CN111598252A

CN111598252A - University computer basic knowledge problem solving method based on deep learning

Info

Publication number: CN111598252A
Application number: CN202010365827.8A
Authority: CN
Inventors: 朱磊; 吕泓瑾; 黑新宏; 冯林林; 张晋源; 刘旭华; 刘尧林; 林泓; 刘瑞
Original assignee: Xian University of Technology
Current assignee: Xian University of Technology
Priority date: 2020-04-30
Filing date: 2020-04-30
Publication date: 2020-08-28
Anticipated expiration: 2040-04-30
Also published as: CN111598252B

Abstract

The invention discloses a university computer basic knowledge problem solving method based on deep learning, which comprises the steps of generating an inference data set containing all basic knowledge points by using an existing university computer basic knowledge graph and an NTN model, then generating a incomplete knowledge graph corresponding to problems by using a PCANET and a GRU network, then matching an approximate subgraph which is closest to the incomplete knowledge graph in the knowledge graph by using a PSQUERY algorithm, and inferring a missing part in the incomplete knowledge graph by using a TransE method, wherein the inferred part is a correct result. The manpower consumption is reduced, and the problem solving efficiency is improved. The invention is easy to realize and can generate answers in batch. The invention improves the efficiency of students self-learning the class and simultaneously reduces the burden of teachers in teaching the class.

Description

University computer basic knowledge problem solving method based on deep learning

Technical Field

The invention belongs to an important direction in the field of artificial intelligence, and particularly relates to a university computer knowledge problem solving method based on deep learning.

Background

With the development of knowledge graph technology, the number of knowledge graphs owned by people is more and more, and the scale of the knowledge graphs is larger and larger. Knowledge maps store a large amount of knowledge in a triplet structure. The extraction of entity relationships is a classic human task in the knowledge graph, and is continuously developed in the past decades, and some stage achievements are obtained. With the coming of the deep learning era, various neural network models also provide a new idea for the study on the aspect of knowledge maps. Compared with the traditional method, the deep learning model has more efficient learning ability, can face more complex text contexts, and can process larger-scale training data. Deep learning therefore also provides a more novel solution for knowledge-graph research in some aspects. With the rapid development of computer technology and the coming of IT era, computers have been integrated into every corner of people's lives. When many students learn the course by self, the situation of low efficiency can occur due to the infirm basic knowledge points. The out-of-class answering time often cannot meet the requirements of a student.

In view of the above, it is important to be able to master the course of "university computer base". In the study of some classmates, the condition that the questions exist but the standard answers do not exist is often found. Therefore, it is very important to invent a system that can obtain answers according to the topics. The systems of this aspect used today are mostly answered manually. Under the condition of lacking more professional knowledge, the method is low in speed and has no guarantee on accuracy. The present invention is a solution to this problem.

Disclosure of Invention

The invention aims to provide a method for solving the problem of basic knowledge of a university computer based on deep learning, which solves the problem of low efficiency of manually giving answers in the prior art.

The method is based on deep learning algorithm and model, uses existing university computer basic knowledge map and NTN model to generate inference data set containing all basic knowledge points, then uses PCANET and GRU network to generate incomplete knowledge map corresponding to problems, then uses PSQUERY algorithm to match approximate subgraph closest to the incomplete knowledge map, and uses TransE method to infer missing part in the incomplete knowledge map, and finally infers the obtained part as correct result.

A university computer knowledge problem solving method based on deep learning is characterized in that a computer basic knowledge graph and an NTN model are used for generating an inference data set containing all basic knowledge points, then a residual knowledge graph corresponding to problems is generated by using a PCANET and GRU network, then a PSQUERY algorithm is used for matching an approximate subgraph which is closest to the residual knowledge graph in the knowledge graph, and a TransE method is used for inferring a missing part in the residual knowledge graph, so that a correct result of a problem is obtained.

The method specifically comprises the following steps:

step 1, carrying out data preprocessing on a basic knowledge graph of a computer to change the basic knowledge graph from a visual form such as a neo4j database to a triple form; simultaneously, screening out knowledge points outside the examination range;

step 2, training the triples in the step 1 by using an NTN network model to generate a reasoning training set of the knowledge points;

step 3, importing a PCANET network model, and training a PCANET network capable of forming a knowledge graph by using the inference training set generated in the step 2;

and 4, training by using the data set processed in the step 1 and a GRU network model to generate a data set capable of identifying the keywords of the knowledge points.

And 5, carrying out data preprocessing on the basic knowledge question bank of the computer to obtain a data set which can be trained by the PCANET network model.

And 6, training the data set which is formed in the step 4 and can identify the key words of the knowledge points and the data set of the basic knowledge graph of the computer formed in the step 5 by using the PCANET network model generated in the step 3 to generate the incomplete knowledge graph for describing the problems.

And 7, matching approximate subgraphs of the knowledge graph in the basic knowledge graph of the university computer, which is the incomplete knowledge graph describing the problem in the step 6, by using a PSQUERY method.

Step 8, using the TransE method, the missing part in the knowledge map used to describe the problem is obtained.

And 9, at this time, the incomplete part in the knowledge graph is a correct answer, and the incomplete part comprises the character concept and the topological relation.

In step 2, the advantages of using NTN and triplets are: the NTN is mainly trained by representing entities in the database as a vector, and a triplet may also be regarded as a vector, i.e., each data in the triplet is regarded as a value in the vector. Therefore, the two can be combined, thereby achieving the effect of training the triad by using the NTN; specifically, NTN is used for training the triplets; the NTN represents each object or individual in the dataset as a vector, operating directly on the vector. The main steps of NTN are: and writing a self-defined layer, initializing tensor shapes, activation functions and tensor parameters, then defining a contrast maximum edge loss function, and finally summarizing data training to obtain a result. These vector carriers can capture facts about the entity and whether it becomes part of relationships, each relationship being defined by the parameters of a new neural tensor network; since NTN may explicitly involve two entity vectors. The effect of training the triplet using NTN will be better.

Step 4, training and generating a data set of the knowledge point keywords by using a GRU network model; the GRU can transmit the current data to be used at the next moment; the GRU is composed of numerous structures called memory blocks, each memory block comprises an input gate, an output gate, a forgetting gate and a memory unit; the input gate judges the data which is selected to pass through according to the output data of the data which is input into the input gate by the activating function; a commonly used activation function is the sigmoid function sigmood (x) ═ 1/(1+ e ^ (-x)), where a number with a value (— ∞) can be mapped to (0, 1); if the number of mappings is greater than a threshold set in advance, then this data can be output, otherwise not;

the activation function comprises a sigmoid activation function, a tanh activation function and a Relu activation function.

In step 7, the optimal approximate subgraph of the result generated by matching the PSQUERY algorithm is used, firstly, the PSQUERT algorithm extracts the characteristics of the query graph, and then coding is respectively carried out according to the extracted characteristics to form node and graph codes; then, an index tree is constructed based on graph coding for filtering; and finally, generating a candidate graph set, and performing sub-graph isomorphism verification to finally obtain a final result set.

In step 9, TransE is used to obtain the incomplete part of the knowledge graph; TransE treats the relationships in each triplet instance as a transformation from a head entity to a tail entity based on a distributed vector representation of the entities and relationships.

The invention has the beneficial effects that:

aiming at the problem that part of question banks in the basic knowledge of a college computer have no standard answers, the invention adopts an NTN network model to train a knowledge graph, uses a GRU network model to train and generate a data set of key words of knowledge points, uses an optimal approximate subgraph of a PSQUERY algorithm matching and generating result, and uses TransE to obtain the incomplete part of the knowledge graph. Compared with the traditional algorithm, the method increases the efficiency of algorithm operation and the accuracy of answers.

Drawings

FIG. 1 is a general flow chart of a university computer basic problem solving method based on deep learning according to the university computer knowledge solving method based on deep learning of the present invention;

FIG. 2 is a process of matching approximate subgraphs based on the PSQUERY algorithm for a university computer knowledge problem solving method based on deep learning according to the present invention;

FIG. 3 is a process of the university computer knowledge problem solving method based on deep learning according to the present invention, in which an inference data set of knowledge points is generated based on NTN;

FIG. 4 is a process of the university computer knowledge problem solving method based on deep learning in the invention, which is based on PCANET to generate incomplete knowledge map corresponding to the problem.

FIG. 5 is a process of the university computer knowledge problem solving method based on deep learning according to the present invention, which generates a keyword data set containing knowledge points based on GRUs.

FIG. 6 is a process of obtaining a missing part in a knowledge graph based on a TransE method by the deep learning-based university computer knowledge problem solving method.

Detailed Description

The following describes embodiments of the present invention with reference to the drawings.

The invention relates to a university computer basic knowledge problem solving method based on deep learning, which specifically comprises the following steps as shown in figure 1:

step 1.1, preprocessing data of the basic knowledge graph of the university computer to convert the basic knowledge graph from a neo4j database and the like into a data form of a triple.

And step 1.2, removing the knowledge points which are not required by the examination in the data set and only need to be known by students.

And 2, training the triples by using an NTN network model to generate an inference data set of the knowledge points. As shown in fig. 3, the tensor shape, activation function, tensor parameters, and custom layer in the NTN model are fixed or may be slightly modified in practice. A maximum edge loss function is also determined. The pseudo code is as follows:

and 3, as shown in fig. 4, importing a PCANET network model, and training the PCANET network model which can generate the knowledge graph by using the inference training set generated in the step 2.

And 4, training by using the data set processed in the step 1 and a GRU network model to generate a data set containing the keywords of the knowledge points. As shown in fig. 5.

And 5.1, reading the corpus data of the basic knowledge question bank of the university computer, and segmenting all data by taking a complete question as a basic unit.

Step 5.2, all topics are numbered for lookup.

And 6.1, inputting the data set generated in the step 4 into the PCANET network model generated in the step 3.

And 6.2, generating a knowledge graph describing the incomplete problem by using the PCANET network model. As shown in fig. 4.

And 7.1, extracting the marking information of the nodes to generate node marking codes, and combining all the node marking codes to generate the node marking codes of the graph.

And 7.2, extracting edge characteristics as adjacent edge information. The encoding of the adjacent edges can be obtained by traversing the hash mapping of the adjacent edges.

And 7.3, extracting the shortest weight path as a third feature. May be implemented using dijkstra's algorithm. Wherein, list represents the weight matrix of the current path. The initial state is a weight between two connected points, and if the two points are not connected, the initial state is considered to be infinite. The path represents the shortest path between two points, and the initial state is the same as the list. The pseudo code is as follows:

and 7.4, taking the minimum value from the weight of each node to obtain the weight path code of the graph D.

And 7.5, generating an N-layer generated graph for each node, calculating a graph to be used for expressing the local topology information of each node, and finally combining to obtain the topology information of the undirected weighted graph. Namely a Laplacian map. The pseudo code is as follows:

and 7.6, solving corresponding characteristic values according to a Jacobian algorithm, and taking the two largest characteristic values as the maps of the nodes. And combining the maps of all the nodes to generate the map of the map.

And 7.7, constructing the index tree by referring to the GCoding method.

And 7.8, filtering the graph and the nodes, and cutting most of the graph by PSQUERY to generate a candidate set, as shown in FIG. 2.

And 7.9, performing sub-graph isomorphism judgment by adopting a VF2 algorithm. The pseudo code is as follows:

step 8, as shown in FIG. 6, a TransE-based approach is used to obtain the missing part of the knowledge-graph describing the problem. In this algorithm, the parameters of hyper-parameters, learning rate, etc. are generally determined and can be slowly modified through practice. The pseudo code of the TransE algorithm is as follows:

TransE treats the relationships in each triplet instance as a transformation from a head entity to a tail entity based on a distributed vector representation of the entities and relationships. For example, the phrase "the originator of the Linux operating system is linnas. bennard. tokwatts" may be expressed as "Linux operating system + originator ═ linnas. bennard. tokwatts". In the incomplete knowledge map, by analogy, as long as the entity ' Linux operating system ' and the relation ' founder ' are known, the entity to be solved can be inferred to be ' linnas. Benne ' kth. Towatz '.

And 9, converting the incomplete part in the knowledge graph into a correct answer described by characters in a manual mode.

The invention reasonably utilizes the characteristics of the basic knowledge map of the computer and the university computer question bank, reduces the consumption of manpower and improves the efficiency of solving the questions. Firstly, the problem is converted into a incomplete knowledge map, then the incomplete knowledge map is matched with the knowledge points, an optimal matching subgraph is found out, and finally, the knowledge map of the incomplete part of the problem is found, namely the answer of the problem. The invention is easy to realize and can generate answers in batch. The invention improves the efficiency of students self-learning the class and simultaneously reduces the burden of teachers in teaching the class.

Claims

1. A university computer knowledge problem solving method based on deep learning is characterized in that a computer basic knowledge graph and an NTN model are used for generating an inference data set containing all basic knowledge points, then a incomplete knowledge graph corresponding to problems is generated by using a PCANET and GRU network, then a PSQUERY algorithm is used for matching an approximate subgraph which is closest to the incomplete knowledge graph in the knowledge graph, and a TransE method is used for inferring a missing part in the incomplete knowledge graph, so that a correct result of the problem is obtained.

2. The university computer knowledge problem solving method based on deep learning according to claim 1, characterized in that the method specifically comprises the following steps:

And 7, matching the optimal approximate subgraph of the knowledge graph which is the incomplete knowledge graph describing the problem in the step 6 in the basic knowledge graph of the university computer by using a PSQUERY method.

3. The university computer knowledge problem solving method based on deep learning of claim 2, wherein the advantage of using NTN and the triplets in step 2 is that: the NTN is mainly trained by representing entities in the database as a vector, and a triplet may also be regarded as a vector, i.e., each data in the triplet is regarded as a value in the vector. Therefore, the two can be combined, thereby achieving the effect of training the triad by using the NTN; specifically, NTN is used for training the triplets; the NTN represents each object or individual in the dataset as a vector, operating directly on the vector. The main steps of NTN are: and writing a self-defined layer, initializing tensor shapes, activation functions and tensor parameters, then defining a contrast maximum edge loss function, and finally summarizing data training to obtain a result. These vector carriers can capture facts about the entity and whether it becomes part of the relationships, each defined by the parameters of a new neural tensor network.

4. The university computer knowledge problem solving method based on deep learning of claim 2, wherein in the step 4, a data set of knowledge point keywords is generated by using GRU network model training; the GRU can transmit the current data to be used at the next moment; the GRU is composed of numerous structures called memory blocks, each memory block comprises an input gate, an output gate, a forgetting gate and a memory unit; the input gate judges the data which is selected to pass through according to the output data of the data which is input into the input gate by the activating function; a commonly used activation function is the sigmoid function sigmood (x) ═ 1/(1+ e ^ (-x)), where a number with a value (— ∞) can be mapped to (0, 1); if the number of mappings is greater than a threshold set in advance, this data can be output, otherwise it is not.

5. The university computer knowledge problem solving method based on deep learning of claim 4, wherein the activation function is sigmoid activation function, tanh activation function, Relu activation function.

6. The university computer knowledge problem solving method based on deep learning according to claim 2, wherein in the step 7, an optimal approximate subgraph of a result generated by matching with a PSQUERY algorithm is used, firstly, the PSQUERT algorithm extracts the features of a query graph, and then, coding is respectively carried out according to the extracted features to form node and graph codes; then, an index tree is constructed based on graph coding for filtering; and finally, generating a candidate graph set, and performing sub-graph isomorphism verification to finally obtain a final result set.

7. The university computer knowledge problem solving method based on deep learning of claim 2, wherein in the step 9, TransE is used to obtain the incomplete part of the knowledge graph; TransE treats the relationships in each triplet instance as a transformation from a head entity to a tail entity based on a distributed vector representation of the entities and relationships.