CN106294505A - A kind of method and apparatus feeding back answer - Google Patents

Abstract

The invention discloses a kind of method and apparatus feeding back answer, belong to field of computer technology.Described method includes: according to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, it is more than the training condition of problem and the semantic nearness of other corresponding answer with the semantic nearness of corresponding optimum answer based on problem, extraction of semantics parameter in default extraction of semantics formula is trained, obtains the trained values of extraction of semantics parameter；When receiving the answer request carrying target problem, according to each answer, described extraction of semantics formula and the trained values of described extraction of semantics parameter in described target problem, answer inquiry storehouse, determine the semantic nearness of described each answer and described target problem respectively；According to the semantic nearness of described each answer Yu described target problem, in described each answer, choose target answer, described answer request is fed back.Use the present invention, server can be improved and carry out the accuracy rate of answer feedback.

Description

A kind of method and apparatus feeding back answer

Technical field

The present invention relates to field of computer technology, particularly to a kind of method and apparatus feeding back answer.

Background technology

Along with computer and the development of information retrieval technique, people increasingly tend to seek by computer Asking the answer of certain problem, accordingly, the use of question answering system is more and more extensive.

The realization of existing community question answering system is usually: user inputs a problem, server by terminal Obtain, from answer inquiry storehouse, all answers prestored, determine problem and certain answer therein that user inputs Publicly-owned vocabulary, calculates the number of times sum that publicly-owned each vocabulary occurs in this answer, answers as this Case and the text nearness of the problem of user's input, in this way, calculate each answer in answer inquiry storehouse With the text nearness of the problem of user's input, the answer maximum with question text nearness is pushed to user.

During realizing the present invention, inventor finds that prior art at least there is problems in that

Implementation method based on above-mentioned community's question answering system, server is when pushing answer to user, mainly It is the text nearness coming between computational problem and answer based on the terminology match degree between problem and answer, But common vocabulary may be there is not with the problem of user's input and (i.e. there is vocabulary in the answer required for user Wide gap), or the number of times that common vocabulary occurs is less, so, is pushed to answer and the user's request of user The probability of coupling is relatively low, thus, cause server to carry out the accuracy rate of answer feedback relatively low.

Summary of the invention

In order to solve problem of the prior art, embodiments provide a kind of method feeding back answer and dress Put.Described technical scheme is as follows:

First aspect, it is provided that a kind of method feeding back answer, described method includes:

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on asking Inscribe the semantic nearness with corresponding optimum answer and be more than the problem semantic nearness with other corresponding answer Training condition, is trained the extraction of semantics parameter in default extraction of semantics formula, obtains described semanteme The trained values of extracting parameter；

When receiving the answer request carrying target problem, according to described target problem, answer inquiry storehouse In each answer, described extraction of semantics formula and the trained values of described extraction of semantics parameter, determine respectively Described each answer and the semantic nearness of described target problem；

According to the semantic nearness of described each answer Yu described target problem, in described each answer, choose target Answer, feeds back described answer request.

Second aspect, it is provided that a kind of device feeding back answer, described device includes:

Training module, for according to problem, optimum answer and other answer right of storage in training sample database Should be related to, based on problem with the semantic nearness of corresponding optimum answer more than problem and other corresponding answer The training condition of semantic nearness, the extraction of semantics parameter in default extraction of semantics formula is trained, Obtain the trained values of described extraction of semantics parameter；

Determine module, for when receiving the answer request carrying target problem, asking according to described target Each answer, described extraction of semantics formula and the instruction of described extraction of semantics parameter in topic, answer inquiry storehouse Practice value, determine the semantic nearness of described each answer and described target problem respectively；

Feedback module, for according to the semantic nearness of described each answer Yu described target problem, described respectively Answer is chosen target answer, described answer request is fed back.

The technical scheme that the embodiment of the present invention provides has the benefit that

In the embodiment of the present invention, according to problem, optimum answer and other answer of storage in training sample database Corresponding relation, answers more than problem and corresponding other based on problem and the semantic nearness of corresponding optimum answer The training condition of the semantic nearness of case, instructs the extraction of semantics parameter in default extraction of semantics formula Practice, obtain the trained values of described extraction of semantics parameter, when receiving the answer request carrying target problem, According to each answer, described extraction of semantics formula and institute's predicate in described target problem, answer inquiry storehouse The trained values of justice extracting parameter, determines the semantic nearness of described each answer and described target problem, root respectively According to the semantic nearness of described each answer Yu described target problem, in described each answer, choose target answer, Described answer request is fed back.So, carry out answer based on semantic nearness choosing, it is to avoid problem The vocabulary Gap existed with answer, it is thus possible to improve the accuracy of the answer for problem feedback.

Accompanying drawing explanation

For the technical scheme being illustrated more clearly that in the embodiment of the present invention, institute in embodiment being described below The accompanying drawing used is needed to be briefly described, it should be apparent that, the accompanying drawing in describing below is only the present invention Some embodiments, for those of ordinary skill in the art, on the premise of not paying creative work, Other accompanying drawing can also be obtained according to these accompanying drawings.

Fig. 1 is a kind of method flow diagram feeding back answer that the embodiment of the present invention provides；

Fig. 2 is a kind of schematic diagram training process that the embodiment of the present invention provides；

Fig. 3 is a kind of apparatus structure schematic diagram feeding back answer that the embodiment of the present invention provides；

Fig. 4 is the structural representation of a kind of server that the embodiment of the present invention provides.

Detailed description of the invention

For making the object, technical solutions and advantages of the present invention clearer, below in conjunction with accompanying drawing to the present invention Embodiment is described in further detail.

Embodiment one

Embodiments provide a kind of method feeding back answer, as it is shown in figure 1, the process stream of the method Journey can comprise the following steps that

Step 101, according to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, Connect more than the semanteme of problem with other corresponding answer with the semantic nearness of corresponding optimum answer based on problem The training condition of recency, is trained the extraction of semantics parameter in default extraction of semantics formula, obtains language The trained values of justice extracting parameter.

Step 102, when receiving the answer request carrying target problem, looks into according to target problem, answer Ask each answer, extraction of semantics formula and the trained values of extraction of semantics parameter in storehouse, determine respectively and respectively answer Case and the semantic nearness of target problem.

Step 103, according to the semantic nearness of each answer Yu target problem, chooses target answer in each answer, Answer request is fed back.

In the embodiment of the present invention, according to problem, optimum answer and other answer of storage in training sample database Corresponding relation, answers more than problem and corresponding other based on problem and the semantic nearness of corresponding optimum answer The training condition of the semantic nearness of case, instructs the extraction of semantics parameter in default extraction of semantics formula Practice, obtain the trained values of described extraction of semantics parameter, when receiving the answer request carrying target problem, According to each answer, described extraction of semantics formula and institute's predicate in described target problem, answer inquiry storehouse The trained values of justice extracting parameter, determines the semantic nearness of described each answer and described target problem, root respectively According to the semantic nearness of described each answer Yu described target problem, in described each answer, choose target answer, Described answer request is fed back.So, carry out answer based on semantic nearness choosing, it is to avoid problem The vocabulary Gap existed with answer, it is thus possible to improve the accuracy of the answer for problem feedback.

Embodiment two

Embodiments providing a kind of method feeding back answer, the executive agent of the method can be service Device, this server can be question and answer website, community or the server of application, can be provided with place in this server Reason device, memorizer, transceiver, processor may be used for the training to extraction of semantics parameter and anti-for problem The process of feedback answer, memorizer may be used for storing the number of data and the generation needed in following processing procedure According to, transceiver may be used for receiving and sending data,.Below in conjunction with detailed description of the invention, to shown in Fig. 1 Handling process be described in detail, content can be such that

Step 101, according to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, Connect more than the semanteme of problem with other corresponding answer with the semantic nearness of corresponding optimum answer based on problem The training condition of recency, is trained the extraction of semantics parameter in default extraction of semantics formula, obtains language The trained values of justice extracting parameter.

Wherein, when big data process, the semanteme of statement (such as problem, answer etc.) can be quantified, Extraction of semantics formula could be for the semantic formula of extraction problem or answer.Extraction of semantics parameter is permissible It is the constant coefficient in extraction of semantics formula, can be determined by training process.Semantic nearness can be to ask Topic and answer are in the degree of closeness of semantic (i.e. the expression meaning of statement) aspect.

In force, server can obtain the answer of some problems and correspondence thereof from the Internet, is deposited Storage in training sample database, such as, can obtain answering of some problems and correspondence in some community's question answering system Case, wherein, for each problem in training sample database to there being a number of answer, including Corresponding to the optimum answer (can be typically by the answer selected of user proposing problem) of each problem and its His answer.For each word present in dictionary to there being term vector (being properly termed as distribution vector), wherein, Term vector can be the vector (d can be 50) of d dimension, and certain one-dimensional numerical value therein can be used to represent and is somebody's turn to do The value of the corresponding a certain semantic item of word, such as, the term vector of BMW one word can be [0.5；0.8；...], Wherein, the semantic item that the first dimension of term vector is corresponding can be " this word is for representing the probability of animal ", 0.5 Represent that the numerical value of this probability, the semantic item of the second dimension correspondence can be that " this word is for representing the possibility of vehicle Property ", 0.8 represents the numerical value of this probability.Server can obtain the word of each word in problem and answer to Amount.For each problem in training sample database and the answer of correspondence thereof, server can obtain its correspondence The matrix (can be described as word matrix) comprising term vector, the word of each word in every string equivalent storehouse of word matrix Vector.Such as, having V word, the dimension of word matrix in current dictionary can be d × V, and server gets In training sample database after certain problem or answer, the word of the word occurred in the problem that obtains or answer can be obtained Vector, and place it in the corresponding position in word matrix, other position of word matrix could be arranged to zero (i.e. Problem or answer do not have the word occurred is corresponding is classified as zero in word matrix), so, each different problem Or answer has the word matrix of its correspondence.

Word matrix-vector corresponding to each problem in training sample database or answer can be obtained by server Characterization problems or the vector of answer, can use E_xRepresenting, subscript x can be that problem (can represent with q and ask Topic) or answer (answer can be represented with a), i.e. E_qAfter the word matrix-vector that problem of representation is corresponding to Amount, E_aRepresent the vector after the word matrix-vector that answer is corresponding.It is every that server obtains in training sample database Individual problem and the E of answer_q、E_aAfter, it is carried out a certain proportion of damage, i.e. can be by some of which value Pressure is set to zero (can be random choose some values wherein), obtain E～_x, and then can utilize following The semanteme that extraction of semantics formulas Extraction problem or answer are characterized:

$z=f(W*\tilde{E}x+b)......\left(1\right)$

Wherein, z can be described as semantic vector, can be with characterization problems or the semanteme of answer, and W is properly termed as weighting square Battle array, for rightExtraction of semantics and dimensionality reduction, b is properly termed as bias vector, is provided commonly for extraction with W and asks Topic or the semanteme that characterized of answer, W, b can be described as extraction of semantics parameter, and f () is nonlinear function, uses In the semanteme that extraction problem or answer are characterized, S function, hyperbolic function or rectification letter can be chosen for Number etc., f () is as a example by rectification function herein, i.e. f (), W and b jointly act on the problem of extraction or The semanteme that answer is characterized.

For each problem in training sample database and the answer of correspondence thereof, after server obtains vector z, can With the semantic nearness between computational problem and answer, problem of representation can be carried out with the cosine angle between vector And the semantic nearness between answer, it is more than problem according to problem with the semantic nearness of corresponding optimum answer With the condition of the semantic nearness of other corresponding answer, the weighting matrix in training formula (1) and biasing square Battle array, obtains final trained values.

Optionally, problem can be deducted with right by increase problem with the semantic nearness of corresponding optimum answer The mode of the difference summation of the semantic nearness of other answer answered, the semanteme obtained in above-mentioned formula (1) carries Taking parameter, accordingly, the processing procedure of step 101 can be such that according to asking of storing in training sample database Topic, optimum answer and the corresponding relation of other answer, based on the semanteme increasing problem and corresponding optimum answer Nearness deducts the training condition of problem and the difference summation of the semantic nearness of other corresponding answer, in advance If extraction of semantics formula in extraction of semantics parameter be trained, obtain the training of described extraction of semantics parameter Value.

In force, obtain a problem in sample training storehouse and the optimum answer of correspondence thereof and other answers, Wherein, the optimum answer that problem is corresponding can be designated as a⁺, other answers corresponding to problem can be designated as a_j ^—, j Other answers of jth that problem of representation is corresponding, can be 1 in the sum of other answers corresponding to this problem Any integer, such as, other answers outside the optimum answer that problem is corresponding have N number of, j=1, and 2 ... N. A problem in the sample training storehouse that will obtain and the optimum answer of correspondence and other answers thereof are as training number According to, set up object function, and the object function set up is trained, obtain the training of extraction of semantics parameter Value.

With the optimum answer of a problem in sample training storehouse and correspondence thereof and other answers as training data Training process is as follows: problem and answer according to formula (1) carry out extraction of semantics obtain semanteme corresponding to problem to Amount, can be designated as z respectively_qAnd z_a, after the semantic vector of all answers obtaining problem and its correspondence, can To calculate the semantic nearness between answer and problem according to formula (2),

$\mathrm{sim}(q,a)=\cos (z_q,z_a)=\frac{z_q\·z_a}{\left|\right|z_q\left|\right|\·\left|\right|z_a\left|\right|}......\left(2\right)$

Wherein, sim (q, a) the anticipated nearness of all answers of problem of representation and correspondence thereof, the formula used herein (2) It is that the cosine angle between employing problem and semantic vector corresponding to answer carrys out all of problem of representation and correspondence thereof The semantic nearness of answer.Server can set up loss function according to formula (3),

$L(q,a)=\underset{j}{\mathrm{\Σ}}(\mathrm{sim}\left({q,a}^{+}\right)-\mathrm{sim}(q,a_j^{-}))......\left(3\right)$

Wherein, (q, a) problem of representation deducts problem and answers with corresponding other L with the semantic nearness of corresponding optimum answer The difference summation of the semantic nearness of case, sim (q, a⁺) problem of representation is close with the semanteme of corresponding optimum answer Degree,The semantic nearness of problem of representation and corresponding other answer each, formula (3) i.e. enters The first object function of row training.The extraction of semantics parameter comprised in formula (3) is arranged initial value, utilizes Gradient descent method, trains first object function, obtains extraction of semantics parameter W, the b comprised in formula (3) Trained values, now, using the answer of a problem in sample training storehouse and correspondence thereof as training data, The training process of the object function set up based on this training data is terminated, after terminating, has obtained extraction of semantics ginseng The trained values of number W, b.

Server obtains the next problem in sample training storehouse and the optimum answer of correspondence thereof and other answer, As training data, and set up first object function according to above-mentioned training process, and utilize BP algorithm (Back Propagation, back propagation) by extraction of semantics parameter W obtained above, the trained values of b, As the initial value of this first object function of training, train first object function, obtain this extraction of semantics Parameter W, the trained values of b, as the initial value trained next time, recursion successively, until sample is instructed Practice all problems in storehouse and the optimum answer of correspondence thereof and complete, whole training process is trained in other answer Terminate, and obtain final extraction of semantics parameter W, the trained values of b, and stored.

Additionally, for reducing the complexity during gradient removing object function, during above-mentioned training, mesh Scalar functions can also use the formula shown in formula (4),

$L\left({q,a}^{+}\right)=\underset{j}{\mathrm{\Σ}}\max \{0,\mathrm{\δ}-\mathrm{sim}\left({q,a}^{+}\right)+\mathrm{sim}\left({q,a}_j^{-}\right)\}......\left(4\right)$

Wherein, the physical meaning that formula (4) represents is approximately the same with formula (3), and the foundation of training is all to make to ask Inscribe the semantic nearness more than problem with other corresponding answers of the semantic nearness with corresponding optimum answer, $\mathrm{\δ}=1/2-\mathrm{sim}\left({q,a}_j^{-}\right).$

Optionally, it is also possible to determined the initial value of above-mentioned training process by another training process, accordingly, Training process can be such that step one, according in training sample database storage each problem and each answer, based on Before reducing the statement and extraction of semantics obtained after statement sequence is performed the inversely processing of extraction of semantics and extraction of semantics The training condition of diversity factor of statement, the extraction of semantics parameter in default extraction of semantics formula is instructed Practice, obtain the middle trained value of described extraction of semantics parameter；Wherein, described statement is problem or answer；Step Rapid two, according to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on Problem is more than the semantic nearness of problem and other corresponding answer with the semantic nearness of corresponding optimum answer Training condition, using the middle trained value of described extraction of semantics parameter as initial input value, to described semanteme The extraction of semantics parameter extracted in formula is trained, and obtains the trained values of described extraction of semantics parameter.

Wherein, order performs extraction of semantics can be to carry out the semanteme of statement according to the formula shown in formula (1) Extracting, be known as the cataloged procedure to statement, the inversely processing of extraction of semantics can be to formula (1) Inversely processing, the E before can obtaining and encoding_xThere is the E ' of same dimension_x, this process can referred to as decode Journey, it is possible to use denoising automatic coding machine realizes cataloged procedure and the whole process of decoding process, can will go Automatic coding machine of making an uproar regards a kind of special neutral net as.

In force, as in figure 2 it is shown, the processing procedure of step one is as follows: obtain in sample training storehouse is each Some problem in each answer of problem or answer, obtain the E of its correspondence_x, afterwards it is carried out certain proportion Damage, can by some of which value force be set to zero, obtainAccording to formula (1), it is carried out Extraction of semantics, obtains the semantic vector z of each problem and each answer, wherein z comprise extraction of semantics parameter W, B is the most rightObtaining corresponding semantic vector z after encoding, z is carried out inverse by server by utilizing g (z) afterwards Conversion, i.e. carries out inverse transformation to the semantic vector z obtained after coding and obtains g (f ()), and this process is decoding process. Based on reducing the E ' obtained after decoding_xWith the E before coding_xThe training condition of diversity factor, set up equation below (5) it is the second object function:

L(g(f()),E_x)=| | g (f ())-E_x||₂……(5)

Wherein, formula (5) represents the extraction of semantics parameter pair utilizing selectionCoding, then decode it, The E ' arrived_xWith the E before damage_xThe mould of difference vector, the value of formula (5) is the least, the extraction of semantics obtained Parameter more can accurately express the semanteme of statement, and the extraction of semantics parameter comprised in formula (5) is arranged initial value, Utilize gradient descent method to train the second object function, obtain the trained values of extraction of semantics parameter.

Server obtains the other problem in sample training storehouse or answer, as training data, and according to Above-mentioned training process sets up the second object function, and utilizes BP algorithm by extraction of semantics parameter obtained above W, b trained values, as the initial value of this second object function of training, trains the second object function, passs successively Pushing away, until all problems and answer training in sample training storehouse are complete, whole training process terminates, and Obtain final extraction of semantics parameter W, the trained values of b.

The trained values of the extraction of semantics parameter that step one is obtained as the middle trained value of whole training process, As the initial value of the training process of step 2, obtain final semanteme according to step 2 continuation training and carry Take parameter W, b, and stored, wherein the training process of step 2 can be in above-mentioned steps 101 based on Increase problem deducts the semanteme of problem and other corresponding answer and connects with the semantic nearness of corresponding optimum answer The training process that the training condition of the difference sum of recency is trained, corresponding processing mode may refer to step Specific descriptions in 101, are not repeated herein.

Optionally, during training extraction of semantics parameter, the problem in sample training storehouse and answer can be entered respectively Row training, obtains problem and answer respective extraction of semantics parameter, accordingly, the process of above-mentioned steps one Journey can be such that according to each problem of storage in training sample database, based on reducing, question order is performed semanteme The training bar of the diversity factor of the statement before the statement extracted and obtain after the inversely processing of extraction of semantics and extraction of semantics Part, is trained the problem extraction of semantics parameter in default problem extraction of semantics formula, obtain described in ask The middle trained value of topic extraction of semantics parameter；According to each answer of storage in training sample database, right based on reducing Statement before the statement obtained after the inversely processing of statement sequence execution extraction of semantics and extraction of semantics and extraction of semantics The training condition of diversity factor, the answer extraction of semantics parameter in default answer extraction of semantics formula is carried out Training, obtains the middle trained value of described answer extraction of semantics parameter；The handling process of above-mentioned steps two is permissible As follows: according to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on Problem is more than the semantic nearness of problem and other corresponding answer with the semantic nearness of corresponding optimum answer Training condition, by middle trained value and the described answer extraction of semantics parameter of described problem extraction of semantics parameter Middle trained value as initial input value, the problem extraction of semantics in described problem extraction of semantics formula is joined Answer extraction of semantics parameter in number and described answer extraction of semantics formula is trained, and obtains described problem language The trained values of justice extracting parameter and the trained values of described answer extraction of semantics parameter.

In force, during the training of above-mentioned steps one, problem or answer in sample training storehouse are permissible Use different W and b, carry out the problem in extraction of semantics, i.e. sample training storehouse according to formula (1) respectively A pair W and b can be used (W can be designated as₁、b₁) carry out extraction of semantics, the problem in sample training storehouse Corresponding answer can use another that (W and b can be designated as W₂、b₂) carry out extraction of semantics, and according to Mode described in step one is set up object function respectively and trains, and obtains W₁、b₁And W₂、b₂Trained values, As the middle trained value of whole training process, and as the initial value of step 2, according to step 2 institute The handling process stated continues training and obtains final extraction of semantics parameter W₁、b₁And W₂、b₂, and deposited Storage, wherein, during the training of step 2, the optimum answer of computational problem and correspondence and other answer Semantic vector, and according to the semantic vector obtained according to formula (2) computing semantic similarity time, semantic phase W is comprised in degree₁、b₁And W₂、b₂Four extraction of semantics parameters, corresponding processing mode may refer to Step one, specific descriptions in two, be not repeated herein.

Step 102, when receiving the answer request carrying target problem, looks into according to target problem, answer Ask each answer, extraction of semantics formula and the trained values of extraction of semantics parameter in storehouse, determine respectively and respectively answer Case and the semantic nearness of target problem.

Wherein, target problem can be the problem wanting to know answer that user is inputted by terminal, and answer is looked into Ask storehouse can be above-mentioned sample training storehouse, it is also possible to be storage server obtain from the Internet some answer The storehouse of case, therefrom chooses the answer of coupling target problem for server.

In force, user inputs target problem and after server sends answer request by terminal, service Device can receive the answer request that user sends, and then it is resolved by server, and acquisition is wherein carried Target problem, the trained values of the extraction of semantics parameter stored by server substitutes in formula (1), can be according to Formula (1) calculates the semantic vector of each answer in target problem and answer inquiry storehouse, obtains target problem And after the semantic vector of each answer in answer inquiry storehouse, answer inquiry storehouse can be calculated according to formula (2) In each answer respectively with the semantic nearness of target problem.

Optionally, situation about being trained respectively for the problems referred to above and answer；Accordingly, connect when server Processing procedure when receiving the answer request of terminal transmission can be such that to work as to receive and carries target problem During answer request, according to each answer in described target problem, answer inquiry storehouse, described problem extraction of semantics Formula, described answer extraction of semantics formula and the trained values of described problem extraction of semantics parameter and described answer The trained values of case semanteme extracting parameter, determines the semantic nearness of described each answer and described target problem respectively.

In force, after server obtains problem and answer extraction of semantics parameter respectively, receive and carry When the answer of target problem is asked, the most corresponding extraction of semantics parameter can be distinguished according to problem with answer, according to Formula (1) calculates the semantic vector of each answer in target problem and answer inquiry storehouse respectively, it is determined that each From semantic vector after, each answer and target problem in answer inquiry storehouse can be calculated according to formula (2) Semantic similarity.

Step 103, according to the semantic nearness of each answer Yu target problem, chooses target answer in each answer, Answer request is fed back.

Wherein, target answer can be answer inquiry storehouse in each answer matches in the answer of target problem, can To be one of them answer, it is also possible to be several answer therein.

In force, each answer during server obtains answer inquiry storehouse is close with the semanteme of target problem respectively After degree, according to order from big to small, the semantic nearness obtained can be ranked up, can be by maximum Answer corresponding to semantic nearness is chosen for target answer, or will after sequence before several semantic nearness pair The answer answered is chosen for target answer, after choosing target answer, target answer is given by terminal feedback and uses Family.

Optionally, it is also possible to the semantic nearness obtained is combined with some features based on terminology match, Accordingly, the handling process of step 103 can be such that the language according to described each answer Yu described target problem Justice nearness, and the text nearness of described each answer and described target problem, select in described each answer Take target answer, described answer request is fed back.

Wherein, text nearness can be each answer and target problem nearness based on terminology match.

In force, after server obtains the semantic nearness of each answer in answer inquiry storehouse and target problem, Stored, and calculated each answer in answer inquiry storehouse and mesh according to the formula shown in formula (6)-(16) Mark problem text based on terminology match nearness,

$F_1={\mathrm{\Σ}}_{q_i\∈q\∩a}c(q_i,a)......\left(6\right)$

$F_2={\mathrm{\Σ}}_{q_i\∈q\∩a}\log (c(q_i,a)+1)......\left(7\right)$

$F_3={\mathrm{\Σ}}_{q_i\∈q\∩a}c(q_i,a)/\left|a\right|......\left(8\right)$

$F_4={\mathrm{\Σ}}_{q_i\∈q\∩a}(c(q_i,a)/|a|+1)......\left(9\right)$

$F_5={\mathrm{\Σ}}_{q_i\∈q\∩a}\log \left(\right|C|/\mathrm{df}\left(q_i\right))......\left(10\right)$

$F_6={\mathrm{\Σ}}_{q_i\∈q\∩a}\left(\log \left(\right|C|/\mathrm{df}\left(q_i\right))\right)......\left(11\right)$

$F_7={\mathrm{\Σ}}_{q_i\∈q\∩a}\log \left(\right|C|/\mathrm{df}\left(q_i\right))+1)......\left(12\right)$

$F_8={\mathrm{\Σ}}_{q_i\∈q\∩a}\log (c(q_i,a)/|a\left|\right)\log \left(\right|C|/\mathrm{df}\left(q_i\right)+1)......\left(13\right)$

$F_9={\mathrm{\Σ}}_{q_i\∈q\∩a}c(q_i,a)\log \left(\right|C|/\mathrm{df}\left(q_i\right))......\left(14\right)$

$F_{10}={\mathrm{\Σ}}_{q_i\∈q\∩a}\log (c(q_i,a)/|a|\·|C|/(c(q_i,C)+1)+1)......\left(15\right)$

$L_1={\mathrm{\Σ}}_{q_i\∈q\∩a}\log \left(\right|a|-c(q_i,a)+0.5)/(c(q_i,a)+0.5)c(q_i,a)*(k_1+1)/(c(q_i,a)+k_1*(1-b+b*|a|/\mathrm{avg}|C|)......\left(16\right)$

Wherein, c (q_i, a) can be q_iThe number of times occurred in a, df (q_i) can be q_iRespectively answering in answer inquiry storehouse The number of times occurred in case, | a | can be the number of the word comprised in answer a, and | C | can be in answer inquiry storehouse The number of word that comprises of each answer, C can be each answer in answer inquiry storehouse, k₁∈ [1.2,2.0], b=0.75, Avg | C | can be the meansigma methods of the number of word that each answer in answer inquiry storehouse comprises, obtain each answer with After target problem text similarity, it is total to the semantic similarity of the above-mentioned each answer determined with target problem With putting in study sequence framework, e.g. SVM sort algorithm, obtain each answer and mesh in answer inquiry storehouse Shown in mark problem integrated ordered, i.e. comprehensive utilization semantic similarity feature and above-mentioned 11 formula based on word Converge the text similarity feature of coupling and the similarity of each answer of obtaining and target problem, wherein these 12 spies The weight levied can carry out artificial assignment based on experience value, it is also possible to utilize sample in sample training storehouse according to SVM sort algorithm is trained obtaining the weight that each feature is corresponding, by answer corresponding for maximum similarity By terminal feedback to user, it is also possible to the answer that before in sorting, several similarities are corresponding is anti-by terminal Feed user.

In the embodiment of the present invention, according to problem, optimum answer and other answer of storage in training sample database Corresponding relation, answers more than problem and corresponding other based on problem and the semantic nearness of corresponding optimum answer The training condition of the semantic nearness of case, instructs the extraction of semantics parameter in default extraction of semantics formula Practice, obtain the trained values of described extraction of semantics parameter, when receiving the answer request carrying target problem, According to each answer, described extraction of semantics formula and institute's predicate in described target problem, answer inquiry storehouse The trained values of justice extracting parameter, determines the semantic nearness of described each answer and described target problem, root respectively According to the semantic nearness of described each answer Yu described target problem, in described each answer, choose target answer, Described answer request is fed back.So, carry out answer based on semantic nearness choosing, it is to avoid problem The vocabulary Gap existed with answer, it is thus possible to improve the accuracy of the answer for problem feedback.

Embodiment three

Based on identical technology design, the embodiment of the present invention additionally provides a kind of device feeding back answer, such as Fig. 3 Shown in, this device includes:

Training module 310, for according to problem, optimum answer and other answer of storage in training sample database Corresponding relation, answers more than problem and corresponding other based on problem and the semantic nearness of corresponding optimum answer The training condition of the semantic nearness of case, instructs the extraction of semantics parameter in default extraction of semantics formula Practice, obtain the trained values of described extraction of semantics parameter；

Determine module 320, for when receiving the answer request carrying target problem, according to described target Each answer, described extraction of semantics formula and described extraction of semantics parameter in problem, answer inquiry storehouse Trained values, determines the semantic nearness of described each answer and described target problem respectively；

Feedback module 330, for the semantic nearness according to described each answer Yu described target problem, described Each answer is chosen target answer, described answer request is fed back.

Optionally, described training module 310, it is used for:

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on increasing It is close with the semanteme of other corresponding answer that big problem deducts problem with the semantic nearness of corresponding optimum answer The training condition of the difference summation of degree, is trained the extraction of semantics parameter in default extraction of semantics formula, Obtain the trained values of described extraction of semantics parameter.

Optionally, described training module 310, it is used for:

According to each problem stored in training sample database and each answer, based on reducing, statement sequence is performed semanteme The training bar of the diversity factor of the statement before the statement extracted and obtain after the inversely processing of extraction of semantics and extraction of semantics Part, is trained the extraction of semantics parameter in default extraction of semantics formula, obtains described extraction of semantics ginseng The middle trained value of number；Wherein, described statement is problem or answer；

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on asking Inscribe the semantic nearness with corresponding optimum answer and be more than the problem semantic nearness with other corresponding answer Training condition, using the middle trained value of described extraction of semantics parameter as initial input value, carries described semanteme The extraction of semantics parameter taken in formula is trained, and obtains the trained values of described extraction of semantics parameter.

Optionally, described training module 310, it is used for:

According to each problem of storage in training sample database, based on reducing, question order is performed extraction of semantics and language The training condition of the diversity factor of the statement before the statement obtained after the inversely processing that justice is extracted and extraction of semantics, in advance If problem extraction of semantics formula in problem extraction of semantics parameter be trained, obtain described problem semanteme and carry Take the middle trained value of parameter；

According to each answer of storage in training sample database, based on reducing, statement sequence is performed extraction of semantics and language The training condition of the diversity factor of the statement before the statement obtained after the inversely processing that justice is extracted and extraction of semantics, in advance If answer extraction of semantics formula in answer extraction of semantics parameter be trained, obtain described answer semanteme and carry Take the middle trained value of parameter；

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on asking Inscribe the semantic nearness with corresponding optimum answer and be more than the problem semantic nearness with other corresponding answer Training condition, by the middle trained value of described problem extraction of semantics parameter and described answer extraction of semantics parameter Middle trained value is as initial input value, to the problem extraction of semantics parameter in described problem extraction of semantics formula It is trained with the answer extraction of semantics parameter in described answer extraction of semantics formula, obtains described problem semantic The trained values of extracting parameter and the trained values of described answer extraction of semantics parameter；

Described determine module 320, be used for:

When receiving the answer request carrying target problem, according to described target problem, answer inquiry storehouse In each answer, described problem extraction of semantics formula, described answer extraction of semantics formula and described problem The trained values of extraction of semantics parameter and the trained values of described answer extraction of semantics parameter, determine respectively and described respectively answer Case and the semantic nearness of described target problem.

Optionally, described feedback module 330, it is used for:

According to the semantic nearness of described each answer Yu described target problem, and described each answer and described mesh The text nearness of mark problem, chooses target answer in described each answer, carries out described answer request instead Feedback.

In the embodiment of the present invention, according to problem, optimum answer and other answer of storage in training sample database Corresponding relation, answers more than problem and corresponding other based on problem and the semantic nearness of corresponding optimum answer The training condition of the semantic nearness of case, instructs the extraction of semantics parameter in default extraction of semantics formula Practice, obtain the trained values of described extraction of semantics parameter, when receiving the answer request carrying target problem, According to each answer, described extraction of semantics formula and institute's predicate in described target problem, answer inquiry storehouse The trained values of justice extracting parameter, determines the semantic nearness of described each answer and described target problem, root respectively According to the semantic nearness of described each answer Yu described target problem, in described each answer, choose target answer, Described answer request is fed back.So, carry out answer based on semantic nearness choosing, it is to avoid problem The vocabulary Gap existed with answer, it is thus possible to improve the accuracy of the answer for problem feedback.

It should be understood that above-described embodiment provide feedback answer device feed back answer time, only more than The division stating each functional module is illustrated, in actual application, and can be as desired by above-mentioned functions Distribution is completed by different functional modules, the internal structure of equipment will be divided into different functional modules, with Complete all or part of function described above.It addition, the device feeding back answer that above-described embodiment provides Belonging to same design with the embodiment of the method for feedback answer, it implements process and refers to embodiment of the method, this In repeat no more.

Embodiment four

Fig. 4 is the structural representation of the server that the embodiment of the present invention provides.This server 1900 can be because of configuration Or performance is different and produce bigger difference, one or more central processing units (central can be included Processing units, CPU) 1922 (such as, one or more processors) and memorizeies 1932, The storage medium 1930 of one or more storage application programs 1942 or data 1944 (such as one or More than one mass memory unit).Wherein, memorizer 1932 and storage medium 1930 can be of short duration storages Or persistently store.The program being stored in storage medium 1930 can include one or more modules (diagram Do not mark), each module can include a series of command operatings in statistical server.Further, Central processing unit 1922 could be arranged to communicate with storage medium 1930, performs on statistical server 1900 A series of command operatings in storage medium 1930.

Server 1900 can also include one or more power supplys 1926, one or more wired or Radio network interface 1950, one or more input/output interfaces 1958, one or more keyboards 1956, and/or, one or more operating systems 1941, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM etc..

Server 1900 can include memorizer, and one or more than one program, one of them Or more than one program is stored in memorizer, and it is configured to be held by one or more than one processor Row one or more than one program comprise the instruction for carrying out following operation:

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on asking Inscribe the semantic nearness with corresponding optimum answer and be more than the problem semantic nearness with other corresponding answer Training condition, is trained the extraction of semantics parameter in default extraction of semantics formula, obtains described semanteme The trained values of extracting parameter；

When receiving the answer request carrying target problem, according to described target problem, answer inquiry storehouse In each answer, described extraction of semantics formula and the trained values of described extraction of semantics parameter, determine respectively Described each answer and the semantic nearness of described target problem；

According to the semantic nearness of described each answer Yu described target problem, in described each answer, choose target Answer, feeds back described answer request.

Optionally, described according to the correspondence of problem, optimum answer and other answer of storage in training sample database Relation, is more than problem and other corresponding answer based on problem with the semantic nearness of corresponding optimum answer The training condition of semantic nearness, is trained the extraction of semantics parameter in default extraction of semantics formula, Obtain the trained values of described extraction of semantics parameter, including:

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on increasing It is close with the semanteme of other corresponding answer that big problem deducts problem with the semantic nearness of corresponding optimum answer The training condition of the difference summation of degree, is trained the extraction of semantics parameter in default extraction of semantics formula, Obtain the trained values of described extraction of semantics parameter.

Optionally, described according to the correspondence of problem, optimum answer and other answer of storage in training sample database Relation, is more than problem and other corresponding answer based on problem with the semantic nearness of corresponding optimum answer The training condition of semantic nearness, is trained the extraction of semantics parameter in default extraction of semantics formula, Obtain the trained values of described extraction of semantics parameter, including:

According to each problem stored in training sample database and each answer, based on reducing, statement sequence is performed semanteme The training bar of the diversity factor of the statement before the statement extracted and obtain after the inversely processing of extraction of semantics and extraction of semantics Part, is trained the extraction of semantics parameter in default extraction of semantics formula, obtains described extraction of semantics ginseng The middle trained value of number；Wherein, described statement is problem or answer；

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on asking Inscribe the semantic nearness with corresponding optimum answer and be more than the problem semantic nearness with other corresponding answer Training condition, using the middle trained value of described extraction of semantics parameter as initial input value, carries described semanteme The extraction of semantics parameter taken in formula is trained, and obtains the trained values of described extraction of semantics parameter.

Optionally, described according to each problem stored in training sample database and each answer, based on reducing statement The difference of the statement before the statement obtained after the inversely processing of order execution extraction of semantics and extraction of semantics and extraction of semantics The training condition of different degree, is trained the extraction of semantics parameter in default extraction of semantics formula, obtains institute The middle trained value of predicate justice extracting parameter, including:

According to each problem of storage in training sample database, based on reducing, question order is performed extraction of semantics and language The training condition of the diversity factor of the statement before the statement obtained after the inversely processing that justice is extracted and extraction of semantics, in advance If problem extraction of semantics formula in problem extraction of semantics parameter be trained, obtain described problem semanteme and carry Take the middle trained value of parameter；

According to each answer of storage in training sample database, based on reducing, statement sequence is performed extraction of semantics and language The training condition of the diversity factor of the statement before the statement obtained after the inversely processing that justice is extracted and extraction of semantics, in advance If answer extraction of semantics formula in answer extraction of semantics parameter be trained, obtain described answer semanteme and carry Take the middle trained value of parameter；

Described according to the corresponding relation of problem, optimum answer and other answer of storage, base in training sample database Close with the semanteme of other corresponding answer more than problem with the semantic nearness of corresponding optimum answer in problem The training condition of degree, using the middle trained value of described extraction of semantics parameter as initial input value, to institute's predicate The extraction of semantics parameter that justice is extracted in formula is trained, and obtains the trained values of described extraction of semantics parameter, bag Include:

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on asking Inscribe the semantic nearness with corresponding optimum answer and be more than the problem semantic nearness with other corresponding answer Training condition, by the middle trained value of described problem extraction of semantics parameter and described answer extraction of semantics parameter Middle trained value is as initial input value, to the problem extraction of semantics parameter in described problem extraction of semantics formula It is trained with the answer extraction of semantics parameter in described answer extraction of semantics formula, obtains described problem semantic The trained values of extracting parameter and the trained values of described answer extraction of semantics parameter；

Described when receiving the answer request carrying target problem, look into according to described target problem, answer Ask each answer, described extraction of semantics formula and the trained values of described extraction of semantics parameter in storehouse, respectively Determine the semantic nearness of described each answer and described target problem, including:

When receiving the answer request carrying target problem, according to described target problem, answer inquiry storehouse In each answer, described problem extraction of semantics formula, described answer extraction of semantics formula and described problem The trained values of extraction of semantics parameter and the trained values of described answer extraction of semantics parameter, determine respectively and described respectively answer Case and the semantic nearness of described target problem.

Optionally, the described semantic nearness according to described each answer Yu described target problem, respectively answer described Case is chosen target answer, described answer request is fed back, including:

According to the semantic nearness of described each answer Yu described target problem, and described each answer and described mesh The text nearness of mark problem, chooses target answer in described each answer, carries out described answer request instead Feedback.

In the embodiment of the present invention, according to problem, optimum answer and other answer of storage in training sample database Corresponding relation, answers more than problem and corresponding other based on problem and the semantic nearness of corresponding optimum answer The training condition of the semantic nearness of case, instructs the extraction of semantics parameter in default extraction of semantics formula Practice, obtain the trained values of described extraction of semantics parameter, when receiving the answer request carrying target problem, According to each answer, described extraction of semantics formula and institute's predicate in described target problem, answer inquiry storehouse The trained values of justice extracting parameter, determines the semantic nearness of described each answer and described target problem, root respectively According to the semantic nearness of described each answer Yu described target problem, in described each answer, choose target answer, Described answer request is fed back.So, carry out answer based on semantic nearness choosing, it is to avoid problem The vocabulary Gap existed with answer, it is thus possible to improve the accuracy of the answer for problem feedback.

One of ordinary skill in the art will appreciate that all or part of step realizing above-described embodiment can be passed through Hardware completes, it is also possible to instructing relevant hardware by program and complete, described program can be stored in In a kind of computer-readable recording medium, storage medium mentioned above can be read only memory, disk or CD etc..

The foregoing is only presently preferred embodiments of the present invention, not in order to limit the present invention, all the present invention's Within spirit and principle, any modification, equivalent substitution and improvement etc. made, should be included in the present invention's Within protection domain.

Claims (10)

1. the method feeding back answer, it is characterised in that described method includes:

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on asking Inscribe the semantic nearness with corresponding optimum answer and be more than the problem semantic nearness with other corresponding answer Training condition, is trained the extraction of semantics parameter in default extraction of semantics formula, obtains described semanteme The trained values of extracting parameter；

When receiving the answer request carrying target problem, according to described target problem, answer inquiry storehouse In each answer, described extraction of semantics formula and the trained values of described extraction of semantics parameter, determine respectively Described each answer and the semantic nearness of described target problem；

According to the semantic nearness of described each answer Yu described target problem, in described each answer, choose target Answer, feeds back described answer request.

Method the most according to claim 1, it is characterised in that described according to storage in training sample database The corresponding relation of problem, optimum answer and other answer, semanteme based on problem with corresponding optimum answer Nearness is more than problem and the training condition of the semantic nearness of other corresponding answer, carries default semanteme The extraction of semantics parameter taken in formula is trained, and obtains the trained values of described extraction of semantics parameter, including:

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on increasing It is close with the semanteme of other corresponding answer that big problem deducts problem with the semantic nearness of corresponding optimum answer The training condition of the difference summation of degree, is trained the extraction of semantics parameter in default extraction of semantics formula, Obtain the trained values of described extraction of semantics parameter.

Method the most according to claim 1, it is characterised in that described according to storage in training sample database The corresponding relation of problem, optimum answer and other answer, semanteme based on problem with corresponding optimum answer Nearness is more than problem and the training condition of the semantic nearness of other corresponding answer, carries default semanteme The extraction of semantics parameter taken in formula is trained, and obtains the trained values of described extraction of semantics parameter, including:

According to each problem stored in training sample database and each answer, based on reducing, statement sequence is performed semanteme The training bar of the diversity factor of the statement before the statement extracted and obtain after the inversely processing of extraction of semantics and extraction of semantics Part, is trained the extraction of semantics parameter in default extraction of semantics formula, obtains described extraction of semantics ginseng The middle trained value of number；Wherein, described statement is problem or answer；

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on asking Inscribe the semantic nearness with corresponding optimum answer and be more than the problem semantic nearness with other corresponding answer Training condition, using the middle trained value of described extraction of semantics parameter as initial input value, carries described semanteme The extraction of semantics parameter taken in formula is trained, and obtains the trained values of described extraction of semantics parameter.

Method the most according to claim 3, it is characterised in that described according to storage in training sample database Each problem and each answer, after reducing the inversely processing that statement sequence is performed extraction of semantics and extraction of semantics The training condition of the diversity factor of the statement before the statement obtained and extraction of semantics, to default extraction of semantics formula In extraction of semantics parameter be trained, obtain the middle trained value of described extraction of semantics parameter, including:

According to each problem of storage in training sample database, based on reducing, question order is performed extraction of semantics and language The training condition of the diversity factor of the statement before the statement obtained after the inversely processing that justice is extracted and extraction of semantics, in advance If problem extraction of semantics formula in problem extraction of semantics parameter be trained, obtain described problem semanteme and carry Take the middle trained value of parameter；

According to each answer of storage in training sample database, based on reducing, statement sequence is performed extraction of semantics and language The training condition of the diversity factor of the statement before the statement obtained after the inversely processing that justice is extracted and extraction of semantics, in advance If answer extraction of semantics formula in answer extraction of semantics parameter be trained, obtain described answer semanteme and carry Take the middle trained value of parameter；

Described according to the corresponding relation of problem, optimum answer and other answer of storage, base in training sample database Close with the semanteme of other corresponding answer more than problem with the semantic nearness of corresponding optimum answer in problem The training condition of degree, using the middle trained value of described extraction of semantics parameter as initial input value, to institute's predicate The extraction of semantics parameter that justice is extracted in formula is trained, and obtains the trained values of described extraction of semantics parameter, bag Include:

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on asking Inscribe the semantic nearness with corresponding optimum answer and be more than the problem semantic nearness with other corresponding answer Training condition, by the middle trained value of described problem extraction of semantics parameter and described answer extraction of semantics parameter Middle trained value is as initial input value, to the problem extraction of semantics parameter in described problem extraction of semantics formula It is trained with the answer extraction of semantics parameter in described answer extraction of semantics formula, obtains described problem semantic The trained values of extracting parameter and the trained values of described answer extraction of semantics parameter；

Described when receiving the answer request carrying target problem, look into according to described target problem, answer Ask each answer, described extraction of semantics formula and the trained values of described extraction of semantics parameter in storehouse, respectively Determine the semantic nearness of described each answer and described target problem, including:

When receiving the answer request carrying target problem, according to described target problem, answer inquiry storehouse In each answer, described problem extraction of semantics formula, described answer extraction of semantics formula and described problem The trained values of extraction of semantics parameter and the trained values of described answer extraction of semantics parameter, determine respectively and described respectively answer Case and the semantic nearness of described target problem.

Method the most according to claim 1, it is characterised in that described according to described each answer with described The semantic nearness of target problem, chooses target answer in described each answer, carries out described answer request Feedback, including:

According to the semantic nearness of described each answer Yu described target problem, and described each answer and described mesh The text nearness of mark problem, chooses target answer in described each answer, carries out described answer request instead Feedback.

6. the device feeding back answer, it is characterised in that described device includes:

Training module, for according to problem, optimum answer and other answer right of storage in training sample database Should be related to, based on problem with the semantic nearness of corresponding optimum answer more than problem and other corresponding answer The training condition of semantic nearness, the extraction of semantics parameter in default extraction of semantics formula is trained, Obtain the trained values of described extraction of semantics parameter；

Determine module, for when receiving the answer request carrying target problem, asking according to described target Each answer, described extraction of semantics formula and the instruction of described extraction of semantics parameter in topic, answer inquiry storehouse Practice value, determine the semantic nearness of described each answer and described target problem respectively；

Feedback module, for according to the semantic nearness of described each answer Yu described target problem, described respectively Answer is chosen target answer, described answer request is fed back.

Device the most according to claim 6, it is characterised in that described training module, is used for:

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on increasing It is close with the semanteme of other corresponding answer that big problem deducts problem with the semantic nearness of corresponding optimum answer The training condition of the difference summation of degree, is trained the extraction of semantics parameter in default extraction of semantics formula, Obtain the trained values of described extraction of semantics parameter.

Device the most according to claim 6, it is characterised in that described training module, is used for:

According to each problem stored in training sample database and each answer, based on reducing, statement sequence is performed semanteme The training bar of the diversity factor of the statement before the statement extracted and obtain after the inversely processing of extraction of semantics and extraction of semantics Part, is trained the extraction of semantics parameter in default extraction of semantics formula, obtains described extraction of semantics ginseng The middle trained value of number；Wherein, described statement is problem or answer；

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on asking Inscribe the semantic nearness with corresponding optimum answer and be more than the problem semantic nearness with other corresponding answer Training condition, using the middle trained value of described extraction of semantics parameter as initial input value, carries described semanteme The extraction of semantics parameter taken in formula is trained, and obtains the trained values of described extraction of semantics parameter.

Device the most according to claim 8, it is characterised in that described training module, is used for:

According to each problem of storage in training sample database, based on reducing, question order is performed extraction of semantics and language The training condition of the diversity factor of the statement before the statement obtained after the inversely processing that justice is extracted and extraction of semantics, in advance If problem extraction of semantics formula in problem extraction of semantics parameter be trained, obtain described problem semanteme and carry Take the middle trained value of parameter；

According to each answer of storage in training sample database, based on reducing, statement sequence is performed extraction of semantics and language The training condition of the diversity factor of the statement before the statement obtained after the inversely processing that justice is extracted and extraction of semantics, in advance If answer extraction of semantics formula in answer extraction of semantics parameter be trained, obtain described answer semanteme and carry Take the middle trained value of parameter；

According to the corresponding relation of problem, optimum answer and other answer of storage in training sample database, based on asking Inscribe the semantic nearness with corresponding optimum answer and be more than the problem semantic nearness with other corresponding answer Training condition, by the middle trained value of described problem extraction of semantics parameter and described answer extraction of semantics parameter Middle trained value is as initial input value, to the problem extraction of semantics parameter in described problem extraction of semantics formula It is trained with the answer extraction of semantics parameter in described answer extraction of semantics formula, obtains described problem semantic The trained values of extracting parameter and the trained values of described answer extraction of semantics parameter；

Described determine module, be used for:

When receiving the answer request carrying target problem, according to described target problem, answer inquiry storehouse In each answer, described problem extraction of semantics formula, described answer extraction of semantics formula and described problem The trained values of extraction of semantics parameter and the trained values of described answer extraction of semantics parameter, determine respectively and described respectively answer Case and the semantic nearness of described target problem.

Device the most according to claim 6, it is characterised in that described feedback module, is used for:

According to the semantic nearness of described each answer Yu described target problem, and described each answer and described mesh The text nearness of mark problem, chooses target answer in described each answer, carries out described answer request instead Feedback.