CN110033281A - A kind of method and device that intelligent customer service is converted to artificial customer service - Google Patents
A kind of method and device that intelligent customer service is converted to artificial customer service Download PDFInfo
- Publication number
- CN110033281A CN110033281A CN201810024959.7A CN201810024959A CN110033281A CN 110033281 A CN110033281 A CN 110033281A CN 201810024959 A CN201810024959 A CN 201810024959A CN 110033281 A CN110033281 A CN 110033281A
- Authority
- CN
- China
- Prior art keywords
- customer service
- user
- value
- intelligent customer
- information
- 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.)
- Granted
Links
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
- 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
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/14—Session management
- H04L67/148—Migration or transfer of sessions
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Mathematical Physics (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Marketing (AREA)
- Signal Processing (AREA)
- General Business, Economics & Management (AREA)
- Finance (AREA)
- Economics (AREA)
- Development Economics (AREA)
- Computer Networks & Wireless Communication (AREA)
- Strategic Management (AREA)
- Artificial Intelligence (AREA)
- Human Computer Interaction (AREA)
- Accounting & Taxation (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- General Engineering & Computer Science (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The problem of the invention discloses the method and devices that a kind of intelligent customer service is converted to artificial customer service, are related to field of communication technology, and method includes: by inputting to user information carries out sentiment analysis, obtains the affective state value of the user;During the current question answering system of intelligent customer service serves the user, by analyzing asking a question in user's predetermined time, the satisfaction value of answer of the user to current question answering system is obtained;According to the described problem information that the user inputs, calculates described problem information and the problem of intelligent customer service knowledge base matches angle value;Angle value, which is matched, according to the affective state value, the satisfaction value and described problem carries out the conversion of intelligent customer service to artificial customer service.
Description
Technical field
The present invention relates to field of communication technology, in particular to method and dress that a kind of intelligent customer service to artificial customer service is converted
It sets.
Background technique
Current automatically request-answering system mainly includes that (Frequently Asked Questions, that often asks asks FAQ
Topic) retrieval type question answering system and knowledge based map question answering system, but the question answering system of mesh first two model all can not be accurate
Solve the problems, such as that user is ever-changing, so usually intelligent Answer System can all have threshold value control question answering system, if it is less than
Certain threshold value, general question answering system can all refuse the methods of manual service of answering or transfer, but existing question answering system is usual
It is only matched according to the problems in question and answer and question and answer library, similarity is calculated by similarity algorithm, is then asked less than certain threshold value
The system of answering will be abandoned answering.Existing threshold value calculation method is too simple, does not make full use of the information of user, such as user
Emotion information and user intention comprehensive expressed by user whithin a period of time.These information are all question answering systems about answering
The information that should be made full use of during the threshold calculations of case accuracy rate.So the present invention propose one based on it is more strategy fusion
The method of automatic conversion manual service in intelligent customer service, takes full advantage of in the emotion information of user and a period of time of user
Be satisfied with information.
Summary of the invention
The technical issues of scheme provided according to embodiments of the present invention solves is can not to answer use in intelligent Answer System
It cannot automatically switch operator attendance when the problem of family.
The method that a kind of intelligent customer service provided according to embodiments of the present invention is converted to artificial customer service, comprising:
The information of the problem of by inputting to user carries out sentiment analysis, obtains the affective state value of the user;
During the current question answering system of intelligent customer service serves the user, by being mentioned in user's predetermined time
Problem is analyzed, and the satisfaction value of answer of the user to current question answering system is obtained;
According to the described problem information that the user inputs, described problem information is calculated and the problem of intelligent customer service knowledge base
Match angle value;
Angle value, which is matched, according to the affective state value, the satisfaction value and described problem carries out intelligent customer service to people
The conversion of work customer service.
Preferably, described that angle value progress is matched according to the affective state value, the satisfaction value and described problem
The conversion of intelligent customer service to artificial customer service includes:
Angle value is matched according to the affective state value, the satisfaction value and described problem, calculates the intelligence visitor
Take the service conversion value of question answering system;
The service conversion value is compared with the service switching threshold prestored;
If the service conversion value is greater than the service switching threshold, current intelligent customer service is converted to artificial visitor
Clothes;
If the service conversion value is not more than the service switching threshold, current intelligent customer service is kept.
Preferably, described that angle value is matched according to the affective state value, the satisfaction value and described problem, it calculates
The service conversion value of the intelligent customer service question answering system includes:
It services conversion value=affective state value * X+ satisfaction value * Y+ problem and matches angle value * Z;
Wherein, described X, Y and Z are positive number, and X+Y+Z=1.
Preferably, it is described by being inputted to user the problem of information carry out sentiment analysis, obtain the emotion shape of the user
State value includes:
The problem of information of the problem of by inputting to user carries out text vector processing, obtains described problem information feature
Vector;
According to the corresponding relationship of preset problem feature vector and emotion mark value, obtain and described problem feature vector phase
Corresponding emotion mark value, and using the emotion mark value as the affective state value of the user.
Preferably, described by analyzing asking a question in user's predetermined time, the user is obtained to working as
The satisfaction value of the answer of preceding question answering system includes:
By carrying out text vector processing respectively to asking a question in user's predetermined time, each putd question to is obtained
The problem of topic feature vector;
According to the corresponding relationship of preset problem feature vector and satisfaction value, obtain and described problem feature vector phase
Corresponding satisfaction value;
Using the satisfaction value all in the predetermined time, answer of the user to current question answering system is obtained
Satisfaction value.
Preferably, the described problem information inputted according to the user, calculates described problem information and intelligent customer service
Knowledge base the problem of matching angle value include:
Using normalization algorithm, the described problem information that the user inputs is normalized to the intelligent customer service knowledge base
In one problem category;
In the intelligent customer service knowledge base, Lucene inspection is carried out by the described problem information inputted to the user
Rope obtains in a Candidate Set of the intelligent customer service knowledge base;
Using described problem classification and the Candidate Set, the problem of calculating described problem information and intelligent customer service knowledge base
With angle value.
Preferably, utilization described problem classification and the Candidate Set, calculating active user's problem and intelligent customer service
The problem of knowledge base, matches angle value
If described problem classification belongs to the Candidate Set, using described problem classification as active user's problem and intelligence visitor
The problem of knowledge base of clothes, matches angle value;
It is similar by being carried out to all answers in the Candidate Set if described problem classification is not belonging to the Candidate Set
Degree calculates, and chooses with the highest candidate answers of active user's problem similarity as active user's problem and intelligent customer service
The problem of knowledge base, matches angle value.
The device that a kind of intelligent customer service provided according to embodiments of the present invention is converted to artificial customer service, comprising:
Affective state module is obtained, information carries out sentiment analysis the problem of for by inputting to user, obtains the use
The affective state value at family;
Satisfaction module is obtained, for during the current question answering system of intelligent customer service serves the user, by right
It asks a question and analyzes in user's predetermined time, obtain the satisfaction of answer of the user to current question answering system
Value;
Computational problem matching degree module, the described problem information for being inputted according to the user calculate described problem letter
The problem of breath and intelligent customer service knowledge base, matches angle value;
Conversion module, for according to the affective state value, the satisfaction value and described problem match angle value into
Row intelligent customer service to artificial customer service conversion.
The equipment that a kind of intelligent customer service provided according to embodiments of the present invention is converted to artificial customer service, the equipment include:
Processor, and the memory with processor coupling;Be stored on the memory to run on the processor
The program that intelligent customer service is converted to artificial customer service, the program that the intelligent customer service to artificial customer service is converted are executed by the processor
Shi Shixian includes:
The information of the problem of by inputting to user carries out sentiment analysis, obtains the affective state value of the user;
During the current question answering system of intelligent customer service serves the user, by being mentioned in user's predetermined time
Problem is analyzed, and the satisfaction value of answer of the user to current question answering system is obtained;
According to the described problem information that the user inputs, described problem information is calculated and the problem of intelligent customer service knowledge base
Match angle value;
Angle value, which is matched, according to the affective state value, the satisfaction value and described problem carries out intelligent customer service to people
The conversion of work customer service.
A kind of computer storage medium provided according to embodiments of the present invention is stored with intelligent customer service to artificial customer service and converts
Program, when the intelligent customer service to the program that artificial customer service is converted is executed by processor realize include:
The information of the problem of by inputting to user carries out sentiment analysis, obtains the affective state value of the user;
During the current question answering system of intelligent customer service serves the user, by being mentioned in user's predetermined time
Problem is analyzed, and the satisfaction value of answer of the user to current question answering system is obtained;
According to the described problem information that the user inputs, described problem information is calculated and the problem of intelligent customer service knowledge base
Match angle value;
Angle value, which is matched, according to the affective state value, the satisfaction value and described problem carries out intelligent customer service to people
The conversion of work customer service.
The scheme provided according to embodiments of the present invention, in automatically switching algorithm, change is commonly simply set according to threshold value
The problem of judging whether to turn operator attendance calmly, user in the emotion information and a period of time of user is made full use of to put question to included
Implicit information, while use normalization algorithm and a variety of strategies of similarity algorithm, more accurately judge current question answering system
Whether user is successfully solved the problems, such as.
Detailed description of the invention
Fig. 1 is the method flow diagram that a kind of intelligent customer service provided in an embodiment of the present invention is converted to artificial customer service;
Fig. 2 is the schematic device that a kind of intelligent customer service provided in an embodiment of the present invention is converted to artificial customer service;
Fig. 3 is the bulk flow of the blending algorithm provided in an embodiment of the present invention based on sentence pattern template and integrated multidimensional similarity
Cheng Tu;
Fig. 4 is the flow chart of entire sentiment analysis module off-line training provided in an embodiment of the present invention;
Fig. 5 is the flow chart of the training term vector of word2vec provided in an embodiment of the present invention;Present design has many places
Term vector training is used, then illustrates the training process of term vector by figure.
Fig. 6 is the structure chart of the single neuron of LSTM provided in an embodiment of the present invention;
Fig. 7 is normalization algorithm off-line training provided in an embodiment of the present invention and online normalized whole flow process figure.
Specific embodiment
Below in conjunction with attached drawing to a preferred embodiment of the present invention will be described in detail, it should be understood that described below is excellent
Select embodiment only for the purpose of illustrating and explaining the present invention and is not intended to limit the present invention.
Fig. 1 is the method flow diagram that a kind of intelligent customer service provided in an embodiment of the present invention is converted to artificial customer service, such as Fig. 1 institute
Show, comprising:
Step S101: information carries out sentiment analysis the problem of by inputting to user, obtains the affective state of the user
Value;
Step S102: during the current question answering system of intelligent customer service serves the user, by predetermined to the user
It asks a question and is analyzed in time, obtain the satisfaction value of answer of the user to current question answering system;
Step S103: the described problem information inputted according to the user calculates described problem information and intelligent customer service is known
The problem of knowing library matches angle value;
Step S104: angle value is matched according to the affective state value, the satisfaction value and described problem and carries out intelligence
Can customer service to artificial customer service conversion.
Wherein, described that angle value progress intelligence is matched according to the affective state value, the satisfaction value and described problem
The conversion of energy customer service to artificial customer service includes: to be matched according to the affective state value, the satisfaction value and described problem
Angle value calculates the service conversion value of the intelligent customer service question answering system;The service conversion value is converted into threshold with the service prestored
Value is compared;If the service conversion value is greater than the service switching threshold, current intelligent customer service is converted to artificial
Customer service;If the service conversion value is not more than the service switching threshold, current intelligent customer service is kept.
Specifically, described match angle value, meter according to the affective state value, the satisfaction value and described problem
The service conversion value for calculating the intelligent customer service question answering system includes: service conversion value=affective state value * X+ satisfaction value * Y+
Problem matches angle value * Z;Wherein, described X, Y and Z are positive number, and X+Y+Z=1;The X is affective state weight;The Y
Satisfaction weight;The Z is problem matching degree weight.
Wherein, it is described by being inputted to user the problem of information carry out sentiment analysis, obtain the affective state of the user
The problem of the problem of value includes: by inputting to user information carries out text vector processing, obtains described problem information feature
Vector;According to the corresponding relationship of preset problem feature vector and emotion mark value, obtain opposite with described problem feature vector
The emotion mark value answered, and using the emotion mark value as the affective state value of the user.
Wherein, described by analyzing asking a question in user's predetermined time, the user is obtained to current
The satisfaction value of the answer of question answering system include: by ask a question in user's predetermined time respectively carry out text to
Quantification treatment obtains the problem of each asked a question feature vector;According to preset problem feature vector and satisfaction value
Corresponding relationship obtains satisfaction value corresponding with described problem feature vector;Using all described full in the predetermined time
Meaning degree value, obtains the satisfaction value of answer of the user to current question answering system.
Wherein, the described problem information inputted according to the user, calculating described problem information and intelligent customer service
It includes: that the described problem information that the user inputs is normalized to institute using normalization algorithm that the problem of knowledge base, which matches angle value,
It states in a problem category of intelligent customer service knowledge base;In the intelligent customer service knowledge base, pass through what is inputted to the user
Described problem information carries out Lucene retrieval, obtains in a Candidate Set of the intelligent customer service knowledge base;Utilize described problem
The problem of classification and Candidate Set, calculating described problem information and intelligent customer service knowledge base, matches angle value.
Specifically, described utilize described problem classification and the Candidate Set, active user's problem and intelligent customer service are calculated
Knowledge base the problem of matching angle value include: if that described problem classification belongs to the Candidate Set, using described problem classification as
The problem of knowledge base of active user's problem and intelligent customer service, matches angle value;If described problem classification is not belonging to the Candidate Set,
Then by carrying out similarity calculation to all answers in the Candidate Set, choose and active user's problem similarity highest
Candidate answers match angle value as the problem of knowledge base of active user's problem and intelligent customer service.
Fig. 2 is the schematic device that a kind of intelligent customer service provided in an embodiment of the present invention is converted to artificial customer service, such as Fig. 2 institute
Show, comprising: obtain affective state module 201, information carries out sentiment analysis the problem of for by inputting to user, obtains described
The affective state value of user;Satisfaction module 202 is obtained, for serving the user in the current question answering system of intelligent customer service
Period obtains the user and answers current question answering system by analyzing asking a question in user's predetermined time
The satisfaction value of case;Computational problem matching degree module 203, the described problem information for being inputted according to the user calculate
The problem of described problem information and intelligent customer service knowledge base, matches angle value;Conversion module 204, for according to the affective state
Value, the satisfaction value and described problem matching angle value carry out the conversion of intelligent customer service to artificial customer service.
The equipment that a kind of intelligent customer service provided in an embodiment of the present invention is converted to artificial customer service, the equipment include: processing
Device, and the memory with processor coupling;The intelligence that can be run on the processor is stored on the memory
The program that customer service is converted to artificial customer service, it is real when the program that the intelligent customer service to artificial customer service is converted is executed by the processor
Now include:
The information of the problem of by inputting to user carries out sentiment analysis, obtains the affective state value of the user;
During the current question answering system of intelligent customer service serves the user, by being mentioned in user's predetermined time
Problem is analyzed, and the satisfaction value of answer of the user to current question answering system is obtained;
According to the described problem information that the user inputs, described problem information is calculated and the problem of intelligent customer service knowledge base
Match angle value;
Angle value, which is matched, according to the affective state value, the satisfaction value and described problem carries out intelligent customer service to people
The conversion of work customer service.
A kind of computer storage medium provided in an embodiment of the present invention is stored with the journey that intelligent customer service is converted to artificial customer service
Sequence, when the intelligent customer service to the program that artificial customer service is converted is executed by processor realize include:
The information of the problem of by inputting to user carries out sentiment analysis, obtains the affective state value of the user;
During the current question answering system of intelligent customer service serves the user, by being mentioned in user's predetermined time
Problem is analyzed, and the satisfaction value of answer of the user to current question answering system is obtained;
According to the described problem information that the user inputs, described problem information is calculated and the problem of intelligent customer service knowledge base
Match angle value;
Angle value, which is matched, according to the affective state value, the satisfaction value and described problem carries out intelligent customer service to people
The conversion of work customer service.
Fig. 3 is the bulk flow of the blending algorithm provided in an embodiment of the present invention based on sentence pattern template and integrated multidimensional similarity
Cheng Tu, as shown in Figure 3, comprising:
Step 1: customer problem being normalized into a certain classification using normalization algorithm when user starts first round dialogue
A。
Step 2: retrieving to obtain answer Candidate Set B by Lucene.
Step 3: if classification A in Candidate Set B, returns to the answer of classification A, if classification A not in selected works B,
By carrying out similarity calculation to Candidate Set B, the highest candidate answers of similarity are returned.
Step 4: after first round dialogue, sentiment analysis is obtained the problem of by sentiment analysis technology by active user
The current affective state of user out.
Step 5: by analyzing to obtain active user to question answering system to the problem of enquirement in user's a period of time recently
Answer satisfaction.
Step 6: the matching of the problems in active user's problem and knowledge base is calculated by normalization algorithm and similarity algorithm
Degree.
Step 7: result being analyzed by problem matching degree, psycho-emotional and the satisfaction of current question answering system is judged
Whether transfer operator attendance.
Wherein, the problem of by sentiment analysis technology by active user, sentiment analysis obtained the current emotion shape of user
State includes:
Sentiment analysis technology has been technology very mature in natural language processing field, traditional sentiment analysis technology master
It will be by sentiment dictionary and some other affective characteristics extraction feature, finally by the class of conventional machines learning algorithm computing statement
Not, as shown in Figure 4.
401, corpus is collected first and corpus is labeled, and is mainly labeled as three kinds of labels: neutral, commendation, derogatory sense.Example
Such as: " my credit balance is how many? ", it is labeled as neutral label;" I want to look into be how much telephone expenses also remain ", is labeled as demoting
Adopted label;" good, I also wants to may I ask how much telephone expenses also remain ", is labeled as commendation emotion.
402, sentiment analysis module extracts text vector and Feature Selection method to language material feature first, then basis
Label passes through SVM (Support Vector Machine, support vector machines) algorithm classification.
1) sentiment analysis module segments corpus by Words partition system first;
2) then sentiment analysis module is using word as minimum semantic granularity, using word frequency as weight, by text vector space
Change, is converted into the vector of [0,1,1,0,3] form.One word feature of every one-dimensional representation, when this word occurs simultaneously in the text
N times so this dimensional feature numerical value occur is n.Such as second dimensional feature word be " good ", then " good, my telephone expenses are also surplus for text
How much ", then the second dimensional feature value is 1.
3) then sentiment analysis module passes through Chi-square Test method for feature ordering, selects the highest some features of correlation
As final feature.Whether Chi-square Test predominantly detects mutually indepedent between feature and target variable.Such as investigate " happy "
The correlation of " commendation " classification.It can be used there are four observed value: comprising " happy " and belonging to the number of files of " commendation " classification,
It is named as A;Comprising " happy " but it is not belonging to the number of files of " commendation " classification, is named as B;Not comprising " happy " but belong to and " praise
The number of files of justice " classification, is named as C;Both do not included " happy " or be not belonging to the number of files of " commendation " classification, and be named as D, total
Number of files is N.
So how to calculate the desired value of feature " happy "? because A+B is the article number comprising " happy ", divided by total text
Gear number is exactly the probability of " happy " appearance, and the article number for belonging to commendation class is A+C, in this few document, it should have
At this point, the desired value of " happy " this feature has been obtained, be next exactly calculate actual value and desired value it
Between deviation.As follows
It is formula in substitution, available:
The biggish K feature of its value is finally selected according to D.
4) selection sort device: the stronger SVM of generalization ability is used, SVM is a kind of two sorting algorithms for having supervision.Due to this
Task is three classification, so sentiment analysis module will be used as a classification task two-by-two, trains three classifiers, finally by
Ballot obtains final class label.
403, last sentiment analysis module is online using the emotion information current by trained model prediction.
Wherein, by answering question answering system to analyzing to obtain active user the problem of enquirement in user's a period of time recently
Case satisfaction includes:
User satisfaction module mainly utilizes LSTM (Long Short-Term Memory, shot and long term memory network) locating
The superior performance of time serial message is managed, while utilizing CNN (Convolutional Neural Network, convolutional Neural net
Network) network progress feature extraction, the input text of user in a period of time is handled, predicts it during this period of time to question answering system
Satisfaction.
501, it first and collects corpus and satisfaction mark is carried out to corpus, mark four kinds of label: it is satisfied, it is substantially full
Meaning is unsatisfied with, very dissatisfied.For example, " there are also how many for my telephone expenses | credit balance also how many " is dissatisfied;" if me
Expense remaining sum is how much | and my set meal is with how many " be satisfied in the main.
502, then it is corresponding to obtain all words using Word2vec to word vectors in corpus for user satisfaction module
Term vector (dimension value is 50 here), as shown in table 1, the word vectors process of Word2vec is as shown in Figure 5.
The word vectors table of table 1:Word2vec
503, then user satisfaction module is extracted profound using the term vector of each word as input using LSTM network
Word information, as shown in Figure 6.
LSTM parameter setting:
Input layer: 20 inadequate zero paddings of term vector, the i.e. term vector of I dimension;Here value is I=50;
Hidden layer: number of plies N, every layer includes H neuron, and value is N=1, H=50 here;
Output layer: O dimensional vector, O=20 identical as the input word number of sentence;
LSTM training goal: 4 weight matrix and 4 offset vectors are asked:
Input gate
Wi: dimension is 400*200 bi: dimension 200
Forget door
Wf: dimension is 400*200 bf: amount dimension is 200
Cell updates transformation
Wc: dimension is 400*200 bc: dimension 200
Out gate
Wo: dimension is 400*200 bo: dimension 200
Linear operation
W: dimension is 200*40 b: dimension 40
LSTM calculation formula:
Input gate: the current input of control and back export the information content for entering new cell
it=σ (Wi·[ht-1, xt]+bi)
Forget door: deciding whether state that is clear or keeping single part
ft=σ (Wf·[ht-1, xt+bf)
Cell updates transformation: transformation output and previous state to last state
Cell state updates step: the state for calculating next timestamp uses previous state and input through processing of moving into one's husband's household upon marriage
Out gate: the output of cell is calculated
ot=σ (Wo[ht-1, xt]+bo)
The output of final LSTM: the tanh transformation of current state is converted again using one
Orea=Wht+b
ht=ot*tanh(Ct)
It by repetition training, for example trains 1000 times, obtains 5 weight matrix and 5 offset vectors.This is arrived, LSTM mould
Type training terminates.
504, it is that an above-mentioned LSTM network structure extracts profound word information for each sentence, then user
Satisfaction module extracts the feature of the time series of adjacent two word using CNN network again.CNN network parameter characteristic window 2*
1, characteristic type 1 exports identical with single input dimension.Softmax operation is finally used, final classification is obtained.
505, user satisfaction module evaluates and tests the satisfied of active user using trained model.
Wherein, the problems in active user's problem and knowledge base are calculated by normalization algorithm and similarity algorithm and matches journey
Degree includes:
It is mainly searched from question answering system knowledge base by normalization algorithm and Similarity algorithm and customer problem is most like
Answer.
601, customer problem is normalized in a certain classification A using normalizing algorithm first, as shown in Figure 7.
1) principle of normalization algorithm is using RNN network by sentence vectorization, and concrete principle is as follows:
The principle for preparing word vector sum term vector is the same, using the form that industry is general, trained vector length here
Degree is S (such as S=50, i.e., the word vector of 50 dimensions);
Word vector shaped like
Upper [0.1525,0.3658,0.1634,0.2510 ...]
Under [0.1825,0.2658,0.1834,0.2715 ...]
Training pattern
It can be determined according to the structure chart of RNN Recognition with Recurrent Neural Network:
Input layer
For example input X word, i.e. I=X* word vector dimension;
Work as X=1, word vector dimension is 50, then I=50;
Work as X=2, word vector dimension is 50, then I=100;
Here it is trained with individual character, i.e. X=1, I=50;
Note: X is bigger, and trained model is more accurate, but the workload of training is bigger;
Hidden layer
The neuronal quantity of hidden layer number and each hidden layer;
For example design has the RNN neural network of 3 hidden layers;
H1=200;H2=800;H3=2000;
H can be rule of thumb arranged;Here provided with 1 hidden layer, the number of neuron is H=200;
Note: the number of hidden layer is more, and trained model is more accurate, but the workload of training is bigger;
Output layer
The number of output layer neuron quantity K and word information is consistent, and is set as K=5000 here.
Trained purpose is to ask N to weight matrix, and the size of N is related with the number of RNN hidden layer;
When there is 1 hidden layer, it will obtain N=3 weight matrix;
When there is 2 hidden layers, it will obtain N=5 weight matrix;
When there is 3 hidden layers, it will obtain N=7 weight matrix;
For example, 1 hidden layer H=200 of setting, then obtain 3 weight matrix:
Wih input layer-hidden layer weight matrix such as 50&200
Whk hidden layer-output layer weight matrix such as 200&5000
Whh hidden layer-hidden layer weight matrix such as 200&200
Due to using two-way training, therefore final weight matrix number is 2N;
Word information is the use when RNN calculates error, and word vector is to be converted to computer capacity in the word of " training dataset "
It is used when the digital information of identification.
The calculation formula of hidden layer neuron:
θhIt is activation primitive I input node number H hidden layer neuron number
The calculation formula of output layer neuron:
θkIt is Softmax function H be hidden layer neuron number K is output layer neuron number
Training goal and result: it by repetition training, for example trains 2000 times, it is therefore an objective to obtain N to weight matrix, i.e. W
Weight file.
This is arrived, RNN sentence indicates that network training finishes, the network is mainly used for following purposes:
The probability of next word of sentence is predicted according to division statement, it can calculate the next word probability of sentence.
Such as:
Ancestral's () that I likes
0.6012 > of < state
0.2017 > of < elder generation
0.0254 > of < ancestor
……
Since sentence library takes full advantage of sentence contextual information, thus its probability statistics can according to the variation of specific context and
It is different:
I graduates from university, on I will go daily
<class 0.2412>
<learning 0.1017>
My, on I will go daily in eight years old this year
<class 0.1016>
<learning 0.1517>
It is similar that sentence fully-connected network weight calculation and last softmax to two different vectorizations calculate its
Degree.Similarity is calculated according to all sentences in read statement and normalization library, returns to the high classification of overall similarity.
602, customer problem library is retrieved by Lucene and obtains Candidate Set B.
603, classification A is checked whether in Candidate Set B, the A when candidate answers if, if not passing through similarity operator if
Method calculates the answer in Candidate Set B and the similarity of customer problem, returns to the highest answer of similarity, similarity algorithm and normalizing
It is similar to change algorithm principle, only normalization and all sentences calculate similarity, and similarity is similar with the calculating of a certain sentence
Angle value.
Wherein, described by problem matching degree, psycho-emotional analysis result and to the satisfaction of current question answering system
Judge whether that switching operator attendance includes:
The weighted sum of computational problem matching degree, psycho-emotional analysis result and the satisfaction to current question answering system,
If weighted sum is greater than threshold value, answer is returned to, artificial customer service of transferring if weighted sum is less than threshold value.
The scheme provided according to embodiments of the present invention analyzes result by problem matching degree, psycho-emotional and to current
The satisfaction of question answering system judges whether operator attendance of transferring, and can more accurately judge whether current question answering system successfully solves
Certainly the problem of user.
Although describing the invention in detail above, but the invention is not restricted to this, those skilled in the art of the present technique
It can be carry out various modifications with principle according to the present invention.Therefore, all to be modified according to made by the principle of the invention, all it should be understood as
Fall into protection scope of the present invention.
Claims (10)
1. a kind of method that intelligent customer service is converted to artificial customer service, comprising:
The information of the problem of by inputting to user carries out sentiment analysis, obtains the affective state value of the user;
During the current question answering system of intelligent customer service serves the user, by being asked a question in user's predetermined time
It is analyzed, obtains the satisfaction value of answer of the user to current question answering system;
According to the described problem information that the user inputs, calculates described problem information and the problem of intelligent customer service knowledge base matches
Angle value;
Angle value, which is matched, according to the affective state value, the satisfaction value and described problem carries out intelligent customer service to artificial visitor
The conversion of clothes.
2. according to the affective state value, the satisfaction value and described being asked according to the method described in claim 1, described
Topic matching angle value carry out intelligent customer service to artificial customer service conversion include:
Angle value is matched according to the affective state value, the satisfaction value and described problem, the intelligent customer service is calculated and asks
Answer the service conversion value of system;
The service conversion value is compared with the service switching threshold prestored;
If the service conversion value is greater than the service switching threshold, current intelligent customer service is converted to artificial customer service;
If the service conversion value is not more than the service switching threshold, current intelligent customer service is kept.
3. according to the affective state value, the satisfaction value and described being asked according to the method described in claim 2, described
Topic matching angle value, the service conversion value for calculating the intelligent customer service question answering system include:
It services conversion value=affective state value * X+ satisfaction value * Y+ problem and matches angle value * Z;
Wherein, described X, Y and Z are positive number, and X+Y+Z=1.
4. according to the method described in claim 1, it is described by being inputted to user the problem of information carry out sentiment analysis, obtain institute
The affective state value for stating user includes:
The information of the problem of by inputting to user carries out text vector processing, the problem of obtaining described problem information feature to
Amount;
According to the corresponding relationship of preset problem feature vector and emotion mark value, obtain corresponding with described problem feature vector
Emotion mark value, and using the emotion mark value as the affective state value of the user.
5. obtaining according to the method described in claim 1, described by analyzing asking a question in user's predetermined time
Include: to satisfaction value of the user to the answer of current question answering system
By carrying out text vector processing respectively to asking a question in user's predetermined time, obtain each asked a question
Problem characteristic vector;
According to the corresponding relationship of preset problem feature vector and satisfaction value, obtain corresponding with described problem feature vector
Satisfaction value;
Using the satisfaction value all in the predetermined time, the satisfaction of answer of the user to current question answering system is obtained
Degree value.
6. according to the method described in claim 1, the described problem information inputted according to the user, calculates described problem
The problem of knowledge base of information and intelligent customer service, matches angle value
Using normalization algorithm, the described problem information that the user inputs is normalized to one of the intelligent customer service knowledge base
In problem category;
In the intelligent customer service knowledge base, Lucene retrieval is carried out by the described problem information inputted to the user, is obtained
Into a Candidate Set of the intelligent customer service knowledge base;
Using described problem classification and the Candidate Set, the problem of calculating described problem information and intelligent customer service knowledge base matching degree
Value.
7. calculating active user according to the method described in claim 6, described utilize described problem classification and the Candidate Set and asking
The problem of knowledge base of topic and intelligent customer service, matches angle value
If described problem classification belongs to the Candidate Set, using described problem classification as active user's problem and intelligent customer service
The problem of knowledge base, matches angle value;
If described problem classification is not belonging to the Candidate Set, by carrying out similarity meter to all answers in the Candidate Set
It calculates, chooses and knowledge of the highest candidate answers of active user's problem similarity as active user's problem and intelligent customer service
The problem of library, matches angle value.
8. the device that a kind of intelligent customer service is converted to artificial customer service, comprising:
Affective state module is obtained, information carries out sentiment analysis the problem of for by inputting to user, obtains the user's
Affective state value;
Satisfaction module is obtained, for during the current question answering system of intelligent customer service serves the user, by described
It asks a question and analyzes in user's predetermined time, obtain the satisfaction value of answer of the user to current question answering system;
Computational problem matching degree module, the described problem information for being inputted according to the user, calculate described problem information and
The problem of intelligent customer service knowledge base, matches angle value;
Conversion module carries out intelligence for matching angle value according to the affective state value, the satisfaction value and described problem
Can customer service to artificial customer service conversion.
9. the equipment that a kind of intelligent customer service is converted to artificial customer service, the equipment includes: processor, and with the processor coupling
The memory connect;The journey that the intelligent customer service that be stored on the memory to run on the processor is converted to artificial customer service
Sequence, when the program that the intelligent customer service to artificial customer service is converted is executed by the processor realize include:
The information of the problem of by inputting to user carries out sentiment analysis, obtains the affective state value of the user;
During the current question answering system of intelligent customer service serves the user, by being asked a question in user's predetermined time
It is analyzed, obtains the satisfaction value of answer of the user to current question answering system;
According to the described problem information that the user inputs, calculates described problem information and the problem of intelligent customer service knowledge base matches
Angle value;
Angle value, which is matched, according to the affective state value, the satisfaction value and described problem carries out intelligent customer service to artificial visitor
The conversion of clothes.
10. a kind of computer storage medium is stored with the program that intelligent customer service is converted to artificial customer service, the intelligent customer service to people
The program of work customer service conversion is realized when being executed by processor
The information of the problem of by inputting to user carries out sentiment analysis, obtains the affective state value of the user;
During the current question answering system of intelligent customer service serves the user, by being asked a question in user's predetermined time
It is analyzed, obtains the satisfaction value of answer of the user to current question answering system;
According to the described problem information that the user inputs, calculates described problem information and the problem of intelligent customer service knowledge base matches
Angle value;
Angle value, which is matched, according to the affective state value, the satisfaction value and described problem carries out intelligent customer service to artificial visitor
The conversion of clothes.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810024959.7A CN110033281B (en) | 2018-01-11 | 2018-01-11 | Method and device for converting intelligent customer service into manual customer service |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810024959.7A CN110033281B (en) | 2018-01-11 | 2018-01-11 | Method and device for converting intelligent customer service into manual customer service |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110033281A true CN110033281A (en) | 2019-07-19 |
CN110033281B CN110033281B (en) | 2023-07-21 |
Family
ID=67234129
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810024959.7A Active CN110033281B (en) | 2018-01-11 | 2018-01-11 | Method and device for converting intelligent customer service into manual customer service |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110033281B (en) |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110428267A (en) * | 2019-07-31 | 2019-11-08 | 联想(北京)有限公司 | Content processing method and device, computer system and readable storage medium storing program for executing |
CN110991178A (en) * | 2019-11-08 | 2020-04-10 | 苏宁金融科技(南京)有限公司 | Intelligent customer service and artificial customer service switching method and device and computer equipment |
CN111090721A (en) * | 2019-11-25 | 2020-05-01 | 出门问问(苏州)信息科技有限公司 | Question answering method and device and electronic equipment |
CN111369080A (en) * | 2020-05-27 | 2020-07-03 | 支付宝(杭州)信息技术有限公司 | Intelligent customer service solution rate prediction method and system and multi-service prediction model |
CN111754061A (en) * | 2019-11-22 | 2020-10-09 | 北京沃东天骏信息技术有限公司 | Method and device for controlling man-machine distribution, server equipment and storage medium |
CN112417842A (en) * | 2019-08-19 | 2021-02-26 | 阿里巴巴集团控股有限公司 | Service satisfaction determining method and device, storage medium and computer equipment |
CN113051386A (en) * | 2021-04-30 | 2021-06-29 | 中国银行股份有限公司 | Adjustment method and device for manual service conversion, electronic equipment and storage medium |
CN113282737A (en) * | 2021-07-21 | 2021-08-20 | 中信建投证券股份有限公司 | Man-machine cooperation intelligent customer service dialogue method and device |
CN113793099A (en) * | 2021-09-13 | 2021-12-14 | 深圳众投互联信息技术有限公司 | Intelligent customer service interaction management method and equipment based on deep learning and computer storage medium |
CN113946673A (en) * | 2021-12-20 | 2022-01-18 | 深圳市一号互联科技有限公司 | Semantic-based intelligent customer service routing processing method and device |
CN114860912A (en) * | 2022-05-20 | 2022-08-05 | 马上消费金融股份有限公司 | Data processing method and device, electronic equipment and storage medium |
CN115099727A (en) * | 2022-08-25 | 2022-09-23 | 深圳市星网信通科技有限公司 | Customer service routing method and device and computer readable storage medium |
CN117332066A (en) * | 2023-10-27 | 2024-01-02 | 天讯瑞达通信技术有限公司 | Intelligent agent text processing method based on large model |
CN118013009A (en) * | 2024-02-19 | 2024-05-10 | 广东融益数字科技有限公司 | Intelligent customer service interaction method and system |
CN118485447A (en) * | 2024-05-30 | 2024-08-13 | 浙江英集动力科技有限公司 | Heat supply digital person and customer service processing method based on heat supply digital person |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102789496A (en) * | 2012-07-13 | 2012-11-21 | 携程计算机技术(上海)有限公司 | Method and system for implementing intelligent response |
CN105631022A (en) * | 2015-12-29 | 2016-06-01 | 上海智臻智能网络科技股份有限公司 | Method and device for processing information |
US9471668B1 (en) * | 2016-01-21 | 2016-10-18 | International Business Machines Corporation | Question-answering system |
CN107315766A (en) * | 2017-05-16 | 2017-11-03 | 广东电网有限责任公司江门供电局 | A kind of voice response method and its device for gathering intelligence and artificial question and answer |
-
2018
- 2018-01-11 CN CN201810024959.7A patent/CN110033281B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102789496A (en) * | 2012-07-13 | 2012-11-21 | 携程计算机技术(上海)有限公司 | Method and system for implementing intelligent response |
CN105631022A (en) * | 2015-12-29 | 2016-06-01 | 上海智臻智能网络科技股份有限公司 | Method and device for processing information |
US9471668B1 (en) * | 2016-01-21 | 2016-10-18 | International Business Machines Corporation | Question-answering system |
CN107315766A (en) * | 2017-05-16 | 2017-11-03 | 广东电网有限责任公司江门供电局 | A kind of voice response method and its device for gathering intelligence and artificial question and answer |
Cited By (24)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110428267B (en) * | 2019-07-31 | 2023-01-06 | 联想(北京)有限公司 | Content processing method and device, computer system and readable storage medium |
CN110428267A (en) * | 2019-07-31 | 2019-11-08 | 联想(北京)有限公司 | Content processing method and device, computer system and readable storage medium storing program for executing |
CN112417842A (en) * | 2019-08-19 | 2021-02-26 | 阿里巴巴集团控股有限公司 | Service satisfaction determining method and device, storage medium and computer equipment |
CN110991178A (en) * | 2019-11-08 | 2020-04-10 | 苏宁金融科技(南京)有限公司 | Intelligent customer service and artificial customer service switching method and device and computer equipment |
CN110991178B (en) * | 2019-11-08 | 2024-06-28 | 苏宁金融科技(南京)有限公司 | Intelligent customer service and manual customer service switching method and device and computer equipment |
CN111754061A (en) * | 2019-11-22 | 2020-10-09 | 北京沃东天骏信息技术有限公司 | Method and device for controlling man-machine distribution, server equipment and storage medium |
CN111090721A (en) * | 2019-11-25 | 2020-05-01 | 出门问问(苏州)信息科技有限公司 | Question answering method and device and electronic equipment |
CN111090721B (en) * | 2019-11-25 | 2023-09-12 | 出门问问(苏州)信息科技有限公司 | Question answering method and device and electronic equipment |
CN111369080A (en) * | 2020-05-27 | 2020-07-03 | 支付宝(杭州)信息技术有限公司 | Intelligent customer service solution rate prediction method and system and multi-service prediction model |
CN113051386B (en) * | 2021-04-30 | 2024-02-02 | 中国银行股份有限公司 | Adjustment method and device for transferring manual service, electronic equipment and storage medium |
CN113051386A (en) * | 2021-04-30 | 2021-06-29 | 中国银行股份有限公司 | Adjustment method and device for manual service conversion, electronic equipment and storage medium |
CN113282737A (en) * | 2021-07-21 | 2021-08-20 | 中信建投证券股份有限公司 | Man-machine cooperation intelligent customer service dialogue method and device |
CN113282737B (en) * | 2021-07-21 | 2021-11-12 | 中信建投证券股份有限公司 | Man-machine cooperation intelligent customer service dialogue method and device |
CN113793099B (en) * | 2021-09-13 | 2022-07-15 | 深圳众投互联信息技术有限公司 | Intelligent customer service interaction management method and equipment based on deep learning and computer storage medium |
CN113793099A (en) * | 2021-09-13 | 2021-12-14 | 深圳众投互联信息技术有限公司 | Intelligent customer service interaction management method and equipment based on deep learning and computer storage medium |
CN113946673A (en) * | 2021-12-20 | 2022-01-18 | 深圳市一号互联科技有限公司 | Semantic-based intelligent customer service routing processing method and device |
CN114860912A (en) * | 2022-05-20 | 2022-08-05 | 马上消费金融股份有限公司 | Data processing method and device, electronic equipment and storage medium |
CN114860912B (en) * | 2022-05-20 | 2023-08-29 | 马上消费金融股份有限公司 | Data processing method, device, electronic equipment and storage medium |
CN115099727A (en) * | 2022-08-25 | 2022-09-23 | 深圳市星网信通科技有限公司 | Customer service routing method and device and computer readable storage medium |
CN117332066B (en) * | 2023-10-27 | 2024-05-28 | 天讯瑞达通信技术有限公司 | Intelligent agent text processing method based on large model |
CN117332066A (en) * | 2023-10-27 | 2024-01-02 | 天讯瑞达通信技术有限公司 | Intelligent agent text processing method based on large model |
CN118013009A (en) * | 2024-02-19 | 2024-05-10 | 广东融益数字科技有限公司 | Intelligent customer service interaction method and system |
CN118013009B (en) * | 2024-02-19 | 2024-06-18 | 广东融益数字科技有限公司 | Intelligent customer service interaction method and system |
CN118485447A (en) * | 2024-05-30 | 2024-08-13 | 浙江英集动力科技有限公司 | Heat supply digital person and customer service processing method based on heat supply digital person |
Also Published As
Publication number | Publication date |
---|---|
CN110033281B (en) | 2023-07-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110033281A (en) | A kind of method and device that intelligent customer service is converted to artificial customer service | |
CN108363690A (en) | Dialog semantics Intention Anticipation method based on neural network and learning training method | |
CN104699763B (en) | The text similarity gauging system of multiple features fusion | |
CN109558487A (en) | Document Classification Method based on the more attention networks of hierarchy | |
CN109241255A (en) | A kind of intension recognizing method based on deep learning | |
CN106326984A (en) | User intention identification method and device and automatic answering system | |
CN108664632A (en) | A kind of text emotion sorting algorithm based on convolutional neural networks and attention mechanism | |
CN108399158A (en) | Attribute sensibility classification method based on dependency tree and attention mechanism | |
CN103309953B (en) | Method for labeling and searching for diversified pictures based on integration of multiple RBFNN classifiers | |
CN106779087A (en) | A kind of general-purpose machinery learning data analysis platform | |
CN109325231A (en) | A kind of method that multi task model generates term vector | |
CN107451278A (en) | Chinese Text Categorization based on more hidden layer extreme learning machines | |
CN108509573A (en) | Book recommendation method based on matrix decomposition collaborative filtering and system | |
CN108427665A (en) | A kind of text automatic generation method based on LSTM type RNN models | |
CN110377727A (en) | A kind of multi-tag file classification method and device based on multi-task learning | |
CN110990711A (en) | WeChat public number recommendation algorithm and system based on machine learning | |
CN110851593A (en) | Complex value word vector construction method based on position and semantics | |
CN111914553A (en) | Financial information negative subject judgment method based on machine learning | |
CN114443846B (en) | Classification method and device based on multi-level text different composition and electronic equipment | |
CN111783464A (en) | Electric power-oriented domain entity identification method, system and storage medium | |
CN108021565A (en) | A kind of analysis method and device of the user satisfaction based on linguistic level | |
CN101226521A (en) | Machine learning method for ambiguity data object estimation modeling | |
CN116976283A (en) | Language processing method, training method, device, equipment, medium and program product | |
CN114912460A (en) | Method and equipment for identifying transformer fault through refined fitting based on text mining | |
Xiong et al. | L-RBF: A customer churn prediction model based on lasso+ RBF |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |