CN106991161A - A kind of method for automatically generating open-ended question answer - Google Patents
A kind of method for automatically generating open-ended question answer Download PDFInfo
- Publication number
- CN106991161A CN106991161A CN201710205299.8A CN201710205299A CN106991161A CN 106991161 A CN106991161 A CN 106991161A CN 201710205299 A CN201710205299 A CN 201710205299A CN 106991161 A CN106991161 A CN 106991161A
- Authority
- CN
- China
- Prior art keywords
- answer
- project
- item
- title
- rule
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/3331—Query processing
- G06F16/334—Query execution
- G06F16/3344—Query execution using natural language analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/332—Query formulation
- G06F16/3329—Natural language query formulation or dialogue systems
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Human Computer Interaction (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a kind of method for automatically generating open-ended question answer, including:The project obtained in content library, the content library at least includes the attribute in item source, the contents of a project and item-title;Set up answer rule and obtained from the content library and meet the project of answer rule to set up answer storehouse, the answer rule include at least with a kind of corresponding screening conditions in the item source, the contents of a project, item-title;Project carried out participle and found the destination item for possessing all participles in answer storehouse the problem of to input;Calculate the synthesis pertinence of described problem project and destination item and find out synthesis pertinence highest destination item to export.The open-ended question that the present invention can be directed in network application provides the autonomous reply of network.
Description
Technical field
The present invention relates to Computer Applied Technology field, more particularly to a kind of side for automatically generating open-ended question answer
Method.
Background technology
With the development of internet, Ask-Answer Community is increasingly becoming a kind of deep internet product welcome by netizen, such as domestic
" knowing ", " top question and answer ", the Quora in the U.S., user can propose various problems in community, and other users see oneself
The problem of familiar field, can just issue the answer of oneself.These answers are understood and seen by all users for browsing the problem, thumb up
Or point is stepped on, welcome answer can be discharged to forward position by system.One challenge of this kind of community is, with the increasing of customer volume
It is long, seldom it is concerned about the problem of more and more by other users, causes unmanned answer, the enthusiasm of quizmaster overwhelms greatly.Have
A little communities, which are provided, answers the main function of inviting automatically, can be according to the answer history of user problem invite automatically suitable user come
Answer.But can so cause it is active answer master and receive excessive invitation, have no idea in the case where ensureing to answer quality and
When answer a question.Therefore, this current problem is still without effective solution.
In terms of automatic answer, existing technology is focused primarily in the answer generation for solving closed question, such as " difficult to understand
Bar horse is which president of the U.S.”.Problem typically can be being converted into a structure by some existing systems using NLP technologies
The query language of change and the knowledge mapping return answer built in advance by inquiry.But existing automatic answer generation technique without
The answer that method solves open-ended question automatically generates problem, such as:" ecology to LeEco is anti-war slightly how to see for you" this kind of ask
What the quizmaster of topic needed is not intellectual answer, but the elaboration of the analysis and viewpoint to problem, so as to reach thought
Exchange.Current network Ask-Answer Community can not have open specific network problem to be replied automatically to above-mentioned, also not have
Realize the function of automatically generating answer of network problem.The shortcoming of above-mentioned functions causes a large amount of wastes of Internet resources, it is impossible to whole
Internet resources are closed, network self-adapting can not be realized interrogation responsor system are handled so that a large number of users problem can not obtain anti-in time
Feedback is with solving.
The content of the invention
The technical problem that technical solution of the present invention is solved is:The open-ended question how being directed in network application provides net
The autonomous reply of network.
In order to solve the above-mentioned technical problem, technical solution of the present invention, which is provided, a kind of automatically generates open-ended question answer
Method, including:
The project obtained in content library, the content library at least includes in item source, the contents of a project and item-title
One attribute;
Set up answer rule and obtained from the content library and meet the project of the answer rule to set up answer storehouse, institute
State answer rule include at least with a kind of corresponding screening conditions in the item source, the contents of a project, item-title;
Project carried out participle and found the destination item for possessing all participles in answer storehouse the problem of to input;
Calculate the synthesis pertinence of described problem project and destination item and find out synthesis pertinence highest destination item
To export.
Optionally, the acquisition content library includes:Content is set up according to the document or video of internal offer or outside crawl
Project in storehouse.
Optionally, the answer rule of setting up includes:
Screening meets the project of at least one in following condition from content library:
The technorati authority of the item source is legal;
The timeliness of the contents of a project is legal;
Entity word is included in the item-title;
The item-title is interrogative sentence.
Optionally, the answer rule of setting up includes:
Meet the project of the answer rule described in artificial screening from the content library and do not meet the answer rule
Project using as the first positive sample and the first negative sample;
The attribute of the project, the first positive sample and the first negative sample input answer rule are set up into model, train described
Answer rule sets up model to set up the answer rule.
Optionally, the training answer rule sets up model is included with setting up the answer rule:
Obtain first positive sample/first negative sample true relative to the first True Data/the second of the item attribute
Real data;
The negative sample of the first positive sample described in model prediction/first is set up using the answer rule relative to the project to belong to
The prediction data of first prediction data of property/second;
The model parameter that the answer rule sets up model is adjusted, when first prediction data and the first True Data one
Cause and the second prediction data is consistent with the second True Data, export the model parameter and its answer rule sets up model.
Optionally, the answer rule sets up model and is based on machine learning algorithm.
Optionally, the machine learning algorithm is in NB Algorithm, GBDT algorithms, SVM algorithm and RNN algorithm
It is a kind of.
Optionally, project carries out participle and finds the target item for possessing all participles in answer storehouse the problem of described pair of input
Mesh includes:
The project the problem of input is divided into the sequence of the single word composition of order;
If item-title possesses each word in the sequence in the answer storehouse, the project of the item-title is selected
It is used as the destination item.
Optionally, the described problem project and the synthesis pertinence of destination item of calculating includes:
Respectively the single degree of correlation between described problem project and destination item is calculated using some degree of correlation functions;
Weight is distributed to the single degree of correlation that calculating is obtained and adds the accumulation of the weight and the corresponding single degree of correlation
To obtain the synthesis pertinence.
Optionally, it is described to calculate single between described problem project and destination item using some degree of correlation functions respectively
The degree of correlation at least includes the one or more for performing following several calculations:
Participle is carried out to destination item title and problem project, by word2vec instruments obtain the numerical value of each word to
Amount, then obtains the addition of vectors of these words the semantic vector and problem project semantic vector of item-title, finally calculates
Two vectorial cosine distances are stated, to obtain the first single degree of correlation;
Destination item title carries out participle with problem project, and the quantity for calculating the two shared word accounts for the ratio of the two all word
Example, to obtain the second single degree of correlation;
The keyword of destination item title is obtained, the numerical value vector of each keyword is obtained by word2vec instruments, so
These vectors are added and are used as the semantic vector of destination item title afterwards;Participle is carried out to problem project, then passed through
Word2vec instruments obtain the numerical value vector of each word and add and as the semantic vector of problem project, finally calculate destination item
Cosine distances between title and problem project semantic vector, to obtain the 3rd single degree of correlation.
Optionally, the method for automatically generating open-ended question answer also includes:
Meet the item-title of described problem project described in artificial screening from the answer storehouse and do not meet described problem
The item-title of project is using as the second positive sample and the second negative sample;
It is individual character sequence to split described problem project, the second positive sample and the second negative sample, and is based on obtaining after RNN is handled
The semantic vector of each sequence;
Whether training RNN models meet true correlation degree with the cosine distances between the above-mentioned each semantic vector of Accurate Prediction,
To obtain RNN model parameters;
The described problem project and the synthesis pertinence of destination item of calculating includes:
Destination item title and problem project are carried out participle to obtain individual character sequence;
Based on the RNN model parameters using mark item-title and problem project individual character sequence described in RNN model predictions it
Between the prediction degree of correlation.
Optionally, the method for automatically generating open-ended question answer also includes:At least extract synthesis pertinence highest
Destination item in object content issued.
The beneficial effect of technical solution of the present invention at least includes:
Technical solution of the present invention can solve the problem that in the prior art the problem of community's question and answer can not be replied provides the user
A kind of technical scheme that answer is effectively automatically generated with reference to network homegrown resource, not only solves network configuration resources idle
Problem, also improves the validity and enthusiasm of community's question-answering environment.
Technical solution of the present invention effectively confirms that optimum network is answered by setting up answer rule and project indicator for coherence
Case, and can select to be combined there is provided a kind of effective technical scheme for obtaining problem answers with user, improve answer and problem
The degree of accuracy of adaptation.
Technical solution of the present invention can be greatly reduced in Ask-Answer Community zero in the case where input resource is less and answer a question
Ratio, while the answer automatically generated can to the user that answer a question is prepared as reference, also reduce write it is high-quality
The threshold of answer, improves Consumer's Experience.
Brief description of the drawings
A kind of method flow schematic diagram for automatically generating open-ended question answer that Fig. 1 provides for technical solution of the present invention;
A kind of method flow schematic diagram for setting up answer rule that Fig. 2 provides for technical solution of the present invention;
Fig. 3 shows for the method flow that a kind of training answer rule that technical solution of the present invention is provided sets up model
It is intended to;
Fig. 4 for technical solution of the present invention provide it is a kind of to input the problem of project carry out participle and find in answer storehouse to have
The method flow schematic diagram of the destination item of standby all participles;
Fig. 5 a kind of calculates described problem project and destination item for what technical solution of the present invention was provided based on artificial rule
The method flow schematic diagram of synthesis pertinence;
Another method flow signal for automatically generating open-ended question answer that Fig. 6 provides for technical solution of the present invention
Figure;
Fig. 7 is based on machine learning method for one kind that technical solution of the present invention is provided and calculates described problem project and target item
The method flow schematic diagram of purpose synthesis pertinence;
Fig. 8 illustrates for another method flow for automatically generating open-ended question answer that technical solution of the present invention is provided
Figure;
Fig. 9 illustrates for another method flow for automatically generating open-ended question answer that technical solution of the present invention is provided
Figure.
Embodiment
The automatic answer for open network problem can not be realized in the prior art, if there can be one kind to be effectively directed to
The answer automatic generation method of open-ended question, it is possible to which the problem of alleviating many users in Ask-Answer Community significantly cannot be timely
The phenomenon of answer.If the accuracy of the automatic answer generation method can not also directly reply quizmaster, reference can also be used as
The system that official documents and correspondence is sent to thinks the user for having the ability to answer former problem, and answer user just can be with output high-quality by simple editing
Answer, is higher by much than writing an answer efficiency one's very own, and ratio also can accordingly decline the problem of without any answer.This
Inventive technique scheme proposes a kind of method for automatically generating answer, can include root in article, the content library of video from one
Candidate answers are extracted according to customer problem, quizmaster both can be directly replied, be invited to potential can also be sent to as material
Master is answered, its cost for writing answer is improved.
According to the overall invention thinking of technical solution of the present invention, it sets up answer firstly the need of by obtaining a content library
Storehouse, answer storehouse foundation need from content library selection meet answer rule project, i.e., " quasi- answer project ", hereafter further according to
The problem of user proposes project, realizes the process of answer Auto-matching.Technical solution of the present invention firstly the need of getting content library,
Its content-form can include article and short-sighted frequency, secondly need to excavate in content library and be suitable as the content of answer as answering
Case storehouse, the problem of then being given according to user, finds most suitable candidate answers by way of retrieval in answer storehouse;To waiting
Select answer that predetermined certainty factor algebra is set to calculate the confidence level of candidate answers, selection confidence level is higher than the candidate answers of threshold value
Directly reply answer or using candidate answers as material be sent to it is potential answer it is main for its reference.
The specific implementation process of technical solution of the present invention is elaborated below in conjunction with accompanying drawing.
A kind of method for automatically generating open-ended question answer according to Fig. 1, comprises the following steps:
Step S100, the project obtained in content library, the content library at least includes item source, the contents of a project and project
An attribute in title.
Step S101, sets up answer rule and is obtained from the content library and meet the project of answer rule to set up
Answer storehouse, the answer rule includes screening bar at least corresponding with one kind in the item source, the contents of a project, item-title
Part;
Step S102, project carries out participle and finds the target item for possessing all participles in answer storehouse the problem of to input
Mesh;
Step S103, calculates the synthesis pertinence of described problem project and destination item and finds out synthesis pertinence highest
Destination item is to export.
According to step S100, the acquisition content library includes:Built according to the document or video of internal offer or outside crawl
Project in vertical content library.In order to set up answer storehouse, it is necessary first to there is a content library comprising article or short-sighted frequency.These
Content can be captured from internet, such as the software based on some crawl internet datas is in specific internet sites
Content carries out matching crawl, and the content library is constantly updated to set up new content library, and be updated the data according to internet.It is another
Situation is that the operator of Ask-Answer Community also runs content platform simultaneously, there is legal content obtaining channel.Such as top question and answer are real
Product under today's tops on border, itself have tens million of legal short-sighted frequencies and article can as answer storehouse source, i.e.,
The amalgamation of content of platform can be directly based upon and content library is got based on platform database.In general, content library needs to contain
Lid field is more wide better, and so giving problem has more maximum probability to find suitable answer or material.There is content library, the present invention
Technical scheme goes out the content item possible as problem answers i.e. according to step S101 by default answer Rules Filtering,
Answer storehouse is set up with this.
Under a kind of change case of technical solution of the present invention, according to step S101, the answer rule of setting up includes:From interior
Screening meets the project of at least one in following condition in Rong Ku:
The technorati authority of the item source is legal;
The timeliness of the contents of a project is legal;
Entity word is included in the item-title;
The item-title is interrogative sentence.
The change case is provided in a kind of relatively simple scheme for setting up answer rule, such scheme, the condition of screening
Can be the screening content of a variety of item attributes arranged side by side, to form the answer rule as defined in one group of screening conditions, screening
Condition can also be a kind of screening content of item attribute, and the project thus set up in a variety of answer rules, answer storehouse needs
Satisfaction answer as defined in one group of screening conditions is regular or meets a variety of answers rules simultaneously.
It is given below a kind of fairly simple to set up answer in the case of answer rule is as defined in one group of screening conditions
The application examples of rule:
The screening process of this group of answer rule settings can be:
Whether identify the ageing whether legal of the contents of a project, the i.e. contents of a project is intellectual long timeliness content;
The type for recognizing item-title is interrogative sentence or declarative sentence;
Recognize in the item-title and whether include entity word;
Recognize whether the length of item-title exceedes predetermined length;
Whether the technorati authority in contents of a project source is higher than predetermined value;
Whether recognition result meets predetermined the selection result, if so, then regarding as screening programme to meet the answer rule
Project then, and it is attributed to answer library item mesh.Such as only screening meets the content of one of following condition as candidate answers, i.e.,
The predetermined the selection result of setting is as follows:
Predetermined the selection result one:(title includes query to (long timeliness content) AND (clear and definite entity is included in title) AND
Sentence);
Predetermined the selection result two:(long timeliness content) AND (clear and definite entity is included in title) AND (title is declarative sentence)
AND (source technorati authority is higher than predetermined value).
For example, the item-title of wherein one article is in content library《Do then mobile phone company all go where3 points
Clock is understood them and lived by what》, judge it is that entity " mobile phone " is included in long timeliness content, title through ageing identification model, and
Have a question sentence, meets predetermined the selection result one, therefore can be screened as candidate answers, meets answer rule, and add and answer
Case storehouse.If user asks some that such as " company for making mobile phone then is a lot, now all in What for" " mobile phone industry has those
The case changed one's profession", this article is exactly a valuable answer.
Recognize that article is ageing according to content of text in the above method, recognize the Entity recognition in the clause of title, title
All it is ripe text analysis technique, will not be repeated here.Technorati authority of originating is usually artificial setting, if what content library was related to
Source is more, it is necessary to certain workload is counted and is classified to the source that content library is related to, a kind of divided originating
The mode of level may be referred to be specified below, but it should be recognized that it is only a kind of example to be specified below, according to different sources and
Classification is required, can there is other different hierarchical approaches:
System has made grading to technorati authority of being originated in internet project in advance, such as is commented for the source degree of governmental site
Level is highest, i.e., 10 grade, and degree grading in public institution's website source is taken second place, i.e., 9 grades, official's speech in social media site
The grading of source degree again, i.e., 8 grades, assert from Media Burst and by official, is rated 7 by of a mass character in social media site
Level, of a mass character in social media site is from Media Burst and hop count and number of visits reach high pre-determined number, is rated 6
Level, of a mass character in social media site is from Media Burst and hop count and number of visits reach middle pre-determined number, is rated 5
Level, of a mass character in social media site is from Media Burst and hop count and number of visits reach low pre-determined number, is rated 4
Level, of a mass character in social media site is from Media Burst and number of visits reaches pre-determined number, is rated 3 grades, it is of a mass character from
Media site and number of visits reach pre-determined number, are rated 2 grades, are otherwise rated 1 grade.
According to step S101, the answer rule of setting up can be using a kind of complex but higher side of screening accuracy
Formula, i.e., train the model of an automatic screening candidate answers using the method for machine learning, then with this model in content library
The contents of a project are made whether to meet the prediction of answer rule, and judge whether the project in content library meets according to predicting the outcome
Answer rule.
Under another change case of technical solution of the present invention, as shown in Fig. 2 the answer rule of setting up includes following step
Suddenly:
Step S200, the project of the answer rule is met from the content library and is not met described described in artificial screening
The project of answer rule is using as the first positive sample and the first negative sample;
Step S201, by the attribute of the project, the first positive sample and the first negative sample input answer rule set up model,
Train the answer rule to set up model to set up the answer rule.
According to step S200, the process of wherein Screening Samples includes:It is first artificial from content library in advance to mark certain amount
Can as candidate answers content (i.e. the first positive sample) and be less properly used as candidate answers (i.e. the first negative sample), one
As each class sample in sample set, i.e. the first positive sample and the first negative sample, at least need each class respectively to mark thousands of individuals
The sample of work screening.
After above-mentioned sample is obtained, further according to step S201, these samples are inputted machine learning model, namely described answered
Case rule sets up model, such as model-naive Bayesian, GBDT models, SVM models etc., how allows model oneself study according to interior
Every attribute of appearance distinguishes positive sample and negative sample.These models can be for differentiating whether the content can be with after study terminates
It is used as the candidate answers of some problems.Contents attribute mentioned here includes the first based on referring in artificial regular method
Explicit attribute, such as, ageing (whether be intellectual long timeliness content) of content, title type (be interrogative sentence or
Declarative sentence, if include univocal entity), the length of title, the technorati authority in source.The category being understood by except these
Property outside, may be incorporated into some do not allow it is readily understood, but may to machine learning model strengthen recognition capability it is helpful multiple
Miscellaneous attribute.Content title is such as converted into numerical value vector with RNN (Recognition with Recurrent Neural Network), individual layer is used as with this numerical value vector
The input of neutral net, obtains the probability P of " title may answer a problem ", and this probability P can also be as interior
The attribute held.For above-mentioned attribute, training pattern is exported to attribute forecast result, can be to above-mentioned prediction knot in application
Fruit is compared with predetermined result, to obtain the content library project for meeting answer rule, so as to set up answer storehouse.Specifically
, according to step S201, as shown in figure 3, the training answer rule is set up model and included with setting up the answer rule
Following steps:
Step S300, obtains first truly number of the negative sample of first positive sample/first relative to the item attribute
According to the/the second True Data;
Step S301, using the answer rule set up the negative sample of the first positive sample described in model prediction/first relative to
The prediction data of first prediction data of the item attribute/second;
Step S302, adjusts the model parameter that the answer rule sets up model, when first prediction data and first
True Data is consistent and the second prediction data is consistent with the second True Data, exports the model parameter and its answer rule is built
Formwork erection type.
Wherein, it is described to obtain first truly number of the negative sample of first positive sample/first relative to the item attribute
Include according to the/the second True Data:First True Data of first positive sample relative to the item attribute is obtained, i.e., such as
Item attribute for it is above-mentioned the first based on the explicit attribute referred in artificial regular method, first True Data is directed to it
In predetermined the selection result one be:Be, be, being (be by binary data:111);And obtain the first negative sample phase
For the second True Data of the item attribute, second True Data is for predetermined the selection result one therein:
It is no, no, no (to be by binary data:000).
It is described to set up the negative sample of the first positive sample described in model prediction/first relative to the item using the answer rule
The prediction data of first prediction data of mesh attribute/second includes:First is being set up described in model prediction using the answer rule just
Sample is relative to the first prediction data of the item attribute, such as the first prediction data is for predetermined the selection result therein
One is:It is, is, no (is by binary data:110);And, set up described in model prediction first using the answer rule negative
Sample is relative to the second prediction data of the item attribute, and second prediction data is for predetermined the selection result therein
One is:It is no, be, no (be by binary data:010).
The adjustment answer rule sets up the model parameter of model, according to examples detailed above, i.e., including being answered described in adjustment
The model parameter that case rule sets up model makes 110 prediction numerical value accurately can be output as 111 in prediction, makes 010 prediction number
Value accurately can be output as 000 in prediction.What above-mentioned training process was just as each sample data, finally in all samples
First prediction data of notebook data is consistent with the first True Data and the second prediction data is consistent with the second True Data, then exports
The model parameter and its answer rule set up model.
Technical solution of the present invention can be at internal operation or outside for step S300~S302 flow
What reason was obtained.According to above-mentioned steps S300~S302, the answer rule sets up model and is based on machine learning algorithm.Specifically,
The machine learning algorithm is one kind in NB Algorithm, GBDT algorithms, SVM algorithm and RNN algorithms.In given one group
Hold attribute and a large amount of positive samples, negative sample, how training machine learning model obtains an effective candidate answers identification model,
A ripe problem being adequately addressed in machine learning field, thus this process particular technique details herein not
Repeating.
According to the above of technical solution of the present invention, it has been described that candidate sets up the flow in candidate answers storehouse.Need
Bright is:
If the operator of Ask-Answer Community has possessed, a class is more complete, and abundant in content content platform is (such as
The operator of top question and answer product possesses daily newly-increased hundreds of thousands article simultaneously, and the top news number of video is from media platform), then Fig. 1
In obtaining step S100 can be directly based upon platform database and acquire, substantial amounts of article is otherwise then captured from internet
The project come with the video (can be short-sighted frequency) with description information in constitution content storehouse, and realize self-renewing.
Technical solution of the present invention additionally provides screening and meets the content library project of answer rule to set up the technology in answer storehouse
Means, including the technological means for directly designing one or more candidate answers rules and by training machine learning model come
The technological means of answer rule is set up, two kinds of means can be used alone in technical scheme, can also use parallel respectively.
In addition, in the technological means for directly designing one or more candidate answers rules, if there are a variety of candidate answers
When regular, it can also be needed according to the screening effect under different situations to different candidate answers rule setting weights, using each
The accumulation result of answer rule judgment result and the product of weight is selected to carry out reality as the judged result of final content library project
Screening.
The means that technical solution of the present invention sets up answer storehouse are not limited by above-mentioned technical proposal.
According to the above of technical solution of the present invention, after candidate answers storehouse is established, give a user and propose
The problem of, most suitable candidate answers can just be searched in answer storehouse according to step S102, specifically, according to step S102,
Technical solution of the present invention provides a kind of feasible answer automatic Matching means, as shown in figure 4, described ask input
Topic project, which carries out participle and finding possesses all participles destination item in answer storehouse, includes step:
Step S400, the project the problem of input is divided into the sequence of the single word composition of order;
Step S401, if item-title possesses each word in the sequence in the answer storehouse, selects the project
The project of title is used as the destination item.
It should be noted that according to step S400, the number of characters one that its general length of project is included the problem of the input
As can be shorter, i.e., it is similar the problem of be a word or the problem of a few words, may be with input problem project length in length
Similar meeting is the item-title in answer storehouse, therefore the process of answer Auto-matching may be limited to problem project and project herein
Title directly matches hand.But it is understood that, if comparison of item is long the problem of input, its content covered can also compare
Many, the problem of typically entering project can also have title, and the title and item-title that can now use input problem project are made
Matching.That is step S400 meaning input the problem of project input the problem of content-length exceed predetermined content length when, choose
Problem project title as project the problem of the input, if but input the problem of content-length not less than predetermined content length
When, then it regard described problem length as project the problem of the input.
, can be according to calculating the comprehensive of described problem project and destination item after destination item is found according to step S103
Close the degree of correlation to speculate the match condition of problem project and destination item, so as to realize Auto-matching answer.The technology of the present invention
Scheme provides the following two kinds numerical procedure to get above-mentioned synthesis pertinence.Specially:
The first numerical procedure is artificial regular method, as shown in figure 5, the calculating described problem project and target item
Purpose synthesis pertinence includes step:
Step S500, calculates the single-phase between described problem project and destination item using some degree of correlation functions respectively
Guan Du;
Step S501, distributes the obtained single degree of correlation of calculating weight and by the weight and the corresponding single degree of correlation
Accumulation obtained the synthesis pertinence.
It is described to be calculated respectively using some degree of correlation functions between described problem project and destination item according to step S500
The single degree of correlation at least include the one or more for performing following several calculations, such as:
The first calculation:Participle is carried out to destination item title and problem project, obtained by word2vec instruments
The numerical value vector of each word, the addition of vectors of these words is then obtained item-title semantic vector and problem project it is semantic to
Amount, finally calculates the cosine distances of above-mentioned two vector, to obtain the first single degree of correlation;
Second of calculation:Destination item title carries out participle with problem project, and the quantity for calculating the two shared word is accounted for
The ratio of the two all word, to obtain the second single degree of correlation;
The third calculation:The keyword of destination item title is obtained, obtains each crucial by word2vec instruments
The numerical value vector of word, then adds these vectors and be used as the semantic vector of destination item title;Participle is carried out to problem project,
Then the numerical value vector of each word is obtained by word2vec instruments and adds and as the semantic vector of problem project, finally calculate
Cosine distances between destination item title and problem project semantic vector, to obtain the 3rd single degree of correlation.
Wherein, Word2vec instruments are that a kind of popular term vector chemical industry has in the prior art, are developed by Google, can
So that word is represented a numerical value vector, and ensure the word of semantic similarity its numerical value vector distance also should calculate closer to, the instrument
Here is omitted for principle.
In the first extension example according to this numerical procedure, order:
Rel (c, q)=w1×f1(c,q)+w2×f2(c,q)+…+wn×fn(c,q)
Wherein, c and q represent destination item and problem project respectively, and rel (c, q) represents c and q synthesis pertinence, f1、
f2、…、fnSingle correlation function is represented respectively, and n is the natural number more than 2, w1、w2、…、wnRepresent correlation function f1、
f2、…、fnWeight, is manual setting.f1、f2、…、fnCalculation can using it is above-mentioned the first to the third calculating side
Formula, it would however also be possible to employ the calculation of other degree of correlation functions of the prior art.Those skilled in the relevant art can also set
More effective degree of correlation functions are counted out, as space is limited, be will not enumerate herein.
In the method that second of numerical procedure of technical solution of the present invention is machine learning, one kind as shown in Figure 6 is automatic
The method (method flow shown in Fig. 6 is based on Fig. 1) of open-ended question answer is generated, in addition to including step S100~S103,
Also comprise the following steps:
Step S600, meets the item-title of described problem project described in artificial screening from the answer storehouse and does not meet
The item-title of described problem project is using as the second positive sample and the second negative sample;
Step S601, it is individual character sequence to split described problem project, the second positive sample and the second negative sample, and based on RNN
The semantic vector of each sequence is obtained after processing;
Whether step S602, training RNN models meet true with the cosine distances between the above-mentioned each semantic vector of Accurate Prediction
The real degree of correlation, to obtain RNN model parameters.
Technical solution of the present invention can be at internal operation or outside for step S600~S602 flow
What reason was obtained.RNN (Recurrent Neutral Networks) models namely Recognition with Recurrent Neural Network model, are a kind of popular
Neural network structure, it inputs a character string, can obtain the potential language that a vector is used for representing this character string
Justice, there is more careful elaboration to the model in the prior art.Technical solution of the present invention carrys out training problem with the RNN models
Matching prediction between mesh and destination item title, so as to improve the degree of accuracy of matching prediction.
Step S103 in Fig. 6, with reference to Fig. 7, the calculating described problem project is comprehensive related to destination item
Degree includes step:
Step S700, carries out participle to obtain individual character sequence to destination item title and problem project;
Step S701, mark item-title described in RNN model predictions and problem project list are used based on the RNN model parameters
The prediction degree of correlation between word sequence.
According to step S700 to S701, it is known that, in technical solution of the present invention, for a customer problem q, can first it mark
Note content { c } (the desired title c matched with the problem the collection that this question answering is suitable as in a collection of candidate answers storehouse
Close), a collection of content { c ' } for being not suitable as this question answering is then marked (i.e. with the unmatched desired title c ' of the problem
Set).The pairing of q and all c compositions can be used as second as the second positive sample, q and all c ' compositions pairing
Negative sample.Generally require and thousands of different q are carried out with this operation, collect number with ten thousand grades of positive negative sample, then can just instruct
Practicing machine learning model, how study calculates rel (c, q) automatically.
The degree of correlation of candidate answers title and customer problem is learnt using popular RNN models in this example, here
Problem and candidate answers title all as individual character sequence (such as problem q for " why Sony, the Japanese enterprises such as Panasonic are in recent years
Decline" it is converted into following individual character sequence<For, it is assorted, rope, Buddhist nun, pine, under, etc., day, this, enterprise, industry closely, in year, is come,
Decline, fall,>), the semantic vector of sequence of question and the semantic vector of candidate answers title are obtained after RNN model treatments.If
The two is related, and the cosine distances between two vectors should be 1, if the two is uncorrelated, and its cosine distance should be 0.
Using classical BPTT algorithms, that is, consider cosine distances between the vector that is obtained using RNN model predictions and
The true error for calculating the cosine distances between obtained vector, RNN models are adjusted with this error in turn in sample
Model parameter, constantly adjustment model parameter make it that resulting predicted value and actual value error are less and less, it is possible to find
Suitable model parameter make it that RNN models are most strong to the predictive ability of (c, q) paired sample correlation.
Find after suitable model parameter, model just trains completion, i.e., just can be using suitable RNN models for any
(c, q) combination calculate the semantic vector and the semantic vector of candidate answers title of correlation, i.e. prediction sequence the problem of obtain
Between cosine distances, so as to judge match condition according to predicted value:First using c title and q as individual character sequence, then
Respective semantic vector is obtained with RNN model treatments respectively, then calculates cosine distances.The distance calculated is bigger, the degree of correlation
It is higher, so this computational methods can be maximally related for being found out from some candidate answers.The prediction degree of correlation is to refer to
According to the cosine distances between the semantic vector of RNN model computational problem sequences and the semantic vector of candidate answers title.
If the degree of accuracy of matching algorithm is higher (to need actual assessment, because the effect of artificial rule needs continuous adjustment
Advise to attempt to improve, it is not easy to obtain satisfied effect, the effect of machine learning method is by sample quality and the shadow of quantity
Ring), correlation can be directly returned to be published to by robot account behind problem higher than the candidate answers of specific threshold and asked
Community is answered, certainly such way has certain risk, safer way is that candidate answers are issued to be adapted to answer the problem
Real user, allows the user to judge whether candidate answers answer problem well.The other user can also be candidate answers
As material, a more preferable answer is replied in modification on this basis, can so be started anew to write efficiency than oneself and be carried out get Geng Gao.
Under another change case of technical solution of the present invention, (Fig. 8 is based on scheme shown in Fig. 1, implements certainly at other as shown in Figure 8
Example in can also be deformed based on scheme shown in Fig. 6), a kind of method for automatically generating open-ended question answer, except including
Outside step S100~S103, also comprise the following steps:Step S800, is at least extracted in synthesis pertinence highest destination item
Object content is issued.
The object of above-mentioned issue can be the user that the user of proposition problem or work done in the manner of a certain author go out answer.According to Fig. 8 institutes
Show an application examples for automatically generating the method for open-ended question answer to operate, as shown in figure 9, wherein customer problem comes from
Ask-Answer Community, is generally not that (such as " whom present US President is to factoid questions to simple factoid questions"), but need
(such as " Donald Trump can bring anything to become to the open problem for wanting more word or video to illustrate after taking up an official post to international situation
Change"), this method comprises the following steps:
It is to excavate candidate answers from a content library comprising magnanimity article or video content in advance and set up candidate first
Answer storehouse, next to that receiving after user's enquirement, the answer of matching is found from candidate answers storehouse and answer is issued.
With reference to Fig. 9, the answer of matching is found from candidate answers storehouse and is further comprised the step of issuing answer:
Step S900, participle is carried out to customer problem;
Step S901, according to the word set after participle, this finds the candidate answers that title includes these words in candidate answers storehouse;
Step S902, for each candidate answers, calculates the correlation of its title and customer problem;
Step S903, system needs to judge whether the answer degree of correlation automatically generated is higher than specific threshold;Higher than threshold value
Candidate answers are then considered as Top k answers;
For Top K answers, the user for being adapted to answer a question can be sent to according to step S904, i.e., it is potential to take main confession
It is referred to, and is write answer for it and is provided material;Can also directly it be issued in Ask-Answer Community using system account according to step S905
Answer.
Although the present invention is disclosed as above with preferred embodiment, it is not for limiting the present invention, any this area
Technical staff without departing from the spirit and scope of the present invention, may be by the methods and techniques content of the disclosure above to this hair
Bright technical scheme makes possible variation and modification, therefore, every content without departing from technical solution of the present invention, according to the present invention
Any simple modifications, equivalents, and modifications made to above example of technical spirit, belong to technical solution of the present invention
Protection domain.
Claims (12)
1. a kind of method for automatically generating open-ended question answer, it is characterised in that including:
The project obtained in content library, the content library at least includes one kind in item source, the contents of a project and item-title
Attribute;
Set up answer rule and the project for meeting the answer rule is obtained from the content library to set up answer storehouse, it is described to answer
Case rule include at least with a kind of corresponding screening conditions in the item source, the contents of a project, item-title;
Project carried out participle and found the destination item for possessing all participles in answer storehouse the problem of to input;
Calculate the synthesis pertinence of described problem project and destination item and find out synthesis pertinence highest destination item with defeated
Go out.
2. the method as claimed in claim 1 for automatically generating open-ended question answer, it is characterised in that the acquisition content library
Including:The project set up according to the document or video of internal offer or outside crawl in content library.
3. the method as claimed in claim 1 for automatically generating open-ended question answer, it is characterised in that described to set up answer rule
Then include:
Screening meets the project of at least one in following condition from content library:
The technorati authority of the item source is legal;
The timeliness of the contents of a project is legal;
Entity word is included in the item-title;
The item-title is interrogative sentence.
4. the method as claimed in claim 1 for automatically generating open-ended question answer, it is characterised in that described to set up answer rule
Then include:
Meet the project of the answer rule described in artificial screening from the content library and do not meet the item of the answer rule
Mesh is using as the first positive sample and the first negative sample;
The attribute of the project, the first positive sample and the first negative sample input answer rule are set up into model, the answer is trained
Rule sets up model to set up the answer rule.
5. the method as claimed in claim 4 for automatically generating open-ended question answer, it is characterised in that answered described in the training
Case rule sets up model to be included with setting up the answer rule:
Obtain first positive sample/first negative sample truly several relative to the first True Data/the second of the item attribute
According to;
The negative sample of the first positive sample described in model prediction/first is set up relative to the item attribute using the answer rule
The prediction data of first prediction data/second;
Adjust the model parameter that the answer rule sets up model, when first prediction data is consistent with the first True Data and
Second prediction data is consistent with the second True Data, exports the model parameter and its answer rule sets up model.
6. the method as claimed in claim 4 for automatically generating open-ended question answer, it is characterised in that the answer rule is built
Formwork erection type is based on machine learning algorithm.
7. the method as claimed in claim 6 for automatically generating open-ended question answer, it is characterised in that the machine learning is calculated
Method is one kind in NB Algorithm, GBDT algorithms, SVM algorithm and RNN algorithms.
8. the method as claimed in claim 1 for automatically generating open-ended question answer, it is characterised in that described to ask input
Topic project carries out participle and finds the destination item for possessing all participles in answer storehouse and include:
The project the problem of input is divided into the sequence of the single word composition of order;
If item-title possesses each word in the sequence in the answer storehouse, the project conduct of the item-title is selected
The destination item.
9. the method as claimed in claim 1 for automatically generating open-ended question answer, it is characterised in that asked described in the calculating
The synthesis pertinence of topic project and destination item includes:
Respectively the single degree of correlation between described problem project and destination item is calculated using some degree of correlation functions;
Weight is distributed to the single degree of correlation that calculating is obtained and is subject to the accumulation of the weight and the corresponding single degree of correlation
To the synthesis pertinence.
10. the method as claimed in claim 9 for automatically generating open-ended question answer, it is characterised in that described to use respectively
It is following several that the single degree of correlation that some degree of correlation functions are calculated between described problem project and destination item at least includes execution
The one or more of calculation:
Participle is carried out to destination item title and problem project, the numerical value vector of each word is obtained by word2vec instruments, so
The addition of vectors of these words is obtained afterwards the semantic vector and problem project semantic vector of item-title, finally calculate above-mentioned two
The cosine distances of vector, to obtain the first single degree of correlation;
Destination item title carries out participle with problem project, and the quantity for calculating the two shared word accounts for the ratio of the two all word, with
Obtain the second single degree of correlation;
Obtain the keyword of destination item title, the numerical value vector of each keyword obtained by word2vec instruments, then
These vectors add and are used as the semantic vector of destination item title;Participle is carried out to problem project, then passes through word2vec works
Tool obtains the numerical value vector of each word and adds and as the semantic vector of problem project, finally calculate destination item title and problem
Cosine distances between project semantic vector, to obtain the 3rd single degree of correlation.
11. the method as claimed in claim 1 for automatically generating open-ended question answer, it is characterised in that also include:
Meet the item-title of described problem project described in artificial screening from the answer storehouse and do not meet described problem project
Item-title using as the second positive sample and the second negative sample;
It is individual character sequence to split described problem project, the second positive sample and the second negative sample, and is based on obtaining each after RNN is handled
The semantic vector of sequence;
Whether training RNN models meet true correlation degree with the cosine distances between the above-mentioned each semantic vector of Accurate Prediction, to obtain
Obtain RNN model parameters;
The described problem project and the synthesis pertinence of destination item of calculating includes:
Destination item title and problem project are carried out participle to obtain individual character sequence;
Marked based on the RNN model parameters using described in RNN model predictions between item-title and problem project individual character sequence
Predict the degree of correlation.
12. the method as claimed in claim 1 for automatically generating open-ended question answer, it is characterised in that also include:At least carry
The object content in synthesis pertinence highest destination item is taken to be issued.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710205299.8A CN106991161B (en) | 2017-03-31 | 2017-03-31 | A method of automatically generating open-ended question answer |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710205299.8A CN106991161B (en) | 2017-03-31 | 2017-03-31 | A method of automatically generating open-ended question answer |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106991161A true CN106991161A (en) | 2017-07-28 |
CN106991161B CN106991161B (en) | 2019-02-19 |
Family
ID=59414626
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710205299.8A Active CN106991161B (en) | 2017-03-31 | 2017-03-31 | A method of automatically generating open-ended question answer |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106991161B (en) |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107766536A (en) * | 2017-10-30 | 2018-03-06 | 江西博瑞彤芸科技有限公司 | The searching method of related information |
CN107832768A (en) * | 2017-11-23 | 2018-03-23 | 盐城线尚天使科技企业孵化器有限公司 | Efficient method to go over files and marking system based on deep learning |
CN108052577A (en) * | 2017-12-08 | 2018-05-18 | 北京百度网讯科技有限公司 | A kind of generic text content mining method, apparatus, server and storage medium |
CN108681610A (en) * | 2018-05-28 | 2018-10-19 | 山东大学 | Production takes turns more and chats dialogue method, system and computer readable storage medium |
CN108717413A (en) * | 2018-03-26 | 2018-10-30 | 浙江大学 | It is a kind of based on the assumption that property semi-supervised learning Opening field answering method |
CN109189894A (en) * | 2018-09-20 | 2019-01-11 | 科大讯飞股份有限公司 | A kind of answer extracting method and device |
CN109284298A (en) * | 2018-11-09 | 2019-01-29 | 上海晏鼠计算机技术股份有限公司 | A kind of contents production system handled based on machine learning and big data |
CN109460453A (en) * | 2018-10-09 | 2019-03-12 | 北京来也网络科技有限公司 | Data processing method and device for positive negative sample |
CN109816111A (en) * | 2019-01-29 | 2019-05-28 | 北京金山数字娱乐科技有限公司 | Reading understands model training method and device |
CN110390050A (en) * | 2019-07-10 | 2019-10-29 | 北京航空航天大学 | A kind of software development question and answer automatic acquisition of scientific information method understood based on deep semantic |
US20200151542A1 (en) * | 2018-11-12 | 2020-05-14 | Alibaba Group Holding Limited | Question and answer matching method, system and storage medium |
CN111382247A (en) * | 2018-12-29 | 2020-07-07 | 深圳市优必选科技有限公司 | Content pushing optimization method, content pushing optimization device and electronic equipment |
CN111985238A (en) * | 2020-06-30 | 2020-11-24 | 联想(北京)有限公司 | Answer generation method and equipment |
CN113841142A (en) * | 2019-04-30 | 2021-12-24 | 艾斯图文有限公司 | Method, apparatus and computer program for providing network security using knowledge graph |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101566998A (en) * | 2009-05-26 | 2009-10-28 | 华中师范大学 | Chinese question-answering system based on neural network |
CN103902652A (en) * | 2014-02-27 | 2014-07-02 | 深圳市智搜信息技术有限公司 | Automatic question-answering system |
CN104102721A (en) * | 2014-07-18 | 2014-10-15 | 百度在线网络技术(北京)有限公司 | Method and device for recommending information |
US20150235130A1 (en) * | 2014-02-19 | 2015-08-20 | International Business Machines Corporation | NLP Duration and Duration Range Comparison Methodology Using Similarity Weighting |
CN105893391A (en) * | 2015-01-26 | 2016-08-24 | 阿里巴巴集团控股有限公司 | Intelligent answering method, apparatus and system, and electronic device |
US20160267383A1 (en) * | 2015-03-10 | 2016-09-15 | International Business Machines Corporation | Enhancement of massive data ingestion by similarity linkage of documents |
-
2017
- 2017-03-31 CN CN201710205299.8A patent/CN106991161B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101566998A (en) * | 2009-05-26 | 2009-10-28 | 华中师范大学 | Chinese question-answering system based on neural network |
US20150235130A1 (en) * | 2014-02-19 | 2015-08-20 | International Business Machines Corporation | NLP Duration and Duration Range Comparison Methodology Using Similarity Weighting |
CN103902652A (en) * | 2014-02-27 | 2014-07-02 | 深圳市智搜信息技术有限公司 | Automatic question-answering system |
CN104102721A (en) * | 2014-07-18 | 2014-10-15 | 百度在线网络技术(北京)有限公司 | Method and device for recommending information |
CN105893391A (en) * | 2015-01-26 | 2016-08-24 | 阿里巴巴集团控股有限公司 | Intelligent answering method, apparatus and system, and electronic device |
US20160267383A1 (en) * | 2015-03-10 | 2016-09-15 | International Business Machines Corporation | Enhancement of massive data ingestion by similarity linkage of documents |
Cited By (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107766536A (en) * | 2017-10-30 | 2018-03-06 | 江西博瑞彤芸科技有限公司 | The searching method of related information |
CN107832768A (en) * | 2017-11-23 | 2018-03-23 | 盐城线尚天使科技企业孵化器有限公司 | Efficient method to go over files and marking system based on deep learning |
CN108052577B (en) * | 2017-12-08 | 2022-06-14 | 北京百度网讯科技有限公司 | Universal text content mining method, device, server and storage medium |
CN108052577A (en) * | 2017-12-08 | 2018-05-18 | 北京百度网讯科技有限公司 | A kind of generic text content mining method, apparatus, server and storage medium |
US11062090B2 (en) | 2017-12-08 | 2021-07-13 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method and apparatus for mining general text content, server, and storage medium |
CN108717413A (en) * | 2018-03-26 | 2018-10-30 | 浙江大学 | It is a kind of based on the assumption that property semi-supervised learning Opening field answering method |
CN108717413B (en) * | 2018-03-26 | 2021-10-08 | 浙江大学 | Open field question-answering method based on hypothetical semi-supervised learning |
CN108681610B (en) * | 2018-05-28 | 2019-12-10 | 山东大学 | generating type multi-turn chatting dialogue method, system and computer readable storage medium |
CN108681610A (en) * | 2018-05-28 | 2018-10-19 | 山东大学 | Production takes turns more and chats dialogue method, system and computer readable storage medium |
CN109189894A (en) * | 2018-09-20 | 2019-01-11 | 科大讯飞股份有限公司 | A kind of answer extracting method and device |
CN109189894B (en) * | 2018-09-20 | 2021-03-23 | 科大讯飞股份有限公司 | Answer extraction method and device |
CN109460453A (en) * | 2018-10-09 | 2019-03-12 | 北京来也网络科技有限公司 | Data processing method and device for positive negative sample |
CN109460453B (en) * | 2018-10-09 | 2021-08-17 | 北京来也网络科技有限公司 | Data processing method and device for positive and negative samples |
CN109284298A (en) * | 2018-11-09 | 2019-01-29 | 上海晏鼠计算机技术股份有限公司 | A kind of contents production system handled based on machine learning and big data |
US11599769B2 (en) * | 2018-11-12 | 2023-03-07 | Alibaba Group Holding Limited | Question and answer matching method, system and storage medium |
US20200151542A1 (en) * | 2018-11-12 | 2020-05-14 | Alibaba Group Holding Limited | Question and answer matching method, system and storage medium |
CN111382247A (en) * | 2018-12-29 | 2020-07-07 | 深圳市优必选科技有限公司 | Content pushing optimization method, content pushing optimization device and electronic equipment |
CN111382247B (en) * | 2018-12-29 | 2023-07-14 | 深圳市优必选科技有限公司 | Content pushing optimization method, content pushing optimization device and electronic equipment |
CN109816111B (en) * | 2019-01-29 | 2022-03-08 | 北京金山数字娱乐科技有限公司 | Reading understanding model training method and device |
CN109816111A (en) * | 2019-01-29 | 2019-05-28 | 北京金山数字娱乐科技有限公司 | Reading understands model training method and device |
CN113841142A (en) * | 2019-04-30 | 2021-12-24 | 艾斯图文有限公司 | Method, apparatus and computer program for providing network security using knowledge graph |
CN113841142B (en) * | 2019-04-30 | 2023-09-08 | 艾斯图文有限公司 | Method, apparatus and computer program for providing network security using knowledge graph |
CN110390050B (en) * | 2019-07-10 | 2021-12-07 | 北京航空航天大学 | Software development question-answer information automatic acquisition method based on deep semantic understanding |
CN110390050A (en) * | 2019-07-10 | 2019-10-29 | 北京航空航天大学 | A kind of software development question and answer automatic acquisition of scientific information method understood based on deep semantic |
CN111985238A (en) * | 2020-06-30 | 2020-11-24 | 联想(北京)有限公司 | Answer generation method and equipment |
Also Published As
Publication number | Publication date |
---|---|
CN106991161B (en) | 2019-02-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106991161B (en) | A method of automatically generating open-ended question answer | |
CN110674840B (en) | Multi-party evidence association model construction method and evidence chain extraction method and device | |
Durairaj et al. | Educational data mining for prediction of student performance using clustering algorithms | |
Jin et al. | An improved ID3 decision tree algorithm | |
CN105808590B (en) | Search engine implementation method, searching method and device | |
CN107368468A (en) | A kind of generation method and system of O&M knowledge mapping | |
CN108717433A (en) | A kind of construction of knowledge base method and device of programming-oriented field question answering system | |
CN110457404A (en) | Social media account-classification method based on complex heterogeneous network | |
Chandra et al. | Biden vs trump: modeling us general elections using bert language model | |
Zhong et al. | Design of a personalized recommendation system for learning resources based on collaborative filtering | |
KR20200047006A (en) | Method and system for constructing meta model based on machine learning | |
CN112926325A (en) | Chinese character relation extraction construction method based on BERT neural network | |
CN111241425A (en) | POI recommendation method based on hierarchical attention mechanism | |
CN115827968A (en) | Individualized knowledge tracking method based on knowledge graph recommendation | |
CN117473034A (en) | Interactive text processing method and device, electronic equipment and storage medium | |
CN117171428B (en) | Method for improving accuracy of search and recommendation results | |
CN117494760A (en) | Semantic tag-rich data augmentation method based on ultra-large-scale language model | |
Chen | Quality evaluation of student education management work based on wireless network data mining | |
Elhebir et al. | A novel ensemble approach to enhance the performance of web server logs classification | |
Saeed et al. | A framework to predict early news popularity using deep temporal propagation patterns | |
CN111143573A (en) | Method for predicting target node of knowledge graph based on user feedback information | |
Gao et al. | Statistics and Analysis of Targeted Poverty Alleviation Information Integrated with Big Data Mining Algorithm | |
CN114942981A (en) | Question-answer query method and device, electronic equipment and computer readable storage medium | |
Wang | Retracted: Multi‐data multiple gray clustering analysis based on layered mining for ubiquitous clouds and social internet of things | |
Ahmed et al. | Assessing public sentiment towards digital India through Twitter sentiment analysis: a comparative study |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CP03 | Change of name, title or address |
Address after: Room 222, Floor 2, Building 1, Yard A23, North Third Ring West Road, Haidian District, Beijing 100098 Patentee after: Beijing Douyin Information Service Co.,Ltd. Address before: 100000 Building 1, AVIC Plaza, 43 North Third Ring West Road, Haidian District, Beijing Patentee before: BEIJING BYTEDANCE TECHNOLOGY Co.,Ltd. |
|
CP03 | Change of name, title or address |