Specific embodiment
Example embodiments are described in detail here, and the example is illustrated in the accompanying drawings.Following description is related to
When attached drawing, unless otherwise indicated, the same numbers in different drawings indicate the same or similar elements.Following exemplary embodiment
Described in embodiment do not represent all embodiments consistent with this specification.On the contrary, they are only and such as institute
The example of the consistent device and method of some aspects be described in detail in attached claims, this specification.
It is only to be not intended to be limiting this explanation merely for for the purpose of describing particular embodiments in the term that this specification uses
Book.The "an" of used singular, " described " and "the" are also intended to packet in this specification and in the appended claims
Most forms are included, unless the context clearly indicates other meaning.It is also understood that term "and/or" used herein is
Refer to and includes that one or more associated any or all of project listed may combine.
It will be appreciated that though various information may be described using term first, second, third, etc. in this specification, but
These information should not necessarily be limited by these terms.These terms are only used to for same type of information being distinguished from each other out.For example, not taking off
In the case where this specification range, the first information can also be referred to as the second information, and similarly, the second information can also be claimed
For the first information.Depending on context, word as used in this " if " can be construed to " ... when " or
" when ... " or " in response to determination ".
Intelligent customer service establishes a kind of efficiently and effectively communication side based on natural language between enterprise and mass users
Formula is widely applied in various industries at present.At present there are mainly two types of intelligent customer service methods of service, one is contact staff to receive
Corresponding answer is found after to customer problem from knowledge base and returns to user, one is answer user by robot customer service
The problem of.Regardless of method of service, knowledge base is to determine a key factor of intelligent customer service service quality.Customer service knowledge
Library is in customer service scene, the problem of user and the set of corresponding answer.Service knowledge base can be used for contact staff and consult reference,
Either autonomous robot, which uses, the enquirement that technological means directly answer user such as matches, searches for.
Traditional method for establishing service knowledge base, needs staff to remove the FAQ of combing business or product content
(Frequently Asked Questions), and collect from questionnaire investigation, application market comment etc. user feedback, then
Service knowledge base is established according to these problems and answer, this method needs to expend a large amount of human cost.Cloud service at present
Quotient then mainly uses the method (new word discovery, keyword extraction, synonym extension) using keyword as core, such as user's input
One keyword then provides the corresponding answer information of the keyword, and this method generalization ability is poor, and coverage is not high, cannot
User is covered to the various describing modes of problem;Also some technically simple visitors by after customer problem in manual service immediately
Clothes, which are answered, is used as answer, and the knowledge base of this method building, the answer for problem is not accurate enough, influences service quality.
In view of the above-mentioned problems, present description provides a kind of method for building up of service knowledge base, from the service log of history
The middle session log for obtaining client and customer service, and determine the dialogue movement of every words in session log, by customer issue and customer service
Answer extracting comes out, and determines the corresponding relationship of customer issue and customer service answer, construct question and answer pair, by the question and answer to establishing
Service knowledge base.Specifically, the method for building up of the service knowledge base is as shown in Figure 1, comprising the following steps:
S102, the session log that client and customer service are obtained from history service log;
S104, the dialogue movement for determining in every session log every words, and extracted according to dialogue movement described
Customer issue and customer service answer in session log;
S106, the corresponding relationship for determining the customer issue and customer service answer construct question and answer pair according to the corresponding relationship,
To generate service knowledge base.
The history service log of general customer service can all store corresponding database, it is possible to visitor is obtained from database
The session log at family and customer service contains the corresponding Role Information of every session log in session log, for example, the words is visitor
The sentence at family or the sentence of customer service can be determined according to Role Information.After obtaining session log, by pair of each logical dialogue
Words record one sequence of composition, and arranged according to dialog sequence.Then each session log in this logical dialogue is determined
Dialogue movement, wherein dialogue acts the intention or property for describing a word, including client questions, customer service answer, query,
Greeting term etc..After the dialogue movement for determining every session log, it can be acted according to dialogue by the visitor in session log
Family problem and customer service Answer extracting come out, as some otiose information, for example, the sentence of " you are good, thanks " etc then may be used
To reject.
In certain embodiments, the dialogue movement of every session log can be using machine learning model, deep learning
The sequence labelling model of textual classification model and/or deep learning determines.It determines.Wherein, general machine learning model all may be used
For determining dialogue movement, such as SVM (Support Vector Machine) model, Fast Text model, KNN (K-
Nearest Neighbor) model, machine learning model is such as used, every session log is determined by machine learning model
Dialogue movement can both determine that the dialogue of client's sentence and customer service sentence was acted using a machine learning model, can also be with
Determine the dialogue movement of client's sentence and customer service sentence respectively using two machine learning models.It is said due to client and customer service
Meaning difference with a word may be very big, thus determine using one machine learning model effect that dialogue acts compared with
Difference respectively can classify to the sentence of client and customer service using two machine learning models, really to improve classifying quality
The dialogue movement of fixed every record.After the every session log often led in dialogue is arranged according to dialogue sequencing, visitor
The input by sentence at family the input by sentence of customer service to a machine learning model, determines each language to a machine learning model
The dialogue movement of sentence, can be improved efficiency in this way.Certainly, it if expecting more accurate classification results, can also use more multiple
The textual classification model of miscellaneous deep learning or the sequence labelling model of deep learning are dynamic come the dialogue for determining every session log
Make, such as the sequence labelling model of deep learning can be used, by the session log sequence of a whole logical dialogue composition and often
The Role Information of session log is input to jointly in the sequence labelling model of deep learning, the sequence labelling model of deep learning
Based on the contextual information of whole logical dialogue, the dialogue movement of every session log can be judged.Due to the sequence mark of deep learning
Injection molding type can determine the dialogue classification of motion based on the contextual information entirely talked with, and this method is available more accurate
Dialogue classification of motion result.
After the dialogue movement for determining every session log in dialogue, it can be acted according to the dialogue of every session log, it will
Customer issue and customer service Answer extracting come out.Then the corresponding relationship of these customer issues and customer service answer is determined again.Such as
The corresponding answer of customer issue A is customer service answer B, and the corresponding answer of customer issue B is customer service answer C, then according to problem with
The corresponding relationship of answer, constructs question and answer pair, these question and answer are to just constituting service knowledge base.For example, customer issue A is answered with customer service
Case B forms one group of question and answer pair, and customer issue B and customer service answer C form a question and answer pair.Some certain situations, a problem can
Multiple answers can be corresponded to or an answer corresponds to multiple problems.For example, customer issue A and customer service answer B forms one group of question and answer
Right, customer issue A and customer service answer D also form one group of question and answer pair.
In certain embodiments, customer issue and the corresponding relationship of customer service answer can by a question and answer Matching Model come
It determines, the question and answer Matching Model can increase income deep learning frame using open source software pytorch to construct.It can be by client
Problem and customer service answer combination of two are input in question and answer Matching Model, and question and answer Matching Model is according to customer issue and customer service answer
Correlation degree choose confidence point to each pair of combination, then determine that the customer issue is answered with the customer service according to confidence point
The corresponding relationship of case, for example confidence point is higher than the combination of certain score value, then it is assumed that there are question and answer relationship between them, which is
The corresponding answer of the problem, confidence point are lower than certain score value, then it is assumed that question and answer relationship is not present between them.For example,
The customer issue extracted has A, B, and the customer service answer extracted has E, F, then problem and answer combination of two are obtained
To AE, AF, BE, BF, question and answer Matching Model can choose confidence point, such as score value point according to the correlation degree of this four groups of combinations
It Wei 90,40,10,70, it is assumed that confidence point thinks problem more than 60 and answer is corresponding, thus can determine A pairs of problem
The answer answered is E, and the corresponding answer of problem B is F, may be constructed two question and answer to AE, BF.Certainly, since a problem can be with
Corresponding multiple answers, an answer can also correspond to multiple problems, in some cases, in order to reduce the feelings of this multi-to-multi
Condition, as far as possible criterion problem and answer can preset some screening conditions according to the interaction habits of actual use scene, screen out one
A little second-rate problems or answer.Certainly, problem and the corresponding relationship of answer can also be by other with similar functions
Algorithm model determines that this specification is with no restriction.
By above method, the corresponding customer service of different customer issues can be extracted from the session log of history and is answered
Case, constructs service knowledge base, and the building of entire knowledge base can be carried out automatically, without expending a large amount of human costs, and can be covered
Cover the various question formulations of user, coverage is higher, by comprehensive analysis to problem and answer, can accurately determine problem and
The corresponding relationship of answer constructs service knowledge base.
Certainly, the service knowledge base problem of above method building and the describing mode of answer are very comprehensive, but many situations
Under, the quality of some possible customer issues and customer service answer is poor, such as sentence is not clear and coherent, it is inaccurate, use to describe
Describing mode, which less meets public, customer service, answers inaccuracy, client and answers the problems such as evaluation is poor to customer service, at this moment also needs pair
The service knowledge base of building is further optimized, and obtains the knowledge base of high quality.In some embodiments it is possible to question and answer pair
In customer issue or customer service answer do a Screening Treatment, filter out second-rate customer issue or customer service answer.It can be with
It is first scored each customer issue of question and answer centering or each customer service answer using preset language model, it then will scoring
Score value is crossed and is filtered lower than the customer issue of preset value or customer service answer, only retains the higher customer issue of scoring score value or customer service is answered
Case.
In certain embodiments, the scoring score value can be obtained based on the weighted value of specified dimension, the specified dimension
Include: length, problem or the temperature of answer of problem or answer, problem or the continuity of answer, the number of keyword, talking with
The frequency of the frequency occurred in record and/or user's affirmative feedback and negative feedback.It can be by a customer issue or customer service
Answer is input in preset language model, and the language model can be using open source software pytorch open source deep learning frame
Frame constructs.It is that each dimension chooses a score value according to pre-set standards of grading, multiplied by preset each
The corresponding weight of dimension obtains a final score value.For example, it is assumed that customer service answer scoring is needed to consider following
The factor of aspect, i.e., following dimension: the length of answer, the continuity of answer sentence, answer occur in session log
The affirmative feedback and negative feedback number of the number of keyword, client to the answer, the weight difference of each dimension in the frequency, answer
It is 0.1,0.2,0.1,0.4,0.2, the score value of each dimension is respectively 80,90,70,85,80, then final point of the customer service answer
Value are as follows: 0.1 × 80+0.2 × 90+0.1 × 70+0.4 × 85+0.2 × 80=83.So the score value of the customer service answer is 83.When
So, customer issue and the dimensions of customer service answer can go to increase or modify according to actual use scene, and weight can also root
It is flexibly set according to usage scenario, this specification is with no restriction.
It is scored, can be filtered out to each customer issue and each customer service answer by preset some specified dimensions
The poor client questions of quality or customer service answer, promote the quality of service knowledge base.
Since the describing mode and query mode of problem are especially more, the describing mode and answer-mode of an answer
It is especially more, thus the customer issue extracted and affirmative include many customer issues equivalent in meaning and customer service equivalent in meaning
Answer, thus customer issue and customer service answer can be sorted out, the customer issue of expression similar import or customer service answer are classified as
One kind can be convenient the customer service problem in knowledge base and the management of customer service answer in this way.As shown in Fig. 2, can be from customer service knowledge
Question and answer pair are obtained in library, and then the customer issue of question and answer centering or customer service answer are clustered respectively using clustering algorithm, obtain one
A or multiple problem clusters and one or more answer clusters.The problem of customer issue in these problems cluster is all semantic similarity,
Answer in answer cluster is also the answer of semantic similarity.For example, after classifying by clustering algorithm " payment how can will be opened
It is precious? " " how open-minded Alipay is? " " the step of opening Alipay be? " this kind of problems equivalent in meaning or close are classified as one
A problem cluster.Wherein, clustering algorithm can be calculated using K-MEANS algorithm, K-MEDOIDS algorithm, CLARANS algorithm, DBSCAN
Method etc. has the algorithm of similar functions, and this specification is with no restriction.Customer issue and customer service answer are being divided into one or more
After problem cluster and/or answer cluster, the customer issue in each problem cluster, i.e. way to put questions can also be ranked up, and according to sequence
Sequence determines the representative problem in described problem cluster and the optimum answer in the answer cluster.The problem of due to the same meaning
Or there are many kinds of the describing modes of answer, the accuracy of description also has a difference, thus can by these problems and answer according to
The sequence of quality height sorts, and the problem of coming foremost as representing problem, using coming the answer of foremost as most
Good answer.In this way, after receiving the enquirement of client, so that it may which all ways to put questions go to carry out with client questions from problem cluster
Matching, finds and the immediate way to put questions of client questions.Again from answer cluster belonging to the corresponding answer of the way to put questions, selection is most preferably answered
Case feeds back to client.
It in one embodiment, can by the customer service answer sequence in the customer issue and answer cluster in the problem cluster after cluster
To sort according to the scoring score value height of the customer service answer in the customer issue or answer cluster in problem cluster;Either according to visitor
The distance of semantic distribution center's distance of the problem of family problem is with where customer issue cluster and customer service answer and customer service answer
The distance of semantic distribution center's distance where the answer cluster at place sorts.It can be according to problem or the length of answer, problem
Or temperature, problem or the continuity of answer of answer, the number of keyword, the frequency occurred in session log and/or user
Certainly the dimensions such as frequency of feedback and negative feedback are each customer issue or a score value is chosen in customer service answer, and according to point
The height of value sorts to customer issue or customer service answer, score value it is high come front, score value is low to be come below.In addition, will
After problem or answer cluster, each problem cluster or answer cluster can have a semantic distribution center, can also be asked according to each
Topic or answer distance at a distance from the semanteme distribution center sorts, closer at a distance from semantic distribution center, illustrates that this is asked
Therefore this problem or answer can be discharged to by the problem of topic or answer more can represent entire cluster or the meaning to be expressed of answer
More front.
Due to by clustering algorithm by question and answer to the problems in or answer cluster, can there is a problem of that classification is inaccurate, i.e.,
The problem of being not belonging to the same meaning or a cluster is assigned in answer, therefore, can also artificially be asked each after cluster
The answer of the problem of inscribing cluster or answer cluster is checked, and checks whether that classification is accurate.In order to reduce the cost of labor of verification, with
And reduce to the operand to the problems in each cluster or answer sequence, can by after cluster problem or answer carry out it is simple
Merging treatment.One simple merging can be carried out to problem or answer by problem or the literal registration of answer, by table
It states essentially identical some problems or answer and synthesizes a problem or answer, reduce the quantity of problem or answer.For example, " you
It is there what problem? " " what problem you have? " " which problem you have? " " there are also what problems for you? " the literal coincidence of these problems
Degree is very high, and only simple several words are different, and the meaning expressed is essentially identical, therefore these problems can be merged
At a problem, the describing mode for selecting one of them more commonly used as merge after problem, such as " what problem you have? " it can
The problem of by calculating the literal registration between different problems or answer, literal registration is greater than preset value or answer
Merge.Problem and answer are carried out after simply merging according to literal registration, some essentially identical ask of looking like can be given up
Topic or answer reduce the quantity of problem and answer in knowledge base, had both reduced the workload of artificial nucleus couple, decrease and subsequent are
The operand of problem and answer sequence.
In order to be explained further this specification offer service knowledge base method for building up, below in conjunction with Fig. 3 and Fig. 4 with
One specific embodiment is illustrated.
For example some insurance company will establish a service knowledge base about insurance products, so that client insures in consulting
When the case where product, contact staff or intelligent robot search answer from service knowledge base and reply client.Customer service knowledge
Library can construct system by the service knowledge base voluntarily constructed and establish automatically, as shown in figure 3, the service knowledge base is built
Erection system 30 includes session log extraction module 31, dialogue movement determining module 32, dialogue dependency resolution module 33, question and answer
Screening module 34, question and answer cluster module 35, question and answer sorting module 36.Service knowledge base building system establishes the specific mistake of knowledge base
Journey as shown in figure 4, firstly, session log extraction module 31 from store obtained in the database that historied customer service records client with
The session log (S401) of customer service, session log include the Role Information of every session log, to identify this session log
It is the sentence of client or the sentence of customer service.Then, dialogue movement determining module 32 using machine learning model determine every it is right
The dialogue of words record acts (S402), in order to improve efficiency, can using two machine learning models respectively to client's sentence and
Customer service sentence, which engages in the dialogue, acts determination, for example, can determine client using SVM (Support Vector Machine) model
The dialogue of sentence acts, and the movement of customer service sentence is determined using KNN (K-Nearest Neighbor) model.For example, it obtains
The session log taken is as follows, can determine that the dialogue of every a word is acted as shown in bracket by machine learning:
Customer service: you are good!(dialogue movement: polite formula term)
Client: you are good!May I ask A money insurance can protect how long? (dialogue movement: client questions)
Customer service: the insurance of A money can be protected lifelong.(dialogue movement: customer service is answered)
Does is client: this insurance premium how many? (dialogue movement: client questions)
Customer service: annual to hand over 10,000 premiums.(dialogue movement: customer service is answered)
Client: it thanks!(dialogue movement: polite formula term)
Customer service: unfriendly.(dialogue movement: polite formula term)
After dialogue acts the dialogue movement that determining module 32 determines every words in session log, extracted according to dialogue movement
Customer issue and customer service answer (S403) in session log, such as customer issue: " may I ask A money insurance can protect how long? (problem
1) " does is this insurance premium how many? (problem 2) ", customer service answer: " insurance of A money can protect lifelong (answer 1)." " annual to hand over 10,000
Premium (answer 2)." extract customer issue and customer service answer after, dialogue dependency resolution module 33 is by customer issue and visitor
It takes answer combination of two to be input in preset language Matching Model, language Matching Model can be according to the association journey of problem and answer
Degree is that each combine chooses confidence point (S404), such as problem 1- answer 1 (90 points), problem 1- answer 2 (20 points), such as
Problem 2- answer 1 (25 points), such as problem 2- answer 2 (95 points) then select confidence point and are higher than the question and answer that preset value 60 divides
Combination constitutes question and answer to (S405).Such as: problem 1- answer 1 is a question and answer pair, and problem 2- answer 2 is an answer pair.
In order to obtain quality compare intellectual with a senior professional title know library, question and answer screening module 34 by language model to question and answer to the problems in and
Answer is scored, to screen the high problem of quality or answer (S406).The standard of scoring can be preset, for example is needed
The dimension of the scoring of consideration has: the temperature of problem or the length of answer, problem or answer, the continuity of problem or answer, key
The number of word, the frequency occurred in session log and user feed back certainly and the frequency of negative feedback.It presets again every
Then the weighted value of a dimension makes a call to a score value in each dimension to problem or answer and finally may be used multiplied by corresponding weighted value
To obtain a comprehensive score of problem and answer, then by comprehensive score be lower than preset value the problem of or answer filter out,
Obtain the higher problem of mass ratio or answer.After screening obtains the higher problem of quality and answer, question and answer cluster module 35
Clustering algorithm can also be used, if K-MEANS algorithm carries out clustering processing to problem and answer, obtains multiple problem clusters and answer
The answer semanteme of cluster (S407), the problems in these problems cluster or answer cluster is more close, the same meaning of primary expression, also
It is the difference description said be to the same problem, or the difference of the same answer is described.Certainly, using clustering algorithm point
There may be the situations of classification inaccuracy for class, accurate in order to ensure classifying, can also manually to after cluster problem cluster and answer
Cluster carries out verification analysis, and the problem of the same semanteme and answer are assigned to one kind as far as possible.After problem and answer classification, question and answer
Cluster module 35 can also do one simply to the answer in the problems in each problem cluster or answer cluster according to literal registration
Merging treatment (S408).Two problems or answer can be compared, calculate the literal registration of the two, set if registration is higher than
Two problems are then merged into a problem by definite value, can choose scoring higher that problem of score value as asking after merging
Topic.After simple merging treatment problem or answer, question and answer sorting module 36 can the scoring based on each problem or answer
Score value sorts the problems in problem cluster or answer cluster or answer according to score value from high to low, and is then come in On The Choice cluster
One the problem of, as representing problem, comes first answer as optimum answer in answer cluster.It, can after the problem of receiving user
Optimum answer is preferentially fed back to user (S409).Question and answer Matching Model, language model and text in the present embodiment is poly-
Open source software pytorch open source deep learning frame can be used in the building of class model.
The method that service knowledge base provided in this embodiment is established is remembered by the dialogue in comprehensive analysis of history record
Record, can extract the more comprehensive problem of ratio and answer of some industry field, and parse the corresponding relationship of problem and answer, obtain
To accurate question and answer pair, then problem and answer are screened, the higher problem of mass ratio and answer is selected, passes through cluster
Algorithm clusters problem and answer, the problem of semantic similarity or answer is divided into one kind, and to the problems in problem cluster or answer
The answer of cluster is sorted, and the priority of offering question and answer selects the problem of representative and optimum answer.This method can obtain ratio
Relatively comprehensively, the service knowledge base of high quality, provides better service for client.
Corresponding with a kind of embodiment of the method for establishing service knowledge base that this specification provides, this explanation additionally provides one
Kind establishes the device of service knowledge base, as shown in figure 5, described device 50 includes:
Talk with modulus block 51;The session log of client and customer service are obtained from history service log;
Talk with classification of motion module 52;Determine the dialogue movement of every words in every session log, and according to described
Dialogue movement extracts customer issue and customer service answer in the session log;
Talk with dependency resolution module 53;The corresponding relationship for determining the customer issue and customer service answer, according to described
Corresponding relationship constructs question and answer pair, to generate service knowledge base.
In one embodiment, dialogue movement based on machine learning model, the textual classification model of deep learning and/
Or the sequence labelling model of deep learning determines.
In one embodiment, the customer issue is matched with the corresponding relationship of the customer service answer based on preset question and answer
Model determines, specifically includes:
The customer issue and the customer service answer combination of two are input to the question and answer Matching Model, so that described ask
Answering Matching Model is that confidence point is chosen in each pair of combination, wherein the confidence divides the pass for describing customer issue Yu customer service answer
Connection degree;
The corresponding relationship of the customer issue and the customer service answer is determined according to the confidence point.
In one embodiment, question and answer are being constructed to later according to the corresponding relationship, further includes:
It is scored respectively the customer issue and customer service answer of the question and answer centering using preset language model;
Filtering scoring score value is obtained lower than the customer issue of preset threshold or customer service answer with updating the service knowledge base
Question and answer pair after to screening.
In one embodiment, the scoring score value is obtained based on the weighted score of specified dimension, the specified dimension packet
Include: the temperature of problem or the length of answer, problem or answer, the number of keyword, is talking with note at the continuity of problem or answer
The frequency of the frequency occurred in record and/or user's affirmative feedback and negative feedback.
In one embodiment, the method also includes:
Obtain the question and answer pair in service knowledge base, based on semantic similarity by the customer issue of the question and answer centering and
The answer cluster of the problem of customer service answer clusters, and obtains customer issue cluster and customer service answer;
By the customer service answer sequence in the customer issue and the answer cluster in described problem cluster, and determine described problem cluster
In representative problem and the optimum answer in the answer cluster.
In one embodiment, the customer service answer sequence in the customer issue in described problem cluster and the answer cluster is successive
It is determined based on the customer issue or the scoring score value height of the customer service answer;Or
Distance based on the customer issue with semantic distribution center's distance of cluster the problem of the customer issue place, with
And the far and near of semantic distribution center's distance where the answer cluster where the customer service answer and the customer service answer determines.
In one embodiment, the customer issue of the question and answer centering and customer service answer are gathered respectively by clustering algorithm
Class, the problem of obtaining customer issue cluster and customer service answer answer cluster after, further includes:
Literal registration in described problem cluster and the answer cluster is higher than to customer issue or the customer service answer for setting certain value
Merging treatment.
The function of each unit and the realization process of effect are specifically detailed in the above method and correspond to step in above-mentioned apparatus
Realization process, details are not described herein.
For device embodiment, since it corresponds essentially to embodiment of the method, so related place is referring to method reality
Apply the part explanation of example.The apparatus embodiments described above are merely exemplary, wherein described be used as separation unit
The unit of explanation may or may not be physically separated, and component shown as a unit can be or can also be with
It is not physical unit, it can it is in one place, or may be distributed over multiple network units.It can be according to actual
The purpose for needing to select some or all of the modules therein to realize this specification scheme.Those of ordinary skill in the art are not
In the case where making the creative labor, it can understand and implement.
For hardware view, as shown in fig. 6, for a kind of hardware for preloading page device place equipment of this specification
Structure chart, it is real other than processor 601 shown in fig. 6, network interface 604, memory 602 and nonvolatile memory 603
Applying the equipment in example where device usually can also include other hardware, such as be responsible for the forwarding chip of processing message;From hardware
The equipment is also possible to be distributed equipment from structure, may include multiple interface cards, to be reported in hardware view
The extension of text processing.
The nonvolatile memory 603 is stored with for storing executable computer instruction, and processor 604 executes institute
It is performed the steps of when stating computer instruction
The session log of client and customer service are obtained from history service log;
Determine in every session log the dialogue movement of every words, and extracted according to dialogue movement it is described right
Customer issue and customer service answer in words record;
The corresponding relationship for determining the customer issue and customer service answer constructs question and answer pair according to the corresponding relationship, with life
At service knowledge base.
Since all or part of this specification the part that contributes to existing technology or the technical solution can be with
The form of software product embodies, which is stored in a storage medium, including some instructions to
So that a terminal device executes all or part of the steps of each embodiment method of this specification.And storage medium packet above-mentioned
It includes: USB flash disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access
Memory, RAM), the various media that can store program code such as magnetic or disk.
The foregoing is merely the preferred embodiments of this specification, all in this explanation not to limit this specification
Within the spirit and principle of book, any modification, equivalent substitution, improvement and etc. done should be included in the model of this specification protection
Within enclosing.