CN105117398A - Software development problem automatic answering method based on crowdsourcing - Google Patents

Abstract

The invention relates to a software development problem automatic answering method based on crowdsourcing; the method uses tag-LDA to do label recommendation for a question submitted by a user, matches with historical question-answer pairs, selects related question-answer pairs, pre-processes all questions and answers, matches answer content with the question proposed by the user, calculates content similarity and credibility, respectively ranks the similarity, credibility and adoption level according to adoption level of the feedback calculation answer, and recommends answers and presenter information for the user. The method can solve the problems that user urgent needs for the answer cannot be directly processed; the method uses questions and answers (crowdsourcing) in historical database to dig out proper answers and presenter information for the user, thus greatly reducing information overwhelm possibility, and recommends the related answer and presenter information for the user from angles of content similarity, presenter credibility and answer adoption level.

Description

A kind of software development problem auto-answer method based on mass-rent

Technical field

The present invention proposes a kind of software development problem auto-answer method based on mass-rent, belong to software data and excavate and analysis field.

Background technology

Along with progress and the arrival in electronization epoch of Information technology, people are more prone on network, seek way to solve the problem and approach.Mass-rent is exactly the new productive organization that internet brings.Mass-rent is a kind of distributed Resolving probiems and production model.Problem is propagated to the solution provider colony of the unknown in the mode of open bidding.User's (referring to " crowd " in mass-rent here) typically forms on-line communities and submits scheme to.Group " crowd " also will examine scheme, finds best.These best schemes are finally owned by the side asked a question at first (mass-rent people, crowdsourcer), and the individual won in group " crowd " is rewarded sometimes.

Before the present invention makes, popular along with mass-rent (crowd) pattern, the software forums such as such as StackOverflow, Sourceforge are subject to liking of increasing developer.The number of users of Q & A website is also increasing, and along with the sharp increase of number of users, the data on website are also doubled and redoubled, and this problem also just causing user's submission can not be solved timely, causes the phenomenon that information is submerged.In research before, concentrate on the research of problem that user is proposed, comprise the quality of problem, not by the feature of problem answered, be intended to help user to propose to be easier to understand and by the problem answered.Research before does not directly solve the active demand that user checks on one's answers, and only improve the quality of problem, and not only there is problem Q & A website, have ignored the research checked on one's answers.

Summary of the invention

Object of the present invention is just to overcome above-mentioned defect, develops a kind of software development problem auto-answer method based on mass-rent.

Technical scheme of the present invention is:

Based on a software development problem auto-answer method for mass-rent, its major technique step is as follows:

(1) with tag-LDA, label recommendations is carried out to the problem that user submits to;

(2) by label with the historical problem answer in problem base StackOverflow to mating, select relevant issues answer pair;

(3) in answer quality evaluation mechanism, pre-service is carried out to all problems and answer, the problem that the content of answer and user propose is carried out content matching, calculates content similarity;

(4) show according to the history of answer presenter, calculate the confidence level of answer;

(5) according to the feedback of other users and problem originator, the degree of adopting of answer is calculated;

(6) according to the similarity that evaluation mechanism calculates, confidence level and degree of adopting three angles sort respectively;

(7) according to ranking results, for user recommends out answer and presenter's information.

The computing formula of described step (2) tag match is as follows:

Similarity=same label number/all label numbers

And by similarity higher than 0.30 problem answers to derivation.

Described step (3) preprocessing process comprises the following steps:

A) numeral is removed;

B) with the portmanteau word having lower stroke short-term to be connected, participle is carried out according to hump rule to some;

C) English stop words is removed;

D) the multi-form of word is normalized;

The computing formula of content similarity is as follows:

$\mathrm{similarity}=\cos \left(\mathrm{\θ}\right)=\frac{A\·B}{\left|\right|A\left|\right|\left|\right|B\left|\right|}=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{A}_i\×B_i}{\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\left(A_i\right)}^2}\×\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\left(B_i\right)}^2}}$

Wherein A, B are the quantization means representing document one and document two.Document one and document two, through participle, remove stop words, remove numeral, the preprocessing process such as root, form vectorial A, B after being quantized in certain sequence by remaining word.In information retrieval, each entry has different degree, and a document represents by there being the proper vector of weights by one, and the frequency that entry occurs in the document is depended in the calculating of weights.Therefore cosine similarity can provide the similarity of two sections of its theme aspects of document.

The computing formula of described step (4) answer confidence level is as follows:

Confidence level=label score/total fame.

The computing formula of described step (5) answer degree of adopting is as follows:

Degree of adopting=vote value/view value.

In addition, it is understood that StackOverflow can measure vote maximum as bestanswer, problem originator can select oneself favorite answer as acceptanswer as in addition, the degree of adopting of these two kinds of answers apparently higher than other answers, so be respectively bestanswer and acceptanswer to add 2 vote.

Advantage of the present invention and effect:

(1) the at present research relevant to mass-rent knowledge has a lot, but does not have for the auto answer of software development problem, and the present invention finally exports is associated answer and presenter's information thereof.

(2) the present invention is from content similarity, and presenter's confidence level and answer degree of adopting three angles provide user associated answer, adapt to different user's requests.

(3) the present invention not only recommends out associated answer, finally also can recommend out presenter's information of answer, so as user do not find satisfied answer or checked on one's answers partial question time, directly link up with presenter, reach the final purpose of dealing with problems.

The present invention is a kind of software development problem auto answer recommended technology based on mass-rent knowledge, utilize existing problem and answer (i.e. mass-rent knowledge) in history library, excavate suitable answer and presenter's information recommendation to user, the possibility that the information of considerably reducing is flooded.

The present invention finally can from content similarity, presenter's confidence level, answer degree of adopting three angles are presenter's information that user recommends associated answer and answer, be conducive to software developer achieve a solution faster when using Q & A website problem method and select to meet the answer of oneself demand, the problem of reducing can not get the possibility answered.

Accompanying drawing explanation

Fig. 1---the overall flow schematic diagram of software development problem auto-answer method.

Fig. 2---the schematic diagram of Tag-LDA model, for recommending label.

Fig. 3---the customer problem example schematic diagram that the upper user of StackOverflow submits to.

Fig. 4---an answer example schematic diagram on StackOverflow.

Fig. 5---an answer example schematic diagram on StackOverflow.

Fig. 6---the homepage of the upper user of StackOverflow, for calculating User reliability schematic diagram.

Fig. 7---the homepage of the upper user of StackOverflow, for calculating User reliability schematic diagram.

Embodiment

Technical thought of the present invention is:

The present invention has applied to Tag-LDA topic model, and to document, word, label carries out modeling.Tag-LDA extends the one of LatentDirichletAllocation model.By Tag-LDA topic model, the multiple labels relevant with document content can be recommended, and the probability of each label and article degree of correlation is estimated, if Fig. 2 is the schematic diagram of Tag-LDA topic model.After Output rusults, sort according to probability height, pick out the conduct expressing document content and recommend label.The present invention uses Tag-LDA to come for label recommended by problem document.

Illustrate the present invention below.

The invention provides a kind of software development problem auto answer recommended technology based on mass-rent knowledge, below in conjunction with accompanying drawing, technical scheme of the present invention be described in detail:

Program as shown in Figure 1:

(1) with tag-LDA, label recommendations is carried out to the problem that user submits to;

(2) by label with the historical problem answer in problem base StackOverflow to mating, select relevant issues answer pair;

(3) in answer quality evaluation mechanism, pre-service is carried out to all problems and answer, the problem that the content of answer and user propose is carried out content matching, calculates content similarity;

(4) show according to the history of answer presenter, calculate the confidence level of answer;

(5) according to the feedback of other users and problem originator, the degree of adopting of answer is calculated;

(6) according to the similarity that evaluation mechanism calculates, confidence level and degree of adopting three angles sort respectively;

(7) according to ranking results, for user recommends out answer and presenter's information.

Illustrate:

(1) according to the customer problem that user submits to, utilize tag-LDA to recommend label, the tag document of generation is submitted to user, select satisfied label by user.As shown in Figure 3, be the problem that a user submits to, problem document is as follows:

After tag-LDA process, recommend out java, the labels such as junit, tdd, private, unit-test, select the most satisfied label by user.(user have selected label java herein, junit, tdd, private, unit-test);

(2) label obtained according to the first step mates the problem in problem base (such as StackOverflow), is compared by the label of label right for each problem answers and customer problem, calculates same label proportion.The Railway Project in problem base as follows:

Problem one:

Problem two:

Problem three:

According to formula: similarity=same label number/all label numbers calculate, and the similarity of problem one is 2/6=0.333, and the similarity of problem two is 1/7=0.143.The similarity of problem three is 3/7=0.428.Setting threshold value be 0.30, by similarity higher than 0.30 problem and answer sequence.

(3) problem and answer that obtain according to sorting, utilize cosine similarity to carry out content matching the problem that the content of answer and user propose.Below several answers of problem one and problem three:

Answer is just like Fig. 4, and answer two is as Fig. 5

By customer problem, after answer one and answer two pre-service (comprise participle, remove stop words, remove numeral, root etc.):

The vector representation of answer one is D1

(<test，5>，<class，4>，〈method，6>，<Junit，6>，<mark，3>，<extend，2>，<framework，3>，<Case，2>，<inherite，1>，<Assert，6>，<version，1>，〈Annotation，1>，〈static，4>，<import，4>，<org，2>，<Redesign，1>，<package，2>)；

The vector representation of answer two is D2

(<class，1>，<method，1>，<Mockito，5>，<JMock，1>，<Mock，6>，〈syntax，2>，<Abridge，1>，<homepage，1>，<import，1>，〈static，2>，<org，1>，<List，5>，<clear，2>)；

The vector representation of customer problem is D

(<test，3>，<class，4>，〈method，3>，<JUnit，1>，<private，2>，<field，2>，<inner，1>，<internal，1>，<neste，1>，〈change，1>，<access，1>，<modifier，1>)。

The content similarity of answer one and problem is calculated: first quantize D1 and D, owing to occurring altogether in D and D1 according to cosine formula:

Test, class, method, JUnit, private, field, inner, internal, neste, change, 1access, modifier, mark, extend, framework, Case, inherite, Assert, version, Annotation, static, import, org, Redesign, package25 word, carries out quantification by this order as follows

D(3，4，3，1，2，2，1，1，1，1，1，1，0，0，0，0，0，0，0，0，0，0，0，0，0)，

D1(5，4，6，6，0，0，0，0，0，0，0，0，3，2，3，2，1，6，1，1，4，4，2，1，2)，

According to cosine formula

$\mathrm{similarity}=\cos \left(\mathrm{\&theta;}\right)=\frac{A\&CenterDot;B}{\left|\right|A\left|\right|\left|\right|B\left|\right|}=\frac{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{A}_i\&times;B_i}{\sqrt{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\left(A_i\right)}^2}\&times;\sqrt{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\left(B_i\right)}^2}}$

The content similarity of the answer one calculated is cos<D, D1>=0.5345.

The same process D and D2, the content similarity of the answer two of calculating is cos<D, D2>=0.0976.

(4) confidence level of answer is calculated according to the history performance of answer presenter, as presenter's homepage that Fig. 6 is answer one, Fig. 7 is presenter's homepage of answer two, according to formula: confidence level=label score/total fame, the confidence level calculating answer one is 0.04287, and the confidence level of answer two is 0.14429.

(5) degree of adopting of answer is calculated according to the feedback of other users and problem originator.

As Fig. 4, the vote value of answer one is 148, is bestanswer and is marked as acceptanswer;

As Fig. 5, the vote value of answer two is 273, is bestanswer and is marked as acceptanswer, is respectively bestanswer and acceptanswer and adds 2 vote.According to formula: degree of adopting=vote value/view value, the degree of adopting calculating answer one is 0.00763, and the degree of adopting of answer two is 0.00263.

(6) content similarity will obtained after completing above-mentioned process, confidence level and degree of adopting, by similarity, confidence level and degree of adopting sort respectively.

	Confidence level
		Answer two	0.14429
Answer one	0.04287
			Degree of adopting
Answer one	0.00763
		Answer two	0.00263

	Similarity
		Answer one	0.5345
Answer two	0.0976

(7) for user recommends out answer and presenter's information:

	Similarity
		The content of answer one and presenter's information	0.5345
The content of answer two and presenter's information	0.0976
			Confidence level
The content of answer two and presenter's information	0.14429
		The content of answer one and presenter's information	0.04287

	Degree of adopting
		The content of answer one and presenter's information	0.00763
The content of answer two and presenter's information	0.00263

Although the present invention illustrates with regard to preferred implementation and describes, only it will be understood by those of skill in the art that otherwise exceed claim limited range of the present invention, variations and modifications can be carried out to the present invention.

Claims (5)

1., based on a software development problem auto-answer method for mass-rent, it is characterized in that comprising the steps:

(1) with tag-LDA, label recommendations is carried out to the problem that user submits to;

(2) by label with the historical problem answer in problem base StackOverflow to mating, select relevant issues answer pair;

(3) in answer quality evaluation mechanism, pre-service is carried out to all problems and answer, the problem that the content of answer and user propose is carried out content matching, calculates content similarity;

(4) show according to the history of answer presenter, calculate the confidence level of answer;

(5) according to the feedback of other users and problem originator, the degree of adopting of answer is calculated;

(6) similarity, confidence level and degree of adopting three angles calculated according to evaluation mechanism sort respectively;

(7) according to ranking results, for user recommends out answer and presenter's information.

2. a kind of software development problem auto-answer method based on mass-rent according to claim 1, is characterized in that the computing formula of step (2) tag match is as follows:

Similarity=same label number/all label numbers

And by similarity higher than 0.30 problem answers to derivation.

3. a kind of software development problem auto-answer method based on mass-rent according to claim 1, is characterized in that step (3) preprocessing process comprises the following steps:

A) numeral is removed;

B) with the portmanteau word having lower stroke short-term to be connected, participle is carried out according to hump rule to some;

C) English stop words is removed;

D) the multi-form of word is normalized;

In addition, the computing formula of content similarity is as follows:

Wherein A, B are the quantization means representing document one and document two; Document one and document two, through participle, remove stop words, remove numeral, root preprocessing process, form vectorial A, B after being quantized in certain sequence by remaining word; In information retrieval, each entry has different degree, a document represents by there being the proper vector of weights by one, and the frequency that entry occurs in the document is depended in the calculating of weights, and therefore cosine similarity can provide the similarity of two sections of its theme aspects of document.

4. a kind of software development problem auto-answer method based on mass-rent according to claim 1, is characterized in that the computing formula of step (4) answer confidence level is as follows:

Confidence level=label score/total fame

。

5. a kind of software development problem auto-answer method based on mass-rent according to claim 1, is characterized in that the computing formula of step (5) answer degree of adopting is as follows:

Degree of adopting=vote value/view value

。