CN111428052A - Method for constructing educational concept graph with multiple relations from multi-source data - Google Patents

Abstract

The invention discloses a method for constructing an educational concept map with multiple relationships from multi-source data, comprising: crawling the multi-source data, using a data mining method, extracting concept texts to form a training data set; obtaining experts' opinions on the training data set Labeling results, extracting concepts and related features between concepts according to the source of the concept and the label of the concept; using the labeled training data set combined with traditional machine learning methods to train the support vector machine for predicting key concepts of education, and training based on training The educational key concepts and the prerequisite relationships and common learning relationships between the pairs of educational key concepts marked in the dataset are combined with traditional machine learning methods to train a hybrid model for predicting the prerequisite relationships and common learning relationships of pairs of educational key concepts; Use trained support vector machines and hybrid models to construct educational concept maps on new datasets. This method can accurately construct educational concept maps with multiple relationships.

Description

A method for constructing educational concept maps with multiple relationships from multi-source data

technical field

The invention relates to the technical field of educational data mining, in particular to a method for constructing an educational concept map with multiple relationships from multi-source data.

Background technique

Concept map, which consists of various concepts and their relationships, is a widely used graphical tool for organizing and representing knowledge. Among the various concept maps, educational concept maps mainly focus on the pedagogical relationships between concepts. Therefore, it facilitates students to organize and acquire knowledge of a subject. The construction of educational concept maps not only helps students to enhance their independent learning strategies, but also helps teachers to improve tasks such as science education, teaching evaluation, and curriculum planning. It can also recommend test questions or learning resources for students based on educational concept maps. Tasks (collectively referred to as follow-up tasks).

Education concept map can help students to learn efficiently and personalized, and is an important cornerstone of intelligent personalized teaching. Automatic and accurate construction of concept maps can help students clearly understand their own learning paths, and at the same time can assist parents and teachers to formulate personalized learning strategies for students. Therefore, how to automatically and accurately construct a concept map has always been an important issue explored in the field of educational data mining.

In the current research work and patents, the methods for constructing educational concept maps mainly include the following methods:

1) An educational concept map method based on artificial construction.

Currently, human-constructed educational concept map methods mainly focus on different disciplines and are provided by teachers or teaching assistants.

2) The construction method of educational concept map based on machine learning.

The method of constructing educational concept maps based on machine learning combines the classification (such as support vector machine) algorithms commonly used in traditional machine learning. Some scholars use this method to extract concept maps in Wikipedia.

Both of the above methods have some shortcomings. The first method is time-consuming, and teachers and teaching assistants can only develop personalized concept maps for students based on their own experience. Therefore, there are inevitably some errors and omissions in manual concept maps. The second method does not consider the help of multi-source information for constructing educational concept maps, and they only focus on one pedagogical relationship, so the constructed maps are imperfect. The educational concept map is used as reference data for subsequent tasks. When the educational concept map is not accurate enough, it will also affect the effect of subsequent tasks.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for constructing an educational concept map with multiple relationships from multi-source data. By performing accurate modeling analysis and processing on different data sources, the accuracy of the prediction results can be improved, and then the accuracy of the prediction results can be accurately constructed. Relationship education concept illustration.

The purpose of this invention is to realize through the following technical solutions:

A method for constructing educational concept maps with multiple relationships from multi-source data, including:

Step 11. Crawling multi-source data, using data mining methods, extracting conceptual texts to form a training data set;

Step 12: Obtain the labeling results of the experts on the training data set. The labeling results include: labels of educational key concepts or non-educational key concepts annotated for each concept according to the importance of the concepts, as well as the prerequisite relationship and common educational key concept pairs. Learning relationships; extracting concepts and related features between concepts according to the source of the concept and the label of the concept;

Step 13: Using the labeled training data set combined with traditional machine learning methods to train a support vector machine for predicting key educational concepts, as well as the educational key concepts marked in the training data set and the prerequisite relationship between the pairs of educational key concepts and co-learning relationships, combined with traditional machine learning methods, to train a hybrid model for predicting prerequisite relationships and co-learning relationships for pairs of key educational concepts;

It can be seen from the technical solution provided by the present invention that the method extracts different features for a variety of different data sources through different data set features; on this basis, for three different tasks, first The feature extracts key concepts, and then extracts two different relationships: prerequisite relationships and co-learning relationships. Through the use of multiple data sources and the extraction of multiple relationships, the problems of single relationship and unsatisfactory classification effect of the existing methods are made up, and the educational concept map is more accurately constructed, which can more accurately realize the personality of students Test questions or recommendations for learning resources.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the following briefly introduces the accompanying drawings used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without any creative effort.

FIG. 1 is a flowchart of a method for constructing an educational concept map with multiple relationships from multi-source data according to an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present invention.

An embodiment of the present invention provides a flowchart of a method for constructing an educational concept map with multiple relationships from multi-source data, as shown in FIG. 1 , which mainly includes the following steps:

Step 11. Crawling multi-source data, using data mining methods, extracting conceptual texts to form a training data set.

In the embodiment of the present invention, the crawled multi-source data includes at least: textbook data and historical answer information of related subjects, and related data in corresponding Wikipedia.

1) The textbook data of related subjects includes n electronic textbooks of the same subject, expressed as: S={B ₁ ,…,B _x ,…,B _n }, where B _x represents the xth electronic textbook; for each Electronic textbook B, which contains _H sub-chapters, expressed as B= _{{ C 1} , . , expressed as C={ct,s ₁ ,...,s _y ,...,s _Y }, where s _y represents the yth sentence of subsection C.

Exemplarily, the electronic textbooks can be downloaded through the Internet, and then the downloaded textbook data (electronic textbooks of elementary, middle and high schools) can be converted into txt format through an OCR tool.

以及题目解析

2) The answer record of the test question includes: the student's answer score, answer time and question information; a test question answer record is a quintuple (u,q,s _uq ,t _uq ,con _q ), where u∈Urepresents the student,U is the set of students; q∈Q represents the test question, Q is the test question set; s _uq represents the answer score; t _uq represents the answer time; con _q represents the test question text, including the test question content

and topic analysis

Exemplarily, each student's test answer record can be obtained from the online learning platform Zhixue.com.

3) The relevant data in Wikipedia corresponds to M pages, expressed as P={p ₁ ,...,p _m ,...,p _M }, where p _m represents the mth page, and each page p contains the title p _t , abstract p _abs and page content, expressed as p=(p _t , p _abs , p _con ).

The text content in the data set is segmented by the word segmentation tool, and then the segmented content is matched with the encyclopedia title, so as to extract different concept texts to form a concept set, and randomly select a specified number of concepts from the concept set (the specific number can be based on the actual number of concepts). need to be set), which constitutes the training data set.

Those skilled in the art can understand that a concept mainly refers to a conceptual form commonly used in mathematics, such as "quadratic equation in one variable", "function", "decimal number" and so on.

Step 12: Obtain the labeling results of the experts on the training data set. The labeling results include: labels of educational key concepts or non-educational key concepts annotated for each concept according to the importance of the concepts, as well as the prerequisite relationship and common educational key concept pairs. Learn relationships; extract concepts and related features between concepts according to their origins and their labels.

In this embodiment of the present invention, the importance of a concept is used as an indicator to measure whether a concept is an educational key concept or a non-educational key concept, and the importance can be determined in various conventional ways, for example, the number of times the concept appears in the title of a mathematics textbook. If it exceeds the specified value, it is considered to be of high importance and belongs to the key concept of education; for example, the “decimal” mentioned above can also be determined by experts based on experience.

In the embodiment of the present invention, the following features are extracted respectively according to the characteristics of the multi-source data set and according to the data source of the concept:

(1) Concept semantic similarity features for each data source, including: title matching feature, which is used to indicate whether the concept appears in the title; concept matching feature, which is used to indicate the relationship between concept pairs; word representation similarity, It is used to represent the similarity and distance of concept pairs in vector space.

(2) Wikipedia link features, including: concept pairs in and out of Wikipedia pages, degree of common neighbors of concept pairs, definition of Wikipedia abstracts, normalized Google page distance, and citation distance.

(3) Textbook structuring features and concept co-occurrence degree, among which, textbook structuring features include: catalogue structuring features and inter-textbook structuring features, and concept co-occurrence degree is used to represent the number of times a concept pair appears simultaneously in a sentence .

(4) The characteristics of test questions and answer records, including: concept frequency characteristics, concept difficulty distance, test content analysis distance, and students' answer record characteristics.

Among the above features, the title matching feature, concept frequency feature, and the degree of discrepancy of concept pairs in Wikipedia pages are for a single concept, so there is no need to distinguish whether a concept is a key concept of education, and the remaining features are for concept pairs. , therefore, only the key concepts of education are extracted (data sources are also considered); for ease of illustration, w _i and w _j are uniformly used below to represent the concepts in the training data set, without distinguishing data sources and corresponding labels.

The features of each type are described in detail below.

1. Conceptual semantic similarity features.

1) Title matching feature.

The title is a summary of the content of the sub-chapter and points out the main points of the sub-chapter. If a concept appears in the title, it is likely to be a key concept. The title matching feature is expressed as:

TM( _wi ,ct)∈{0,1}

Among them, ct∈{CT,p _t ,q′}, q′ represents the title of the test question q, _wi represents a concept, when the concept _wi appears in the corresponding title, then TM( _wi ,ct)=1; Otherwise, TM( _wi ,ct)=0.

2) Concept matching features.

Given a concept pair < _wi , _wj >, if concept _wi appears in concept _wj , then _wi is more likely to have a preconditional relationship with _wj . The concept matching feature is expressed as:

Among them, ||.|| represents the number statistical symbol;

3) The degree of concept co-occurrence.

4) Word representation similarity.

The word representation similarity includes: cosine similarity WEcs( _wi , _wj ) and Euclidean distance WEed( _wi , _wj ):

Cosine similarity WEcs( _wi , _wj ) reflects the semantic association between concept pairs ( _wi , _wj ), which is expressed as:

The Euclidean distance WEed( _wi , _wj ) represents the Euclidean distance of the concept pair ( _wi , _wj ) in the vector space, and is expressed as:

分别表示概念w_i、w_j的词向量，k为向量中元素的序号，P为向量长度。in,

Represent the word vectors of concepts w _i and w _j respectively, k is the serial number of the elements in the vector, and P is the length of the vector.

2. Wikipedia link feature.

1) The in- and out-degree of concept pairs in Wikipedia pages.

Calculate the in-out degree of the concept through the Wikipedia page, and define the in-out degree of the concept pair ( _wi , w _j ) as: IN( _wi ), OUT( _wi ), IN(w _j ), OUT(w _j ) .

2) The degree of common neighbors of concept pairs.

The degree of common neighbors of the concept pair: for the concept pair ( _wi , _wj ), the more common neighbors of the concept pair ( _wi , _wj ), the higher the semantic similarity of the concept pair ( _wi , _wj ), Expressed as:

3) Wikipedia abstract definition.

Wikipedia abstract definition: If concept _wi is in the abstract definition of concept w _j , then concept _wi is a pre-order concept of concept w _j , expressed as:

4) Normalized Google page distance.

Normalized Google page distance: Through the hyperlinks between concepts in Google pages, the degree of association between concepts is obtained, which is expressed as:

5) Reference distance.

Reference distance: If the concepts most associated with _wi all point to w _j , then _wi is more likely to be a preorder concept of w _j , expressed as:

与

均表示维基百科中相应页面的概念；

表示概念

是否指向概念w_i所在维基百科页面，1表示指向，0表示未指向；

表示概念

在概念w_j所在维基百科页面的重要程度，

表示概念

是否指向概念w_i所在维基百科页面；

表示概念

在概念w_i所在维基百科页面的重要程度，

表示概念

是否指向概念w_j所在维基百科页面。Among them, O ₁ represents the number of other concepts in the Wikipedia page where the concept _wi is located, O ₂ represents the number of other concepts in the Wikipedia page where the concept _wi is located by other concepts on the Wikipedia page where the concept w _j is located, and O ₃ represents the number of links The number of other concepts in the Wikipedia page where the concept w _j is located, O ₄ represents the number of other concepts linked by other concepts in the Wikipedia page where the concept _{w j} is located;

and

Both represent the concept of the corresponding page in Wikipedia;

represent concepts

Whether it points to the Wikipedia page where the concept w _i is located, 1 means pointing, 0 means not pointing;

represent concepts

In terms of the importance of the Wikipedia page where the concept w _j is located,

represent concepts

Whether to point to the Wikipedia page where the concept w _i is located;

represent concepts

In terms of the importance of the Wikipedia page where the concept w _i is located,

represent concepts

Whether to point to the Wikipedia page where the concept w _j is located.

3. The structural characteristics of textbooks and the degree of concept co-occurrence.

The degree of concept co-occurrence is used to represent the number of times a concept pair occurs simultaneously in a sentence. The calculation formula is as follows:

Among them, r(s, _wi )∈{0,1} indicates whether the concept _wi appears in the sentence s, if it appears in the sentence s, the value is 1, otherwise, the value is 0. The same is true for the meaning of r(s,w _j ).

The Textbook Table of Contents (TOC) and textbook structure demonstrate the interconnectedness of concepts as teachers' lesson planning is based on this information. The hierarchical structure features of two textbooks are defined, including the catalogued structure feature and the inter-textbook structure feature, to help infer the relationship between concepts.

1) Directory structure features. The relation of concept pair ( _wi , w _j ) in subsection C, expressed as:

Among them, |B| represents the number of textbooks, |S| represents the number of books, f( _wi ,C) refers to the number of sub-chapters C that contain the concept _wi , and the final result is a number; f(w _j ,C) denotes the number of subsections C that contain concept w _j .

2) Structural features between textbooks.

Similar to the structured feature of the catalog, the structured feature between textbooks reflects the relationship between the concept pairs ( _wi , w _j ) in the textbook, and is expressed as:

Among them, f( _wi ,B) refers to the number of textbooks B that contain the concept _wi .

4. Recording features of test questions and answers.

1) Concept frequency features.

If the concept _wi is frequently mentioned in the test item content, then _wi is more likely to be a key concept. On the basis of this assumption, key concepts can be extracted by this feature.

The concept frequency feature is defined as the occurrence frequency of the concept _wi , which is expressed as:

是试题内容中出现的概念w_i的次数。in,

is the number of times the concept _wi appears in the test content.

2) Concept difficulty distance.

The concept difficulty distance represents the distance between the average difficulty of the questions containing the concept w _i and the average difficulty of the questions containing the concept w _j , expressed as:

CDD( _wi , _wj )=CD( _wi )-CD( _wj )

Among them, CD( _wi ) and CD( _wj ) represent the average difficulty of concepts _wi and _wj ; generally speaking, the difficulty of test questions refers to the proportion of students who answered the questions correctly, and the average conceptual difficulty of concept _wi , CD( w _i ) is the average difficulty of the tested questions including concept w _i , CD( _wi ) is calculated as follows:

表示试题内容

中概念w_i出现的次数，反映了试题q中概念w_i的重要程度；dif_q为试题q的难度；L表示试题集合Q中包含概念w_i的试题集合，|L|表示L的数目。in,

Indicates the content of the test

The number of times that the concept _wi appears in the test question q reflects the importance of the concept _wi in the test question q; dif _q is the difficulty of the test question q; L represents the test question set that contains the concept wi _i in the test question set Q, and |L| represents the number of L.

Similarly, CD(w _j ) is also calculated in a similar way, except that the subscript i is replaced by j.

3) Analysis distance of test content: Generally, the concepts that appear in the content of the test will be learned after the concepts that appear in the analysis of the test question. Based on this feature, the analysis distance of the content of the test is used to measure the sequence relationship between the two concepts.

The content analysis distance of the test questions, the calculation formula is:

Qcad( _wi , _wj )=Qcaw( _wj , _wi )-Qcaw( _wi , _wj )

in:

表示试题内容

中概念w_j出现的次数；

表示概念w_j是否出现在试题分析

中，

表示概念w_i是否出现在试题分析

中，出现取值为1，否则取值为0；当然，如果w_i(或者w_j)出现在试题内容中，而w_j(或者w_i)出现在试题分析中，那么Qcaw(w_i,w_j)(Qcaw(w_j,w_i))就会变大，这符合实际的情况。in,

Indicates the content of the test

The number of occurrences of the concept w _j in ;

Indicates whether the concept w _j appears in the test item analysis

middle,

Indicates whether the concept _wi appears in the test item analysis

In , the value of occurrence is 1, otherwise the value is 0; of course, if _wi (or w _j ) appears in the content of the test question, and w _j (or _wi ) appears in the analysis of the test question, then Qcaw( _wi , w _j )(Qcaw(w _j , _wi )) will become larger, which is in line with the actual situation.

4) The characteristics of students' answering records.

Define the test item set of student u as Q, define I(Q; w _i ) as the test item index containing the concept wi in the test item set Q, and I(Q; w _j ) as the test item index in the test item set _Q containing the concept w _j . For example, if _wi appears in the first and third question of the question set Q, then I(Q; _wi )∈{1,3}. Assuming w _j is a pre-order concept of _wi , in student u's answer sequence, if the student answers the question containing the concept wi incorrectly, then the student u is more likely to answer the question containing the concept w _{j incorrectly} . Based on this observation, for a given concept pair < _wi ,w _j >, define S(Q)={(i ₁ ,j ₁ )│i ₁ ∈I(Q; _wi ),j ₁ ∈I( Q; w _j ), i ₁ <j ₁ }, the characteristics of students' answering records are as follows:

分别为学生u在试题i₁、试题j₁上的得分，U为学生集合，|U|表示U的数目。in,

are the scores of student u on question i ₁ and question j ₁ respectively, U is the set of students, and |U| represents the number of U.

Step 13: Using the labeled training data set combined with traditional machine learning methods to train a support vector machine for predicting key educational concepts, as well as the educational key concepts marked in the training data set and the prerequisite relationship between the pairs of educational key concepts and co-learning relationships, combined with traditional machine learning methods, to train a hybrid model for predicting prerequisite and co-learning relationships for pairs of educational key concepts.

Due to the lack of large-scale label data sets in the concept map construction, in the embodiment of the present invention, three binary classifiers are trained based on the traditional machine learning method; The degree of feature and concept pair in the Wikipedia page is used to extract the key educational concept set C'; the other two binary classifiers are used as mixed models to predict the key educational concepts on the basis of obtaining the key educational concept set C'. The precondition relationship and the common learning relationship between the key concept pairs (wi _' ,w _j' ) in the set C', the preferred implementation of the training phase is as follows:

1) Train a support vector machine.

与实际标签X_i间的误差：Using the labeled training data set, according to the labels of each concept, as well as previously extracted concept features, that is, title matching features, and concept frequency features extracted according to the source of concept pairs, and/or concept pairs in Wikipedia pages. , train the support vector machine to obtain the complete parameter W ¹ of the support vector machine, and the first threshold K ^* ; the training goal is to minimize the predicted label

Error from actual label _Xi :

表示支持向量机预测到的第i个概念的标签(即概念为教育关键概念或非教育关键概念)，

为第i个概念的相关特征，

为对于第i个概念的参数，角标T为矩阵转置符号，M₁个参数

构成支持向量机的完整参数W¹；X_i表示专家为第i个概念标注的标签(即实际标签)；λ₁||W¹||²是正则化项，λ₁是手动调节的参数。where M1 represents the number _of concepts in the training dataset,

represents the label of the ith concept predicted by the support vector machine (that is, the concept is an educational key concept or a non-educational key concept),

is the relevant feature of the i-th concept,

is the parameter for the i-th concept, the index T is the matrix transpose symbol, and M is ₁ parameter

The complete parameter W ¹ constituting the support vector machine; X _i represents the label (ie, the actual label) annotated by experts for the ith concept; λ ₁ ||W ¹ || ² is the regularization term, and λ ₁ is a manually adjusted parameter.

2) Train a binary classifier for predicting precondition relations.

The prerequisite relationship between key concept pairs (wi _' , w _j' ) is determined by concept matching feature, word representation similarity, concept difficulty distance, test content analysis distance, student answering record feature, catalog structuring feature, and inter-textbook structure. The predictions are based on normalized features, degree of common neighbors of concept pairs, Wikipedia abstract definitions, normalized Google page distances, and citation distances.

与实际标签X′_l之间的误差：In the training phase, the key educational concepts in the training dataset are selected according to the labels of the concepts in the training data set, and the prerequisite relationship between the key educational concept pairs marked by experts is used, combined with the concept matching feature and the similarity of the word representation between the key educational concept pairs, As well as the concept difficulty distance extracted from the source of the concept pair, the analysis distance of the test question content and the characteristics of the students' answering records, the structured characteristics of the catalog and the structured characteristics of the textbooks, and/or the degree of common neighbors of the concept pairs, Wikipedia abstract definition, normalization The Google page distance and citation distance are used to train a binary classifier for predicting preconditional relations, obtaining the full parameters W ² of the binary classifier and the second threshold P ₁ ; the goal of training is to minimize the predicted labels

Error from actual label X' _l :

表示对于二分类器预测到的第l个教育关键概念对的标签，即教育关键概念对是否存在先决条件关系，

为第l个教育关键概念对的相关特征，W² _l为对于第l个教育关键概念对的参数，M₂和参数W² _l构成了二分类器的完整参数W²；X′_l表示专家为第l个教育关键概念对标注的先决条件关系(即实际标签)，λ₂||W²||²是正则化项，λ₂是手动调节的参数。where _M2 represents the number of educational key concept pairs,

represents the label of the l-th educational key concept pair predicted by the binary classifier, that is, whether there is a prerequisite relationship for the educational key concept pair,

is the relevant feature of the l-th educational key concept pair, W ² _l is the parameter for the l-th educational key concept pair, M ₂ and parameter W ² _l constitute the complete parameter W ² of the binary classifier; X′ _l represents the expert λ ₂ ||W ² || ² is the regularization term, and λ ₂ is a manually tuned parameter.

3) Train a binary classifier for predicting co-learned relations.

If a concept pair ( _wi , _wj ) has a common learning relationship, it should have the following properties:

Semantic similarity: they share the same semantic information;

Co-occurrence: they may appear in the same sentence;

Concept matches: they may contain common words;

Similar difficulty: Problem A containing w _i and problem B containing w _j may have the same difficulty;

Similar neighbors: they may share the same neighbors in the Wikipedia link;

Shared definitions: w _i may appear in the definition of w _j and vice versa.

Based on these assumptions, the common learning relationship between educational key concept pairs (wi _′ ,wj _′ ) is determined by concept matching features, word representation similarity, concept co-occurrence degree, concept difficulty distance, degree of common neighbors of concept pairs, and wiki Encyclopedia abstract definition to predict.

与实际标签X″_l之间的误差：In the training phase, the key educational concepts in the training dataset are selected according to the labels of the concepts in the training data set, and the common learning relationship between the key educational concept pairs marked by experts is used, combined with the concept matching feature and the similarity of the word representation between the key educational concept pairs. As well as training a binary classifier according to the concept co-occurrence degree extracted from the source of the concept pair, the concept difficulty distance, and/or the degree of common neighbors of the concept pair and the Wikipedia abstract definition, to obtain a complete binary classifier for predicting jointly learned relations. parameter W ³ and second threshold P ₃ ; the goal of training is to minimize predicted labels

Error from actual label X″ _l :

表示对于二分类器预测到的第l个教育关键概念对的标签，即教育关键概念对是否存在共同学习关系，

为第l个教育关键概念对的相关特征，W³ _l为对于第l个教育关键概念对的参数，M₂和参数W³ _l构成了二分类器的完整参数W³；X″_l表示专家为第l个教育关键概念对标注的共同学习关系(即实际标签)，λ₃||W³||²是正则化项，λ₃是手动调节的参数。where _M2 represents the number of educational key concept pairs,

represents the label of the l-th educational key concept pair predicted by the binary classifier, that is, whether there is a common learning relationship between the educational key concept pair,

is the relevant feature of the lth educational key concept pair, W ³ _l is the parameter for the l th educational key concept pair, M ₂ and parameter W ³ _l constitute the complete parameter W ³ of the binary classifier; X″ _l represents the expert is the common learning relationship (ie the actual label) of the labels for the l-th educational key concept, λ ₃ ||W ³ || ² is the regularization term, and λ ₃ is a manually adjusted parameter.

In this embodiment of the present invention, the value of the first threshold K ^* can be appropriately adjusted as needed; for example, when you want to filter out more key concepts of education, the value of the first threshold K* can be appropriately reduced; otherwise, the value of the first threshold K ^* can be appropriately increased A threshold value of K ^* .

Those skilled in the art can understand that each feature of a concept pair is calculated according to the relevant information of the data source where the concept pair is located. Therefore, the concept pair mentioned here mainly refers to two concepts in the same data source. In most cases, the same concept pair exists in all three data sources, that is to say, a concept pair with the same content can be calculated according to the relevant information in the three data sources. However, it also considers the case that the concept pair only appears in one or two data sources. At this time, a concept pair with the same content can only extract the two or three types of features mentioned in step 12. Therefore, the above During the training process, an "and/or" description is used between the features extracted from the source according to the concept.

Step 14: Use the trained support vector machine and the hybrid model to construct an educational concept map for the new data set.

For an unpublished new data set, extract each concept text according to step 11, and extract related features between concepts according to step 12; then, use the parameters of the trained support vector machine and hybrid model and related thresholds , construct the concept map G, the steps are as follows:

以及支持向量机的参数W¹以及第一阈值K^*，抽取关键概念集合C′，表示为：First, according to the method of step 11 (that is, based on the word segmentation technology), extract each concept text, form a concept candidate set R, and combine the relevant features of each candidate concept

As well as the parameter W ¹ of the support vector machine and the first threshold K ^* , the key concept set C' is extracted, which is expressed as:

是指第t个概念的特征(与步骤13中的

是类似的含义)，即标题匹配特征、以及根据概念对来源提取的概念频率特征、或概念对在维基百科页面中的出入度，Among them, the relevant features

refers to the feature of the t-th concept (same as in step 13

is a similar meaning), that is, the title matching feature, and the concept frequency feature extracted from the source of the concept pair, or the degree of ins and outs of the concept pair in the Wikipedia page,

On the basis of obtaining the key concept set C', according to the parameters W ² and W ³ of the mixed model, as well as the two thresholds P ₂ and P ₃ , the key concept pair {(wi _' ,w _j' )| _wi is predicted respectively _′ , w _j′ ∈ C′} whether there is a prerequisite relationship and a common learning relationship:

分别表示关键概念集合C′中第l′个概念对(w_i′,w_j′)之间的用于预测先决条件关系、共同学习关系的相关特征，与步骤13中的

是类似的含义，即

包含的特征有：概念匹配特征与词表征相似度，以及根据概念对来源提取的概念难度距离、试题内容分析距离与学生答题记录特征，或者目录结构化特征与课本间结构化特征，或者概念对的公共邻居程度、维基百科摘要定义、归一化的谷歌页面距离与引用距离；

包含的特征有：概念匹配特征与词表征相似度，以及根据概念对来源提取的概念共现程度，或者概念难度距离，或者概念对的公共邻居程度以及维基百科摘要定义；以筛选出的关键概念集合C′中的每一教育关键概念作为节点，根据教育关键概念对之间是否存在先决条件关系与共同学习关系，来构造相应节点之间的连接关系，从而构建教育概念图。Among them, <wi _' ,w _j' >=0 means that there is no precondition and common learning relationship between concept wi _' and concept w _j' , and <wi _' ,w _j' >=1 means that concept wi _' There is a prerequisite relationship with the concept w _j' , <wi _' , w _j' >= 2 means there is a common learning relationship between the concept wi _' and the concept w _j' ;

Respectively represent the related features between the l'th concept pair (wi _' , w _j' ) in the key concept set C' for predicting the prerequisite relationship and the common learning relationship, which are the same as those in step 13.

is a similar meaning, i.e.

The features included are: the similarity between concept matching features and word representations, as well as the concept difficulty distance extracted from the source of the concept pair, the analysis distance of the test question content and the student answering record feature, or the structure feature of the catalog and the structured feature between textbooks, or the concept pair. The degree of common neighbors, Wikipedia abstract definition, normalized Google page distance and citation distance;

The features included are: similarity between concept matching features and word representations, and the degree of concept co-occurrence extracted from the source of the concept pair, or the concept difficulty distance, or the degree of common neighbors of the concept pair, and the definition of the Wikipedia abstract; to filter out the key concepts Each key educational concept in the set C' is used as a node, and the connection relationship between the corresponding nodes is constructed according to whether there is a prerequisite relationship and a common learning relationship between the key educational concept pairs, thereby constructing an educational concept map.

Since unpublished new datasets usually correspond to students, the educational concept map can reflect students' knowledge mastery. After linking the educational concept map with the test questions, test item recommendations can be generated based on the information on the educational concept map. list and recommend to the appropriate student. For example, through the information on the educational concept map, it is found that the students' understanding of the key concept of quadratic function is insufficient, and the corresponding test item recommendation list can be generated to test the students' pre-order concept of quadratic function (linear function) and Whether the common learning concept (quadratic equation) is understood or not, in this way, the students' abilities can be checked layer by layer, and finally the crux of the students' lack of understanding can be found, and then the personalized recommendation of test questions or learning resources can be realized through these cruxes.

The above scheme of the embodiment of the present invention extracts different features according to different data sources and through different data set features; on this basis, for three different tasks, the key concepts are first extracted based on the relevant features, Afterwards, two different relationships are extracted: the prerequisite relationship and the co-learning relationship. Through the use of multiple data sources and the extraction of multiple relationships, the problems of single relationship and unsatisfactory classification effect in the existing methods are made up, so that the educational concept map can be constructed more accurately.

From the description of the above embodiments, those skilled in the art can clearly understand that the above embodiments can be implemented by software or by means of software plus a necessary general hardware platform. Based on this understanding, the technical solutions of the above embodiments may be embodied in the form of software products, and the software products may be stored in a non-volatile storage medium (which may be CD-ROM, U disk, mobile hard disk, etc.), including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present invention.

The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited to this. Substitutions should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (10)

1. A method for constructing an educational concept graph having multiple relationships from multiple sources of data, comprising:

step 11, crawling multi-source data, extracting concept texts by using a data mining method, and forming a training data set;

step 12, obtaining the labeling result of the expert on the training data set, wherein the labeling result comprises the following steps: labels of education key concepts or non-education key concepts labeled for the respective concepts according to the concept importance degrees, and prerequisite relationships and common learning relationships between pairs of the education key concepts; extracting concepts and related features between the concepts according to the source of the concepts and the labels of the concepts;

step 13, training a support vector machine for predicting the education key concepts by using the labeled training data set in combination with a traditional machine learning method, and training a mixed model for predicting the prerequisite relationships and the common learning relationships of the education key concepts in pairs based on the education key concepts labeled in the training data set and the prerequisite relationships and the common learning relationships between the education key concepts in pairs in combination with the traditional machine learning method;

and 14, constructing an educational concept graph for the new data set by using the trained support vector machine and the mixed model.

2. The method of claim 1, wherein the crawled multi-source data comprises at least: textbook data and historical answer information of related disciplines and related data in corresponding wikipedia; wherein:

textbook data of related disciplines contains n electronic textbooks of the same discipline, represented as: s ═ B₁,…,B_x,…,B_n}, wherein B_xRepresenting the xth electronic textbook; for each electronic textbook B, which contains H subsections, denoted B ═ C₁,…,C_h,…,C_H}, wherein C_hRepresents the h-th sub-chapter; for each sub-chapter, the title CT and Y sentences, denoted C ═ CT, s₁,…,s_y,…,s_Y}, wherein ,s_yThe y-th sentence representing sub-section C;

the test question answering record comprises: the student answering score, answering time and question information; one test question answer record is a quintuple (u, q, s)_uq,t_uq,con_q) Wherein U ∈ U represents students, U represents student set, Q ∈ Q represents test questions, Q represents test question set, and s_uqRepresenting an answer score; t is t_uqRepresenting the answering time; con_qRepresenting test question text, including test question content

And topic analysis

The relevant data in wikipedia corresponds to M pages, denoted P ═ P₁,…,p_m,…,p_M}, wherein p_mRepresenting the m-th page, each page p containing a title p_tAbstract p_absAnd page content, denoted as p ═ (p)_t,p_abs,p_con)。；

The method comprises the steps of segmenting text contents in a data source through a segmentation tool, matching the segmented contents with encyclopedia titles to extract different mathematical concepts to form a concept set, and randomly selecting a specified number of concepts from the concept set to form a training data set.

3. The method of claim 2, wherein the features to be extracted according to the source of the concept comprise: the concept semantic similarity characteristic for each data source comprises the following steps: a title matching feature to indicate whether a concept appears in a title; concept matching features for the relationship between pairs of concepts; the word representation similarity is used for representing the similarity and the distance of the concept pair in a vector space;

wikipedia link features including: the degree of entry and exit of concept pairs in Wikipedia pages, the degree of public neighbourhood of concept pairs, Wikipedia abstract definition, normalized Google page distance and reference distance;

the degree of co-occurrence of textbook structural features and concepts, wherein the textbook structural features include: directory structured features and textbook structured features; the concept co-occurrence degree is used for representing the number of times of a concept pair appearing in a sentence simultaneously;

test question answering record characteristics include: concept frequency characteristics, concept difficulty distance, test question content analysis distance and student answer record characteristics;

the title matching features, the concept frequency features and the entrance and exit degree of the concepts in the Wikipedia page are specific to a single concept, and whether the concepts are education key concepts or not is not distinguished; and the rest of the characteristics are extracted only for the education key concept pairs.

4. The method for constructing an educational concept graph having multiple relationships from multiple data according to claim 3,

the title matching features are expressed as:

TM(w_i,ct)∈{0，1}

wherein, CT ∈ { CT, p_tQ '}, q' denotes the title of the test question q, w_iRepresents a concept when the concept w_iAppearing in the corresponding title, TM (w)_iCt) ═ 1; otherwise, TM (w)_i,xt)＝0；

The concept matching features are expressed as:

wherein ,(w_i,w_j) For a conceptual pair, | |. | | represents a number statistics symbol;

the word representation similarity comprises: cosine similarity WEcs (w)_i,w_j) And Euclidean distance WEed (w)_i,w_j)；

Cosine similarity WEcs (w)_i,w_j) Reflects the concept pair (w)_i,w_j) The semantic association between them is expressed as:

euclidean distance WEed (w)_i,w_j) Representing concept pairs (w) in vector space_i,w_j) Expressed as:

wherein ,

respectively represent concepts w_i、w_jK is the sequence number of the element in the vector, and P is the vector length.

5. The method for constructing an educational concept graph having multiple relationships from multiple data according to claim 3,

concept versus degree of entry and exit in wikipedia pages: will concept pair (w)_i,w_j) Is defined as IN (w)_i)、OUT(w_i)、IN(w_j)、OUT(w_j)；

Common neighbor degree of concept pair: for concept pair (w)_i,w_j) Concept pair (w)_i,w_j) The more common neighbors there are, thenConcept pair (w)_i,w_j) The higher the semantic similarity of (a) is, it is expressed as:

wikipedia abstract definition: if the concept w_iAt concept w_jIn the abstract definition of (1), then the concept w_iIs a concept w_jThe precedence concept of (a) is expressed as:

normalized google page distance: through the hyperlink between the concepts in the Google webpage, the association degree between the concepts is obtained and is expressed as:

quote distance, expressed as:

wherein ,O₁Representing a concept w_iNumber of other concepts in Wikipedia Page, O₂Representing a concept w_iOther concepts in the Wikipedia page are conceptualized w_jThe number of links to other concepts in the Wikipedia page, O₃Representing a concept w_jNumber of other concepts in Wikipedia Page, O₄Representing a concept w_jOther concepts in the Wikipedia page are conceptualized w_iThe number of links of other concepts in the Wikipedia page;

and

all represent the concept of the corresponding page in wikipedia;

representing concepts

Whether or not to point to concept w_iIn the Wikipedia page, 1 indicates pointing, and 0 indicates not pointing;

representing concepts

At concept w_jThe importance of the wikipedia page in which it is located,

representing concepts

Whether or not to point to concept w_iThe Wikipedia page where the user is located;

representing concepts

At concept w_iThe importance of the wikipedia page in which it is located,

representing concepts

Whether or not to point to concept w_jThe wikipedia page.

6. The method for constructing an educational concept graph having multiple relationships from multiple data according to claim 3,

directory structuring feature embodying concept pairs (w) in subsection C_i,w_j) Is expressed as:

where | B | represents the number of textbooks, | S | represents the number of books, f (w)_iC) means that the concept w is included_iNumber of sub-sections C, f (w)_jC) indicates that the concept w is included_jThe number of sub-sections C;

the structural characteristics between textbooks embody the concept pair (w) in the textbook_i,w_j) Is expressed as:

wherein ,f(w_iB) means that the concept w is included_iThe number of textbooks B;

the degree of concept co-occurrence is calculated by the following formula:

wherein, r (s, w)_i) ∈ {0,1} represents the concept w_iWhether the sentence appears in the sentence s or not is judged, if the sentence appears in the sentence s, the value is 1, otherwise, the value is 0; r (s, w)_j) ∈ {0,1} represents the concept w_jAnd whether the sentence appears in the sentence s or not is judged, if so, the value is 1, otherwise, the value is 0.

7. The method for constructing an educational concept graph having multiple relationships from multiple data according to claim 3,

concept frequency signature, representing concept w_iIs expressed as:

wherein ,

is a concept w appearing in the content of the test question_iThe number of times of (c);

concept difficulty distance, representing inclusion of concept w_iAverage difficulty of test questions and contained concept w_jThe distance of the average difficulty of the test questions is expressed as:

CDD(w_i,w_j)＝CD(w_i)-CD(w_j)

wherein, CD (w)_i)、CD(w_j) Representing a concept w_i、w_jAverage difficulty of; CD (w)_i) The calculation formula of (a) is as follows:

wherein ,

showing the contents of the test questions

Concept of middle_iThe number of occurrences reflects the concept w in the test question q_iThe degree of importance of; dif_qFor the difficulty of the test question Q, L shows that the concept w is included in the test question set Q_iThe set of questions, | L | represents the number of L;

analyzing the distance of the test question content, wherein the calculation formula is as follows:

Qcad(w_i,w_j)＝Qcaw(w_j,w_i)-Qcaw(w_i,w_j)

wherein ：

wherein ,

showing the contents of the test questions

Concept of middle_jThe number of times of occurrence of the event,

representing a concept w_jWhether it appears in the analysis of test question

In (1),

representing a concept w_iWhether it appears in the analysis of test question

If so, the value is 1, otherwise, the value is 0; to represent

Student answer record characteristics, expressed as:

wherein ,

test questions i for the student u₁Test question j₁Score of (i), (ii) s (q) { (i)₁,j₁)│i₁∈I(Q；w_i),j₁∈I(Q；w_j),i₁＜j₁}、I(Q；w_i)、I(Q；w_j) Each containing a concept w in a test question set Q_i、w_jU is the student set.

8. The method of claim 3, wherein training the support vector machine for predicting educational key concepts comprises:

training the support vector machine by using the labeled training data set according to the label of each concept and the extracted concept characteristics, namely the title matching characteristics, and the concept frequency characteristics extracted from the concept pair source and/or the entrance and exit degree of the concept pair in the Wikipedia page, and obtaining the complete parameter W of the support vector machine¹And a first threshold value K^*(ii) a The goal of the training is to minimize the predictive label

And the actual label X_iError between:

wherein ,M₁Representing the number of concepts in the training dataset,

a label representing the ith concept predicted by the support vector machine,

as a relevant feature of the ith concept,

for parameters for the ith concept, the corner mark T is the matrix transpose symbol, M₁A parameter

Complete parameters W forming a support vector machine¹；X_iA label representing the label of the ith concept labeled by the expert; lambda [ alpha ]₁||W¹||²Is a regularization term, λ₁Is a parameter that is adjusted manually.

9. The method of claim 8, wherein the hybrid model comprises a classifier for predicting prerequisite relationships and a classifier for predicting common learning relationships; wherein:

training a classifier for predicting prerequisite relationships includes:

in the training stage, education key concepts in the training data set are selected according to labels of the concepts in the training data set, the presupposition relationship between expert-labeled education key concept pairs is utilized, the concept matching features and word expression similarity between the education key concept pairs are combined, the concept difficulty distance extracted from the concept pair sources, the analysis distance of test question contents and the student answer record features, the directory structured features and the structured features between textbooks are used for training a binary classifier for predicting the presupposition relationship according to the concept difficulty distance, the test question content analysis distance and the student answer record features of the concept pair sources, and/or the public neighbor degree, the Wikipedia abstract definition, the normalized Google page distance and the reference distance of the concept pairs, and the complete parameter W of the binary classifier is obtained²And a second threshold value P₁(ii) a The goal of the training is to minimize the predictive label

And actual tag X'_lThe error between:

wherein ,M₂Represents the number of pairs of educational key concepts,

a label representing the ith pair of educational key concepts predicted by the second classifier, i.e., whether there is a prerequisite relationship for the pair of educational key concepts,

for the relevant characteristics of the first pair of educational key concepts, W² _lAs a parameter for the first pair of educational key concepts, M₂And a parameter W² _lComplete parameter W forming a classifier²；X′_lDenotes the expert's prerequisite relationship to the annotation for the first educational key concept, λ₂||W²||²Is a regularization term, λ₂Is a manually adjusted parameter;

the way to train the classifiers for predicting the common learning relationship includes:

in the training stage, the education key concepts in the training data set are selected according to the labels of the concepts in the training data set, the common learning relationship between the education key concept pairs labeled by experts is utilized, the similarity between the concept matching features and word characteristics between the education key concept pairs is combined, the co-occurrence degree of the concepts extracted from the concept pair sources, the concept difficulty distance, the common neighbor degree of the concept pairs and the definition of Wikipedia abstract are combined to train a two-classifier, and the complete parameter W of the two-classifier for predicting the common learning relationship is obtained³And a second threshold value P₃(ii) a The goal of the training is to minimize the predictive label

And the actual label X ″)_lThe error between:

wherein ,M₂Represents the number of pairs of educational key concepts,

a label indicating the ith pair of education key concepts predicted by the second classifier, i.e., whether or not there is a common learning relationship with the pair of education key concepts,

for the relevant characteristics of the first pair of educational key concepts, W³ _lAs a parameter for the first pair of educational key concepts, M₂And a parameter W³ _lComplete parameter W forming a classifier³；X″_lDenotes the common learning relationship, lambda, of the expert for the first educational key concept pair labels₃||W³||²Is a regularization term, λ₃Is a parameter that is adjusted manually.

10. The method for building an educational concept graph with multiple relationships from multi-source data according to claim 3, 8 or 9, wherein the building of the educational concept graph for the new data set using the trained support vector machine and the hybrid model comprises:

for a new data set which is not released, extracting each concept text according to the mode of step 11, and extracting the related characteristics between the concepts according to step 12; then, a conceptual diagram G is constructed by using the parameters of the trained support vector machine and the mixed model and the related threshold, and the steps are as follows:

firstly, according to the mode of step 11, extracting each concept text to form concept candidate set R, combining the relevant characteristics of each candidate concept

And support vector machine parameters W¹And a first threshold value K^*Extracting a key concept set C' as follows: (ii) a

On the basis of obtaining the key concept set C', according to the parameters W of the mixed model²And W³And two thresholds P₂And P₃Respectively predictMeasure key concept pairs { (w)_i′,w_j′)|w_i′,w_j′∈ C' } whether there is a prerequisite relationship and a common learning relationship:

wherein ,＜w_i′,w_j′0 denotes the concept w_i′And concept w_j′There is no prerequisite and co-learning relationship between, < w_i′,w_j′1 denotes a concept w_i′And concept w_j′With a prerequisite relationship between, < w_i′,w_j′2 denotes the concept w_i′And concept w_j′Have a common learning relationship;

respectively representing the l' th concept pair (w) in the key concept set C_i′,w_j′) Relative characteristics used for predicting prerequisite relations and common learning relations;

and constructing a connection relation between corresponding nodes by taking each education key concept in the screened key concept set C' as a node according to whether a prerequisite relation and a common learning relation exist between the education key concept pairs, thereby constructing an education concept graph.

Images