputting the question text and knowledge base entries into the BERT model. And calculating the similarity between the problem text and the knowledge base items by using the cosine similarity, and selecting the item with the highest similarity as a matching result. The model is evaluated using the validation dataset, and the accuracy of the problem classification and the accuracy of the knowledge base matching are calculated. And (3) performing model tuning and improvement according to the evaluation result, and improving the performance of the model.
S4, acquiring the professional field of the intelligent customer service of the seat, and comparing the professional field of the problem with the professional field of the intelligent customer service of the seat to obtain a field matching degree evaluation value.
And obtaining keywords and semantics in the problem description, and primarily judging the professional field related to the problem by analyzing the keywords and semantics in the problem description. And constructing a knowledge base of the intelligent customer service of the seat, wherein the knowledge base comprises related knowledge and solutions in various professional fields. And matching the problems with data in a knowledge base by using a vector space model matching algorithm to obtain a matching result, and distributing the problems to corresponding professional fields according to the matching result. And establishing a domain professional agent customer service library according to the related information of each professional domain and the professional agent customer service, and matching the problems with the professional domain of the agent customer service by using a vector space model matching algorithm. And calculating the similarity between the problem and the professional field of the intelligent customer service of the seat by using the cosine similarity to obtain a field matching degree evaluation value.
For example, the question describes how i want to know how to change packages on chinese telecom hall APP, the keywords are chinese telecom hall APP, change packages, and the semantics are how to change packages on chinese telecom hall APP. By analyzing keywords and semantics in the problem description, the professional field related to the problem can be primarily judged to be package migration. The knowledge base in the field of package migration and modification contains contents such as how to replace packages, solve common problems and the like. And extracting features from each entity and constructing a feature matrix by using a vector space model matching algorithm from the problems to the entity China telecom business hall APP, replacement and package. And (3) using a FLANN matching algorithm, setting weights based on distance or characteristic attributes to indicate matching similarity according to service requirements, and matching keywords in the problem description, namely the Chinese telecom business hall APP, replacement and package, with related data in a knowledge base to obtain a matching result of 95%. And distributing the problems to the corresponding professional field, namely the package migration field according to the matching result. The customer service library of the seat in the field of package migration and modification contains customer service information of a professional capable of answering and changing packages. And matching the key words of the Chinese telecom business hall APP, the replacement, the package and the seat customer service professional range in the package migration modification field in the problem description by using a vector space model matching algorithm, so that a matching result is 80%. And calculating the similarity between the problem and the professional field of the intelligent customer service of the seat by using the cosine similarity, and obtaining a field matching degree evaluation value of 8.
And acquiring the professional field of the seat customer service through the seat customer service processing record.
And acquiring the number of the sales problems, the after-sales problems and the technical support problems processed by the seat customer service by recording the processing records of the seat customer service on the customer problems. Comprehensively evaluating customer satisfaction according to product or service quality evaluation, delivery time rate, after-sales service quality evaluation, customer feedback and complaint handling condition attributes; the processing capacity of the seat customer service is judged through the performance of the seat customer service in the aspects of processing sales problems, after-sales problems and technical support, including complex problem processing capacity, response speed and communication skills. And judging the professional field of the customer satisfaction degree and the processing capacity according to the number of customer service processing.
For example, by recording the processing records of customer questions by the customer service, the number of sales questions, after-sales questions, and technical support questions for each customer service process can be obtained. A certain seat customer service has handled 50 sales problems, 30 after-sales problems and 20 technical support problems in the past month. The performance of the seat customer service in the aspect of sales problem processing is calculated, and the performance can be evaluated according to the product or service quality evaluation attribute. The number of orders after consultation of sales problem clients is 10, and the satisfaction degree of the before-market service is evaluated to be 3, so that the seat service score is 10×3=30 in the aspect of sales problem processing. The ability of the seat customer service to handle after-market problems is assessed, and the after-market quality of service assessment attribute can be referenced. If the customer satisfaction with the after-sales service in the after-sales problem is evaluated as 4 points, the seat customer service score in the after-sales problem processing aspect is 30×4=120 points. Judging the processing capacity of the seat customer service in technical support, and considering the customer feedback and complaint processing condition attribute. The feedback evaluation of the customer in the technical support problem is 5 points, and the seat customer service score in the aspect of technical support problem processing is 20×5=100 points. When comprehensively evaluating the customer satisfaction, the calculation can be performed according to the customer satisfaction attribute. Customer satisfaction rating was 80%, the overall rating score was (30+120+100)/3×0.8=67. In summary, the seat customer service score for sales problem processing is 30 points, the score for after-sales problem processing is 120 points, the score for technical support problem processing is 100 points, and the comprehensive evaluation score is 67 points. The seat customer service can be judged to have strong capability in the aspect of after-sales problem processing, and is slightly insufficient in the aspect of sales problem processing, but the overall performance is better.
And S5, distributing tasks to the customer service according to the field matching degree evaluation value and the customer service busyness.
The matching degree of customer service in the field of processing specific tasks is determined by evaluating the expertise, knowledge and experience of the customer service. Judging the current workload and busy degree of the customer service by examining the workload, the processing task number and the processing time of the customer service; the capacity and efficiency of customer service processing tasks are evaluated through the working experience of the customer service and the speed and quality of solving the problems. And determining whether the customer service is available for processing tasks through the working time, vacation and off-duty of the customer service. And (3) integrating the workload, the processing task number and time of customer service, the customer service working experience, the problem solving speed and quality by adopting a weighted average method to obtain the availability score of the customer service. And judging the emergency degree of the task according to the deadline, importance and influence on the client of the task, and sequencing the priority of task processing according to the emergency degree of the task. And distributing the tasks to customer services with field matching degree higher than preset matching degree and availability degree higher than preset grading according to the priority ordering sequence of task processing.
For example, telecommunications agents are responsible for handling complaints and problems from customers. The expertise, knowledge and experience of the customer service can be assessed based on their training and certification in the relevant area. Customer service a has 3 years of relevant working experience and has authentication in the relevant field, while customer service B has only 1 year of working experience and has no authentication in the relevant field. The customer service a has a higher degree of matching when handling the specific task area. To determine the current workload and busyness of customer service, their workload, number of tasks processed, and time can be examined. Customer service a processed 10 tasks on average daily, each task taking 30 minutes, while customer service B processed 5 tasks daily, each task taking 45 minutes. The current workload of the customer service A is judged to be larger, and the busy degree is higher. The ability and efficiency of customer service processing tasks are evaluated, and their working experience, speed of solving the problem and quality can be examined. Customer service A averages 20 minutes for each task, the quality of the solution to the problem gets a customer satisfaction score of 90/100, while customer service B averages 25 minutes for each task, the quality of the solution to the problem gets a customer satisfaction score of 80/100. The capability and efficiency of evaluating customer service A processing tasks are high. To determine if customer service is available for processing tasks, their hours of work, holiday and off duty conditions may be examined. Customer service A works 5 days per week for 8 hours per day without holiday and off-duty records, while customer service B works 4 days per week for 6 hours per day with a holiday record. The time to determine that customer service a is available for processing tasks is longer. The service availability score can be obtained by integrating the workload, the processing task number and time, the working experience, the problem solving speed and quality of the service. The availability score of customer service A was 80/100 and the availability score of customer service B was 70/100 by a weighted average method. The urgency of the task can be determined based on the deadline, importance, and impact on the customer of the task. The deadline for task 1 is tomorrow, the importance rating is high, the influence on the customer is critical, while the deadline for task 2 is the next week, the importance rating is medium, and the influence on the customer is general. The emergency level of task 1 is high. And distributing the tasks to customer services with field matching degree higher than preset matching degree and availability degree higher than preset grading according to the priority ordering sequence of task processing. The field matching degree of customer service A to task 1 is preset to 80/100, the usability degree score is preset to 75/100, the field matching degree of customer service B to task 1 is preset to 70/100, and the usability degree score is preset to 60/100. The field matching degree of customer service A to task 2 is preset to 80/100, the usability degree score is preset to 75/100, the field matching degree of customer service B to task 2 is preset to 75/100, and the usability degree score is preset to 70/100. According to the task priority, firstly, task 1 is allocated, the preset matching degree of the task 1 is 65/100, the matching degree of the field of customer service A and the field of customer service B are higher than the preset matching degree, the field matching degree and the availability grade of the customer service A to the task 1 are higher than those of the customer service B, then the task 1 is allocated to the customer service A, and the availability of the customer service A is changed into 0/100; and (3) distributing a task 2, wherein the preset matching degree of the task 2 is 60/100, the matching degree of the customer service A and the customer service B fields is higher than the preset matching degree, the availability of the customer service B is higher than that of the customer service A, and distributing the task 2 to the customer service B.
And determining that the tasks are allocated to proper customer service according to the professional field, the processing speed and the current load factor of the customer service.
And acquiring the professional field label of each customer service. And determining the processing speed of each customer service through the system log. And acquiring the number of the tasks currently processed by each customer service, and analyzing the complexity of the tasks to judge the actual load of the customer service. If the processing speed of the customer service is greater than the preset speed and the actual load is smaller than the preset load value, the task allocation weight is increased for the customer service. And obtaining the matching degree of the task type and the professional field of the customer service by using the correlation analysis, and if the task type is matched with the professional field of the customer service, adding weight to the customer service. And (3) performing task distribution by using a weighted polling method, ensuring uniform distribution of tasks and distributing according to customer service weights. And dynamically adjusting the weight in the weighted polling method according to experience and historical data of customer service. For emergency tasks, the use of priority queues ensures that they are quickly assigned to customer services whose processing speed is higher than a preset processing speed. And according to the task attribute and the customer service attribute, a decision tree algorithm is applied to carry out task allocation decision. And evaluating the distribution effect through a regular feedback mechanism, and fine-tuning algorithm parameters.
And S6, after the tasks are distributed, judging whether the capability of the customer service meets the task requirements according to the historical performance, training records and user evaluation data of the customer service.
Historical performance data of customer service is obtained, including speed, accuracy and customer satisfaction of solving the problem. By analyzing the historical performance data, the performance level of customer service in past work is determined. And acquiring a training record of the customer service, wherein the training record comprises participated training courses and acquired certification. Judging whether the customer service has necessary expertise and skills or not through the training records of the customer service. User evaluation data is obtained, including evaluation of quality of service, attitude and expertise. And obtaining the satisfaction degree and evaluation of the customer on the customer service by analyzing the user evaluation data. And determining whether the customer service has relevant professional background and experience according to the professional field requirements of the task. And acquiring the complexity of the task, and judging whether the customer service has the capability and experience of processing the complex task. The professional fields of customer service are acquired, and the areas where the customer service has professional knowledge and experience and whether the customer service is competent for related tasks are determined. Matching is carried out according to the professional field and task requirements of customer service. And obtaining overdue unfinished records of the customer service to obtain whether overdue unfinished tasks exist in the customer service. If there is an overdue incomplete record for customer service, evaluation and risk management are required. And judging whether the capability of the customer service meets the task requirement according to the historical performance, training records and user evaluation data of the customer service.
For example, it is desirable to evaluate the speed of a telecommunication agent customer service to solve a problem. Time data for 100 questions processed over the past year of customer service may be collected and an average resolution time calculated. The total customer service time to handle these 100 questions is 3000 minutes, then the average solution time is 3000 minutes divided by 100 questions, i.e., 30 minutes. The accuracy of the problems handled by customer service over the past year is assessed by collecting accuracy data. Customer service has handled 100 problems, of which 95 are correctly resolved, and the accuracy is 95%. And customer satisfaction may be assessed by collecting user rating data for customer service. The customer service evaluation was collected for 100 users, of which 80 gave satisfactory evaluation for quality of service, attitude and expertise, and customer satisfaction was 80%. For the training records of the customer service, the training courses and the obtained certification participated by the customer service can be checked. Customer service was found to be involved in 5 training courses related to customer service skills and 2 related authentications were obtained. Finally, if the customer service needs to be judged whether to have relevant professional background and experience, the education background, the work experience and the professional field of the customer service can be known. Customer service may have a college educational background, 2 years of customer service work experience, and related expertise and experience in the field of electronic products. The speed, accuracy and customer satisfaction of judging customer service in the aspect of processing similar problems are good, and meanwhile, the customer service has a professional background for processing a specific field, so that the customer service is very suitable for processing the task.
S7, decomposing the task according to the complexity of the problem to obtain a plurality of sub-problems which are more suitable for the customer service knowledge field.
Whether a plurality of possible answers or solutions to the problem proposed by the customer exist or not, whether multi-step solutions are needed or whether comprehensive application of knowledge in a plurality of related fields is needed, and fuzzy information or incomplete conditions exist. When a problem presented by a customer has a plurality of possible answers or solutions, or when the problem needs to be solved by using knowledge in a plurality of fields, or fuzzy information or incomplete conditions exist in the problem, more information needs to be acquired through further inquiry or investigation, and the problem is judged to be a complex task. And acquiring professional fields contained in the complex task, including product knowledge, technical knowledge and industry knowledge fields. According to the field contained by the task, decomposing the complex task, and according to the decomposed field, generating a plurality of sub-problems more suitable for the customer service knowledge field.
And S8, determining the priority of the task for processing according to the emergency degree and the importance of the sub-problem, and obtaining a final solution.
A relaxation algorithm is adopted, task weights are distributed according to the emergency degree and the importance, and a comprehensive score is generated for the tasks corresponding to the sub-problems according to a simple weighting method; determining a preliminary priority of each task by using an bubbling sequencing algorithm according to the comprehensive scores; determining the final task priority by using a priority queue method according to the preliminary priorities of all the tasks; selecting a corresponding branch limit algorithm solving strategy according to the final task priority; performing specific processing and operation on the task by using a selected algorithm; verifying whether the task is completed according to the processing result; if the task is not completed, recalculating the comprehensive score and determining the priority; and obtaining a final solution according to the processing results of all the tasks.
Using a relaxation algorithm, task weights are calculated and task processing order is determined based on urgency and importance. The higher the urgency and importance the higher the weight value, the urgency and importance are mapped to the initial weight value using a linear function. The workload or time estimate required for a task is used as a measure of the task size. And analyzing the pre-conditions or the dependent items among the tasks, and judging whether the tasks have other task dependencies or are used as the pre-conditions of other tasks. If the task requires a precondition, the weight value is increased. And analyzing whether the starting time and the finishing time of the task can be flexibly adjusted, and judging the variability degree of the task. The variability of the task is higher than the preset variability, and the weight value is reduced. And calculating the final weight of the task by using the weight to calculate the weight of the task according to the emergency degree, importance, size, dependency relationship and variability of the task. And determining the execution sequence and the priority of the tasks according to the weights of the tasks. And arranging the execution sequence of the tasks according to the sequence from high to low of the weight, and outputting the task execution sequence and the priority to obtain the finally determined task execution sequence and priority.
The invention uses NLP sensitive word detection technology based on a large model to judge whether malicious information or guiding keywords are contained in the problem content. The text is analyzed actively, rather than analyzing the text selected by the user. And extracting keywords by combining a TF-IDF scoring algorithm. And determining the matching degree of the question and each item in the knowledge base through a cosine similarity method. Using TensorFlow tools, a specific class of problems is obtained. And sequencing the evaluation data sources through the Hadoop cluster. The K-means clustering method is used for labeling the problems based on the characteristics of the problems. And transmitting the specific category information of the problems to a subsequent processing module through a RESTful API. And using FastText text classification models to conduct classification discrimination of malicious error guidance on the problem data. And (3) carrying out data labeling on the malicious behaviors, and if the malicious behaviors are judged to be malicious clients or malicious guides, automatically marking the problem as malicious content by using a label propagation algorithm. And according to the field matching degree evaluation value, adopting a task allocation algorithm to solve the customer service capability evaluation problem. And determining the priority of the task by adopting a priority algorithm according to the emergency degree and importance of the sub-problems, and processing to obtain a final solution.
The present invention is not limited to the above embodiments, but is capable of modification and variation in detail, and other modifications and variations can be made by those skilled in the art without departing from the scope of the present invention.
Claims (8)
1. An intelligent agent text processing method based on a large model is characterized by comprising the following steps of: the method comprises the following steps:
S1, decomposing questions asked by a user by using a large language model ChatGLM, judging attributes of the questions of the user, analyzing emotion colors of the questions of the user, identifying intentions and information requirements of the questions of the user, and labeling keywords in the questions;
S2, acquiring a problem type and a complexity evaluation value according to the problem attribute, and integrating the problem type, the emergency degree, the professional vocabulary and the professional problem attribute to determine the complexity of the problem; the problem types comprise professional after-sale problems, technical support problems and complex relation problems;
s3, acquiring domain information of a user question according to the type of the question, judging whether the question is a question attack and a malicious guide of a malicious client, and if the question is not a malicious error guide, determining the professional domain of the question by comparing the question with a professional domain knowledge base;
S4, acquiring the professional field of the intelligent customer service of the seat, and comparing the professional field of the problem with the professional field of the intelligent customer service of the seat to obtain a field matching degree evaluation value;
S5, distributing tasks to proper customer service according to the field matching degree evaluation value and the customer service busyness;
s6, after tasks are distributed, judging whether the capability of the customer service meets the task requirements according to the historical performance of the customer service, training records and user evaluation data;
s7, decomposing the task according to the complexity of the problem to obtain a plurality of sub-problems which are more suitable for the customer service knowledge field;
s8, determining the priority of the task for processing according to the emergency degree and the importance of the sub-problem, and obtaining a final solution;
In the step S7, the task decomposition includes the following steps:
When a plurality of possible answers or solutions exist for the problem proposed by the customer, the problem can be solved by applying knowledge in a plurality of fields, and fuzzy information or incomplete conditions exist in the problem, more information needs to be acquired through further inquiry or investigation, and the complex task is judged; acquiring professional fields contained in a complex task, wherein the professional fields comprise product knowledge, technical knowledge and industry knowledge fields; according to the field contained by the task, decomposing the complex task, and according to the decomposed field, generating a plurality of sub-problems more suitable for the customer service knowledge field.
2. The intelligent agent text processing method based on the large model as claimed in claim 1, wherein: in the step S1, the attributes of the questions include help, advice and complaints; wherein, analyzing the emotion color of the user question adopts an LSTM model.
3. The intelligent agent text processing method based on the large model as claimed in claim 1, wherein: in the step S2, the determining of the problem type includes:
if the questions include the product model, purchase date and warranty status attributes, judging that the questions are professional after-sale questions; if the problem comprises the operating system, the software version, the equipment model and the network environment attribute, judging the problem as a technical support problem; if the problem includes personnel identity, job responsibilities and team organization structure attributes, then the problem is determined to be a complex relationship problem.
4. A large model-based intelligent agent text processing method as claimed in claim 3, wherein: in the step S2, the determining of the complexity evaluation value specifically includes the following steps:
Judging whether a large number of same problems exist in a short time according to the problem description and the history record, if so, recording the number and the frequency of the same problems, judging that the number of questions is higher than a preset number value to be an urgent problem, and if so, recording the urgent degree of the problem and the required solving time;
determining whether a large number of professional terms and domain-specific knowledge are involved in the problem through the problem description and related documents, judging whether the problem belongs to the professional problem according to the professional vocabulary, the problem description and related knowledge involved in the problem, and recording the professionality of the problem and the participation degree of required professional technicians if the problem belongs to the professional problem.
5. The intelligent agent text processing method based on the large model as claimed in claim 1, wherein: the step S3 specifically comprises the following steps:
Using TF-IDF text classification algorithm to obtain specific class of problem and information of user questioning field; detecting the sensitive words by using a DFA algorithm, judging whether malicious information or guiding keywords are contained in the problem content, and marking the problem content as malicious behaviors if the malicious information or the guiding keywords are detected; information retrieval using inverted indexes, acquiring professional domain content related to the problem from a domain knowledge base; calculating the matching degree of the problem and the knowledge base content by adopting cosine similarity; using a TF-IDF text classification algorithm to match a vector machine, marking the problems and comparing the problems with data labels in a domain knowledge base; determining whether the problem is in a certain professional field by using a threshold judgment method according to the data tag comparison result in the field knowledge base; comparing the questions with the knowledge base content, screening out the most matched answers, and returning the most matched answers to the user;
Specifically, judging whether the problem content contains emotion of the problem by adopting a naive Bayesian combined with an Adaboost algorithm, carrying out emotion classification on the problem to be analyzed by using a trained naive Bayesian classifier and an Adaboost classifier, and obtaining the emotion type of the problem and a judging result of whether the problem contains malicious guidance by comparing the emotion type of the problem and the characteristic matching degree of related properties of the malicious guidance;
Specifically, a BERT model is adopted in the professional field of the determined problem, a problem text is input into the trained model, a BERT classification result is obtained, semantic matching is carried out on the new problem text and each item in the knowledge base, the problem text and the knowledge base item are input into the BERT model, an BERT output representation is obtained, then cosine similarity is used for calculating the similarity between the problem text and the knowledge base item, and the item with the highest similarity is selected as a matching result.
6. The intelligent agent text processing method based on the large model as claimed in claim 1, wherein: in the step S4, the step of obtaining the intelligent customer service professional field of the seat includes:
The method comprises the steps of obtaining the number of sales problems, after-sales problems and technical support problems processed by a seat customer service by recording the processing records of the seat customer service on customer problems; comprehensively evaluating customer satisfaction according to product or service quality evaluation, delivery time rate, after-sales service quality evaluation, customer feedback and complaint handling condition attributes; judging the processing capacity of the seat customer service through the performance of the seat customer service in the aspects of processing sales problems, after-sales problems and technical support, including complex problem processing capacity, response speed and communication skills; judging the professional field of the customer satisfaction degree and the processing capacity according to the number of customer service processing;
The step of obtaining the field matching degree evaluation value comprises the following steps:
Obtaining keywords and semantics in the problem description, and primarily judging the professional field related to the problem by analyzing the keywords and semantics in the problem description; constructing a knowledge base of intelligent customer service of the seat, which contains related knowledge and solutions in each professional field; matching the problems with data in a knowledge base by using a vector space model matching algorithm to obtain a matching result, and distributing the problems to corresponding professional fields according to the matching result; establishing a domain professional agent customer service library according to the related information of each professional domain and the professional agent customer service, and matching the problems with the professional domain of the agent customer service by using a vector space model matching algorithm; and obtaining a domain matching degree evaluation value by using the similarity between the professional domain of the cosine similarity calculation problem and the professional domain of the intelligent customer service of the seat.
7. The intelligent agent text processing method based on the large model as claimed in claim 1, wherein: in the step S5, the task allocation includes the following steps:
Judging the current workload and busy degree of the customer service by examining the workload, the processing task number and the processing time of the customer service; the capacity and efficiency of customer service processing tasks are evaluated through the working experience of the customer service and the speed and quality of solving the problems; determining whether the customer service can be used for processing tasks or not according to the working time, vacation and off-duty of the customer service; the weighted average method is adopted, the workload of customer service, the number of tasks and the time of processing are synthesized, the working experience of the customer service, the speed and the quality of solving the problem are synthesized, and the availability score of the customer service is obtained;
Judging the emergency degree of the task according to the deadline, importance and influence on the client of the task, and sequencing the priority of task processing according to the emergency degree of the task; and distributing the tasks to customer services with field matching degree higher than preset matching degree and availability degree higher than preset grading according to the priority ordering sequence of task processing.
8. The intelligent agent text processing method based on the large model as claimed in claim 1, wherein: the step S8 includes the steps of:
Task weights are distributed according to the emergency degree and the importance, and a comprehensive score is generated for the tasks corresponding to the sub-problems according to a simple weighting method; determining a preliminary priority of each task by using an bubbling sequencing algorithm according to the comprehensive scores; determining the final task priority by using a priority queue method according to the preliminary priorities of all the tasks; selecting a corresponding branch limit algorithm solving strategy according to the final task priority;
Performing specific processing and operation on the task by using a selected algorithm; verifying whether the task is completed according to the processing result; if the task is not completed, recalculating the comprehensive score and determining the priority; and obtaining a final solution according to the processing results of all the tasks.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311402059.9A CN117332066B (en) | 2023-10-27 | 2023-10-27 | Intelligent agent text processing method based on large model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311402059.9A CN117332066B (en) | 2023-10-27 | 2023-10-27 | Intelligent agent text processing method based on large model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117332066A CN117332066A (en) | 2024-01-02 |
| CN117332066B true CN117332066B (en) | 2024-05-28 |
Family
ID=89275373
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311402059.9A Active CN117332066B (en) | 2023-10-27 | 2023-10-27 | Intelligent agent text processing method based on large model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117332066B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103686723A (en) * | 2012-09-19 | 2014-03-26 | 镇江诺尼基智能技术有限公司 | System and method for processing complex SMS (short message service) message of mobile customer hotline SMS messaging service hall |
| CN103995858A (en) * | 2014-05-15 | 2014-08-20 | 北京航空航天大学 | Individualized knowledge active pushing method based on task decomposition |
| CN110033281A (en) * | 2018-01-11 | 2019-07-19 | 中兴通讯股份有限公司 | A kind of method and device that intelligent customer service is converted to artificial customer service |
| CN112885348A (en) * | 2021-01-25 | 2021-06-01 | 广州中汇信息科技有限公司 | AI-combined intelligent voice electric marketing method |
| CN115330243A (en) * | 2022-08-25 | 2022-11-11 | 中国电信股份有限公司 | Customer service work order distribution method and device |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3734471A1 (en) * | 2019-04-30 | 2020-11-04 | Tata Consultancy Services Limited | Method and system for utilizing domain knowledge to automatically identify solution to a problem |
-
2023
- 2023-10-27 CN CN202311402059.9A patent/CN117332066B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103686723A (en) * | 2012-09-19 | 2014-03-26 | 镇江诺尼基智能技术有限公司 | System and method for processing complex SMS (short message service) message of mobile customer hotline SMS messaging service hall |
| CN103995858A (en) * | 2014-05-15 | 2014-08-20 | 北京航空航天大学 | Individualized knowledge active pushing method based on task decomposition |
| CN110033281A (en) * | 2018-01-11 | 2019-07-19 | 中兴通讯股份有限公司 | A kind of method and device that intelligent customer service is converted to artificial customer service |
| CN112885348A (en) * | 2021-01-25 | 2021-06-01 | 广州中汇信息科技有限公司 | AI-combined intelligent voice electric marketing method |
| CN115330243A (en) * | 2022-08-25 | 2022-11-11 | 中国电信股份有限公司 | Customer service work order distribution method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117332066A (en) | 2024-01-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109767787B (en) | Emotion recognition method, device and readable storage medium | |
| CN109583966B (en) | High-value customer identification method, system, equipment and storage medium | |
| CN109816483B (en) | Information recommendation method and device and readable storage medium | |
| US12033163B2 (en) | Systems and methods for detecting complaint interactions | |
| CN107833059B (en) | Service quality evaluation method and system for customer service | |
| CN111507573A (en) | Business staff assessment method, system, device and storage medium | |
| CN112632989A (en) | Method, device and equipment for prompting risk information in contract text | |
| CN113051291A (en) | Work order information processing method, device, equipment and storage medium | |
| CN110399473A (en) | The method and apparatus for determining answer for customer problem | |
| CN117151870A (en) | Portrait behavior analysis method and system based on guest group | |
| CN118014622B (en) | Advertisement pushing method and system based on user portrait | |
| CN111179055A (en) | Credit limit adjusting method and device and electronic equipment | |
| EP4338089A1 (en) | System and method of automatic topic detection in text | |
| CN110532374B (en) | Insurance information processing method and device | |
| CN117332066B (en) | Intelligent agent text processing method based on large model | |
| CN116542676A (en) | Intelligent customer service system based on big data analysis and method thereof | |
| CN110929144A (en) | Business data management method, system and readable storage medium | |
| CN111159378A (en) | Method and device for classifying problem description information | |
| CN115203382A (en) | Service problem scene identification method and device, electronic equipment and storage medium | |
| CN115293867A (en) | Financial reimbursement user portrait optimization method, device, equipment and storage medium | |
| CN114037154A (en) | Method and system for predicting scientific and technological achievement number and theme based on attention characteristics | |
| KR102078541B1 (en) | Issue interest based news value evaluation apparatus and method, storage media storing the same | |
| CN111881287B (en) | Classification ambiguity analysis method and device | |
| CN117891898B (en) | Classification retrieval method and system based on GPT large model | |
| US20230134796A1 (en) | Named entity recognition system for sentiment labeling |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| CB03 | Change of inventor or designer information | ||
| CB03 | Change of inventor or designer information |
Inventor after: Cen Zhongman Inventor after: Wu Zhiwen Inventor after: Li Jieming Inventor after: He Jianwei Inventor after: Ye Tao Inventor after: Wu Yang Inventor after: Liang Yun Inventor after: Chen Yukai Inventor after: Dong Fuzhuang Inventor after: Lu Jiarong Inventor before: Cen Manzhong Inventor before: Wu Zhiwen Inventor before: Li Jieming Inventor before: He Jianwei Inventor before: Ye Tao Inventor before: Wu Yang Inventor before: Liang Yun Inventor before: Chen Yukai Inventor before: Dong Fuzhuang Inventor before: Lu Jiarong |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant |