CN109522395A - Automatic question-answering method and device - Google Patents
Automatic question-answering method and device Download PDFInfo
- Publication number
- CN109522395A CN109522395A CN201811192199.7A CN201811192199A CN109522395A CN 109522395 A CN109522395 A CN 109522395A CN 201811192199 A CN201811192199 A CN 201811192199A CN 109522395 A CN109522395 A CN 109522395A
- Authority
- CN
- China
- Prior art keywords
- semantic
- input problem
- phrase
- input
- vector
- 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.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/01—Customer relationship services
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- General Business, Economics & Management (AREA)
- Health & Medical Sciences (AREA)
- Marketing (AREA)
- Finance (AREA)
- Economics (AREA)
- Development Economics (AREA)
- Accounting & Taxation (AREA)
- Strategic Management (AREA)
- Artificial Intelligence (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Machine Translation (AREA)
Abstract
This disclosure relates to field of artificial intelligence, a kind of automatic question-answering method and device are specifically disclosed, comprising: the semantic vector of building input problem；Semantic integrity judgement is carried out to the input problem according to the semantic vector；If the semanteme of the input problem is imperfect, predict to obtain the omission phrase in the input problem from the corresponding corpus above of the input problem by semantic supplement model；According to the semantic incomplete input problem and the return information for omitting phrase and obtaining the input problem from Q & A database obtained.Using the method for deep learning, it carries out omitting phrase prediction for semantic incomplete input problem, user experience is improved according to the return information for omitting the problem of phrase combines input and obtaining input problem that prediction obtains so as to improve the flexibility of automatic question answering.
Description
Technical field
This disclosure relates to field of artificial intelligence, in particular to a kind of automatic question-answering method and device.
Background technique
Currently, the semantic complete problem that customer service robot can only be proposed for user is answered, and for some
The problem of being omitted according to content above, customer service robot can not then be handled.Such as following dialogue:
Client (user): car owner blocks what material needed
Seat (customer service robot): car owner's card needs xxx material
Client (user): how much is annual fee?
Customer service robot can not according to " how much is annual fee " this be based on above omitted the problem of get corresponding answer
It is multiple.
A kind of settling mode for such issues that be directed in the prior art is: go out the theme of current session by template extraction,
Then default and this theme is transmitted to next section of dialogue.Such as following dialogue:
Does what material Client (user): car owner block need? (template can extract " car owner's card " as theme, and under transmitting
It goes.)
Seat (customer service robot): car owner's card needs xxx material.
Client (user): that tourism card? it (inherits theme above " car owner's card ", customer issue becomes that " car owner blocks that
Tourism card? " become the obstructed problem of syntax)
So this kind of method is based on template and rule, still compare very clumsy, covered scene is fewer, it is difficult to expand
Exhibition.
Therefore, how allowing customer service robot to answer the problem of being omitted based on content above, there are also to be solved.
Summary of the invention
In order to solve the problems, such as present in the relevant technologies, present disclose provides a kind of automatic question-answering method and devices.
A kind of automatic question-answering method, which comprises
Construct the semantic vector of input problem；
Semantic integrity judgement is carried out to the input problem according to the semantic vector；
It is corresponding above from the input problem by semantic supplement model if the semanteme of the input problem is imperfect
Prediction obtains the omission phrase in the input problem in corpus；
It is obtained from Q & A database according to the semantic incomplete input problem and the omission phrase obtained
The return information of the input problem.
A kind of automatic call answering arrangement, comprising:
Semantic vector constructs module, is configured as executing: the semantic vector of building input problem；
Semantic integrity judgment module is configured as executing: carrying out language to the input problem according to the semantic vector
Adopted integrality judgement；
Phrase prediction module is omitted, is configured as executing: if the semanteme of the input problem is imperfect, being mended by semanteme
Mold filling type is predicted to obtain the omission phrase in the input problem from the corresponding corpus above of the input problem；
Return information obtains module, is configured as executing: according to the semantic incomplete input problem and obtained
The return information for omitting phrase and obtaining the input problem from Q & A database.
In one embodiment, the semantic vector building module includes:
Participle unit is configured as executing: segmenting to the input problem；
Part-of-speech tagging unit is configured as executing: carrying out part of speech mark to the word in the input problem according to word segmentation result
Note, with weight corresponding to each word in the determination input problem；
Semantic vector construction unit is configured as executing: according to coding corresponding to the word in the input problem and
The corresponding weight constructs to obtain the semantic vector of the input problem.
In one embodiment, the semantic integrity judgment module includes:
Semantic integrity label prediction unit is configured as executing: using Semantic judgement model according to the semantic vector
Prediction obtains the semantic integrity label of the input problem；
Semantic integrity judging unit is configured as executing: judging the input problem according to the semantic complete tag
Semanteme it is whether complete.
In one embodiment, described device further include:
Building of corpus module above, is configured as executing: corpus and reply the problem of according to before the input problem
Corpus constructs the corresponding corpus above of the input problem.
In one embodiment, the omission phrase prediction module includes:
Vector construction unit is configured as executing: the vector of the building corpus above indicates；
The vector prediction unit for omitting phrase, is configured as executing: using the semantic supplement model according to the semanteme
Vector prediction from the expression of the vector of the corpus above obtains vector corresponding to the omission phrase in the input problem；
Phrase determination unit is omitted, is configured as executing: the omission is determined according to vector corresponding to the omission phrase
Phrase.
In one embodiment, described device further include:
Sample acquisition module is configured as executing: obtaining several incomplete sample problems of semanteme and the sample problem
Sample corpus above, and the province that corresponding sample problem is supplemented according to the sample corpus above of the sample problem
Slightly phrase；
Model training module is configured as executing: by the imperfect sample problem of several semantemes and corresponding
Sample corpus above, the omission phrase supplemented carry out the training of the semantic supplement model；
Module is completed in training, is configured as executing: when the semantic supplement model in semantic incomplete problem to omitting
The prediction of phrase reaches designated precision, completes the training of the semantic supplement model.
In one embodiment, the return information acquisition module includes:
Complete semantic vector construction unit, is configured as executing: according to the semantic incomplete input problem and being obtained
The omission phrase obtained constructs the complete semantic vector of the input problem；
Return information matching unit is configured as executing: being matched from Q & A database by the complete semantic vector
Obtain the return information of the input problem.
A kind of automatic call answering arrangement, comprising:
Processor；
Memory for storage processor executable instruction；
Wherein, the processor is configured to executing above-described automatic question-answering method.
A kind of computer readable storage medium, is stored thereon with computer program, and the computer program is held by processor
Above-described automatic question-answering method is realized when row.
The technical scheme provided by this disclosed embodiment can include the following benefits: using the method for deep learning,
For the problem that semantic incomplete input problem carries out omitting phrase prediction, input is combined according to the omission phrase that prediction obtains
The return information for obtaining input problem improves user experience, especially manually so as to improve the flexibility of automatic question answering
Customer service robot in technical field of intelligence.
It should be understood that the above general description and the following detailed description are merely exemplary, this can not be limited
It is open.
Detailed description of the invention
The drawings herein are incorporated into the specification and forms part of this specification, and shows and meets implementation of the invention
Example, and in specification together principle for explaining the present invention.
Fig. 1 is the schematic diagram of the implementation environment according to involved in the disclosure；
Fig. 2 is a kind of block diagram of server for answering question shown according to an exemplary embodiment；
Fig. 3 is a kind of flow chart of automatic question-answering method shown according to an exemplary embodiment；
Fig. 4 is the flow chart of step S110 in embodiment illustrated in fig. 3；
Fig. 5 is the flow chart of step S130 in embodiment illustrated in fig. 3；
Fig. 6 is the flow chart of step S150 in embodiment illustrated in fig. 3；
Fig. 7 is a kind of flow chart of the automatic question-answering method shown according to another exemplary embodiment；
Fig. 8 is the flow chart of step S170 in Fig. 3 corresponding embodiment；
Fig. 9 is a kind of block diagram of automatic call answering arrangement shown according to an exemplary embodiment；
Figure 10 is a kind of block diagram of the automatic call answering arrangement shown according to another exemplary embodiment.
Specific embodiment
Here will the description is performed on the exemplary embodiment in detail, 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 consistented with the present invention.On the contrary, they be only with it is such as appended
The example of device and method being described in detail in claims, some aspects of the invention are consistent.
Fig. 1 is the schematic diagram of the implementation environment according to involved in the disclosure.The implementation environment includes: server for answering question 200
With at least one terminal 100.
Wherein terminal 100 can be smart phone, tablet computer, laptop, desktop computer etc. and can take with question and answer
Business device establishes network connection and can run the electronic equipment of client, herein without specifically limiting.Terminal 100 and question and answer take
Wireless or wired network connection has been pre-established between business device 200, thus, pass through the client run on the terminal 100
Realize that terminal 100 is interacted with server for answering question 200.
Based on the interaction between server for answering question 200 and terminal 100, server for answering question 200 can get user and exist
Then the problem of inputting in terminal 100 carries out semantic vector building for this problem, semantic integrity judges, it is pre- to omit phrase
Survey and match return information etc..Terminal 100 can receive the matched return information of server for answering question institute, and be in by return information
User is now given, to realize that the problem of inputting automatically to user is answered.
Fig. 2 is a kind of block diagram of server for answering question shown according to an exemplary embodiment.Clothes with this hardware configuration
Business device can be used for carrying out automatic question answering and being deployed in implementation environment shown in FIG. 1.
It should be noted that the server for answering question is the example for adapting to the disclosure, must not believe that there is provided
To any restrictions of disclosure use scope.The server for answering question can not be construed to the Fig. 2 that needs to rely on or must have
Shown in one or more component in illustrative server for answering question 200.
The hardware configuration of the server for answering question can generate biggish difference due to the difference of configuration or performance, such as Fig. 2 institute
Show, server for answering question 200 includes: power supply 210, interface 230, at least a memory 250 and an at least central processing unit
(CPU,Central Processing Units)270。
Wherein, power supply 210 is used to provide operating voltage for each hardware device on server for answering question 200.
Interface 230 includes an at least wired or wireless network interface 231, at least a string and translation interface 233, at least one defeated
Enter output interface 235 and at least USB interface 237 etc., is used for and external device communication.
The carrier that memory 250 is stored as resource, can be read-only memory, random access memory, disk or CD
Deng the resource stored thereon includes operating system 251, application program 253 and data 255 etc., and storage mode can be of short duration
It stores or permanently stores.Wherein, operating system 251 be used to manage and control each hardware device on server for answering question 200 with
And application program 253 can be Windows to realize calculating and processing of the central processing unit 270 to mass data 255
ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM etc..Application program 253 be based on operating system 251 it
The upper computer program for completing at least one particular job, may include an at least module (being not shown in Fig. 2), each module
The series of computation machine readable instruction to server for answering question 200 can be separately included.Data 255 can be stored in disk
In Q & A database etc..
Central processing unit 270 may include the processor of one or more or more, and be set as through bus and memory
250 communications, for the mass data 255 in operation and processing memory 250.
As described in detail above, the server for answering question 200 for being applicable in the disclosure will be read by central processing unit 270 to be deposited
The form of the series of computation machine readable instruction stored in reservoir 250 completes automatic question-answering method.
In the exemplary embodiment, server for answering question 200 can be by one or more application specific integrated circuit
At (Application Specific Integrated Circuit, abbreviation ASIC), digital signal processor, digital signal
Manage equipment, programmable logic device, field programmable gate array, controller, microcontroller, microprocessor or other electronic components
It realizes, for executing following methods.Therefore, realize that the present invention is not limited to any specific hardware circuit, software and the two
Combination.
Fig. 3 is a kind of flow chart of automatic question-answering method shown according to an exemplary embodiment.The automatic question-answering method
It can be adapted in the server for answering question 200 of implementation environment shown in Fig. 1, executed by server for answering question 200, may include following step
It is rapid:
Step S110 constructs the semantic vector of input problem.
User inputs problem on the terminal 100, and the input that then server for answering question 200 obtains user from terminal 100 is asked
Topic, and for the semantic vector of acquired input problem building input problem.
Wherein input problem is the problem of user is inputted on the terminal 100, for different application scenarios, inputs problem
Can be aiming at the problem that different field, such as in the application scenarios of bank, input problem can be about bank card, storage,
The problem of the correlation such as loan, interest；In the application scenarios of insurance, input problem, which can be, the phases such as handles about insurance, settles a claim
The problem of pass, naturally it is also possible to be the comprehensive problem of multiple fields, be student such as user, input problem can be pass
In the comprehensive encyclopaedic knowledge the problem of.Herein without specifically limiting.
Wherein, it is worth noting that, scene of the automatic question-answering method of the disclosure especially suitable for more taking turns question and answer.
Semantic vector refers to by constructed by the corresponding coding of word each in input problem and the corresponding weight of each word
Vector, which is used to indicate the semanteme of input problem.In a sentence, each word is to the semantic of sentence in sentence
Percentage contribution is different.Such as in " handle safety car owner and block which material needed " this problem, " handling " " car owner's card " " material
Material " is bigger to the semantic percentage contribution of entire problem, so weight corresponding to three words is bigger, and " safety ", " needs "
" which " these words are slightly smaller to the semantic percentage contribution of entire problem, so the corresponding weight of those words is smaller.
For a text processing system, a text database can be constructed.Such as it can be constructed in disclosed method
The return information (answer) of one Q & A database, the problem of may being asked which includes several users and problem.Certainly, should
Q & A database is for constructed by scene applied by automatic question answering, such as the application scenarios of insurance, the question and answer number
According to library the problem of included and the return information of problem is for insurance related information.So for different application scenarios,
The problem of including in question and answer data has differences.
It include a dictionary in Q & A database for these problems and return information, which is used for these problems
With the word in return information encoded and problem or return information in each word configure corresponding weight.To connect
, can be according to the dictionary of Q & A database after the input problem for having received user, the coding of word and input are asked in input problem
The semantic vector of the weight building input problem of each word in topic.
In a particular embodiment, each word is corresponding is encoded to real number, i.e., indicates the word in sentence, word institute with number
Corresponding weight is also real number, thus, it is finally a real vector for semantic vector constructed by input problem.
In one exemplary embodiment, as shown in figure 4, step S110 includes:
Step S111 segments input problem.
Wherein, participle refers to the word for being divided into several to be arranged in order input problem, such as " handles safety car owner
Which material card needs " it can segment as " handle ^ safety ^ car owner and block ^ needs which ^ material of ^ ".
Participle can using segmentation methods carry out, such as can using based on string matching segmenting method, based on reason
The segmenting method of solution or segmenting method scheduling algorithm based on statistics, herein without limiting.
Step S112 carries out part-of-speech tagging to the word in input problem according to word segmentation result, to determine in input problem
Weight corresponding to each word.
Part-of-speech tagging refers to the syntactic position according to each word in input problem and the mark of attribute progress, such as locates
Subject position word, be in the word of predicate position, be in the word of object position；Such as noun (refer to name word, refer to name word,
Time noun, place noun etc.), verb, the parts of speech such as negative word, it is certainly above be only it is exemplary enumerate, be not construed as pair
The limitation of disclosure use scope.
Dictionary in Q & A database mentioned above, to the word weight in each problem or answer in the dictionary
Configuration be according to this problem or in answer in syntactic position and the part of speech of the word configured.To basis
Part-of-speech tagging as a result, weight corresponding to each word in input problem can be determined.
Step S113 is constructed to obtain to input and be asked according to coding corresponding to the word in input problem and corresponding weight
The semantic vector of topic.
Obtaining word segmentation result, so as to according to word segmentation result from the dictionary of Q & A database the corresponding coding of middle word
And the semantic vector of the corresponding weight building input problem of word.
Step S130 carries out semantic integrity judgement to input problem according to semantic vector.
Semantic integrity judgement judges whether the semanteme of the input problem is complete by semantic vector.It is wherein semantic endless
Whole input problem include for the input problem omitted above, such as subject omissions, object deletion, attribute omission and
Other sentence elements omission etc. is answering " vehicle for example, if having asked " car owner blocks what material needed " before the input problem
Main card needs ××× material " after, user is also possible to ask " that tourism card? " certainly, with regard to this simple problem " that tourism
Card ", the semanteme of the problem is incomplete, if corpus information not above, server for answering question is can not to be asked according to this
What topic was replied.
In one exemplary embodiment, can be sentenced by the semantic integrity that the method for deep learning carries out input problem
It is disconnected, i.e., Semantic judgement model is constructed using neural network, and the training of Semantic judgement model is carried out using several sample problems, from
And Semantic judgement model can be for the semantic whether complete of an input Problem judgment input problem after the completion of training.Tool
Body, it may include steps of as shown in figure 5, Semantic judgement model carries out semantic integrity judgement:
Step S131 predicts to obtain the semantic integrity mark of input problem according to semantic vector using Semantic judgement model
Label.
Step S132 judges whether the semanteme of input problem is complete according to semantic complete tag.
Wherein semantic integrity label includes for semantic complete label " complete " and for semantic incomplete label
" imperfect ".In a specific embodiment, semantic integrity label can be is embodied by encoding, such as is indicated with number 0
Semantic incomplete label " complete " indicates semantic complete label " imperfect " with number 1.So as to according to Semantic judgement
The semantic integrity label of model output judge whether the semanteme of input problem is complete.
In one embodiment, it before carrying out semantic integrity judgement using Semantic judgement model, needs to semanteme
Judgment models are trained, and are specifically included:
It obtains several sample problems and the semantic integrity of sample problem is marked.Wherein sample problem includes semantic complete
Whole sample problem and semantic incomplete sample problem, such as the complete semantic sample problem that do not omit, are saved based on above
The semantic sample problem of slightly imperfect.Corresponding, the semantic integrity mark to sample problem includes for semantic endless
The mark of whole sample problem and mark for semantic full sample problem.
The training of Semantic judgement model is carried out according to acquired sample problem and corresponding mark.I.e. by sample problem and
Corresponding mark is input in Semantic judgement model, and Semantic judgement model can export a semanteme for each sample input problem
Integrality label, if semantic integrity corresponding to sample problem indicated by the semantic integrity label exported and mark institute
The semantic integrity of instruction is different, then adjusts the parameter of Semantic judgement model, until semantic integrity indicated by the two is identical.
Semantic judgement model after the completion of training can be used in the disclosure inputting the semantic integrity judgement of problem.Having
In body embodiment, whether Semantic judgement model trains completion, can be determined by testing, i.e., is different from used in training with several
Sample problem tested, if Semantic judgement model is after training directed to semanteme corresponding to test sample problem used
The accuracy rate of integrality judgement reaches permissible accuracy, then can stop the training of Semantic judgement model, i.e. Semantic judgement model
Training is completed.If accuracy rate does not reach permissible accuracy, continue the training of Semantic judgement model.Certainly, it arranges
In addition to the influence of the model structure of Semantic judgement model, the sample problem of training is more, then the standard of Semantic judgement model prediction
True rate is higher.
The semantic integrity judgement that input problem is carried out by constructed Semantic judgement model, improves automatic question answering
Efficiency, moreover, semantic integrity judgement can be carried out aiming at the problem that different type after Semantic judgement model training,
Improve treatment effeciency.
In one exemplary embodiment, if judging that the semanteme of input problem is complete according to Semantic judgement model, root
The return information of input problem is matched from Q & A database according to the semantic vector of input problem.
Step S150, if input problem semanteme it is imperfect, by semantic supplement model from input problem it is corresponding on
Prediction obtains the omission phrase in input problem in literary corpus.
Wherein semantic supplement model is carried out that is, by the way of deep learning defeated by model constructed by neural network
Enter to omit the prediction of phrase in problem, so as to avoid carrying out the automatic question answering mode of theme transmitting, Ke Yiti according to template
The flexibility of high automatic question answering.
In one exemplary embodiment, before executing step S150, further includes:
The problem of according to before input problem corpus with reply the corresponding corpus above of corpus building input problem.
As indicated above, the automatic question-answering method of the disclosure is suitable for carrying out the scenes for taking turns question and answer more, is taking turns more
In question and answer scene, if there is omitting in input problem, generally be directed to being omitted above, so being directed to an input problem has pair
The corpus above answered.The corpus above is the problem of proposition by user before input problem and server for answering question is according to user
The problem of before the problem of return information that institute's proposition problem is made is constituted, and user proposes before input problem, that is, input problem language
Material, the reply corpus before the return information i.e. input problem that server for answering question is made according to the proposed problem of user.
In one embodiment, under conditions of guaranteeing the precision of prediction to the omission phrase in input problem, in order to
Reduce the operand of semantic supplement model, general upper two problems and the corresponding reply of upper two problems for utilizing input problem
Construct the corresponding corpus above of input problem.Thus after the return information for being matched to input problem in Q & A database,
Corpus above is updated, i.e., will reply the input problem completed and corresponding return information adds in corpus above, and
The problem of distance input problem proposes earliest in corpus above and corresponding return information, are removed from corpus above,
Using updated corpus above as the corpus above of next input problem.
In one embodiment, use deep learning model carry out omit phrase prediction can be as shown in fig. 6, step S150
Include:
Step S151, the vector for constructing corpus above indicate.
As described above, the problem of corpus is by before input problem above corpus and reply corpus are constituted.Such as follows
Dialogue in:
Client (user): car owner blocks what material needed
Seat (customer service robot): car owner's card needs ××× material
Client (user): how much is annual fee
It is so directed to " how much is annual fee " of user this problem, the corpus above of the problem can be by " it is assorted that car owner blocks needs
Material " and " car owner's card needs ××× material " this basket corpus and reply corpus are constituted.Certainly in more wheels
In question and answer, the corpus above for inputting problem can be by multiple groups question and answer corpus (problem corpus and the reply language before input problem
Material) it constitutes.
Using mode identical with the building semantic vector of input problem, the vector for constructing corpus above is indicated i.e. to upper
The problems in literary corpus corpus and reply corpus construct complete semantic vector.
Step S152, using semantic supplement model according to semantic vector from the vector of corpus above expression in predict to obtain
Vector corresponding to omission phrase in input problem.
Semantic vector of the semantic supplement model according to input problem, the thinking understood using reading, by the language for the problem that inputs
Adopted vector combines with the expression of the vector of corpus above, related to the word in input problem from being matched in corpus above
The part of connection, so that prediction obtains omitting vector corresponding to phrase in input problem.
In one exemplary embodiment, semantic supplement model can use R-net model.Wherein R-net is mentioned by Microsoft
The neural network model that understanding is read for machine out.In the disclosure, R-net model is creatively used for automatic question answering
The prediction of omission phrase in the process, so that the flexibility and efficiency of automatic question answering are improved, so that customer service robot can be to base
It is replied in the input problem omitted above.
R-net model includes four layers, and wherein tier I is used to construct the semantic vector of input problem and corpus above, i.e.,
The semantic vector of input problem mentioned above and the vector of corpus above indicate；Tier ii is used to input the semanteme of problem
Vector is compared with the expression of the vector of corpus above, to find out or phase associated with input problem in corpus above
As phrase；The associated or similar phrase that layer III compares tier ii is placed on above by attention mechanism
It is compared in corpus, so that navigating to may be several phrases for omitting phrase；If Section IV layer is according to being navigated to
Vector corresponding to dry phrase carries out classification prediction, i.e. classification prediction obtains vector corresponding to each phrase in several phrases
For the probability for omitting vector corresponding to phrase, then the probability of vector corresponding to each phrase in several phrases is compared
Compared with using vector corresponding to the phrase of maximum probability as vector corresponding to omission phrase.
Step S153 determines omission phrase according to vector corresponding to phrase is omitted.
Certainly, semantic supplement model also needs to be trained semantic supplement model before being used to omit the prediction of phrase,
So in one exemplary embodiment, as shown in fig. 7, before step S150, further includes:
Step S011 obtains the sample corpus above of the incomplete sample problem of several semantemes and sample problem, and
The omission phrase that corresponding sample problem is supplemented according to the sample of sample problem corpus above.
Step S012 by the imperfect sample problem of several semantemes and corresponding sample corpus above, is supplemented
Omit phrase carry out semantic supplement model training.
Wherein the training process in step S012 includes:
Each sample problem and corresponding sample corpus above are input in semantic supplement model, prediction obtains sample
Omission phrase in this problem.
According to the omission phrase supplemented sample problem, the model parameter of semantic supplement model is adjusted, until measuring in advance
To omission phrase it is identical as the omission phrase supplemented.
Step S013, when semantic supplement model reaches specified essence to the prediction for omitting phrase in semantic incomplete problem
Degree completes the training of semantic supplement model.
Step S170 is obtained from Q & A database according to semantic incomplete input problem and omission phrase obtained
The return information of input problem.
Some problem corpus is stored in Q & A database and replies corpus, according to input problem and omission obtained
Phrase can correspond to the complete semanteme for the problem of determining user is inputted, and then according to complete and match from Q & A database
Corpus is replied as the return information to input problem.In a particular embodiment, if what the creation of Q & A database can collect
Dry problem and corresponding answer to create, or collection problem on the basis of, problem is converted to obtain multiple open up
Exhibition problem, because user is different, the mode putd question to for same problem is also different.Certainly, for disclosure automatic question answering
The problem of field of system application is different, collected and corresponding answer be not also identical, for example, in the application scenarios of insurance
In, the problems in Q & A database corpus and reply corpus for insuring related service, such as the classification of insurance, risk,
Claims Resolution, expense of insurance of insurance etc..
In one exemplary embodiment, as shown in figure 8, step S170 includes:
Step S171 constructs the complete of input problem according to semantic incomplete input problem and the phrase obtained that omits
Semantic vector.
Step S172 is matched from Q & A database by complete semantic vector and is obtained the return information of input problem.
By the technical solution of the disclosure, by the way of deep learning, the omission phrase in input problem is divided
Class prediction, so as to understand the complete semanteme of input problem according to the omission phrase that prediction obtains during automatic question answering,
It is replied, the flexibility and efficiency of automatic question answering is improved, so as to improve user experience.
Following is embodiment of the present disclosure, can be used for executing the automatic of the above-mentioned execution of server for answering question 200 of the disclosure
Answering method embodiment.For those undisclosed details in the apparatus embodiments, it is real to please refer to disclosure automatic question-answering method
Apply example.
A kind of automatic call answering arrangement, as shown in Figure 9, comprising:
Semantic vector constructs module 110, is configured as executing: the semantic vector of building input problem.
Semantic integrity judgment module 130, the module connect with semantic vector building module 110, are configured as executing: root
Semantic integrity judgement is carried out to input problem according to semantic vector.
Phrase prediction module 150 is omitted, which connect with semantic integrity judgment module 130, be configured as executing: such as
The semanteme that fruit inputs problem is imperfect, predicts to obtain from the corresponding corpus above of input problem by semantic supplement model defeated
Enter the omission phrase in problem.
Return information obtains module 170, which connect with phrase prediction module 150 is omitted, be configured as executing: according to
Semantic incomplete input problem and the return information obtained for omitting phrase and obtaining input problem from Q & A database.
The automatic call answering arrangement of the disclosure can be applied to the customer service robot of different application scene, ask automatically user
Topic is replied, and so as to improve the flexibility of customer service robot, improves user experience.
In one embodiment, semantic vector building module 110 includes:
Participle unit is configured as executing: segmenting to input problem.
Part-of-speech tagging unit is configured as executing: part-of-speech tagging is carried out to the word in input problem according to word segmentation result, with
Determine weight corresponding to each word in input problem.
Semantic vector construction unit is configured as executing: right according to coding corresponding to the word in input problem and institute
The weight answered constructs to obtain the semantic vector of input problem.
In one embodiment, semantic integrity judgment module 130 includes:
Semantic integrity label prediction unit is configured as executing: being predicted using Semantic judgement model according to semantic vector
Obtain the semantic integrity label of input problem.
Semantic integrity judging unit is configured as executing: the semanteme for judging input problem according to semantic complete tag is
It is no complete.
In one embodiment, automatic call answering arrangement further include:
Building of corpus module above, is configured as executing: corpus and replying corpus the problem of according to before input problem
Construct the corresponding corpus above of input problem.
In one embodiment, omitting phrase prediction module 150 includes:
Vector construction unit is configured as executing: the vector for constructing corpus above indicates.
The vector prediction unit for omitting phrase, is configured as executing: using semantic supplement model according to semantic vector from upper
Prediction obtains vector corresponding to the omission phrase in input problem in the vector expression of literary corpus.
Phrase determination unit is omitted, is configured as executing: determining omission phrase according to vector corresponding to phrase is omitted.
In one embodiment, automatic call answering arrangement further include:
Sample acquisition module is configured as executing: obtaining the sample of the incomplete sample problem of several semantemes and sample problem
This corpus above, and the omission phrase that corresponding sample problem is supplemented according to the sample corpus above of sample problem.
Model training module is configured as executing: passing through the imperfect sample problem of several semantemes and corresponding sample
The training of corpus, the omission phrase progress semantic supplement model supplemented above.
Module is completed in training, is configured as executing: when semantic supplement model is to omitting phrase in semantic incomplete problem
Prediction reach designated precision, complete the training of semantic supplement model.
In one embodiment, return information acquisition module 170 includes:
Complete semantic vector construction unit, is configured as executing: according to semantic incomplete input problem and obtained
Omit the complete semantic vector of phrase building input problem.
Return information matching unit is configured as executing: being matched and is obtained from Q & A database by complete semantic vector
The return information of input problem.
Modules/unit function and the realization process of effect are specifically detailed in above-mentioned automatic question-answering method in above-mentioned apparatus
The realization process of middle corresponding step, details are not described herein.
It is appreciated that these modules can by hardware, software, or a combination of both realize.When realizing in hardware
When, these modules may be embodied as one or more hardware modules, such as one or more specific integrated circuits.When with software side
When formula is realized, these modules may be embodied as the one or more computer programs executed on the one or more processors, example
The program being stored in as performed by the central processing unit 270 of Fig. 2 in memory 250.
Optionally, the disclosure also provides a kind of automatic call answering arrangement, which can be used for implementing shown in Fig. 1
In the server for answering question 200 of environment, for executing all or part of step of any of the above automatic question-answering method embodiment.Such as
Shown in Figure 10, which includes:
Processor 1001, and
Memory 1002 for 1001 executable instruction of storage processor.
Wherein, processor 1001 is configured as executing the method in the above automatic question-answering method embodiment.Executable instruction
It can be computer-readable instruction, processor 1001 is read from memory 1002 at work, by bus/data line 1003
Computer-readable instruction executes all or part of step in any of the above automatic question-answering method embodiment.
Modules/unit function and the realization process of effect are specifically detailed in above-mentioned automatic question-answering method in above-mentioned apparatus
The realization process of middle corresponding step, details are not described herein.
In the exemplary embodiment, a kind of computer readable storage medium is additionally provided, computer program is stored thereon with,
The automatic question-answering method in any of the above embodiment is realized when computer program is executed by processor.The computer-readable storage medium
Memory 250 of the matter for example including instruction, above-metioned instruction can be executed by the central processing unit 270 of server for answering question 200 to complete
Above-mentioned automatic question-answering method.
It should be understood that the present invention is not limited to the precise structure already described above and shown in the accompanying drawings, and
And various modifications and change can executed without departing from the scope.The scope of the present invention is limited only by the attached claims.
Claims (10)
1. a kind of automatic question-answering method, which is characterized in that the described method includes:
Construct the semantic vector of input problem；
Semantic integrity judgement is carried out to the input problem according to the semantic vector；
If the semanteme of the input problem is imperfect, by semantic supplement model from the corresponding corpus above of the input problem
Prediction obtains the omission phrase in the input problem in library；
Described in being obtained from Q & A database according to the semantic incomplete input problem and the omission phrase obtained
The return information of input problem.
2. the method according to claim 1, wherein the semantic vector of the building input problem, comprising:
The input problem is segmented；
Part-of-speech tagging is carried out to the word in the input problem according to word segmentation result, with each word in the determination input problem
Corresponding weight；
It constructs to obtain the input according to coding corresponding to the word in the input problem and the corresponding weight to ask
The semantic vector of topic.
3. the method according to claim 1, wherein it is described according to the semantic vector to the input problem into
The judgement of row semantic integrity, comprising:
It is predicted to obtain the semantic integrity label of the input problem according to the semantic vector using Semantic judgement model；
Judge whether the semanteme of the input problem is complete according to the semantic complete tag.
4. the method according to claim 1, wherein it is described by semantic supplement model from the input problem pair
Before prediction obtains the omission phrase in the input problem in the corpus above answered, further includes:
The problem of according to before the input problem, corpus with corpus is replied constructed the corresponding corpus above of the input problem.
5. the method according to claim 1, wherein it is described by semantic supplement model from the input problem pair
Prediction obtains the omission phrase in the input problem in the corpus above answered, comprising:
The vector for constructing the corpus above indicates；
Using the semantic supplement model according to the semantic vector from the vector of the corpus above expression in predict to obtain
Vector corresponding to omission phrase in the input problem；
The omission phrase is determined according to vector corresponding to the omission phrase.
6. the method according to claim 1, wherein it is described by semantic supplement model from the input problem pair
Before prediction obtains the omission phrase in the input problem in the corpus above answered, further includes:
The sample corpus above of the incomplete sample problem of several semantemes and the sample problem is obtained, and according to the sample
The omission phrase that the sample of this problem corpus above supplements corresponding sample problem；
Pass through several imperfect sample problems of semanteme and corresponding sample corpus above, the omission phrase supplemented
Carry out the training of the semantic supplement model；
When the semantic supplement model reaches designated precision to the prediction for omitting phrase in semantic incomplete problem, described in completion
The training of semantic supplement model.
7. the method according to claim 1, wherein described according to the semantic incomplete input problem and institute
The omission phrase obtained obtains the return information of the input problem from Q & A database, comprising:
The complete of the input problem is constructed according to the semantic incomplete input problem and the omission phrase obtained
Semantic vector；
It is matched from Q & A database by the complete semantic vector and obtains the return information of the input problem.
8. a kind of automatic call answering arrangement characterized by comprising
Semantic vector constructs module, is configured as executing: the semantic vector of building input problem；
Semantic integrity judgment module is configured as executing: being carried out according to the semantic vector to the input problem semantic complete
Whole property judgement；
Phrase prediction module is omitted, is configured as executing: if the semanteme of the input problem is imperfect, passing through semantic supplement mould
Type is predicted to obtain the omission phrase in the input problem from the corresponding corpus above of the input problem；
Return information obtains module, is configured as executing: according to the semantic incomplete input problem and obtained described
Omit the return information that phrase obtains the input problem from Q & A database.
9. a kind of automatic call answering arrangement characterized by comprising
Processor；
Memory for storage processor executable instruction；
Wherein, the processor is configured to method described in any one of perform claim requirement 1 to 7.
10. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the computer program
The method as described in any one of claims 1 to 7 is realized when being executed by processor.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811192199.7A CN109522395A (en) | 2018-10-12 | 2018-10-12 | Automatic question-answering method and device |
| PCT/CN2018/125252 WO2020073533A1 (en) | 2018-10-12 | 2018-12-29 | Automatic question answering method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811192199.7A CN109522395A (en) | 2018-10-12 | 2018-10-12 | Automatic question-answering method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109522395A true CN109522395A (en) | 2019-03-26 |
Family
ID=65772381
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811192199.7A Withdrawn CN109522395A (en) | 2018-10-12 | 2018-10-12 | Automatic question-answering method and device |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN109522395A (en) |
| WO (1) | WO2020073533A1 (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110457450A (en) * | 2019-07-05 | 2019-11-15 | 平安科技（深圳）有限公司 | Answer generation method and relevant device based on neural network model |
| CN110717023A (en) * | 2019-09-18 | 2020-01-21 | 平安科技（深圳）有限公司 | Method and device for classifying interview answer texts, electronic equipment and storage medium |
| CN110956962A (en) * | 2019-10-17 | 2020-04-03 | 中国第一汽车股份有限公司 | Reply information determination method, device and equipment for vehicle-mounted robot |
| CN111078853A (en) * | 2019-12-13 | 2020-04-28 | 上海智臻智能网络科技股份有限公司 | Question-answer model optimization method and device, computer equipment and storage medium |
| CN111104503A (en) * | 2019-12-24 | 2020-05-05 | 华中科技大学 | Construction engineering quality acceptance standard question-answering system and construction method thereof |
| WO2020242383A1 (en) * | 2019-05-28 | 2020-12-03 | Active Intelligence Pte Ltd | Conversational diaglogue system and method |
| CN112035636A (en) * | 2020-08-28 | 2020-12-04 | 康键信息技术（深圳）有限公司 | Question-answer management method, device, equipment and storage medium of medical inquiry system |
| CN112183114A (en) * | 2020-08-10 | 2021-01-05 | 招联消费金融有限公司 | Model training and semantic integrity recognition method and device |
| CN113282733A (en) * | 2021-06-11 | 2021-08-20 | 上海寻梦信息技术有限公司 | Customer service problem matching method, system, device and storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105589844A (en) * | 2015-12-18 | 2016-05-18 | 北京中科汇联科技股份有限公司 | Missing semantic supplementing method for multi-round question-answering system |
| CN107632979A (en) * | 2017-10-13 | 2018-01-26 | 华中科技大学 | The problem of one kind is used for interactive question and answer analytic method and system |
| CN107798140A (en) * | 2017-11-23 | 2018-03-13 | 北京神州泰岳软件股份有限公司 | A kind of conversational system construction method, semantic controlled answer method and device |
| CN108334487A (en) * | 2017-07-14 | 2018-07-27 | 腾讯科技（深圳）有限公司 | Lack semantics information complementing method, device, computer equipment and storage medium |
-
2018
- 2018-10-12 CN CN201811192199.7A patent/CN109522395A/en not_active Withdrawn
- 2018-12-29 WO PCT/CN2018/125252 patent/WO2020073533A1/en active Application Filing
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105589844A (en) * | 2015-12-18 | 2016-05-18 | 北京中科汇联科技股份有限公司 | Missing semantic supplementing method for multi-round question-answering system |
| CN108334487A (en) * | 2017-07-14 | 2018-07-27 | 腾讯科技（深圳）有限公司 | Lack semantics information complementing method, device, computer equipment and storage medium |
| CN107632979A (en) * | 2017-10-13 | 2018-01-26 | 华中科技大学 | The problem of one kind is used for interactive question and answer analytic method and system |
| CN107798140A (en) * | 2017-11-23 | 2018-03-13 | 北京神州泰岳软件股份有限公司 | A kind of conversational system construction method, semantic controlled answer method and device |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020242383A1 (en) * | 2019-05-28 | 2020-12-03 | Active Intelligence Pte Ltd | Conversational diaglogue system and method |
| CN110457450A (en) * | 2019-07-05 | 2019-11-15 | 平安科技（深圳）有限公司 | Answer generation method and relevant device based on neural network model |
| CN110457450B (en) * | 2019-07-05 | 2023-12-22 | 平安科技（深圳）有限公司 | Answer generation method based on neural network model and related equipment |
| CN110717023A (en) * | 2019-09-18 | 2020-01-21 | 平安科技（深圳）有限公司 | Method and device for classifying interview answer texts, electronic equipment and storage medium |
| CN110717023B (en) * | 2019-09-18 | 2023-11-07 | 平安科技（深圳）有限公司 | Method and device for classifying interview answer text, electronic equipment and storage medium |
| CN110956962A (en) * | 2019-10-17 | 2020-04-03 | 中国第一汽车股份有限公司 | Reply information determination method, device and equipment for vehicle-mounted robot |
| CN111078853B (en) * | 2019-12-13 | 2023-05-02 | 上海智臻智能网络科技股份有限公司 | Question-answering model optimization method, device, computer equipment and storage medium |
| CN111078853A (en) * | 2019-12-13 | 2020-04-28 | 上海智臻智能网络科技股份有限公司 | Question-answer model optimization method and device, computer equipment and storage medium |
| CN111104503A (en) * | 2019-12-24 | 2020-05-05 | 华中科技大学 | Construction engineering quality acceptance standard question-answering system and construction method thereof |
| CN112183114A (en) * | 2020-08-10 | 2021-01-05 | 招联消费金融有限公司 | Model training and semantic integrity recognition method and device |
| CN112183114B (en) * | 2020-08-10 | 2024-05-14 | 招联消费金融股份有限公司 | Model training and semantic integrity recognition method and device |
| CN112035636A (en) * | 2020-08-28 | 2020-12-04 | 康键信息技术（深圳）有限公司 | Question-answer management method, device, equipment and storage medium of medical inquiry system |
| CN113282733A (en) * | 2021-06-11 | 2021-08-20 | 上海寻梦信息技术有限公司 | Customer service problem matching method, system, device and storage medium |
| CN113282733B (en) * | 2021-06-11 | 2024-04-09 | 上海寻梦信息技术有限公司 | Customer service problem matching method, system, equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| WO2020073533A1 (en) | 2020-04-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109522395A (en) | Automatic question-answering method and device | |
| WO2022007823A1 (en) | Text data processing method and device | |
| CN107273503B (en) | Method and device for generating parallel text in same language | |
| CN111291570B (en) | Method and device for realizing element identification in judicial documents | |
| CN110033022A (en) | Processing method, device and the storage medium of text | |
| CN110287489A (en) | Document creation method, device, storage medium and electronic equipment | |
| CN109815487A (en) | Text quality detecting method, electronic device, computer equipment and storage medium | |
| CN107220220A (en) | Electronic equipment and method for text-processing | |
| CN108090218B (en) | Dialog system generation method and device based on deep reinforcement learning | |
| CN112988963B (en) | User intention prediction method, device, equipment and medium based on multi-flow nodes | |
| CN112800203B (en) | Question-answer matching method and system fusing text representation and knowledge representation | |
| CN111695354A (en) | Text question-answering method and device based on named entity and readable storage medium | |
| CN113821622B (en) | Answer retrieval method and device based on artificial intelligence, electronic equipment and medium | |
| CN110969023B (en) | Text similarity determination method and device | |
| CN110245232A (en) | File classification method, device, medium and calculating equipment | |
| CN112463989A (en) | Knowledge graph-based information acquisition method and system | |
| CN113723077B (en) | Sentence vector generation method and device based on bidirectional characterization model and computer equipment | |
| CN110399547A (en) | For updating the method, apparatus, equipment and storage medium of model parameter | |
| CN110532562A (en) | Neural network training method, Chinese idiom misuse detection method, device and electronic equipment | |
| CN111931503B (en) | Information extraction method and device, equipment and computer readable storage medium | |
| CN112749556B (en) | Multi-language model training method and device, storage medium and electronic equipment | |
| CN113705207A (en) | Grammar error recognition method and device | |
| CN111241843A (en) | Semantic relation inference system and method based on composite neural network | |
| CN111723186A (en) | Knowledge graph generation method based on artificial intelligence for dialog system and electronic equipment | |
| JP2017538226A (en) | Scalable web data extraction |
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 | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20190326 |
|
| WW01 | Invention patent application withdrawn after publication |