CN114969260A - Automatic subjective question evaluation and reading deep learning method combining test question classification and evaluation learning - Google Patents
Automatic subjective question evaluation and reading deep learning method combining test question classification and evaluation learning Download PDFInfo
- Publication number
- CN114969260A CN114969260A CN202210597773.7A CN202210597773A CN114969260A CN 114969260 A CN114969260 A CN 114969260A CN 202210597773 A CN202210597773 A CN 202210597773A CN 114969260 A CN114969260 A CN 114969260A
- Authority
- CN
- China
- Prior art keywords
- test
- questions
- test question
- subjective
- answer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/3331—Query processing
- G06F16/334—Query execution
- G06F16/3344—Query execution using natural language analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/35—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention discloses an automatic subjective question evaluation deep learning method combining test question classification and evaluation learning, and relates to the crossing field of artificial intelligence and intelligent education. According to the invention, aiming at the characteristics that different types of subjective questions have different evaluation modes, the automatic evaluation neural network model for the subjective questions can master the evaluation method for the different types of the test questions through the combined test question classification and evaluation learning, so that the automatic evaluation problem of the subjective questions is solved by a more effective method. Firstly, the subjective questions to be evaluated are classified through an integrating neural network, and the types of the subjective questions to be evaluated are obtained. Then, the test question type embedding of the subjective questions to be evaluated is connected with each character embedding of the student answer input sequence and the standard answer input sequence to obtain student answer input and standard answer input of test question type perception, and high-precision scoring is carried out through a heterozygous model consisting of convolution and a bidirectional long-short term memory network.
Description
Technical Field
The invention relates to the crossing field of artificial intelligence and intelligent education, in particular to a subjective question automatic scoring deep learning method combining test question classification and scoring learning, which can be widely applied to a subjective question computer automatic scoring system of each subject.
Background
The examination questions in the examination paper are generally divided into two categories, objective questions and subjective questions, in terms of the form of answer composition. The test questions such as the single-choice question, the multiple-choice question, and the judgment question, whose answers are expressed by the choice numbers, are called objective questions, and the test questions such as the short answer question, the noun explanation question, and the discussion question, whose answers are expressed by natural language, are called subjective questions. Because the answers of objective questions such as single-choice questions, multiple-choice questions, judgment questions and the like are all expressed by option numbers, when the current computer automatically scores the questions, only simple matching operation needs to be carried out on the option numbers of the standard answers and the option numbers of the student answers, and the answers are correct if the matching is successful, so that the processing technology has achieved better results. However, the automatic scoring technology of the subjective questions with answers expressed by natural language is as follows: for automatic examination paper of simple answer questions, noun explanation and discussion questions and the like, the effect is not ideal because the examination paper is influenced by theories and technical bottlenecks such as natural language understanding, artificial intelligence and the like.
With the continuous development of artificial neural network technology, many deep learning models such as LSTM-based models, CNN & LSTM-based models, and transform-based models are applied to subjective question review. These deep learning models automatically extract semantic features from the answer text using different neural networks, thereby providing an end-to-end approach that does not require any artificial feature engineering. However, the deep learning method of automatic subjective question review still has many challenges. Among these, one important challenge is: the answers of some question types of the subjective questions have the orderliness, but the answers of some question types of the subjective questions do not have the orderliness, for example, for the test question, "please write the units of the storage capacity in sequence from small to large", the answer key points "B, KB, MB, and GB" have the orderliness; however, for the test question "brief description of the advantages of twisted pair", the answer key points "high flexibility, oil resistance, water resistance, wear resistance, acid resistance, cold resistance, bending resistance" have no sequence, that is, the answer of the student "water resistance, cold resistance, bending resistance, oil resistance, wear resistance, acid resistance, high flexibility" is also completely correct. Therefore, how to automatically learn different review methods for different question types is one of the problems to be solved by the subjective question automatic review deep learning method.
Disclosure of Invention
The invention discloses an automatic subjective question evaluation deep learning method combining test question classification and grading learning.
In order to achieve the purpose, the technical scheme of the invention is as follows:
a subjective question automatic evaluation deep learning method combining test question classification and evaluation learning is characterized by comprising the following steps:
s1, classifying the test questions of the subjective questions to be evaluated through a CNN test question classification submodel with residual connection to obtain the types of the test questions of the subjective questions to be evaluated;
s2, embedding and connecting the test question type of the subjective question to be evaluated with each character of the student answer input sequence and the standard answer input sequence to obtain student answer perception input and standard answer perception input of test question type perception, and then sending the student answer perception input and the standard answer perception input into a subjective question test paper grading sub-model consisting of CNN and Bi-LSTM to grade to obtain the grading grade of the student answer;
s3, training a subjective question automatic evaluation neural network model for combined test question classification and evaluation learning through the loss functions of the combined calculation step S1 and the step S2;
s4, automatically evaluating the neural network model by using the formed subjective questions trained in the step S3, and evaluating any student answers;
CNN is an abbreviation for the English conditional Neural Networks, representing a Convolutional Neural network, and is described by the documents "Lawrence S, Giles CL, Tsio AC et al. face repetition: A conditional Neural Networks, IEEE Transactions on 1997; 8(1) 98-113';
the Bi-LSTM is an abbreviation of English Directional Long Short-Term Memory and represents a two-way Long-Short Term Memory Neural network, wherein the LSTM Neural network is proposed by the documents "Sepp Hochreiter and Juregen Schmidhuber.1997 Long Short-Term Memory, 9(8): 1735-;
the subjective question types comprise definition subjective question test questions, sequence subjective question test questions and general subjective question test questions, wherein the definition subjective question test questions refer to subjective questions requiring students to explain definition of a certain concept, the sequence subjective question test questions refer to subjective questions with answers including a plurality of key points given in sequence, and the general subjective question test questions refer to other subjective questions except the definition and the sequence.
Further, the step S1 specifically includes:
s1.1 character sequence of test question to be classifiedWord embedding sequence T converted into test questions to be classified through table look-up operation e The calculation process is as follows:
wherein n represents the sentence length of the test question to be classified, onehot (T) represents the conversion of the test question word sequence T into a unique heat matrix, T o Is a one-hot matrix of the test word sequence T, | V | represents the size of a dictionary V, which is a dictionary established by a model and containing all words,a matrix is embedded for the pre-trained words corresponding to dictionary V,representing a word embedding vector corresponding to the ith word in the test question word sequence T, and h representing the dimension of the word embedding vector;
s1.2 embedding words of test questions to be classified into sequence T e Inputting the data into a convolutional neural network adopting zero filling to obtain shallow semantic representation T of the test question to be classified c The calculation process is as follows:
wherein PCNN (-) represents a convolutional neural network with zero padding, (T) c ) ij Represents T c The ith row and the jth column of (c), k is the convolution kernel size, d c Is the number of the convolution kernels and is,word-embedded sequence T representing test questions to be classified e As a result of the zero-padding being performed,to representThe j (th) column to the j + (k-1) column of the matrix,the weight of the ith convolution kernel is represented,is shown asThe offset of the i convolution kernels is such that,representing a convolution operation;
s1.3 shallow semantic representation T of test questions to be classified c Sending the test question into a feedforward neural network consisting of two full-connection layers for enhancement to obtain enhanced semantic representation of the test question to be classifiedThe calculation process is as follows:
wherein the content of the first and second substances,respectively the weight matrices of the two fully-connected layers,bias vectors of two fully-connected layers respectively, ReLU is an abbreviation of English 'Rectified Linear Unit', and ReLU (·) represents an activation function adopting modified Linear Unit operation;
s1.4, representing T by shallow semantics of test questions to be classified c With its enhanced semantic representation T f Residual error connection is carried out to obtain deep semantic representation T of the test question to be classified u The calculation process is as follows:
wherein LayerNorm (·) denotes layer normalization;
s1.5 deep semantic representation T of test questions to be classified u Performing maximum pooling to obtain a classification vector Z of the test questions to be classified c The calculation process is as follows:
wherein maxPooling (·) represents a max pooling operation;
s1.6 Classification vector Z of test questions to be classified c Executing linear transformation of softmax, performing probability calculation of the test question types, and obtaining the final test question types, wherein the calculation process is as follows:
o=M c Z c +b o (8)
wherein the content of the first and second substances,is a matrix of the test question type representation,is an offset vector, d y Is the number of the categories of the test question types,represents Z c Confidence score vectors over all test question types, y being one test question type, o y Represents Z c Confidence score, o, labeled test question type y i Represents Z c Confidence score, Ρ (yz | Z) on the ith test question type c ) Classification vector Z representing test questions to be classified c The prediction probability on the question type Y, which is the set of all question types,the representation returns such that Pp (y | Z) c ) Test question type of maximum value, y test question type of final classification of test questions to be classified, exp(. cndot.) represents an exponential function based on a natural constant e.
Further, the step S2 specifically includes:
s2.1, performing table look-up operation on the student answer word sequence and the standard answer word sequence according to the step S1.1 to respectively obtain student answer word embedded sequencesStandard answer word embedding sequenceThen embedding the test words into the sequence T e Respectively embedding the answer words with studentsStandard answer word embedding sequenceSplicing the head and the tail to obtain an answer input sequence A of the student e And a standard answer input sequence S e The calculation process is as follows:
wherein, ": "denotes the end-to-end concatenation operation of the sequence, n1 is the sentence length of student answers, n2 is the sentence length of standard answers, m1 is the student answer input sequence A e M1 ═ n + n1, m2 is the standard answer input sequence S e And m2 ═ n + n 2;
s2.2 converting the test question type y obtained in the step S1 into the corresponding test question type embedding through table lookup operationThen v is measured t Respectively input with student's answer input sequence A e And a standard answer input sequence S e Each character in the Chinese character string is embedded and connected to obtain student answer perception input A t And standard answer perception input S t The calculation process is as follows:
wherein the content of the first and second substances,respectively, student's answer input sequence A e And student answer perception input A t Word-embedding vector of the ith column, d t Is the dimension of the embedding of the test question types,respectively, standard answer input sequence S e And standard answer perception input S t The word in column j embeds a vector, "; "represents a concatenation operation of vectors;
s2.3 inputting student answer perception A t And standard answer perception input S t Respectively sending the data to two convolutional neural networks which adopt zero padding and share parameters to obtain shallow semantic representation A of student answers c And a standard answer shallow semantic representation S c The calculation process is as follows:
wherein sharedPCNN1(·) and sharedPCNN2 denote that two take zeroConvolutional neural networks populated and sharing parameters, d c The number of convolution kernels;
s2.4 shallow semantic representation A of student answers c And the standard answer shallow semantic representation S c Sending the data into an interactive network consisting of Bi-LSTM to obtain a scoring vector Z of student answers g As follows:
wherein Interactive (-) represents an Interactive network composed of Bi-LSTM, d L Dimension of hidden state of Bi-LSTM in interactive network;
s2.5 Scoring vector Z to student' S answer g Performing linear transformation of softmax, performing probability calculation of the grade, and obtaining a final grade, wherein the calculation process is as follows:
x=M g Z g +b g (18)
wherein the content of the first and second substances,is a matrix of the score level representation,is an offset vector, d g Is the number of the rating levels to be scored,is represented by Z g Vector of confidence scores of all the rating levels, g being one rating level, x g Represents Z g Marked as score level gConfidence score, x i Represents Z g Confidence score on the i-th scoring level, Ρ (g | Z) g ) Representing a score vector Z g The prediction probability on the score scale G, G being the set of all score scales, G being the final rated score scale,denotes return such that Pp (g | Z) g ) The score rating of the maximum value.
Further, the step S2.4 specifically includes:
s2.4.1 use Bi-LSTM of two shared parameters to respectively represent A to the shallow semantic of student answer c And the standard answer shallow semantic representation S c Encoding to obtain student answer context expression A r And a standard answer context representation S r The calculation process is as follows:
wherein ShardedBilSTM 1 (-), ShardedBilSTM 2 (-), represent Bi-LSTM sharing two parameters,respectively represent A c And S c The connection result of the Bi-directional hidden state corresponding to the ith time step in the Bi-LSTM,respectively represent A c And S c From left to rightHidden state at the ith time step in the network,respectively represent A c And S c In a right to left directionHidden state at ith time step in the network;
s2.4.2 context representation of student answers by attention mechanism A r And a standard answer context representation S r Interact to get mutual attentionAndthe calculation process is as follows;
wherein the content of the first and second substances,is composed ofThe (c) th element of (a),the (c) th element of (a),respectively represent pairAndthe transposition operation is carried out and the operation is carried out,to refer to a logical quantity of "all",respectively representing hidden state tuplesAndattention weight of (1);
s2.4.3 respectively will beAndwith student answer context representation A r And a standard answer context representation S r Fusing to obtain student answer fusion expression A f Fused representation S with standard answers f The calculation process is as follows:
wherein the content of the first and second substances,is A f The (c) th element of (a),is S f The ith element in (1), Fusion (-), represents a Fusion operation, and an element-by-element multiplication operation of the matrix;
s2.4.4 fusion of student answers using two shared parameters Bi-LSTM respectively f Fused representation S with standard answers f Performing enhancement processing to obtain deep semantic representation A of student answers u And standard answer deep semantic representation S u The calculation process is as follows:
wherein the content of the first and second substances,respectively represent A f And S f The connection result of the Bi-directional hidden state corresponding to the ith and jth time steps in the Bi-LSTM;
s2.4.5 deep semantic representation A of student answers respectively u And standard answer deep semantic representation S u Carrying out average pooling and maximum pooling, and connecting all pooled vectors to obtain score vector Z of student answer g The calculation process is as follows:
wherein maxPooling (. cndot.) represents the maximum pooling operation, avePooling (. cndot.) represents the average pooling operation,is A u The maximum pooling vector of (a) is,is A u The average pooling vector of (a) is,is S u The maximum pooling vector of (a) is,is S u The average pooling vector of "; "represents a concatenation operation of vectors.
Further, the step S3 specifically includes:
s3.1, calculating a loss function of the test question classification submodel and a loss function of the subjective question automatic scoring submodel by using the cross entropy loss error respectively, wherein the calculation process is as follows:
wherein, omega is a training set of the subjective question automatic evaluation task of the joint test question classification and evaluation learning, | omega | represents the size of the training set omega,is the true test question classification of the ith training sample in omega,is the true score level of the ith training sample in omega,is the classification vector of the test question of the ith training sample in omega,is the score vector of student answer of the ith training sample in omega, gamma 1 Is a loss function used in the training of the test question classification submodel, gamma 2 Is a loss function used when carrying out the automatic scoring submodel training of the subjective questions;
s3.2, calculating a joint loss function of the subjective question automatic evaluation neural network model for joint test question classification and evaluation learning by using the following formula
Where λ and β are two weight parameters;
s3.3 the training goal of the neural network model for automatic review of subjective questions in combination of test question classification and score learning is to minimize the combined loss error of the formula (40).
Further, the table lookup operation in step S2.2 specifically includes:
wherein, OneHot (y) * ) Label y for indicating to-be-tested question type * Is converted into a one-hot vector,for test question type y * The one-hot vector of (a) is,embedding matrices for question types, d y Is the number of categories of test question types, d t Is the dimension of the question type embedding.
The invention provides an automatic subjective question evaluation deep learning method combining test question classification and grading learning aiming at the characteristic that different types of subjective questions have different evaluation modes. Firstly, classifying the subjective questions to be evaluated through a CNN test question classification submodel with residual connection to obtain the types of the subjective questions to be evaluated. Then, the test question type embedding of the subjective questions to be evaluated is connected with each character embedding of the student answer input sequence and the standard answer input sequence to obtain student answer perception input and standard answer perception input of test question type perception, and the examination paper is evaluated on the basis.
The invention has the following advantages:
(1) classifying the test questions of the subjective questions to be evaluated through a CNN test question classification submodel with residual connection, so that the model can master the types of the test questions;
(2) embedding and connecting the test question type of the subjective question to be evaluated with each character of the student answer input sequence and the standard answer input sequence, so that the grading sub-module can sense the test question type;
(3) the test question classification and the test question scoring are jointly learned, so that the precision of the test question classification and the adaptivity of the test question scoring can be further improved;
(4) the scoring is carried out through a hybrid neural network consisting of a convolutional neural network and a bidirectional long-short term memory neural network, so that the scoring precision can be greatly improved.
Drawings
FIG. 1 is a schematic flow diagram of the process of the present invention.
FIG. 2 is a schematic diagram of the Bi-LSTM interaction network of the present invention.
Detailed Description
The present invention is further illustrated by the following specific examples, but the scope of the present invention is not limited to the following examples.
