CN116719902A - Essay off-topic detection and scoring system and method based on latent semantic indexing - Google Patents

Abstract

The invention relates to the technical field of natural language processing, in particular to a composition running question detection scoring system and method based on potential semantic indexes, wherein a data collection and preprocessing module is used for collecting compositions and corresponding questions thereof and preprocessing collected articles; the term matrix module is used for establishing a term-document matrix and normalizing the result; the latent semantic indexing module is used for carrying out singular value decomposition on the matrix according to the latent semantic index and calculating the similarity between the article and the template; the running question detection module is used for comprehensively detecting whether the article runs questions or not by using the similarity results of the TF-IDF algorithm and the potential semantic indexing module; the scoring module is used for comparing according to the detection result and a preset scoring standard and giving the score of the composition; the output module is used for outputting the article running question detection result and the scoring result of the article to the user.

Description

Essay off-topic detection and scoring system and method based on latent semantic indexing

Technical field

The present invention relates to the technical field of natural language processing, specifically to an essay off-topic detection and scoring system and method based on latent semantic indexing.

Background technique

Composition is one of the important contents in teaching, and it is also an effective way to assess students' language expression ability and thinking level. However, essay review is a time-consuming and labor-intensive task that requires teachers to read, grade and mark each essay carefully. In addition, due to teachers' subjectivity, experience and individual differences, there is often a certain degree of unfairness and inconsistency in essay scoring. Therefore, how to use computer technology to automatically review essays and improve review efficiency and quality is an important issue facing the current education field.

At present, some automatic essay review systems based on natural language processing technology have been proposed and applied. These systems usually use supervised learning methods, using a small number of manually rated samples to learn a prediction model to describe and evaluate the quality of compositions based on different features. For example, traditional methods use the results of shallow analysis of natural language processing to construct features, such as the length of the article, the number of paragraphs, and the richness of vocabulary. In recent years, end-to-end learning methods based on deep learning have also been applied to essay scoring, and essays are abstractly represented as distributed vectors.

However, these methods have some shortcomings. First, they often ignore the relationship between the composition and the given topic, and cannot effectively detect whether the composition is off-topic or off-topic. Secondly, they often lack an in-depth understanding of the composition's content and structure, and cannot accurately evaluate the composition's strengths and weaknesses in terms of themes, arguments, arguments, and chapter structure. Third, they often lack interpretability and cannot provide specific scoring basis and improvement suggestions.

Contents of the invention

The purpose of the present invention is to provide an essay off-topic detection and scoring system and method based on latent semantic indexing to solve the problems raised in the above background technology.

In order to solve the above technical problems, the present invention provides the following technical solution: an essay off-topic detection and scoring system based on latent semantic indexing. The system includes a data collection and preprocessing module, a term matrix module, a latent semantic index module, an off-topic detection module, and a scoring module. and output modules,

The data collection and preprocessing module is used to collect compositions and their corresponding topics, and preprocess the collected articles;

The term matrix module is used to establish a term-document matrix and normalize the results;

The latent semantic index module is used to perform singular value decomposition on the matrix based on the latent semantic index and calculate the similarity between the article and the sample article;

The off-topic detection module is used to use the similarity results of the TF-IDF* algorithm and the latent semantic index module to comprehensively detect whether the article is off-topic;

The scoring module is used to compare the test results with the preset scoring standards and give the score of the composition;

The output module is used to output the article off-topic detection results and the article scoring results to the user.

This invention uses latent semantic indexing to reduce the dimensionality of the term-document matrix, improves the efficiency and effect of text processing, can identify the relationship between terms and documents that are not obvious in the traditional vector space model, and improves the recognition degree. Identify and utilize the underlying structure in the document collection to improve the accuracy of retrieval. It can handle synonyms and polysemy. It uses the TF-IDF* algorithm to improve the accuracy of article off-topic detection. It can better judge whether the article is off-topic and optimize the keywords. extract.

Further, the data collection and preprocessing module includes a data collection unit and a preprocessing unit. The data collection unit is composed of a data capture unit and a data storage unit. The data capture unit is used to analyze compositions and composition topics. Carry out crawling and collection. The data storage unit is used to store the data obtained by the data crawling unit. The preprocessing unit is composed of a data cleaning unit and an English conversion unit. The data cleaning unit is used to deactivate the article. Words and punctuation marks are deleted, and the English conversion unit is used to convert all English words in the article into lowercase English.

By collecting articles and their corresponding topics, the present invention can better match the relevance of articles and topics, and store data in a data storage unit, which can make preprocessing more convenient. By preprocessing articles, it can reduce Redundant information improves the efficiency of text processing. Removing punctuation marks can make the text more standardized and easier to process, improving the accuracy of the results. Converting English to lowercase eliminates the interference of uppercase and lowercase letters, improves the consistency of data, and makes the text The format is more unified and regular, making data processing more efficient and convenient.

Further, the term matrix module includes a matrix creation unit and a normalization unit. The matrix creation unit is used to create a term-document matrix. The normalization unit is used to normalize the representation method of the matrix. to facilitate similarity calculation.

The present invention creates a term-document matrix, normalizes the matrix, facilitates similarity calculation, expresses documents as vectors, and improves calculation efficiency.

Further, the latent semantic index module includes a singular value decomposition unit and a similarity detection unit. The singular value decomposition unit is used to perform singular value decomposition on the matrix to reduce the dimension of the data. The similarity detection unit is used to use cosine Similarity is used to calculate the similarity between the article and the sample article.

This invention performs singular value decomposition on the term-document matrix, reduces the data dimension, improves the efficiency and effect of text processing, improves retrieval accuracy by identifying and utilizing the underlying structure in the collection, and uses cosine similarity to calculate the number of articles and topics. Good fit, does not require a lot of matrix calculations, improves calculation speed, has higher accuracy, and is suitable for high-dimensional vectors.

Further, the off-topic detection module includes a TF-IDF* algorithm unit and a detection unit. The TF-IDF* algorithm unit is used to calculate whether the article is off-topic. The detection unit is used to comprehensively consider the TF-IDF* algorithm and The similarity calculation results of latent semantic index are used to detect whether the article is off-topic.

The present invention uses the TF-IDF* algorithm, improves the original TF-IDF algorithm, comprehensively considers the similarity results of the TF-IDF* algorithm and the latent semantic index, detects whether the article is off-topic, and improves the efficiency of text similarity calculation. Accuracy, optimized keyword extraction, and more accurate detection results can be obtained.

Further, the scoring module includes a similarity comparison unit and a scoring unit. The similarity comparison unit is used to compare the similarity of articles with system standards and divide them into different levels. The scoring unit is used to compare articles based on similarity. Units are graded based on grade results and rubrics.

The present invention divides different grades by comparing the similarity of the article with the system standards, and compares the grades with the scoring standards to obtain the score of the article. No complicated calculations are required. The scoring method is simpler and clearer, and the evaluation results can be obtained quickly. .

Further, the output module includes a visualization unit and a report generator, the visualization unit is used to convert the article similarity and scoring results into chart form, and the report generator is used to generate a report based on the article similarity and scoring results.

The present invention converts article similarity and scoring results into chart form through visualization, which is more intuitive and clear. It generates reports on article similarity and scoring results, improves the data visualization effect, facilitates monitoring of results, and facilitates teachers to provide better services to students. Give suggestion.

An essay off-topic detection method based on latent semantic indexing, which method includes the following steps:

Step S100: The data collection unit collects articles and topics corresponding to the articles;

Step S200: Preprocess the article, delete stop words and punctuation marks, and convert English to lowercase;

Step S300: Create a term-document matrix, with each row corresponding to a term and each column corresponding to a document;

Let A_m×n be a term-document matrix whose rows correspond to terms and whose columns correspond to documents. Let d be the ih column vector of A, i=1,2,…,n. To compare the similarity between document i and document j, we simply do the following:

D·d, or dd in matrix notation, to normalize the results, we divide the results by their norm and call this normalized result cosine similarity. cos(θ)＝d·d/‖d‖‖d‖where -1≤cos(θ)≤1θ is the angle between vector d and vector d. The closer the cosine similarity value is to 1, it means that document i and document The higher the similarity of j, the

Step S400: The latent semantic index module performs singular value decomposition on the matrix and performs cosine similarity calculation;

The singular value decomposition method:

Through singular value decomposition (SVD), A can be decomposed into: A_m×n=U_m×m·Σ_m×n·V_n×n; U=[u_1,u_2,···,u_m],Σ=diag{σ_1,σ_2 ,···,σ_p}, V=[v_1, v_2,···,v_n], where U is the orthogonal matrix of AA, satisfying AA=U∧ ₁ U, where ∧ ₁ is the diagonal and V is AA is an orthogonal matrix that satisfies AA=V∧ ₂ V, where ∧ ₂ is the diagonal, and Σ is a diagonal matrix that contains AA or the square root of the eigenvalues of AA, arranged in descending order on its diagonal,

The singular value decomposition of A can also be written in a more compact form as follows:

A_m×n=U_m×p, Σ_p×p, V_p×n, U=[u_1, u_2,···,u_p], Σ=diag{σ_1,σ_2,···,σ_p}, V=[v_1, v_2,···,v_p], p=minimum value (m, n),

Next, select the k largest σ(σ_1,σ_2,···,σ_k) from Σ, and their corresponding vectors u(u_1,u_2,···,u_k) and v(v_1 from U and V , v_2,···,v_k) to obtain the minimum error (Frobenius norm) of the rank k approximation of A. Write this approximation as: A*_m×n=U_m×k·Σ*_k×k·V_k×n, U=[u_1,u_2,···,u_k], Σ*=diag{σ_1,σ_2,···,σ_k}, V=[v_1, v_2,···,v_k], k≤min(m, n)=p,

Do not use the results contained in AA, i.e. d·d, Instead, use the approximate result contained in A*A*, which is d*·d*/> To compare the similarity between documents i and document j, d is the ih-th column vector of A, i=1,…,n and d* is the ih-th column vector of A*, i=1,…,n, ( Σ*_k×kw)·(Σ*_k×kw), or using matrix notation (Σ*_k×kw)(Σ*_k×kw),

where w is the ih column vector of V,

A*A*＝(U_m×m·Σ*_m×n·V_n×n)·(U_m×m·Σ*_m×n·V_n×n),

Σ*_m×n＝diag{σ_1,σ_2,···,σ_k}, Σ*_k×k＝diag{σ_1,σ_2,···,σ

_k},

The similarity calculation method:

q*=U_m×kq, in order to compare the similarity between the new document query and document i, calculate q* and Σ*_k×kw

The cosine similarity between them is as follows: cos(θ)=q*·(Σ*_k×kw)/‖q*‖‖Σ*_k×kw‖where -1≤cos(θ)≤1, θ is the angle between vector q* and vector Σ*_k×kw,

Step S500: The off-topic detection module uses the TF-IDF* algorithm and the similarity results of the latent semantic index module to comprehensively detect whether the article is off-topic;

The TF-IDF* algorithm is improved from the traditional TF-IDF algorithm.

The traditional TF-IDF algorithm formula is: TF-IDF=TF·IDF,

where n _i,j represents the number of times the term ti appears in the document dj, TFi,j represents the frequency of the term ti appearing in the document dj,

where |D| represents the number of all documents, |j：ti∈dj| represents the number of documents containing the term ti,

The improved TF-IDF* algorithm formula is:

Among them, TF represents the frequency of the term appearing in the text, N represents the total number of documents in the corpus, IDF represents the number of documents containing the term, L represents the average length of the document, and I represents the length of the term.

Improve:

(1) Take the logarithm of TF to smooth the weight of high-frequency words in the article;

(2) Add a constant 0.5 to the IDF to reduce the weight of low-frequency words in the article;

Formula principle:

(1) The more times a word appears in a document, the more it reflects the topic of the document, so TF is positively correlated;

(2) The fewer documents a word appears in the corpus, the better it can distinguish between different documents, so IDF is negatively correlated;

(3) The longer the length of an entry, the more characteristic it is, so the length of the entry is positively correlated;

(4) The longer a document is, the more information it contains, so document length is negatively correlated;

The larger the absolute value of the calculation result of TF-IDF*, the higher the importance, and the better it can reflect the topic, that is, it is more relevant to the topic. The smaller the absolute value, the lower the importance, which cannot reflect the topic, that is, it is off-topic.

Off-topic detection method:

P＝|0.4cosθ|+0.6|(TF-IDF*)|,

Among them, P represents the off-topic detection index of the article, cosθ represents the cosine similarity, TF-IDF* represents the inverse document frequency of word frequency. When P≥a, the article is not off-topic. When P<a, the article is judged to be off-topic. a is the system predetermined Set the threshold value,

Step S600: The scoring module scores according to the standard corresponding to the similarity;

Step S700: Convert the similarity and scoring results of the article into charts for display, and generate a report.

Compared with the prior art, the beneficial effects achieved by the present invention are:

(1) This invention uses latent semantic indexing to reduce the dimensionality of the term-document matrix, improve the efficiency and effect of text processing, and improve recognition by identifying the relationship between terms and documents that are not obvious in the traditional vector space model. , identify and exploit underlying structures in document collections to improve retrieval accuracy;

(2) The present invention preprocesses the article, reduces redundant information, improves text processing efficiency, deletes punctuation marks to make the text more standardized and easier to process, improves the accuracy of the results, unifies the text format, and makes the data Processing is more efficient and convenient;

(3) The present invention uses singular value decomposition to reduce the data dimension and improve text processing efficiency. It uses cosine similarity to calculate the fit between the article and the title, which improves the calculation speed and has higher accuracy;

(4) The present invention uses the TF-IDF* algorithm. By improving the original TF-IDF algorithm and comprehensively considering the similarity results of the TF-IDF* algorithm and the latent semantic index, it detects whether the article is off-topic and improves text similarity. The accuracy of the calculation is optimized, and the extraction of keywords is optimized to obtain more accurate detection results;

(5) The present invention displays similarity and scoring results through visualization and reporting, which improves the data visualization effect, facilitates monitoring of results, and facilitates teachers to better provide suggestions to students.

Description of the drawings

The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. In the attached picture:

Figure 1 is a schematic structural diagram of the essay off-topic detection and scoring system based on latent semantic indexing according to the present invention;

Figure 2 is a schematic flow chart of the essay off-topic detection and scoring method based on latent semantic indexing according to the present invention.

Detailed ways

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

Please refer to Figure 1. The present invention provides a technical solution: an essay off-topic detection and scoring system based on latent semantic indexing. The system includes a data collection and preprocessing module, a term matrix module, a latent semantic index module, an off-topic detection module, a scoring module and an output module,

The data collection and preprocessing module is used to collect compositions and their corresponding topics, and preprocess the collected articles;

The term matrix module is used to establish a term-document matrix and normalize the results;

The latent semantic index module is used to perform singular value decomposition on the matrix based on the latent semantic index and calculate the similarity between the article and the sample article;

The off-topic detection module is used to use the similarity results of the TF-IDF* algorithm and the latent semantic index module to comprehensively detect whether the article is off-topic;

The scoring module is used to compare the test results with the preset scoring standards and give the score of the composition;

The output module is used to output the article off-topic detection results and the article scoring results to the user.

This invention uses latent semantic indexing to reduce the dimensionality of the term-document matrix, improves the efficiency and effect of text processing, can identify the relationship between terms and documents that are not obvious in the traditional vector space model, and improves the recognition degree. Identify and utilize the underlying structure in the document collection to improve the accuracy of retrieval. It can handle synonyms and polysemy. It uses the TF-IDF* algorithm to improve the accuracy of article off-topic detection. It can better judge whether the article is off-topic and optimize the keywords. extract.

The data collection and preprocessing module includes a data collection unit and a preprocessing unit. The data collection unit is composed of a data capture unit and a data storage unit. The data capture unit is used to capture compositions and composition topics. Collect, the data storage unit is used to store the data obtained by the data capture unit, the pre-processing unit is composed of a data cleaning unit and an English conversion unit, the data cleaning unit is used to convert the stop words and punctuation of the article symbols are deleted, and the English conversion unit is used to convert all English words in the article into lowercase English.

By collecting articles and their corresponding topics, the present invention can better match the relevance of articles and topics, and store data in a data storage unit, which can make preprocessing more convenient. By preprocessing articles, it can reduce Redundant information improves the efficiency of text processing. Removing punctuation marks can make the text more standardized and easier to process, improving the accuracy of the results. Converting English to lowercase eliminates the interference of uppercase and lowercase letters, improves the consistency of data, and makes the text The format is more unified and regular, making data processing more efficient and convenient.

The term matrix module includes a matrix creation unit and a normalization unit. The matrix creation unit is used to create a term-document matrix. The normalization unit is used to normalize the representation method of the matrix to facilitate processing. Similarity calculation.

The present invention creates a term-document matrix, normalizes the matrix, facilitates similarity calculation, expresses documents as vectors, and improves calculation efficiency.

The latent semantic index module includes a singular value decomposition unit and a similarity detection unit. The singular value decomposition unit is used to perform singular value decomposition on the matrix to reduce the dimension of the data. The similarity detection unit is used to use cosine similarity to Calculate the similarity between the article and the sample article.

This invention performs singular value decomposition on the term-document matrix, reduces the data dimension, improves the efficiency and effect of text processing, improves retrieval accuracy by identifying and utilizing the underlying structure in the collection, and uses cosine similarity to calculate the number of articles and topics. Good fit, does not require a lot of matrix calculations, improves calculation speed, has higher accuracy, and is suitable for high-dimensional vectors.

The off-topic detection module includes a TF-IDF* algorithm unit and a detection unit. The TF-IDF* algorithm unit is used to calculate whether the article is off-topic. The detection unit is used to comprehensively consider the TF-IDF* algorithm and latent semantic index. The similarity calculation results are used to detect whether the article is off-topic.

The present invention uses the TF-IDF* algorithm, improves the original TF-IDF algorithm, comprehensively considers the similarity results of the TF-IDF* algorithm and the latent semantic index, detects whether the article is off-topic, and improves the efficiency of text similarity calculation. Accuracy, optimized keyword extraction, and more accurate detection results can be obtained.

The scoring module includes a similarity comparison unit and a scoring unit. The similarity comparison unit is used to compare the similarity of articles with system standards and divide them into different grades. The scoring unit is used to compare the grades of units based on similarity. Results and scoring criteria are scored.

The present invention divides different grades by comparing the similarity of the article with the system standards, and compares the grades with the scoring standards to obtain the score of the article. No complicated calculations are required. The scoring method is simpler and clearer, and the evaluation results can be obtained quickly. .

The output module includes a visualization unit and a report generator, the visualization unit is used to convert article similarity and scoring results into chart form, and the report generator is used to generate a report from the article similarity and scoring results.

The present invention converts article similarity and scoring results into chart form through visualization, which is more intuitive and clear. It generates reports on article similarity and scoring results, improves the data visualization effect, facilitates monitoring of results, and facilitates teachers to provide better services to students. Give suggestion.

Please refer to Figure 2, an essay off-topic detection method based on latent semantic indexing. The method includes the following steps:

Step S100: The data collection unit collects articles and topics corresponding to the articles;

Step S200: Preprocess the article, delete stop words and punctuation marks, and convert English to lowercase;

Step S300: Create a term-document matrix, with each row corresponding to a term and each column corresponding to a document;

Let A_m×n be a term-document matrix whose rows correspond to terms and whose columns correspond to documents. Let d be the ih column vector of A, i=1,2,…,n. To compare the similarity between document i and document j, we simply do the following:

D·d, or d in matrix notation, to normalize the results, we divide the results by their norm and call this normalized result cosine similarity. cos(θ)＝d·d/‖d‖‖d‖where -1≤cos(θ)≤1θ is the angle between vector d and vector d. The closer the cosine similarity value is to 1, it means that document i and document The higher the similarity of j, the

Step S400: The latent semantic index module performs singular value decomposition on the matrix and performs cosine similarity calculation;

The singular value decomposition method:

Through singular value decomposition (SVD), A can be decomposed into: A_m×n=U_m×m·Σ_m×n·V_n×n;

U=[u_1,u_2,···,u_m], Σ=diag{σ_1, σ_2,···,σ_p}, V=[v_1,

v_2,···,v_n], where U is the orthogonal matrix of AA, satisfying AA=U∧ ₁ U, where ∧ ₁ is the diagonal, V is the orthogonal matrix of AA, satisfying AA=V∧ ₂ V, where ∧ ₂ is the diagonal, Σ is a diagonal matrix containing AA or the square roots of the eigenvalues of AA in descending order on its diagonal,

The singular value decomposition of A can also be written in a more compact form as follows:

A_m×n=U_m×p, Σ_p×p, V_p×n, U=[u_1, u_2,···,u_p], Σ=diag{σ_1,σ_2,···,σ_p}, V=[v_1, v_2,···,v_p], p=minimum value (m, n),

Next, select the k largest σ(σ_1,σ_2,···,σ_k) from Σ, and their corresponding vectors u(u_1,u_2,···,u_k) and v(v_1 from U and V , v_2,···,v_k) to obtain the minimum error (Frobenius norm) of the rank k approximation of A. Write this approximation as: A*_m×n=U_m×k·Σ*_k×k·V_k×n, U=[u_1,u_2,···,u_k], Σ*=diag{σ_1,σ_2,···,σ_k}, V=[v_1, v_2,···,v_k], k≤min(m, n)=p,

Do not use the result contained in AA, i.e. d·d, Instead, use the approximate result contained in A*A*, which is d*·d*,/> To compare the similarity between documents i and document j, d is the ih-th column vector of A, i=1,…,n and d* is the ih-th column vector of A*, i=1,…,n, ( Σ*_k×kw)·(Σ*_k×kw), or using matrix notation (Σ*_k×kw)(Σ*_k×kw),

where w is the ih column vector of V,

A*A*＝(U_m×m·Σ*_m×n·V_n×n)·(U_m×m·Σ*_m×n·V_n×n),

Σ*_m×n=diag{σ_1,σ_2,···,σ_k}, Σ*_k×k=diag{σ_1,σ_2,···,σ_k},

The similarity calculation method:

q*=U_m×kq, in order to compare the similarity between the new document query and document i, calculate q* and Σ*_k×kw

The cosine similarity between them is as follows: cos(θ)=q*·(Σ*_k×kw)/‖q*‖‖Σ*_k×kw ‖where -1≤cos(θ)≤1, θ is the angle between vector q* and vector Σ*_k×kw,

Step S500: The off-topic detection module uses the TF-IDF* algorithm and the similarity results of the latent semantic index module to comprehensively detect whether the article is off-topic;

The TF-IDF* algorithm is improved from the traditional TF-IDF algorithm.

The traditional TF-IDF algorithm formula is: TF-IDF=TF·IDF,

where n _i,j represents the number of times the term ti appears in the document dj, TFi,j represents the frequency of the term ti appearing in the document dj,

where |D| represents the number of all documents, |j：ti∈dj| represents the number of documents containing the term ti,

The improved TF-IDF* algorithm formula is:

Among them, TF represents the frequency of the term appearing in the text, N represents the total number of documents in the corpus, IDF represents the number of documents containing the term, L represents the average length of the document, and I represents the length of the term.

Improve:

(1) Take the logarithm of TF to smooth the weight of high-frequency words in the article;

(2) Add a constant 0.5 to the IDF to reduce the weight of low-frequency words in the article;

Formula principle:

(1) The more times a word appears in a document, the more it reflects the topic of the document, so TF is positively correlated;

(2) The fewer documents a word appears in the corpus, the better it can distinguish between different documents, so IDF is negatively correlated;

(3) The longer the length of an entry, the more characteristic it is, so the length of the entry is positively correlated;

(4) The longer a document is, the more information it contains, so document length is negatively correlated;

The larger the absolute value of the calculation result of TF-IDF*, the higher the importance, and the better it can reflect the topic, that is, it is more relevant to the topic. The smaller the absolute value, the lower the importance, which cannot reflect the topic, that is, it is off-topic.

Off-topic detection method:

P＝|0.4cosθ|+0.6|(TF-IDF*)|,

Among them, P represents the off-topic detection index of the article, cosθ represents the cosine similarity, TF-IDF* represents the inverse document frequency of word frequency. When P≥a, the article is not off-topic. When P<a, the article is judged to be off-topic. a is the system predetermined Set the threshold value,

Example:

Create a term-document matrix for the article, perform singular value decomposition on the matrix, and use cosine similarity to calculate it. The result is that cosθ is 0.4, the frequency of term entries in the article TF is 0.2, and the total number of documents in the corpus N is 3, including The number of documents of the term IDF is 2, the average length of the document L is 6, and the length of the term I is 2.

Substitute data into The article's off-topic indicator P=|0.4*0.4|+0.6*|-0.58|=0.5. The preset threshold a in the system is 0.4. Compare the article's off-topic indicator and the threshold. P>a indicates that the article is not off-topic.

Step S600: The scoring module scores according to the standard corresponding to the similarity;

Step S700: Convert the similarity and scoring results of the article into charts for display, and generate a report.

It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that these entities or operations are mutually exclusive. any such actual relationship or sequence exists between them. Furthermore, the terms "comprises," "comprises," or any other variations thereof are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that includes a list of elements includes not only those elements, but also those not expressly listed other elements, or elements inherent to the process, method, article or equipment.

Finally, it should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, it is still The technical solutions described in the foregoing embodiments may be modified, or some of the technical features may be equivalently replaced. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims (10)

1. The composition running question detection scoring system based on the potential semantic index is characterized by comprising a data collection and preprocessing module, a term matrix module, a potential semantic index module, a running question detection module, a scoring module and an output module,

the data collection and preprocessing module is used for collecting compositions and corresponding topics thereof and preprocessing the collected articles;

the term matrix module is used for establishing a term-document matrix and normalizing the result;

the latent semantic indexing module is used for carrying out singular value decomposition on the matrix according to the latent semantic indexes and calculating the similarity between the article and the template;

the running question detection module is used for comprehensively detecting whether the article runs questions or not by using the similarity results of the TF-IDF algorithm and the potential semantic indexing module;

the scoring module is used for comparing according to the detection result and a preset scoring standard and giving the score of the composition;

and the output module is used for outputting the article running question detection result and the scoring result of the article to the user.

2. The underlying semantic indexing based composition running question detection scoring system of claim 1, wherein: the data collection and preprocessing module comprises a data collection unit and a preprocessing unit, wherein the data collection unit consists of a data grabbing unit and a data storage unit, the data grabbing unit is used for grabbing and collecting the titles of the compositions and the works, the data storage unit is used for storing the data acquired by the data grabbing unit, the preprocessing unit consists of a data cleaning unit and an English conversion unit, the data cleaning unit is used for deleting the stop words and punctuation marks of the articles, and the English conversion unit is used for converting all English words in the articles into lower-case English.

3. The underlying semantic indexing based composition running question detection scoring system of claim 1, wherein: the term matrix module comprises a matrix creation unit and a normalization unit, wherein the matrix creation unit is used for creating a term-document matrix, and the normalization unit is used for performing normalization processing on a representation method of the matrix so as to perform similarity calculation.

4. The underlying semantic indexing based composition running question detection scoring system of claim 1, wherein: the latent semantic indexing module comprises a singular value decomposition unit and a similarity detection unit, wherein the singular value decomposition unit is used for performing singular value decomposition on the matrix and reducing the dimension of data, and the similarity detection unit is used for calculating the similarity between the article and the norm by using cosine similarity.

5. The underlying semantic indexing based composition running question detection scoring system of claim 1, wherein: the running problem detection module comprises a TF-IDF algorithm unit and a detection unit, wherein the TF-IDF algorithm unit is used for calculating whether the article runs problems or not, and the detection unit is used for detecting whether the article runs problems or not by comprehensively considering similarity calculation results of the TF-IDF algorithm and potential semantic indexes.

6. The underlying semantic indexing based composition running question detection scoring system of claim 1, wherein: the scoring module comprises a similarity comparison unit and a scoring unit, wherein the similarity comparison unit is used for comparing the article similarity with the system standard and dividing different grades, and the scoring unit is used for scoring according to the grade result and the scoring standard of the similarity comparison unit.

7. The underlying semantic indexing based composition running question detection scoring system of claim 1, wherein: the output module comprises a visualization unit and a report generator, wherein the visualization unit is used for converting the similarity and the scoring result of the articles into a chart form, and the report generator is used for generating a report of the similarity and the scoring result of the articles.

8. A composition running question detection method based on latent semantic index applied to the composition running question detection scoring system based on latent semantic index as claimed in any one of claims 1-7, characterized in that: the method comprises the following steps:

step S100: the data acquisition unit is used for collecting the articles and the topics corresponding to the articles;

step S200: preprocessing the article, deleting stop words and punctuation marks, and converting English into lowercase;

step S300: creating a term-document matrix, wherein each row corresponds to a term, and each column corresponds to a document;

step S400: the latent semantic indexing module carries out singular value decomposition on the matrix and carries out cosine similarity calculation;

step S500: the running question detection module uses the similarity result of the TF-IDF algorithm and the potential semantic indexing module to comprehensively detect whether the article runs questions or not;

step S600: the scoring module scores according to the corresponding standards of the similarity;

step S700: and converting the similarity and scoring results of the articles into charts for display, and generating reports.

9. The method for detecting and scoring a composition running question based on potential semantic indexes as claimed in claim 8, wherein: the TF-IDF algorithm in step S500 is modified from the conventional TF-IDF algorithm,

the improved TF-IDF algorithm formula is:

where TF represents the frequency of occurrence of the term in the text, N represents the total number of documents in the corpus, IDF represents the number of documents containing the term, L represents the average length of the documents, and I represents the length of the term.

10. The method for detecting and scoring a composition running question based on potential semantic indexes as claimed in claim 8, wherein: the running problem detection method in the step S500 is as follows:

P＝|0.4cosθ|+0.6|(TF-IDF*)|，

wherein P represents the running problem detection index of the article, cos theta represents cosine similarity, TF-IDF represents word frequency inverse document frequency, when P is more than or equal to a, the article does not run problems, when P is less than a, the article judges that the article runs problems, and a is a threshold preset by a system.