A Short Answer Automatic Scoring Method Based on Multi-Feature Fusion

Abstract

The present invention discloses a short answer automatic scoring method based on multi-feature fusion, which collects texts in a subject field and uses a word vector model to obtain a subject field word vector model; inputs the reference answer and the candidate's answer text into a BERT model to extract the semantic features of the answer text; performs syntactic parsing on the two answer texts to obtain the syntactic dependency graph in the text; calculates the noun similarity of the two answer texts to obtain the entity graph in the text; uses semantic features to supplement the lacking semantic information to obtain an extended syntactic dependency graph and an extended entity graph, and dynamically fuses them to obtain an overall feature fusion graph; obtains the global information of the answer text through the overall feature fusion graph; uses the attention mechanism to obtain the local information of the answer text; combines the global information and the local information as the final representation of the answer text, and calculates the similarity of the two answer texts using the cosine similarity formula to obtain the score of the student's answer. The present invention is applicable to the marking of examination papers in the field of computer science.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular, to a method for automatically grading short answers based on multi-feature fusion. Background Art

[0002] Automatic Short Answer Grading (ASAG) is a popular research direction in the current field of computer-aided instruction. Improving the accuracy of the calculation result of the similarity between the student's answer and the reference answer is the key to realizing the problem of automatic short answer grading. Implementing automatic grading of short answers can not only liberate teachers from the tedious manual marking work and ensure the objectivity and fairness of students' grades, but also provide quick feedback and improve students' learning ability.

[0003] The existing methods for automatically grading short answers and their existing problems are as follows:

[0004] (1) Rule-based method for automatically grading short answers. This method establishes regular expression rules based on the reference answer, and each rule is associated with a scoring point. When the student's answer matches the rule, the corresponding score is obtained. Due to the limited accuracy and expression ability of rule acquisition, the generalization ability of this method is poor.

[0005] (2) Feature engineering-based method for automatically grading short answers. This method cannot effectively encode text sequences and cannot understand the student's answer at the semantic level. Secondly, the feature engineering-based method cannot avoid problems such as time-consuming, laborious, and feature redundancy in the feature construction process.

[0006] (3) Deep learning-based method for automatically grading short answers. This method uses a neural network to automatically extract local and global information from the answer text converted into word embedding vectors, so as to realize automatic grading of the student's answer in an end-to-end manner. However, the deep learning-based method still faces a series of problems. First, students use different free texts to answer the same question when answering questions, and there may be significant differences in the syntactic structures of different students' answers, which greatly affects the effectiveness of the constructed word vector model. Secondly, each discipline involves different domain knowledge, and students may use different professional terms or expressions to answer the same question. How to measure the accuracy of the student's answer is also a huge challenge. Summary of the Invention

[0007] The purpose of the present invention is to provide a method for automatically grading short answers based on multi-feature fusion.

[0008] The technical solution adopted by the present invention is:

[0009] A short answer automatic scoring method based on multi-feature fusion, which includes the following steps:

[0010] Step 1, collect texts within the subject field and obtain the subject field word vector model using the word vector model;

[0011] Step 2, input the short answer question information, where the short answer question information includes the question, the reference answer, and the score;

[0012] Step 3, obtain the answer text answered by the candidate, input the reference answer in the short answer question information and the candidate's answer text into the BERT model for encoding, and extract the semantic features of the answer text;

[0013] Step 4, perform syntactic parsing on the reference answer text and the student's answer text to obtain the syntactic dependency graph in the text;

[0014] Step 5, calculate the similarity of nouns in the reference answer text and the student's answer text to obtain the entity graph in the text;

[0015] Step 6, use the semantic features extracted in Step 3 to supplement the lacking semantic information in the syntactic dependency graph and the entity graph, and then obtain the extended syntactic dependency graph and the extended entity graph respectively;

[0016] Step 7, dynamically fuse the extended syntactic dependency graph and the extended entity graph to obtain the overall feature fusion graph;

[0017] Step 8, input the overall feature fusion graph as a whole into the graph convolutional neural network, and capture the global word co-occurrence and global interaction between answer texts through graph convolution;

[0018] Step 9, represent the two groups of vectors in the output result of the graph convolutional neural network that represent the reference answer and the student's answer as R A and R B , and use the attention mechanism to obtain the local information of R A and R B ;

[0019] Step 10, combine the global information and the local information as the final representation of the answer text, and use the cosine similarity formula to calculate the similarity between the reference answer and the student's answer, and then obtain the score of the student's answer.

[0020] Furthermore, in Step 1, texts within the subject field are collected through web data crawling and Wikipedia, and the domain word vector model is obtained using the word vector model Word2Vec.

[0021] Further, in step 3, the answers submitted by the candidates are directly obtained through an online question answering system, or the answers submitted by the candidates from multiple sources such as other systems or databases are read through data import.

[0022] Further, in step 4, a syntactic parser is used to extract the syntactic dependency relationships between words.

[0023] Further, the specific steps of step 5 are as follows:

[0024] Step 5-1, extract all nouns from the answer text to obtain a noun set M;

[0025] Step 5-2, use the constructed domain word vector model to calculate the similarity between different nouns in the noun set M;

[0026] Regarding each noun in the noun set M as a node, calculate the similarity of the edges between nouns. The calculation formula for the similarity of the edges between nouns is as follows:

[0027]

[0028] Among them, E(n i ,n j ) represents an entity graph composed of the similarities between two nouns. n i represents the i-th noun in the noun set M, and n j represents the j-th noun in the noun set M. s(n i ,n j ) represents the similarity between two noun vectors n i and n j calculated by cosine similarity. i and j respectively represent the subscripts of different nouns in the noun set M;

[0029] Step 5-3, use the similarity of the edges between nouns as the weights of the edges between nodes to construct an entity graph of the answer text.

[0030] Further, the expression formula for the extended syntactic dependency graph in step 6 is as follows:

[0031]

[0032] Among them, C(w i ,w j ) is the syntactic dependency graph of the answer; is the extended syntactic dependency graph of the answer; is the enhancement coefficient based on the gate structure syntactic dependency graph, is the vector form of the semantic feature, and respectively represent the weights and biases of the gates corresponding to the syntactic dependency graph of the answer,

[0033] The expression formula of the extended entity graph is as follows:

[0034]

[0035] where, is the extended entity graph of the answer; E(n i ,n j ) is the entity graph of the answer; is the enhancement coefficient based on the gate structure entity graph, σ is the sigmoid function, is the vector form of the semantic feature, and respectively represent the weights and biases of the gates corresponding to the entity graph of the answer.

[0036] Furthermore, the fusion formula in step 7 is as follows:

[0037]

[0038]

[0039]

[0040] where, α is 's trainable weight coefficient; β is 's trainable weight coefficient, W 1 , W 2 are weights, b 1 , b 2 are biases, and X (i,j) is the overall feature fusion graph.

[0041] Furthermore, the convolution operation of the graph convolutional neural network in step 8 is as follows:

[0042]

[0043] where, θ is the convolution kernel, * δ is the convolution operator, X is the feature matrix, is the normalized Laplacian matrix, represents the self-loop, that is, the features of the current node and neighbor nodes are simultaneously involved in the convolution operation to enhance the aggregation effect of the adjacency matrix A on node information, I n is the N-order identity matrix, is 's degree matrix;

[0044] On the basis of the convolution operation, a neural network activation layer is introduced and a multi-layer model is constructed by stacking graph convolutional neural networks. The specific formula is as follows:

[0045]

[0046] Where H (l) is the activation matrix in the l-th layer, and the initial feature matrix H (0) = X, σ(·) is the activation function, and W (l) is the learnable weight.

[0047] Furthermore, the expressions for the local information of R A and R B obtained by the attention mechanism in step 9 are as follows:

[0048]

[0049]

[0050] Where V A is the final representation of the reference answer, V B is the final representation of the student's answer, w i is the word in the answer text, R A represents the reference answer vector, R B represents the student's answer vector, u A represents the attention probability of the word in the reference answer, u B represents the attention probability of the word in the student's answer, W A represents the weight parameter of the k×d dimensional model of the reference answer, w A represents the weight parameter of the 1×d dimensional model of the reference answer; W B represents the weight parameter of the k×d dimensional model of the student's answer, w B is the weight parameter of the 1×d dimensional model of the student's answer in the 1×d dimension, k represents the number of rows of the weight parameter matrix, and d represents the number of columns of the weight parameter matrix; P represents the affinity matrix, is the affinity matrix, W c represents the weight parameter of the d×d dimension; represents the i-th element in u A , represents the i-th element in u B ; represents the i-th column in R A , represents the i-th column in R B ; T represents the transpose of the matrix.

[0051] Furthermore, the specific steps of Step 10 are as follows:

[0052] Step 10-1: Calculate the similarity between the reference answer and the student answer using the cosine similarity formula. The cosine similarity calculation formula is as follows:

[0053]

[0054] where A i represents the final vector representation of the reference answer, and B i represents the final vector representation of the student answer;

[0055] Step 10-2: Standardize the cosine similarity from 0 to 5;

[0056] Step 10-3: Use the standardized value as the final score of the student answer.

[0057] The present invention adopts the above technical solutions, introduces domain knowledge into the short answer automatic scoring process, introduces syntactic dependency features into the short answer automatic scoring process by constructing a syntactic dependency graph, and uses a graph convolutional neural network to encode the overall feature fusion graph to effectively extract the global information of the answer text. The attention mechanism calculates the attention probability of words to obtain the local information of the answer text.

[0058] The present invention has the following advantages compared with the prior art: (1) For students who use different free texts to answer the same question when answering questions, the present invention introduces syntactic dependency features into the short answer automatic scoring process by constructing a syntactic dependency graph, and solves the problem that there are significant differences in the syntactic structures of different student answers. (2) Combining the characteristics of the field discipline, the present invention introduces domain knowledge into the short answer automatic scoring process, and solves the problem of insufficient semantic understanding of specific domain technical terms. (3) Taking advantage of the ability of the graph convolutional neural network to extract the global information of the text, the present invention uses the graph convolutional neural network to encode the overall feature fusion graph to effectively extract the global information of the answer text. (4) The present invention calculates the attention probability of words through the attention mechanism to obtain the local information of the answer text, and solves the problem of insufficient ability of the graph convolutional neural network to extract local information. (5) The present invention effectively introduces semantic features, syntactic dependency features, and domain knowledge features into the short answer automatic scoring method, and can be applied to the scoring of test papers in the computer professional field. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The following further describes the present invention in detail with reference to the drawings and specific embodiments;

[0060] Figure 1 It is a schematic flowchart of a short answer automatic scoring method based on multi-feature fusion according to the present invention;

[0061] Figure 2 Schematic diagram of the process for extracting semantic features of the answer text of the present invention;

[0062] Figure 3 Schematic diagram of the syntactic dependency relationship of the answer text;

[0063] Figure 4 Schematic diagram of the syntactic dependency graph of the reference answer text;

[0064] Figure 5 Schematic diagram of the syntactic dependency graph of the student answer text;

[0065] Figure 6 Schematic diagram of the entity graph of the answer text;

[0066] Figure 7 Schematic diagram of the process for obtaining the extended syntactic dependency graph and the extended entity graph;

[0067] Figure 8 Schematic diagram of the process for fusing the overall feature fusion graph;

[0068] Figure 9 Schematic diagram of the scoring process using the overall feature fusion graph. Detailed implementation manners

[0069] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application.

[0070] Adaptive testing and assessment are key components of Intelligent Tutoring Systems (ITSs), which can capture the current cognitive level of students and provide a basis for the system to formulate personalized learning routes. For the simplicity of implementation, standardized questions (multiple-choice questions or true / false questions) are widely used in adaptive assessment. However, standardized questions have the following two obvious disadvantages. (1) Standardized questions usually only provide a few alternative items and cannot list all possible student answers; (2) Students may randomly select answers from the listed items without thinking, so the cognitive level of students cannot be accurately captured. Compared with standardized questions, short-answer questions require students to answer questions using text, which can activate more complex reconstructive cognitive processes and further promote students' learning.

[0071] As Figures 1 to 9 shown in one of them, the present invention discloses a short-answer automatic scoring method based on multi-feature fusion, including the following steps:

[0072] Step 1, constructing a domain word vector model;

[0073] Collect texts within the subject field through methods such as web data crawling and Wikipedia, and on this basis, use a word vector model (such as Word2Vec, etc.) to obtain a domain word vector model.

[0074] For example, obtain different terms and their description texts in the field of computer science through relevant entries in Wikipedia (such as Computer science-Wikipedia), and use the collected texts as the input of Word2Vec to construct a word vector model in the field of computer science.

[0075] Step 2: Enter the short-answer question information, which mainly includes the question, the reference answer, and the score.

[0076] The teacher enters the short-answer question information, which mainly includes the question, the reference answer, and the score. The question is a description of a short-answer question in the field of computer science, such as "What is a variable?" (What is a variable?). The reference answer is the answer preset by the teacher for each exam question. The score is the score that the candidate can get after correctly solving the problem, usually a number between 0 and 5.

[0077] For example, the question can be set as follows:

[0078] Question: What is a variable?

[0079] Reference Answer: A location in memory that can store a value.

[0080] Score: 5

[0081] Step 3: Obtain the answer text answered by the candidate, and input both into the BERT model for encoding and then extract the semantic features of the answer text.

[0082] The answer submitted by the candidate can be directly obtained through an online answering system, or the answers submitted by the candidate from various sources such as other systems or databases can be read through data import. Assume that the answer submitted by a certain student at this stage is: It is a location in memory where value can be stored.

[0083] Merge the reference answer text and the student answer text. In this process, the [CLS] flag is placed at the beginning of the first sentence, and the representation vector obtained by BERT can be used for subsequent tasks. The [SEP] flag is used to separate the two input sentences. For example, for input sentences A and B, the [SEP] flag should be added after sentences A and B. Therefore, the merged form of the reference answer text and the student answer text is as follows:

[0084] [CLS]A location in memory that can store a value[SEP]It is a locationin memory where value can be stored[SEP]

[0085] Input the answer text in the above form into the BERT model, and extract the semantic features of the answer text by obtaining semantic vectors, such as Figure 2 shown.

[0086] Step 4, perform syntactic parsing on the reference answer text and the student answer text to obtain the syntactic dependency graph in the text;

[0087] For the syntactic structures of the reference answer text and the student answer text, the present invention uses a syntactic parser (such as Stanford CoreNLP) to extract the syntactic dependency relationships between words, such as Figure 3 shown.

[0088] Take words as the nodes of the graph, and define two types of directed edges as shown in formula (1), thereby constructing the syntactic dependency graph of the answer text, and obtaining the dependency relationships of each word in the answer text through formula (1). In formula (1), C(w i , w j ) represents the dependency relationship between two words, 1 indicates the existence of a dependency, and 0 indicates the non-existence of a dependency. Among them, w represents a word, i and j represent the positions where the word appears in the sentence, and ω represents the set of dependency syntactic arcs in the sentence. Since the extracted dependency syntactic relationships are unidirectional, in order to enable syntactic information to flow in the same and reverse directions along the syntactic arcs, the present invention regards all the edges in the graph as undirected edges, and constructs a self-loop edge for each node, so as to retain the syntactic information of each word during the transfer iteration process.

[0089]

[0090] The syntactic dependency graph of the reference answer text of the embodiment is as Figure 4 shown, and the syntactic dependency graph of the student answer text of the embodiment is as Figure 5 shown.

[0091] Step 5, calculate the similarity of the nouns in the reference answer text and the student answer text to obtain the entity graph in the text;

[0092] Extract all nouns from the answer text to obtain the noun set M, and calculate the similarity between different nouns in M using the domain word vector model constructed in Step 1. Take each noun in M as a node, calculate the similarity of the edges between nouns using formula (2), and use this as the weight of the edges between nodes, thereby constructing the entity graph of the answer text.

[0093] Specifically, in formula (2), E(n i ,n j ) represents the entity graph composed of the similarity between two nouns, n represents a noun, and s(n i ,n j ) represents the similarity of two noun vectors calculated by cosine similarity, and i, j represent the subscripts of the nouns in M.

[0094]

[0095] For the reference answer “A location in memory that can store a value”, the noun set can be obtained as M = {location, memory, value}. Through the domain word vector model, the word vectors corresponding to value, memory, and location are [0.08465, 0.07275, -0.25968, …, -0.14847], [0.58646, 0.16764, -0.07353, …, 0.58633], [0.21454, -0.21605, 0.26466, …, 0.15583] respectively. As Figure 6 shown, the similarity between each noun in M can be calculated through formula (2).

[0096] Similarly, for the student answer “It is a location in memory where value can be stored”, the noun set [location, memory, value] can be obtained, and the similarity between each noun in the noun set can be calculated through formula (2).

[0097] Step 6, obtain the extended syntactic dependency graph and the extended entity graph;

[0098] The present invention uses the semantic features extracted in Step 3 to supplement the semantic information lacking in the syntactic dependency graph and the entity graph, thereby obtaining the extended syntactic dependency graph and the extended entity graph respectively. Formulas (3) and (4) represent the semantic features based on the gate structure. Formulas (5) and (6) represent the processes of obtaining the extended syntactic dependency graph and the extended entity graph respectively. Among them, σ is the sigmoid function, It is the vector form of semantic features. and are the weights and biases of the gate.

[0099]

[0100]

[0101]

[0102]

[0103] The processes of obtaining the extended syntactic dependency graph and the extended entity graph are as Figure 7 shown.

[0104] Specifically, the vectorized representation of the reference answer of the embodiment in the extended syntactic dependency graph is [[0.8498, 0.6071, 0.7295, …, 0.2767], …, [0.3786, 0.7983, 0.4689, …, 0.3401]], and the vectorized representation of the student answer of the embodiment in the extended syntactic dependency graph is [[0.3426, 0.5361, 0.3518, ..., 0.3252], …, [0.5058, 0.2544, 0.3952, …, 0.3537]]. The vectorized representation of the reference answer of the embodiment in the extended entity graph is [[0.7068, 0.4743, 0.3974, …, 0.3534], …, [0.5058, 0.2544, 0.3952, …, 0.3537]], and the vectorized representation of the student answer of the embodiment in the extended entity graph is [[0.1790, 0.9323, 0.4456, …, 0.8262], …, [0.2986, 0.5756, 0.7106, …, 0.1963]].

[0105] Step 7, obtain the overall feature fusion graph;

[0106] To effectively and dynamically fuse the extended syntactic dependency graph and the extended entity graph to obtain the overall feature fusion graph, the present invention designs the adaptive fusion formulas (7) - (9). Among them, W 1 , W 2 are weights, b 1 , b 2 are biases, and X (i,j) is the overall feature fusion graph.

[0107]

[0108]

[0109]

[0110] The overall feature fusion graph fusion process is as follows Figure 8 shown. Specifically, the vectorized representation of the reference answer in the overall feature fusion graph is [[0.1589, 0.3737, 0.5399, …, 0.1409], …, [0.1180, 0.5574, 0.5531, …, 0.9772]], and the vectorized representation of the student answer in the overall feature fusion graph is [[0.4783, 0.6546, 0.5468, …, 0.3943], …, [0.4472, 0.7668, 0.1310, …, 0.4340]].

[0111] Step 8, obtaining the global information of the answer text;

[0112] The present invention inputs the overall feature fusion graph as a whole into a graph convolutional neural network, and captures the global word co-occurrence and global interaction between answer texts through graph convolution. Its convolution operation is as shown in Equation (10):

[0113]

[0114] where θ is the convolution kernel, * δ is the convolution operator, X is the feature matrix, and the normalized Laplacian matrix Adding self-loops is equivalent to simultaneously participating the current node and neighbor node features in the convolution operation, which can enhance the aggregation effect of the adjacency matrix A on node information. I n is the N-order identity matrix, is 's degree matrix. To enrich its representation ability, a neural network activation layer is introduced based on Equation (10), and a multi-layer model is constructed by stacking graph convolutional neural networks, specifically as shown in Equation (11):

[0115]

[0116] where H (l) is the activation matrix in the l-th layer, the initial feature matrix H (0) = X, σ(·) is the activation function, and W (l) is the learnable weight.

[0117] Specifically, the vector representation of the global information of the reference answer text extracted by using the graph convolutional neural network is [[0.8046, 0.5662, 0.1287, …, 0.1020], …, [0.8018, 0.8518, 0.4473, …, 0.5164]]. The vector representation of the global information of the student answer text is [[0.8916, 0.8860, 0.8693, …, 0.3955], …, [0.8257, 0.7642, 0.8021, …, 0.3478]].

[0118] Step 9, introduce the attention mechanism, and extract the local information of the answer text by focusing on the attention probabilities of the words;

[0119] In natural language processing tasks, the attention mechanism obtains the local information of the answer text by calculating the attention probabilities of the words. In the present invention, two groups of vector representations representing the reference answer and the student answer respectively in the output result of the graph convolutional neural network are denoted as R A and R B , and the attention mechanism is used to obtain the local information of R A and R B .

[0120]

[0121]

[0122] Among them, W A , W B , w A , w B are model parameters, is the affinity matrix, and u A , u B calculate the attention probabilities of the words in the reference answer and the student answer respectively. Finally, the vector representation of the answer is calculated by the following formula.

[0123]

[0124]

[0125] Among them, u i represents the i-th element in u, and R i represents the i-th column in R.

[0126] Specifically, the vector representation of the local information of the reference answer text extracted by using the attention mechanism is [[0.2809, 0.9882, 0.2225, …, 0.7584], …, [0.6434, 0.5676, 0.1194, …, 0.3117]]. The vector representation of the local information of the student answer text is [[0.3274, 0.7420, 0.6136, …, 0.6011], …, [0.1950, 0.5518, 0.4787, …, 0.8228]].

[0127] Step 10: Combine the global information and the local information as the final representation of the answer text, and use the cosine similarity formula to calculate the similarity between the reference answer and the student answer, thereby obtaining the score of the student answer.

[0128] In the similarity calculation, the cosine value of the vector angle between the reference answer text and the student answer text can be used to represent their similarity. Suppose there are two vectors in space, and their cosine similarity calculation formula is as shown in Equation (16). Next, normalize the cosine similarity from 0 to 5, and then use the normalized value as the final score of the student answer. Among them, the overall feature fusion graph scoring process in Steps 8 - 10 is as Figure 9 shown.

[0129]

[0130] Among them, A i represents the final vector representation of the reference answer, and B i represents the final vector representation of the student answer.

[0131] Specifically, the vectorized representation of A i is [0.3469, 0.3533, 0.6142, …, 0.1280], and the vectorized representation in B i is [0.4094, 0.3875, 0.6270, …, 0.1025]. It can be calculated that the similarity is 0.91, and through normalization, the similarity is converted into the score obtained by the student, which is 4.775.

[0132] The present invention adopts the above technical solutions and has the following advantages compared with the prior art: (1) Aiming at the situation that students use different free texts to answer the same question when answering questions, the present invention introduces syntactic dependency features into the short answer automatic scoring process by constructing a syntactic dependency graph, solving the problem that there are significant differences in syntactic structures among different students' answers. (2) Combining the characteristics of the field discipline, the present invention introduces domain knowledge into the short answer automatic scoring process, solving the problem of insufficient semantic understanding of specific domain-specific terms. (3) Taking advantage of the ability of the graph convolutional neural network to extract global information of the text, the present invention encodes the overall feature fusion graph using the graph convolutional neural network and effectively extracts the global information of the answer text. (4) The present invention calculates the attention probability of words through the attention mechanism to obtain the local information of the answer text, solving the problem of insufficient ability of the graph convolutional neural network to extract local information. (5) The present invention effectively introduces semantic features, syntactic dependency features, and domain knowledge features into the short answer automatic scoring method and can be applied to the test paper scoring in the computer professional field.

[0133] Obviously, the described embodiments are part of the embodiments of the present application, rather than all embodiments. Without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. Usually, the components of the embodiments of the present application described and illustrated in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application is not intended to limit the scope of the present application claimed, but merely represents the selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

Claims

1. A short answer automatic scoring method based on multi-feature fusion, characterized in that: It includes the following steps: Step 1, collect texts within the subject field and obtain the subject field word vector model using the word vector model; Step 2, input the short answer question information, where the short answer question information includes the question, the reference answer, and the score; Step 3, obtain the answer text answered by the candidate, input the reference answer in the short answer question information and the candidate's answer text into the BERT model for encoding, and extract the semantic features of the answer text; Step 4, perform syntactic parsing on the reference answer text and the student's answer text to obtain the syntactic dependency graph in the text; Step 5, calculate the similarity of nouns in the reference answer text and the student's answer text to obtain the entity graph in the text; Step 6, use the semantic features extracted in Step 3 to supplement the lacking semantic information in the syntactic dependency graph and the entity graph, and then obtain the extended syntactic dependency graph and the extended entity graph respectively; the expression formula of the extended syntactic dependency graph is as follows: Among them, C(w i , w j ) is the syntactic dependency graph of the answer; is the extended syntactic dependency graph of the answer; is the enhancement coefficient based on the gate structure syntactic dependency graph, is the vector form of the semantic feature, and respectively represent the weight and bias of the gate corresponding to the syntactic dependency graph of the answer, The expression formula of the extended entity graph is as follows: Among them, is the extended entity graph of the answer; E9n i ,n j ) is the entity graph of the answer; is the enhancement coefficient based on the entity graph of the gate structure, σ is the sigmoid function, is the vector form of the semantic feature, and respectively represent the weight and bias of the gate corresponding to the entity graph of the answer; Step 7, dynamically fuse the extended syntactic dependency graph and the extended entity graph to obtain the overall feature fusion graph; Step 8, input the overall feature fusion graph as a whole into the graph convolutional neural network, and capture the global word co-occurrence and global interaction between answer texts through graph convolution; Step 9, represent the two groups of vectors in the output of the graph convolutional neural network that represent the reference answer and the student answer as $\mathbf{R}$ A and $\mathbf{R}$ B , and use the attention mechanism to obtain the local information of $\mathbf{R}$ A and $\mathbf{R}$ B ; Step 10, combine the global information and the local information as the final representation of the answer text, and use the cosine similarity formula to calculate the similarity between the reference answer and the student's answer, and then obtain the score of the student's answer.

2. A short answer automatic scoring method based on multi-feature fusion according to claim 1, characterized in that: In Step 1, texts within the subject field are collected through web data crawling and Wikipedia, and the domain word vector model is obtained using the word vector model Word2Vec.

3. A short answer automatic scoring method based on multi-feature fusion according to claim 1, characterized in that: In Step 3, the answer submitted by the candidate is directly obtained through the online answering system, or the answers submitted by the candidate from multiple sources in other systems or databases are read through data import.

4. A short answer automatic scoring method based on multi-feature fusion according to claim 1, characterized in that: In Step 4, a syntactic parser is used to extract the syntactic dependency relationship between words.

5. A short answer automatic scoring method based on multi-feature fusion according to claim 1, characterized in that: The specific steps of Step 5 are as follows: Step 5-1, extract all nouns from the answer text to obtain the noun set M; Step 5-2, calculate the similarity between different nouns in the noun set M using the constructed domain word vector model; Regarding each noun in the noun set M as a node, calculate the similarity of the edges between nouns. The calculation formula of the similarity of the edges between nouns is as follows: Among them, E(n i ,n j ) represents an entity graph composed of the similarities between two nouns, n i represents the i-th noun in the noun set M, n j represents the j-th noun in the noun set M, s(n i ,n j ) represents the similarity between two noun vectors n i and n j calculated by cosine similarity, where i and j respectively represent the subscripts of different nouns in the noun set M; Step 5-3, use the similarity of the edges between nouns as the weight of the edges between nodes to construct the entity graph of the answer text.

6. A short answer automatic scoring method based on multi-feature fusion according to claim 1, characterized in that: The fusion formula in step 7 is as follows: Among them, α is the trainable weight coefficient; β is the trainable weight coefficient, W 1 and W 2 are weights, b 1 and b 2 are biases, and X (i,j) is the overall feature fusion graph.

7. A short-answer automatic scoring method based on multi-feature fusion according to claim 1, characterized in that: The convolution operation of the graph convolutional neural network in step 8 is as follows: where θ is the convolution kernel, * δ is the convolution operator, X is the feature matrix, is the normalized Laplacian matrix, represents a self-loop, that is, the features of the current node and neighbor nodes are simultaneously involved in the convolution operation to enhance the aggregation effect of the adjacency matrix A on node information, I n is the N - order identity matrix, is the degree matrix; On the basis of the convolution operation, a neural network activation layer is introduced and a multi-layer model is constructed by stacking graph convolutional neural networks. The specific formula is as follows: where H (l) is the activation matrix in the l-th layer, and the initial feature matrix H (0) = X, σ(·) is the activation function, and W (l) is the learnable weight.

8. A short-answer automatic scoring method based on multi-feature fusion according to claim 1, characterized in that: The expressions of the local information of R A and R B obtained by the attention mechanism in step 9 are as follows: W A ,W B has dimensions k×d, w A ,w B has dimensions 1×d Among them, V A is the final representation of the reference answer, V B is the final representation of the student's answer, w i is the word in the answer text, R A represents the reference answer vector, R B represents the student answer vector, u A represents the attention probability of the word in the reference answer, u B represents the attention probability of the word in the student's answer, W A represents the weight parameter of the k×d dimensional model of the reference answer, w A represents the weight parameter of the 1×d dimensional model of the reference answer; W B represents the weight parameter of the k×d dimensional model of the student's answer, w B is the weight parameter of the 1×d dimensional model of the student's answer in the 1×d dimension, k represents the number of rows of the weight parameter matrix, d represents the number of columns of the weight parameter matrix; P represents the affinity matrix, is the affinity matrix, W c represents the weight parameter of the d×d dimension; represents the A ith element in u, represents the B ith element in u; represents the A ith column in R, represents the B ith column in R; T represents the transpose of the matrix.

9. A short-answer automatic scoring method based on multi-feature fusion according to claim 1, characterized in that: The specific steps of step 10 are as follows: Step 10-1, calculate the similarity between the reference answer and the student answer using the cosine similarity formula. The cosine similarity calculation formula is as follows: Among them, A i represents the final vector representation of the reference answer, and B i represents the final vector representation of the student answer; Step 10-2, standardize the cosine similarity from 0 to 5; Step 10-3, use the standardized value as the final score of the student answer.

Images