According to the method flow chart shown in fig. 1, the automatic subjective question evaluation deep learning method combining test question classification and evaluation learning is realized through the following steps:
s1, classifying the test questions of the subjective questions to be evaluated through a CNN test question classification submodel with residual connection to obtain the types of the test questions of the subjective questions to be evaluated;
s2, embedding and connecting the test question type of the subjective question to be evaluated with each character of the student answer input sequence and the standard answer input sequence to obtain student answer perception input and standard answer perception input of test question type perception, and then sending the student answer perception input and the standard answer perception input into a subjective question test paper grading sub-model consisting of CNN and Bi-LSTM to grade to obtain the grading grade of the student answer;
s3, training a subjective question automatic evaluation neural network model for combined test question classification and evaluation learning through the loss functions of the combined calculation step S1 and the step S2;
s4, automatically evaluating the neural network model by using the formed subjective questions trained in the step S3, and evaluating any student answers;
CNN is an abbreviation of the English conditional Neural Networks, representing a Convolutional Neural network, and is known from the literature "Lawrence S, Giles CL, Tsio AC et al. face recognition: A conditional Neural-network approach. Neural Networks, IEEE Transactions on 1997; 8(1) 98-113';
the Bi-LSTM is an abbreviation of English Directional Long Short-Term Memory and represents a two-way Long-Short Term Memory Neural network, wherein the LSTM Neural network is proposed by the documents "Sepp Hochreiter and Juregen Schmidhuber.1997 Long Short-Term Memory, 9(8): 1735-;
the subjective question types comprise definition subjective question test questions, sequence subjective question test questions and general subjective question test questions, wherein the definition subjective question test questions refer to subjective questions requiring students to explain definition of a certain concept, the sequence subjective question test questions refer to subjective questions with answers comprising a plurality of key points given in sequence, and the general subjective question test questions refer to other subjective questions except the definition and the sequence.
Further, the step S1 specifically includes:
s1.1 character sequence of test question to be classifiedWord embedding sequence T converted into test question to be classified through table look-up operation e The calculation process is as follows:
wherein n represents the sentence length of the test question to be classified, onehot (T) represents the conversion of the test question word sequence T into a unique heat matrix, T o Is a one-hot matrix of the test word sequence T, | V | represents the size of a dictionary V, which is a dictionary established by a model and containing all words,a matrix is embedded for the pre-trained words corresponding to dictionary V,representing a word embedding vector corresponding to the ith word in the test question word sequence T, and h representing the dimension of the word embedding vector;
s1.2 embedding words of test questions to be classified into sequence T e Inputting the data into a convolutional neural network adopting zero filling to obtain shallow semantic representation T of the test question to be classified c The calculation process is as follows:
wherein PCNN (. cndot.) represents a convolutional neural network with zero padding, (T) c ) ij Represents T c The ith row and the jth column of (c), k is the convolution kernel size, d c Is the number of the convolution kernels and is,word-embedded sequence T representing test questions to be classified e As a result of the zero-padding being performed,to representThe j (th) column to the j + (k-1) column of the matrix,the weight of the ith convolution kernel is represented,represents the offset of the ith convolution kernel,representing a convolution operation;
s1.3 shallow semantic representation T of test questions to be classified c Sending the test question into a feedforward neural network consisting of two full-connection layers for enhancement to obtain enhanced semantic representation of the test question to be classifiedThe calculation process is as follows:
wherein the content of the first and second substances,respectively the weight matrices of the two fully-connected layers,bias vectors of two fully-connected layers respectively, ReLU is an abbreviation of English 'Rectified Linear Unit', and ReLU (·) represents an activation function adopting modified Linear Unit operation;
s1.4, expressing T by shallow semantic of test questions to be classified c With its enhanced semantic representation T f Residual error connection is carried out to obtain deep semantic representation T of the test question to be classified u The calculation process is as follows:
wherein LayerNorm (·) denotes layer normalization;
s1.5 deep semantic representation T of test questions to be classified u Performing maximum pooling to obtain a classification vector Z of the test questions to be classified c The calculation process is as follows:
wherein maxPooling (·) represents a max pooling operation;
s1.6 Classification vector Z of test questions to be classified c Executing linear transformation of softmax, performing probability calculation of the test question types, and obtaining the final test question types, wherein the calculation process is as follows:
o=M c Z c +b o (8)
wherein, the first and the second end of the pipe are connected with each other,is a matrix of the test question type representation,is an offset vector, d y Is the number of the categories of the test question types,represents Z c Confidence score vectors over all test question types, y being one test question type, o y Represents Z c Confidence score, o, labeled test question type y i Represents Z c Confidence score, Ρ (yz | Z) on the ith test question type c ) Classification vector Z representing test questions to be classified c In trialThe prediction probability on topic type Y, which is the set of all question types,the representation returns such that Pp (y | Z) c ) Type of question at maximum, y * For the finally classified test question type of the test questions to be classified, exp (-) represents an exponential function with a natural constant e as the base.
Further, the step S2 specifically includes:
s2.1, performing table look-up operation on the student answer word sequence and the standard answer word sequence according to the step S1.1 to respectively obtain student answer word embedded sequencesStandard answer word embedding sequenceThen embedding the test question words into the sequence T e Respectively embedding the answer words with studentsStandard answer word embedding sequenceSplicing the head and the tail to obtain an answer input sequence A of the student e And a standard answer input sequence S e The calculation process is as follows:
wherein, ": "denotes the end-to-end concatenation operation of the sequence, n1 is the sentence length of student answers, n2 is the sentence length of standard answers, m1 is the student answer input sequence A e And m1 is n + n1, m2 is the standard answerInput sequence S e And m2 ═ n + n 2;
s2.2 converting the test question type y obtained in the step S1 into the corresponding test question type embedding through table lookup operationThen v is measured t Respectively input with student's answer input sequence A e And a standard answer input sequence S e Each word in the Chinese character string is embedded and connected to obtain student answer perception input A t And standard answer perception input S t The calculation process is as follows:
wherein the content of the first and second substances,respectively, student's answer input sequence A e And student answer perception input A t Word-embedding vector of the ith column, d t Is the dimension of the embedding of the test question types,respectively, standard answer input sequence S e And standard answer perception input S t The word in column j embeds a vector, "; "represents a concatenation operation of vectors;
s2.3 inputting student answer perception A t And standard answer perception input S t Respectively sending the data to two convolutional neural networks which adopt zero padding and share parameters to obtain shallow semantic representation A of student answers c And the standard answer shallow semantic representation S c The calculation process is as follows:
wherein sharedPCNN1(·), sharedPCNN2 represent two convolutional neural networks with zero padding and shared parameters, d c The number of convolution kernels;
s2.4 shallow semantic representation A of student answers c And the standard answer shallow semantic representation S c Sending the data into an interactive network consisting of Bi-LSTM to obtain a scoring vector Z of student answers g As follows:
where Interactive (. cndot.) represents an Interactive network consisting of Bi-LSTM, d L Dimension of hidden state of Bi-LSTM in interactive network;
s2.5 Scoring vector Z to student' S answer g Performing linear transformation of softmax, performing probability calculation of the grade, and obtaining a final grade, wherein the calculation process is as follows:
x=M g Z g +b g (18)
wherein the content of the first and second substances,is a matrix of the score level representation,is an offset vector, d g Is the number of the rating levels to be scored,is represented by Z g Vector of confidence scores of all the rating levels, g being one rating level, x g Represents Z g Confidence score, x, labeled as rating g i Represents Z g Confidence score on the i-th scoring level, Ρ (g | Z) g ) Representing a score vector Z g Prediction probability on a score level G, G being the set of all score levels, G * In order to achieve the final rating of the rating scale,denotes return such that Pp (g | Z) g ) The score rating of the maximum value.
Further, the step S2.4 specifically includes:
s2.4.1 shallow semantic representation A of student answers using two Bi-LSTMs sharing parameters respectively c And the standard answer shallow semantic representation S c Encoding to obtain student answer context expression A r And a standard answer context representation S r The calculation process is as follows:
wherein ShardedBilSTM 1 (-), ShardedBilSTM 2 (-), represent Bi-LSTM sharing two parameters,respectively represent A c And S c The connection result of the Bi-directional hidden state corresponding to the ith time step in the Bi-LSTM,respectively represent A c And S c From left to rightHidden state at the ith time step in the network,respectively represent A c And S c In a right to left directionHidden state at ith time step in the network;
s2.4.2 representation of student answer context A through attention mechanism r And a standard answer context representation S r Interact to get mutual attentionAndthe calculation process is as follows;
wherein the content of the first and second substances,is composed ofThe (c) th element of (a),is composed ofThe (c) th element of (a),respectively represent pairAndthe transposition operation is carried out and the operation is carried out,to refer to a logical quantity of "all",respectively representing hidden state tuplesAndattention weight of (a);
s2.4.3 respectively will beAndwith student answer context representation A r And a standard answer context representation S r Fusing to obtain student answer fusion expression A f Fused representation S with standard answers f The calculation process is as follows:
wherein the content of the first and second substances,is A f The (i) th element of (a),is S f The ith element in (1), Fusion (-), represents a Fusion operation, and an element-by-element multiplication operation of the matrix;
s2.4.4 fusion of student answers using two shared parameters Bi-LSTM respectively f Fused representation S with standard answers f Performing enhancement processing to obtain deep semantic representation A of student answers u And standard answer deep semantic representation S u The calculation process is as follows:
wherein, the first and the second end of the pipe are connected with each other,respectively represent A f And S f The connection result of the Bi-directional hidden state corresponding to the ith and jth time steps in the Bi-LSTM;
s2.4.5 deep semantic representation A of student answers respectively u And standard answer deep semantic representation S u Carrying out average pooling and maximum pooling, and connecting all pooled vectors to obtain score vector Z of student answer g The calculation process is as follows:
wherein maxPolling (. cndot.) represents the maximum pooling operation, avePolling (. cndot.) represents the average pooling operation,is A u The maximum pooling vector of (a) is,is A u The average pooling vector of (a) is,is S u The maximum pooling vector of (a) is,is S u The average pooling vector of "; "represents a concatenation operation of vectors.
Further, the step S3 specifically includes:
s3.1, calculating a loss function of the test question classification submodel and a loss function of the subjective question automatic scoring submodel by using the cross entropy loss error respectively, wherein the calculation process is as follows:
wherein, omega is a training set of the subjective question automatic evaluation task of the joint test question classification and evaluation learning, | omega | represents the size of the training set omega,is the real test question classification of the ith training sample in omega,is the true score level of the ith training sample in omega,is the classification vector of the test question of the ith training sample in omega,is the score vector of student answer of the ith training sample in omega, gamma 1 Is a loss function used during the training of the test question classification submodel, upsilon 2 The method is a loss function used during the sub-model training of automatic scoring of the subjective questions;
s3.2 calculating the automatic review god of the subjective questions of the combined test question classification and the scoring learning by using the following formulaJoint loss function via network model
Where λ and β are two weight parameters;
s3.3 subjective question automatic review neural network model training objective of joint test question classification and scoring learning is to minimize the joint loss error of equation (40).
Further, the table lookup operation in step S2.2 specifically includes:
wherein, OneHot (y) * ) Label y for indicating to-be-tested question type * Is converted into a one-hot vector,as test question type y * The one-hot vector of (a) is,embedding matrices for question types, d y Is the number of categories of test question types, d t Is the dimension of the question type embedding.
Examples of applications
1. Example hyper-parameters
TABLE 1 example model hyper-parameter settings
2. Data set
The data of this example comes from end-of-term examination questions of computer network of the department of the university software engineering specialty of the same year, to obtain 3948 chinese subjective question questions, each question is provided with 1 standard answer and 4 student answers, and finally, a question classification data set containing 3948 chinese subjective question questions and an automatic scoring data set of 15792 chinese subjective questions with answers are obtained, as shown in tables 2 and 3:
TABLE 2 question classification dataset
TABLE 3 automatic scoring dataset partitioning
3. Examples comparative results
Table 4 shows the classification result comparison results of the test question types of the various models, and table 5 shows the classification comparison results of the student answer score intervals of the various models:
TABLE 4 comparison of the results of the classification of test question types for various models
Model (model) | Accuracy | Recall | Precision | F1 value |
Test question classification model for joint scoring | 96.03 | 96.1 | 95.60 | 95.85 |
Independent test question classification model | 94.56 | 94.21 | 94.56 | 94.38 |
TABLE 5 results of classification comparison of student answer score intervals for various models
The experimental results in tables 4 and 5 show that the automatic deep learning method for subjective question review by combined test question classification and scoring learning provided by the invention is obviously superior to the independent review model and the independent test question classification model, and is obviously superior to the review model without the CNN layer and the review model without the Bi-LSTM interaction layer, which fully proves that the method of the invention is feasible and excellent.
4. Examples of the invention
Table 6 gives examples of student answer scores for various test question types:
TABLE 6 student answer Scoring example for various test question types
Claims (5)
1. A subjective question automatic evaluation deep learning method combining test question classification and evaluation learning is characterized by comprising the following steps:
s1, classifying the test questions of the subjective questions to be evaluated through a CNN test question classification submodel with residual connection to obtain the types of the test questions of the subjective questions to be evaluated;
s2, embedding and connecting test question types of subjective questions to be evaluated with each character of a student answer input sequence and a standard answer input sequence to obtain student answer perception input and standard answer perception input of test question type perception, and then sending the student answer perception input and the standard answer perception input into a subjective question test paper scoring sub-model consisting of CNN and Bi-LSTM to score, so as to obtain scoring grades of the student answers;
s3, training a subjective question automatic evaluation neural network model for combined test question classification and evaluation learning through the loss functions of the combined calculation step S1 and the step S2;
s4, automatically evaluating the neural network model by using the formed subjective questions trained in the step S3, and evaluating any student answers;
CNN is an abbreviation for the English conditional Neural Networks, representing a Convolutional Neural network, and is described by the documents "Lawrence S, Giles CL, Tsio AC et al. face repetition: A conditional Neural Networks, IEEE Transactions on 1997; 8(1) 98-113';
the Bi-LSTM is an abbreviation of English Directional Long Short-Term Memory and represents a two-way Long-Short Term Memory Neural network, wherein the LSTM Neural network is proposed by the documents "Sepp Hochreiter and Juregen Schmidhuber.1997 Long Short-Term Memory, 9(8): 1735-;
the subjective question types comprise definition subjective question test questions, sequence subjective question test questions and general subjective question test questions, wherein the definition subjective question test questions refer to subjective questions requiring students to explain definition of a certain concept, the sequence subjective question test questions refer to subjective questions with answers comprising a plurality of key points given in sequence, and the general subjective question test questions refer to other subjective questions except the definition and the sequence.
2. The method for automatic deep learning of subjective questions by combined test question classification and score learning according to claim 1, wherein:
the step S1 specifically includes:
s1.1 character sequence of test question to be classifiedWord embedding sequence T converted into test question to be classified through table look-up operation e The calculation process is as follows:
wherein n represents the sentence length of the test question to be classified, onehot (T) represents the conversion of the test question word sequence T into a unique heat matrix, T o Is a one-hot matrix of the test word sequence T, | V | represents the size of a dictionary V, which is a dictionary established by a model and containing all words,a matrix is embedded for the pre-trained words corresponding to dictionary V,representing a word embedding vector corresponding to the ith word in the test question word sequence T, and h representing the dimension of the word embedding vector;
S1.2 embedding the words of the test questions to be classified into the sequence T e Inputting the data into a convolutional neural network adopting zero filling to obtain shallow semantic representation T of the test question to be classified c The calculation process is as follows:
wherein PCNN (. cndot.) represents a convolutional neural network with zero padding, (T) c ) ij Represents T c I row and j column of (c), k is the convolution kernel size, d c Is the number of the convolution kernels and is,word-embedded sequence T representing test questions to be classified e As a result of the zero-padding being performed,to representThe j (th) column to the j + (k-1) column of the matrix,the weight of the ith convolution kernel is represented,represents the offset of the ith convolution kernel,representing a convolution operation;
s1.3 shallow semantic representation T of test questions to be classified c Is sent into a full-connected twoThe feedforward neural network formed by the connected layers is enhanced to obtain the enhanced semantic representation of the test questions to be classifiedThe calculation process is as follows:
wherein, W 1 f ,Respectively the weight matrices of the two fully-connected layers,bias vectors of two fully-connected layers respectively, ReLU is an abbreviation of English 'Rectified Linear Unit', and ReLU (·) represents an activation function adopting modified Linear Unit operation;
s1.4, expressing T by shallow semantic of test questions to be classified c With its enhanced semantic representation T f Residual error connection is carried out to obtain deep semantic representation T of the test question to be classified u The calculation process is as follows:
wherein LayerNorm (·) denotes layer normalization;
s1.5 deep semantic representation T of test questions to be classified u Performing maximum pooling to obtain a classification vector Z of the test questions to be classified c The calculation process is as follows:
wherein maxPooling (·) represents a max pooling operation;
s1.6 Classification vector Z of test questions to be classified c Executing linear transformation of softmax, performing probability calculation of the test question types, and obtaining the final test question types, wherein the calculation process is as follows:
o=M c Z c +b o (8)
wherein the content of the first and second substances,is a matrix of the test question type representation,is an offset vector, d y Is the number of the categories of the test question types,represents Z c Confidence score vectors over all test question types, y being one test question type, o y Represents Z c Confidence score, o, labeled test question type y i Represents Z c Confidence score, Ρ (yz | Z) on the ith test question type c ) Classification vector Z representing test questions to be classified c The prediction probability over the test question type Y, which is the set of all test question types,the representation returns such that Pp (y | Z) c ) Type of test question, y, as maximum * For the finally classified test question type of the test questions to be classified, exp (-) represents an exponential function with a natural constant e as the base.
3. The method for automatically deep learning by reviewing subjective questions combined with test question classification and score learning according to claim 1, wherein:
the step S2 specifically includes:
s2.1, performing table look-up operation on the student answer word sequence and the standard answer word sequence according to the step S1.1 to respectively obtain student answer word embedded sequencesStandard answer word embedding sequenceThen embedding the test question words into the sequence T e Respectively embedding the answer words with studentsStandard answer word embedding sequenceSplicing the head and the tail to obtain an answer input sequence A of the student e And a standard answer input sequence S e The calculation process is as follows:
wherein, ": "denotes the end-to-end concatenation operation of the sequence, n1 is the sentence length of student answers, n2 is the sentence length of standard answers, m1 is the student answer input sequence A e M1 ═ n + n1, m2 is the standard answer input sequence S e And m2 ═ n + n 2;
s2.2, the test question types y obtained in the step S1 are passedThe table look-up operation is converted into the corresponding test question type to be embeddedThen v is measured t Respectively input with student's answer input sequence A e And a standard answer input sequence S e Each character in the Chinese character string is embedded and connected to obtain student answer perception input A t And standard answer perception input S t The calculation process is as follows:
wherein, the first and the second end of the pipe are connected with each other,respectively, student's answer input sequence A e And student answer perception input A t Word-embedding vector of the ith column, d t Is the dimension of the embedding of the test question types,respectively, standard answer input sequence S e And standard answer perception input S t The word in column j embeds a vector, "; "represents a concatenation operation of vectors;
s2.3 inputting student answer perception A t And standard answer perception input S t Respectively sending the data to two convolutional neural networks which adopt zero padding and share parameters to obtain shallow semantic representation A of student answers c And the standard answer shallow semantic representation S c The calculation process is as follows:
wherein sharedPCNN1(·), sharedPCNN2 represent two convolutional neural networks with zero padding and shared parameters, d c The number of convolution kernels;
s2.4 shallow semantic representation A of student answers c And the standard answer shallow semantic representation S c Sending the data into an interactive network consisting of Bi-LSTM to obtain a scoring vector Z of student answers g As follows:
wherein Interactive (-) represents an Interactive network composed of Bi-LSTM, d L Dimension of hidden state of Bi-LSTM in interactive network;
s2.5 Scoring vector Z to student' S answer g Performing linear transformation of softmax, performing probability calculation of the grade, and obtaining a final grade, wherein the calculation process is as follows:
x=M g Z g +b g (18)
wherein the content of the first and second substances,is a matrix of the score level representation,is an offset vector, d g Is the number of the rating levels to be scored,is represented by Z g Vector of confidence scores of all the rating levels, g being one rating level, x g Represents Z g Confidence score, x, labeled as rating g i Represents Z g Confidence score on the i-th scoring level, Ρ (g | Z) g ) Representing a score vector Z g Prediction probability on a score level G, G being the set of all score levels, G * In order to achieve the final rating of the rating scale,denotes return such that Pp (g | Z) g ) The score rating of the maximum value.
4. The automatic deep learning method for subjective questions assessment combined with test question classification and assessment learning according to claim 3, wherein:
the step S2.4 specifically includes:
s2.4.1 shallow semantic representation A of student answers using two Bi-LSTMs sharing parameters respectively c And the standard answer shallow semantic representation S c Encoding to obtain student answer context expression A r And a standard answer context representation S r The calculation process is as follows:
wherein ShardedBilSTM 1 (-), ShardedBilSTM 2 (-), represent Bi-LSTM sharing two parameters, respectively represent A c And S c The connection result of the Bi-directional hidden state corresponding to the ith time step in the Bi-LSTM;
s2.4.2 context representation of student answers by attention mechanism A r And a standard answer context representation S r Interact to get mutual attentionAndthe calculation process is as follows;
wherein the content of the first and second substances,is composed ofThe (c) th element of (a),is composed ofThe (c) th element of (a),respectively represent pairAndthe transposition operation is carried out and the operation is carried out,to refer to a logical quantity of "all",respectively representing hidden state tuplesAndattention weight of (1);
s2.4.3 respectively will beAndwith student answer context representation A r And a standard answer context representation S r Fusing to obtain student answer fusion expression A f Fused representation S with standard answers f The calculation process is as follows:
wherein the content of the first and second substances,is A f The (c) th element of (a),is S f The ith element in (1), Fusion (-), represents a Fusion operation, and an element-by-element multiplication operation of the matrix;
s2.4.4 fusion of student answers using two Bi-LSTMs sharing parameters respectively f Fused representation S with standard answers f Performing enhancement processing to obtain deep semantic representation A of student answers u And standard answer deep semantic representation S u The calculation process is as follows:
wherein, the first and the second end of the pipe are connected with each other,respectively represent A f And S f The connection result of the Bi-directional hidden state corresponding to the ith and jth time steps in the Bi-LSTM;
s2.4.5 deep semantic representation A of student answers respectively u And standard answer deep semantic representation S u Carrying out average pooling and maximum pooling, and connecting all pooled vectors to obtain score vector Z of student answer g The calculation process is as follows:
wherein maxPooling (. cndot.) represents the maximum pooling operation, avePooling (. cndot.) represents the average pooling operation,is A u The maximum pooling vector of (a) is,is A u The average pooling vector of (a) is,is S u The maximum pooling vector of (a) is,is S u The average pooling vector of "; "represents a concatenation operation of vectors.
5. The method for automatic deep learning of subjective questions by combined test question classification and score learning according to claim 1, wherein:
the step S3 specifically includes:
s3.1, calculating a loss function of the test question classification submodel and a loss function of the subjective question automatic scoring submodel by using the cross entropy loss error respectively, wherein the calculation process is as follows:
wherein, omega is a training set of the subjective question automatic evaluation task of the joint test question classification and evaluation learning, | omega | represents the size of the training set omega,is the real test question classification of the ith training sample in omega,is the true score level of the ith training sample in omega,is the ith training sample in ΩThe classification vector of the test question of the present invention,the score vector of student answer, y, of the ith training sample in Ω 1 Is a loss function used in the training of the test question classification submodel, gamma 2 Is a loss function used when carrying out the automatic scoring submodel training of the subjective questions;
s3.2 calculating the joint loss function of the automatic evaluation neural network model of the subjective questions of the joint test question classification and evaluation learning by using the following formula
Wherein λ and β are two weight parameters;
s3.3 subjective question automatic review neural network model training objective of joint test question classification and scoring learning is to minimize the joint loss error of equation (40).
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210597773.7A CN114969260A (en) | 2022-05-30 | 2022-05-30 | Automatic subjective question evaluation and reading deep learning method combining test question classification and evaluation learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210597773.7A CN114969260A (en) | 2022-05-30 | 2022-05-30 | Automatic subjective question evaluation and reading deep learning method combining test question classification and evaluation learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114969260A true CN114969260A (en) | 2022-08-30 |
Family
ID=82957945
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210597773.7A Pending CN114969260A (en) | 2022-05-30 | 2022-05-30 | Automatic subjective question evaluation and reading deep learning method combining test question classification and evaluation learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114969260A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115827879A (en) * | 2023-02-15 | 2023-03-21 | 山东山大鸥玛软件股份有限公司 | Low-resource text intelligent review method and device based on sample enhancement and self-training |
CN117034954A (en) * | 2023-10-09 | 2023-11-10 | 华南师范大学 | Text scoring method, device, equipment and storage medium |
CN117252739A (en) * | 2023-11-17 | 2023-12-19 | 山东山大鸥玛软件股份有限公司 | Method, system, electronic equipment and storage medium for evaluating paper |
-
2022
- 2022-05-30 CN CN202210597773.7A patent/CN114969260A/en active Pending
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115827879A (en) * | 2023-02-15 | 2023-03-21 | 山东山大鸥玛软件股份有限公司 | Low-resource text intelligent review method and device based on sample enhancement and self-training |
CN117034954A (en) * | 2023-10-09 | 2023-11-10 | 华南师范大学 | Text scoring method, device, equipment and storage medium |
CN117034954B (en) * | 2023-10-09 | 2024-02-06 | 华南师范大学 | Text scoring method, device, equipment and storage medium |
CN117252739A (en) * | 2023-11-17 | 2023-12-19 | 山东山大鸥玛软件股份有限公司 | Method, system, electronic equipment and storage medium for evaluating paper |
CN117252739B (en) * | 2023-11-17 | 2024-03-12 | 山东山大鸥玛软件股份有限公司 | Method, system, electronic equipment and storage medium for evaluating paper |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111554268B (en) | Language identification method based on language model, text classification method and device | |
CN110245229B (en) | Deep learning theme emotion classification method based on data enhancement | |
CN108549658B (en) | Deep learning video question-answering method and system based on attention mechanism on syntax analysis tree | |
CN114969260A (en) | Automatic subjective question evaluation and reading deep learning method combining test question classification and evaluation learning | |
Uto | A review of deep-neural automated essay scoring models | |
CN111753098A (en) | Teaching method and system based on cross-media dynamic knowledge graph | |
CN113656570A (en) | Visual question answering method and device based on deep learning model, medium and equipment | |
CN110516070B (en) | Chinese question classification method based on text error correction and neural network | |
CN111985239A (en) | Entity identification method and device, electronic equipment and storage medium | |
CN109101490B (en) | Factual implicit emotion recognition method and system based on fusion feature representation | |
CN113204633B (en) | Semantic matching distillation method and device | |
CN113569001A (en) | Text processing method and device, computer equipment and computer readable storage medium | |
CN112131883A (en) | Language model training method and device, computer equipment and storage medium | |
CN112232053A (en) | Text similarity calculation system, method and storage medium based on multi-keyword pair matching | |
CN111897954A (en) | User comment aspect mining system, method and storage medium | |
CN111581364B (en) | Chinese intelligent question-answer short text similarity calculation method oriented to medical field | |
CN113704392A (en) | Method, device and equipment for extracting entity relationship in text and storage medium | |
CN113342958A (en) | Question-answer matching method, text matching model training method and related equipment | |
CN110874392B (en) | Text network information fusion embedding method based on depth bidirectional attention mechanism | |
CN116186237A (en) | Entity relationship joint extraction method based on event cause and effect inference | |
CN112131345A (en) | Text quality identification method, device, equipment and storage medium | |
CN113934835B (en) | Retrieval type reply dialogue method and system combining keywords and semantic understanding representation | |
CN115935991A (en) | Multitask model generation method and device, computer equipment and storage medium | |
CN116975350A (en) | Image-text retrieval method, device, equipment and storage medium | |
CN115017879A (en) | Text comparison method, computer device and computer storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |