Method, system and device for evaluating the degree of aggregation of the theme of an essay based on the main word of the domain of discourse

Abstract

The application provides a composition theme aggregation degree evaluation method, system and device based on a field main word, the method comprising: obtaining a to-be-evaluated composition; preprocessing the text of the to-be-evaluated composition to obtain paragraph boundaries and part-of-speech tagging results of the evaluated composition; extracting field main words from each paragraph according to the part-of-speech tagging results and a preset word class priority order, the field main words comprising pronoun type field main words and subject type field main words; restoring the reference of the pronoun type field main words in the context paragraphs and the sentences where the pronoun type field main words are located according to the paragraph boundaries to determine the semantic content of the pronoun type field main words; constructing a vector space model of the field main words; calculating the similarity between the field main word vectors of each paragraph and the overall field main word vector of the to-be-evaluated composition to obtain the theme matching degrees of each paragraph and the to-be-evaluated composition; and outputting the overall theme aggregation degree of the to-be-evaluated composition according to the theme matching degrees of each paragraph.

Description

Technical Field

[0001] This application relates to the field of essay evaluation technology, and in particular to a method, system and device for evaluating the degree of aggregation of essay topics based on domain subject words. Background Technology

[0002] With the widespread application of artificial intelligence technology and large-scale natural language processing models in education, automated essay assessment has gradually become an important means of reducing teachers' grading burden and improving teaching efficiency. In particular, Chinese writing assessment faces unique challenges: as a semantic language, Chinese lacks rich morphological markers, making traditional automated grading methods based on surface features (such as word frequency, sentence length, and misspelling detection) insufficient to accurately reflect thematic consistency and semantic coherence between paragraphs. Furthermore, the diagnostic feedback needed for essays in classroom teaching requires not only objective scoring but also interpretable diagnostic information that can guide writing improvement. This places higher demands on the semantic understanding and interpretability of automated assessment systems.

[0003] In related technologies, existing research mainly uses topic models, document divergence, or unsupervised methods based on local density to determine whether an essay is "off-topic" or deviates from the theme. In recent years, large models have also been used for discourse evaluation. However, these methods either focus on statistical distributions and are insufficient in modeling Chinese semantic cues, or they are unstable in zero-sample / few-sample situations, making it difficult to provide clear paragraph-level diagnosis and teaching suggestions. Summary of the Invention

[0004] In view of this, it is necessary to provide a method, system and device for evaluating the aggregation degree of essay topics based on the domain subject, which can at least overcome one of the above-mentioned defects.

[0005] In a first aspect, embodiments of this application provide a method for evaluating the aggregation degree of essay topics based on domain keywords, the method comprising: Obtain the essays to be evaluated; The text of the essay to be evaluated is preprocessed to obtain the paragraph boundaries and part-of-speech tagging results of the essay to be evaluated; Based on the part-of-speech tagging results and the preset part-of-speech priority order, the main topic words are extracted from each paragraph. The main topic words include pronoun-type main topic words and subject-type main topic words. Based on the paragraph boundaries, the referential meaning of the pronoun domain subject is restored in the sentence and surrounding paragraphs to determine the semantic content of the pronoun domain subject. The vector space model of the domain term is constructed based on the domain term, the frequency of the domain term in the text, and the word topic degree obtained through corpus training. In the vector space model, each paragraph is represented as a paragraph domain main word vector, and the essay to be evaluated is represented as the overall domain main word vector; Calculate the similarity between the main word vector of each paragraph and the overall main word vector of the essay to be evaluated to obtain the thematic fit between each paragraph and the essay to be evaluated; The overall thematic convergence of the essay to be evaluated is output based on the thematic relevance of each paragraph.

[0006] In one embodiment, the step of extracting the domain subject from each paragraph based on the part-of-speech tagging results and a preset part-of-speech priority order includes: The candidate words in the paragraph are sorted according to the preset word class priority and the part-of-speech tagging results to obtain a candidate word sequence; The candidate word sequence is screened according to the representativeness evaluation rules to obtain the domain main word; The representativeness evaluation rule includes evaluating based on at least one of the frequency of occurrence of the candidate word in the paragraph and the whole text, and the semantic relevance of the candidate word to other words in the paragraph.

[0007] In one embodiment, the method further includes: Obtain the topic degree of the words in the domain of discourse by training on a large-scale classification corpus; Obtain the frequency of the domain term in the essay to be evaluated; Obtain the positional importance factor of the domain subject within the paragraph; The vector space model of the domain subject is constructed based on the topic degree, frequency of occurrence, and positional importance factor of the domain subject.

[0008] In one embodiment, the method further includes: Obtain the statistical distribution characteristics of the thematic fit between each paragraph and the essay to be evaluated; The overall thematic aggregation level of the essay to be evaluated is determined based on the statistical distribution characteristics.

[0009] In one embodiment, the method further includes: Calculate the mean and standard deviation of the thematic relevance of each paragraph; The mean and standard deviation are compared with the grade threshold range determined by the dynamic threshold method; The overall thematic cohesion level of the essay to be evaluated is output based on the comparison results.

[0010] In one embodiment, the dynamic threshold method includes: The dataset containing multiple essays with labeled aggregation levels was divided into several subsets and subjected to five-fold cross-validation. For each level, the combination of mean threshold and standard deviation threshold is iteratively adjusted within a reasonable range with a preset step size, and the corresponding F1 value is calculated. The threshold combination with the largest F1 value is selected as the threshold range for that level. In each cross-validation, the training subset is used to determine the threshold range for each level, and the test subset is used to evaluate the model performance. Based on the results of multiple cross-validations, a five-level classification standard for topic aggregation was determined, where level one indicates highly aggregated topics and level five indicates severely dispersed topics.

[0011] In one embodiment, the method further includes: Generate the topic relevance of each paragraph; Identify off-topic paragraphs whose topic relevance is below a preset threshold; Based on the analysis results of the domain subject, improvement suggestions including keyword supplementation and semantic enhancement are generated for the off-topic paragraph.

[0012] In one embodiment, the preprocessing of the text of the essay to be evaluated to obtain the paragraph boundaries and part-of-speech tagging results of the essay includes: Generate a structured representation of the text of the essay to be evaluated, wherein the structured representation includes at least the paragraph boundaries, terms, and the part-of-speech tags corresponding to the terms.

[0013] Secondly, embodiments of this application provide a system for evaluating the aggregation degree of essay topics based on domain keywords, applied to implement the essay topic aggregation degree evaluation method based on domain keywords as described in the first aspect, characterized in that the system includes: The information acquisition module is used to acquire the essays to be evaluated; The domain subject extraction module is used to preprocess the text of the essay to be evaluated to obtain the paragraph boundaries and part-of-speech tagging results of the essay to be evaluated, and extract the domain subject from each paragraph according to the part-of-speech tagging results and the preset part-of-speech priority order. The domain subject includes pronoun domain subject and subject domain subject. The vector space model construction module is used to perform referential restoration of the pronoun-type domain subject in the sentence and surrounding paragraphs based on the paragraph boundaries to determine the semantic content of the pronoun-type domain subject. It then calculates weights based on the domain subject, its frequency in the text, and the word topic degree obtained through corpus training, and constructs a vector space model for the domain subject. In this model, each paragraph is represented as a paragraph domain subject vector, and the essay to be evaluated is represented as the overall domain subject vector. The result output module is used to calculate the similarity between the main word vector of each paragraph and the overall main word vector of the essay to be evaluated, to obtain the theme fit of each paragraph and the essay to be evaluated, and to output the overall theme aggregation degree of the essay to be evaluated based on the theme fit of each paragraph.

[0014] Thirdly, embodiments of this application provide an electronic device, including: processor; Memory used to store processor-executable instructions; The processor is configured to implement the essay topic aggregation degree evaluation method based on domain subject as described in the first aspect when executing the instructions.

[0015] This application provides a method, system, and device for evaluating essay topic aggregation based on domain subject words. It extracts domain subject words that represent core semantic units from the essay text, restores their referential meaning to clarify semantic content, constructs a semantic relationship model of domain subject words, and finally quantifies topic aggregation based on dynamic thresholds, achieving deep semantic analysis of the essay's content organization logic. This method effectively overcomes the problems of low efficiency and strong subjectivity in traditional manual grading, and also compensates for the shortcomings of existing automatic grading technologies, which are mostly limited to grammatical error correction and have difficulty assessing content fit. Specifically addressing the lack of morphological markers in the semantically cohesive language of Chinese, through precise location and semantic mining of domain subject words, it achieves an objective and accurate measurement of the consistency between each paragraph and the overall theme of the essay at the content level. Compared to the unstable performance of general large language models on this task, the evaluation results of this scheme are more reliable. Furthermore, its dynamic threshold mechanism can adapt to the characteristics of different types of essays, significantly improving the accuracy of the evaluation and its applicability for teaching, providing an efficient and accurate intelligent auxiliary tool for writing instruction. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for evaluating essay topic aggregation based on domain subject words, provided as an embodiment of this application.

[0017] Figure 2 This is a schematic diagram showing the change of the F1 value of the dynamic threshold method provided in an embodiment of this application with the mean of the fit.

[0018] Figure 3 This is a schematic diagram of a module of an essay topic aggregation degree evaluation system based on domain subject words provided in an embodiment of this application.

[0019] Figure 4 A schematic diagram of an electronic device provided in an embodiment of this application.

[0020] Explanation of main component symbols Essay Topic Aggregation Evaluation System Based on Domain Subject 10 Information Acquisition Module 11 Domain Subject Extraction Module 12 Vector Space Model Building Module 13 Result Output Module 14 Electronic devices 20 Processor 21 Memory 22 Method steps S100-800 Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0022] It should be noted that, in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0023] It should be noted that in the embodiments of this application, the terms "first," "second," etc., are used only for descriptive purposes and should not be construed as indicating or implying relative importance, nor as indicating or implying order. Features specified as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0024] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] With the widespread application of artificial intelligence technology and large-scale natural language processing models in education, automated essay assessment has gradually become an important means of reducing teachers' grading burden and improving teaching efficiency. In particular, Chinese writing assessment faces unique challenges: as a semantic language, Chinese lacks rich morphological markers, making traditional automated grading methods based on surface features (such as word frequency, sentence length, and misspelling detection) insufficient to accurately reflect thematic consistency and semantic coherence between paragraphs. Furthermore, the diagnostic feedback needed for essays in classroom teaching requires not only objective scoring but also interpretable diagnostic information that can guide writing improvement. This places higher demands on the semantic understanding and interpretability of automated assessment systems.

[0026] In related technologies, existing research mainly uses topic models, document divergence, or unsupervised methods based on local density to determine whether an essay is "off-topic" or deviates from the theme. In recent years, large models have also been used for discourse evaluation. However, these methods either focus on statistical distributions and are insufficient in modeling Chinese semantic cues, or they are unstable in zero-sample / few-sample situations, making it difficult to provide clear paragraph-level diagnosis and teaching suggestions.

[0027] This application provides a method, system, and device for evaluating essay topic aggregation based on domain subject words. It extracts domain subject words that represent core semantic units from the essay text, restores their referential meaning to clarify semantic content, constructs a semantic relationship model of domain subject words, and finally quantifies topic aggregation based on dynamic thresholds, achieving deep semantic analysis of the essay's content organization logic. This method effectively overcomes the problems of low efficiency and strong subjectivity in traditional manual grading, and also compensates for the shortcomings of existing automatic grading technologies, which are mostly limited to grammatical error correction and have difficulty assessing content fit. Specifically addressing the lack of morphological markers in the semantically cohesive language of Chinese, through precise location and semantic mining of domain subject words, it achieves an objective and accurate measurement of the consistency between each paragraph and the overall theme of the essay at the content level. Compared to the unstable performance of general large language models on this task, the evaluation results of this scheme are more reliable. Furthermore, its dynamic threshold mechanism can adapt to the characteristics of different types of essays, significantly improving the accuracy of the evaluation and its applicability for teaching, providing an efficient and accurate intelligent auxiliary tool for writing instruction.

[0028] Figure 1 This is a flowchart illustrating a method for evaluating the aggregation degree of essay topics based on domain subject words, provided in one embodiment of this application. Figure 1The method for evaluating essay topic aggregation degree based on domain subject words includes at least the following steps: S100: Obtain the essay to be evaluated; S200: Preprocess the text of the essay to be evaluated to obtain the paragraph boundaries and part-of-speech tagging results; S300: Extract domain subject words from each paragraph according to the part-of-speech tagging results and the preset part-of-speech priority order. Domain subject words include pronoun domain subject words and subject domain subject words; S400: Perform referential restoration on pronoun domain subject words in the sentence and surrounding paragraphs according to the paragraph boundaries to determine the semantic content of pronoun domain subject words; S500: Calculate the weights of domain subject words, their frequency in the text, and the topic degree of words obtained through corpus training, and construct a vector space model of domain subject words; S600: Calculate the similarity between the domain subject word vector of each paragraph and the overall domain subject word vector of the essay to be evaluated to obtain the topic fit degree between each paragraph and the essay to be evaluated; S700: Output the overall topic aggregation degree of the essay to be evaluated based on the topic fit degree of each paragraph.

[0029] S100: Obtain the essays to be evaluated.

[0030] Specifically, the acquisition process includes receiving essay text from a specified data source and simultaneously collecting necessary metadata (such as essay identifier, student's grade / semester, writing topic, and submission time), performing unified encoding and format verification on the input text to ensure the usability of subsequent processing, and minimizing existing format noise (such as line breaks, repeated spaces, or non-text symbols introduced by scanning) to avoid affecting paragraph boundary recognition.

[0031] Understandably, this step only completes the standardization and registration of the original materials, providing complete and consistent input data and contextual identifiers for subsequent extraction of domain terms and calculation of aggregation degree, making the evaluation process traceable.

[0032] S200: Preprocess the text of the essay to be evaluated to obtain paragraph boundaries and part-of-speech tagging results.

[0033] Specifically, intermediate representations containing paragraph boundary information and term attributes are generated through text normalization and structuring transformation, and the position index of each term in the paragraph and its association information with surrounding terms are recorded. At the same time, necessary metadata fields (such as term candidate categories, intra-sentence grammatical role indicators and intra-text occurrence positions) are reserved for subsequent semantic processing so that they can be uniformly called by the extraction, reference restoration and weight calculation modules.

[0034] Understandably, the purpose of this step is to transform the raw string into a structured representation that can be processed by a machine—indicating "where the paragraphs are" and preserving semantic / locational clues for each word, providing foundational data for keyword filtering and semantic matching.

[0035] S300: Based on the part-of-speech tagging results and the preset part-of-speech priority order, extract the domain subject from each paragraph. The domain subject includes pronoun domain subjects and subject domain subjects.

[0036] Specifically, candidate words within a paragraph are sorted according to a predefined word class priority, and a representative index is calculated for each candidate (considering the occurrence of the word within the paragraph / the whole text, its relevance to the semantic center of the paragraph, and its semantic consistency with the context). Several words are selected from the top-ranked and highly representative candidates as the domain words of the paragraph, and pronouns and subjects are marked separately for subsequent differential processing.

[0037] Understandably, this step aims to select keywords or entities (i.e., domain terms) from each paragraph that can represent the topic of that paragraph. These domain terms are the core signal sources for subsequent vector construction and fit evaluation.

[0038] S400: Based on paragraph boundaries, the referential meaning of pronoun-type domain subjects in the sentence and surrounding paragraphs is restored to determine the semantic content of pronoun-type domain subjects.

[0039] Specifically, by comparing the semantic compatibility and intra-text positional relationship of potential antecedents within and between paragraphs, the most suitable antecedent is selected to replace the pronoun, and the replacement result is written back into the structured representation. When there is ambiguity or multiple candidate antecedents, the optimal or conservative strategy (e.g., marking multiple candidates or retaining uncertain markers) is adopted based on the semantic consistency score and the recent discourse constraints to ensure the reviewability of the restoration result.

[0040] Understandably, this step maps ambiguous pronouns to specific semantic entities or terms, thereby avoiding topic judgment bias caused by unparsed pronouns and ensuring that subsequent weight calculations use explicit semantic terms.

[0041] S500: Calculate the weights of the domain subject, the frequency of the domain subject in the text, and the topic degree of the words obtained through corpus training, and construct a vector space model of the domain subject.

[0042] Specifically, the topic association score of each term obtained from the large-scale classification corpus is combined with its frequency of occurrence in the current text and paragraph, the importance of the term in the paragraph, and the information role it plays. The final weight is synthesized according to the set rules and the weight set is normalized. Based on the normalized weight, the domain words are vectorized and a vector space model with consistent dimensions is constructed so as to compare the features of the paragraph and the whole text in the same metric space.

[0043] Understandably, this step combines corpus statistical information with intra-text signals to assign numerical weights to each subject word that reflect its "topic contribution," and then transforms these weights into vector representations that can be used to calculate similarity.

[0044] In this embodiment, a word topic weighting calculation model is applied for core calculation. This model determines the topic weighting by comprehensively measuring the word's "neighborhood domain" and "domain unevenness." Specifically, the neighborhood domain is calculated using the formula... This formula effectively filters out mid-frequency words that characterize the topic, while weakening the influence of high-frequency functional words and low-frequency rare words; domain unevenness is addressed through the formula. This quantifies the dispersion of word distribution across different topic categories; a higher value indicates a stronger ability of the word to represent a specific topic. Finally, the topic weight of a word is calculated using the following formula.

[0045]

[0046] in, , Is class The sum of the frequencies of all the words contained. It is a word In class Number of times it appears; ,in Number of categories; This indicates the words that appear in the training corpus. Number of times, It is the sum of the occurrences of all words in the training corpus.

[0047] Understandably, the design of this topic-degree calculation mechanism aims to achieve a balance between two aspects and interpretability: on the one hand, by suppressing high-frequency general words and extreme low-frequency noise through neighborhood domains, words assigned high topic degrees are more likely to be mid-frequency terms that truly carry topic information; on the other hand, by highlighting words with significant biases across several topic categories through domain unevenness, thereby enhancing topic discrimination capabilities. The results obtained by combining these two indicators... The value can be directly used to weight and sort candidate words within a paragraph / the whole text, construct weighted vectors, and participate in paragraph-to-text similarity calculation, thereby improving the robustness and interpretability of aggregation degree determination. In practical applications, the value can be processed by smoothing, normalization, etc., and integrated with the frequency and position factors within the text to adapt to the diverse needs of corpus size, category imbalance, and teaching scenarios.

[0048] In this embodiment of the application, in the vector space model, each paragraph is represented as a paragraph domain main word vector, and the essay to be evaluated is represented as the overall domain main word vector.

[0049] Specifically, the weighted vectors of the selected domain terms in each paragraph are merged into a paragraph vector according to a predetermined aggregation strategy. The weighted or directly merged result of all paragraph vectors in the whole text is used as the overall domain term vector. At the same time, the vectors are standardized and aligned in a necessary manner to avoid numerical bias. If necessary, adjacent paragraphs or paragraphs with similar themes can be smoothed to strengthen the chapter-level theme signal.

[0050] It is understandable that by condensing the thematic information of each paragraph into vector units and then summing these units into a vector representing the theme of the entire text, the subsequent similarity comparison can objectively reflect the thematic relationship between the paragraphs and the entire text.

[0051] S600: Calculate the similarity between the main word vector of each paragraph and the overall main word vector of the essay to be evaluated, and obtain the thematic fit between each paragraph and the essay to be evaluated.

[0052] Specifically, a vector similarity metric is used to compare each paragraph vector with the overall vector to obtain a numerical fit score, and outliers or paragraphs with significantly low scores are marked for subsequent diagnosis. At the same time, supplementary information such as paragraph position and coherence can be combined to locally weight or smooth the fit score, producing an evaluation value that reflects both the semantic fit of a single paragraph and the coherence of the entire text.

[0053] Understandably, this step quantifies whether "a certain paragraph discusses a theme consistent with the whole text." This paragraph-level fit is a direct basis for judging the overall theme coherence of the text and also provides evidence for accurately locating "deviation paragraphs."

[0054] S700: Outputs the overall thematic convergence of the essay to be evaluated based on the thematic relevance of each paragraph.

[0055] Specifically, statistical characteristics of all paragraph fit (such as mean, dispersion, and the proportion of low-scoring paragraphs) are converted into an overall aggregation score and level according to a pre-set or training-determined mapping rule, and a diagnostic report is generated that includes a list of deviating paragraphs, alternative or supplementary keyword suggestions, and targeted writing improvement suggestions. The final evaluation results are output in the form of a structured report (such as machine-readable scoring records and human-readable diagnostic text for teachers to view).

[0056] Understandably, this step summarizes the paragraph-level quantitative results into an overall conclusion that is easy to apply in teaching—providing both an overall evaluation level and actionable directions for improvement, thereby supporting teaching interventions and enhancing students' writing skills.

[0057] In this embodiment, the extraction of domain main words from each paragraph based on part-of-speech tagging results and a preset part-of-speech priority order includes: sorting candidate words within the paragraph according to the preset part-of-speech priority and tagging results to obtain a candidate word sequence; and filtering the candidate word sequence according to representativeness evaluation rules to obtain the domain main words. The representativeness evaluation rules include evaluating based on at least one of the following: the frequency of occurrence of candidate words in the paragraph and the entire text; and the semantic relevance of candidate words to other words within the paragraph.

[0058] Specifically, for the candidate word sequence, the system first calculates several quantifiable representative indicators for each candidate word (such as the occurrence rate within a paragraph, the overall coverage, the semantic similarity with the paragraph's topic word, and the semantic dependence strength with the predicate or main component within the sentence). Then, the indicators are weighted and summarized according to the designed weight combination or priority strategy to obtain the representative score of the candidate word. Subsequently, several words with the highest scores are selected as the main words of the paragraph's domain according to the scores from high to low, and the selection results are checked for consistency (such as checking for duplicate references or semantic conflicts). When the semantic relevance cannot be determined or the data is sparse, a conservative strategy is adopted (such as retaining multiple candidates or using the closest item in the context) to reduce the risk of misselection.

[0059] Understandably, this step aims to filter out potential thematic signals in paragraphs through quantifiable representativeness assessments, ensuring that the selected domain terms are both representative within the paragraph and thematically relevant in the context of the entire text, thus providing robust input for subsequent semantic modeling and aggregation degree calculation.

[0060] In this embodiment of the application, the method further includes: obtaining the topic degree of the domain term trained in a large-scale classification corpus; obtaining the frequency of occurrence of the domain term in the essay to be evaluated; obtaining the positional importance factor of the domain term in the paragraph; and constructing a vector space model of the domain term based on the topic degree, frequency of occurrence, and positional importance factor of the domain term.

[0061] Specifically, the system first retrieves the topic association value of each domain subject from the constructed classification corpus, and performs necessary smoothing or normalization processing on the value to reduce the impact of extreme values; at the same time, it counts the number of times the subject appears in the current text as a whole and within a paragraph, and assigns a position importance factor according to its position in the paragraph (such as the beginning, middle, or end of the paragraph); the topic association value, frequency of occurrence, and position factor are combined into the final weight according to the set fusion rules, and the corresponding semantic embedding vector or discrete feature vector is weighted using this weight as a coefficient, and finally a vector representation that can participate in similarity calculation is generated under a unified dimension.

[0062] Understandably, this step combines corpus-level thematic information with intratextual signals, enabling the same word to obtain a numerical representation reflecting its "thematic contribution" in different texts, thereby improving the ability of vectorized representation to depict real thematic relationships.

[0063] In this embodiment of the application, the method further includes: obtaining the statistical distribution characteristics of the thematic fit between each paragraph and the essay to be evaluated; and determining the overall thematic convergence level of the essay to be evaluated based on the statistical distribution characteristics.

[0064] Specifically, the system collects the fit scores obtained by comparing all paragraphs with the overall vector, calculates descriptive statistics of these scores (such as arithmetic mean, median, standard deviation, minimum, maximum, and proportion of low-scoring paragraphs), and can calculate the segment distribution (such as local mean divided by paragraph position) according to teaching needs. Then, these statistics are mapped to a predefined evaluation index space, and the overall aggregation score and corresponding level label are output through rules or learned mapping functions.

[0065] Specifically, the cohesion assessment uses two core parameters to comprehensively present the thematic cohesion: the mean of the relevance of each paragraph to the main theme of the entire essay, and the standard deviation of these relevance values. Logically, the mean is the first condition; its value must reach a preset basic threshold, indicating that the essay content maintains a fundamental alignment with the theme as a whole. Building upon this, the standard deviation is further considered. This parameter, as the square root of the variance, effectively quantifies the dispersion of the relevance data for each paragraph, thus accurately reflecting the coherence and structural compactness of the thematic argument. Only when both the mean and standard deviation simultaneously meet the preset level threshold conditions can the essay be judged to have appropriate thematic cohesion, indicating its compact structure and clear theme. The threshold ranges for the mean and standard deviation corresponding to each cohesion level were ultimately determined through repeated trials and optimizations on a large number of essay samples.

[0066] Understandably, this step assesses thematic consistency across the entire text by summarizing quantitative evidence at the paragraph level, ensuring that the final grade reflects both general trends (such as the mean) and internal fluctuations (such as standard deviation and the proportion of low-scoring paragraphs), thus providing a more comprehensive and cohesive assessment.

[0067] In this embodiment, the method further includes: calculating the mean and standard deviation of the topic relevance of each paragraph; comparing the mean and standard deviation with a grade threshold range determined by a dynamic thresholding method; and outputting the overall topic cohesion grade of the essay to be evaluated based on the comparison results. Please refer to [further details needed]. Figure 2 , Figure 2 This is a schematic diagram showing the change of the F1 value of the dynamic threshold method provided in an embodiment of this application with the mean of the fit.

[0068] Specifically, the system first performs numerical correction and outlier filtering on the fit scores of each paragraph, then calculates the mean and standard deviation of the fit sequence to form a two-dimensional statistical vector. This statistical vector is then compared step by step with a set of level thresholds generated by a dynamic threshold method (for example, first comparing whether the mean falls within the mean range of a certain level, and then checking whether the standard deviation meets the dispersion requirements of that level). If both are met, it is determined to be that level. If a conflict occurs, the final classification is performed according to the set priority rules or by using a trained decision-maker.

[0069] Understandably, by examining both the mean and standard deviation simultaneously, the judgment mechanism can not only identify the overall level of thematic consistency, but also detect the differences between paragraphs, avoiding misjudgments that ignore local deviations based solely on the average level, thus making the cohesion level more discriminative and valuable for teaching reference.

[0070] Specifically, refer to Figure 2 As shown, the 240 labeled essays were first neatly stratified into four types and then divided into five equal parts (48 essays in each part, 12 essays in each of the four types), and five-fold cross-validation was performed: in each round, four parts were used as the training set and the remaining part as the test set. For each essay in the training set, the paragraph-to-text fit sequence was calculated based on the domain subject semantic relationship model, and then the mean and standard deviation of the fit of each essay were obtained as candidate criteria. When determining the threshold of a certain level, the initial lower limit (e.g., ≥0.1) was used as the starting point on the training set, and candidate thresholds were enumerated sequentially along the mean dimension with a fixed step size (0.1 in this embodiment); for each candidate threshold, combined with the corresponding standard deviation constraint, the training set samples were split according to the threshold and compared with the manually labeled levels, and the precision, recall and F1 score at that threshold were calculated. The curve of F1 changing with the threshold was recorded (i.e., Figure 2 (See the curve shown). The threshold combination corresponding to the maximum F1 score on this curve is taken as the optimal threshold for that level. The same strategy is used to determine the threshold range for the remaining levels. After determining the thresholds for the training set, the obtained threshold rules are applied to the corresponding test set for level division, and the macro-average and weighted average of the F1 scores for the five levels are calculated. This process is repeated until all five training / test rounds are completed. Finally, the arithmetic mean of the five macro-averages and weighted averages is used as the overall evaluation index of this method.

[0071] Understandably, by adopting the above-mentioned training-driven threshold search and five-fold cross-validation mechanism, on the one hand, a performance trade-off between precision and recall is achieved through a mesh-based (or approximately optimized) search with F1 as the objective function, thereby selecting the best-performing rank boundary on the training samples; on the other hand, cross-validation minimizes the randomness of a single partition, ensuring the robustness and generalization ability of the threshold under different sample partitions. Figure 2The F1-threshold curve in the model is not only used to select the optimal threshold point, but also to intuitively reflect the threshold range that is relatively sensitive to a certain level of judgment, providing a basis for subsequent confidence interval estimation, robustness analysis or online adjustment. In addition, this method is naturally adapted to the situation of imbalanced sample classes, because the determination of the threshold is based directly on classification performance (rather than a single statistic), which is conducive to obtaining a classification result that is closer to human scoring in teaching scenarios.

[0072] In this embodiment, the dynamic thresholding method includes: dividing a dataset of essays with multiple labeled aggregation levels into multiple subsets and performing five-fold cross-validation. For each level, the mean threshold and standard deviation threshold combination for that level are iteratively adjusted within a reasonable range with a preset step size, and the corresponding F1 score is calculated. The threshold combination with the largest F1 score is selected as the threshold range for that level. In each cross-validation, the training subset is used to determine the threshold range for each level, and the test subset is used to evaluate the model performance. By combining the results of multiple cross-validations, the final five-level classification standard for topic aggregation is determined, where level one indicates highly aggregated topics, and level five indicates severely dispersed topics.

[0073] Specifically, the implementation first involves stratified sampling of the labeled data to ensure that samples of each level are representatively distributed across each fold. During the training phase of each fold, several threshold combinations are evaluated in parallel iteratively on candidate intervals of mean and standard deviation using grid search or Bayesian optimization, and the classification boundary fitting rule calculated on the training fold is used as a temporary threshold. Subsequently, performance metrics such as precision, recall, and F1 score are calculated on the corresponding validation fold, with F1 score as the primary metric for selecting the optimal threshold combination. After all folds have been completed, robustness analysis is performed on the optimal thresholds for each fold (e.g., calculating the mean, confidence interval, or using majority voting) to form the final threshold set, and the confidence interval for each level is recorded for reference during subsequent online adjustments.

[0074] Understandably, the dynamic thresholding method, through cross-validation and performance-driven threshold search, makes the grading rules robust and interpretable under different sample divisions, thereby improving the generalization ability and reliability of grading determination in actual teaching corpora.

[0075] In this embodiment, the method further includes: generating the topic relevance of each paragraph; identifying off-topic paragraphs with a topic relevance below a preset threshold; and generating improvement suggestions for off-topic paragraphs, including keyword supplementation and semantic enhancement, based on the domain keyword analysis results.

[0076] Specifically, the system outputs a fit score for each paragraph, a list of the most contributing domain keywords, and the semantic distance or inconsistencies between these keywords and the main themes of the entire text. For paragraphs below the warning threshold, it automatically lists candidate keywords that are related to the main theme of the entire text but are missing or weakened in that paragraph, and provides feasible rewriting examples or supplementary sentence structures based on syntactic and semantic relationships (such as suggesting adding a topic indicator sentence at the beginning of the paragraph, replacing or adding certain entity words, or adjusting arguments to echo the main theme of the entire text). At the same time, the suggestions are marked with rewriting priorities and potential risk warnings (such as possible semantic conflicts or duplicate information) so that teachers and students can adopt them in a targeted manner.

[0077] Understandably, this step not only indicates "which paragraphs are off-topic," but also provides specific and actionable directions for improvement, thereby transforming the results of the automatic assessment into direct teaching intervention suggestions, which are convenient for students to revise and for teachers to grade.

[0078] In this embodiment of the application, the text of the essay to be evaluated is preprocessed to obtain the paragraph boundaries and part-of-speech tagging results of the essay to be evaluated, including: generating a structured representation of the text of the essay to be evaluated, wherein the structured representation includes at least paragraph boundaries, terms and the part-of-speech tagging information corresponding to the terms.

[0079] Specifically, the preprocessing step converts the original submitted text into a unified intermediate data structure. This structure records the paragraph distribution of the text, the original string of each term, its normalized form and its corresponding attribute labels (including part-of-speech labels, ordinal numbers in paragraphs and grammatical role indicators within sentences, etc.), and establishes an index relationship between each term and other terms in the full text for fast retrieval. This structure also supports the injection of additional meta-information required for subsequent processing (such as uncertainty markers, multiple candidate reference lists or manually corrected annotation fields).

[0080] Understandably, the structured representation generated by preprocessing serves as the foundational data layer for the entire evaluation process. It standardizes data exchange between processing modules and provides complete and traceable input for subsequent keyword extraction, referential reconstruction, and weight calculation, thereby ensuring the reproducibility of the evaluation process. The following describes, with several exemplary embodiments, specific application scenarios of the essay topic aggregation degree evaluation method based on domain subject provided in this application. Example 1: Domain Subject Extraction and Reference Restoration This embodiment details the extraction of domain terms and the restoration of pronoun references. First, candidate terms within a paragraph are prioritized by part of speech (personal pronouns → personal names → common nouns → organization names → noun abbreviations → place name abbreviations → other proper nouns → place names → product names → locative words → time words → quantifiers → adjectives). For each candidate, a representative score is calculated based on its occurrence rate within the paragraph / entire text, semantic relevance to the paragraph's semantic center, and syntactic role. Several terms with higher scores are selected as the paragraph's domain terms. For pronoun terms, reference restoration is performed by comparing the semantic matching degree of potential antecedents and their positional relationship within the same and adjacent paragraphs. When multiple candidates or ambiguities exist, multiple candidate markers can be retained for manual review. The following tables illustrate the extraction and restoration results of common noun and pronoun domain terms (Tables 1 and 2).

[0081]

[0082] Table 1 Extracts the domain subject of common nouns.

[0083]

[0084] Table 2 shows the domain terms extracted for referential reduction. Understandably, this embodiment demonstrates how to connect keywords within a paragraph with the overall context, so that the subject reflects both the local theme and the semantic flow of the entire text, thereby improving the accuracy of subsequent aggregation degree judgment.

[0085] Example 2: Topic Degree Calculation and Vectorized Representation This embodiment uses the appendix essay as an example to illustrate the calculation of keyword frequency, weight, and topicality, and demonstrates how to select full-text keyword terms and their vector representation at the paragraph level. The full-text frequency and topicality obtained from corpus training are calculated for the extracted domain words, and their weights are calculated accordingly, as shown in Table 3. By weighting the domain words, the words with the highest weights and topicality can be selected as full-text keyword terms (in this example, "morning glory" and "tree"). The topicality of these two words in each paragraph is further calculated to form paragraph vectors (e.g., the first paragraph vector is represented as "morning glory, 5.94; tree, 3.06"). Then, the similarity between the paragraph and full-text vectors is calculated (Table 4) to obtain paragraph fit.

[0086] Specifically, for target essays Each word in The calculation of its topic weight follows these steps: First, obtain the global topic weight value of the word obtained through training on a large-scale corpus. Next, statistics were compiled on the frequency of this word in essays. Number of times in And essays Total number of words And calculate its relative word frequency. Finally, the global topic score and local word frequency are expressed using the formula:

[0087] The formula integrates the words, where n is a preset adjustable parameter (n≥1). This formula uses the parameter n to flexibly control the influence of word frequency: when n is small, word frequency has a strong influence on the final weight, which is beneficial to highlighting high-frequency words in the essay; when n is large, the influence of word frequency is weakened, so that keywords that appear infrequently but have high overall topicality can also obtain significant weight.

[0088]

[0089] Table 3 Calculation of overall thematic degree

[0090] Table 4 shows the degree of alignment between the subject of each paragraph and the overall theme. Understandably, the above calculations can be used to identify paragraphs with high relevance and those with low relevance (such as paragraph 4) in the text, and suggestions for improvement or points for classroom discussion can be made based on this.

[0091] Example 3: Method for Determining and Determining the Degree of Aggregation This embodiment illustrates the implementation details of two grading strategies and the dynamic thresholding method. On one hand, the mean thresholding method defines a five-level standard by dividing the mean range of fit into several levels (examples are shown in Table 6). On the other hand, the dynamic thresholding method divides the labeled essay dataset into several subsets and iteratively adjusts the mean and standard deviation threshold combinations for each level using cross-validation, step-size search, or optimization algorithms. It selects the optimal threshold combination using a performance metric (such as F1) as the objective function and performs robustness analysis and confidence interval estimation on the cross-sectional results to determine the final threshold (examples are shown in Table 5). Tables 5 and 6 provide example threshold ranges for the two methods.

[0092]

[0093] Table 5. Dynamic Threshold Method for Defining Topic Aggregation Levels

[0094] Table 6. Classification of Topic Aggregation Levels Using the Mean Threshold Method Understandably, the dynamic thresholding method focuses more on classification performance on real labeled data and determines the boundary through cross-validation and performance-driven search, thereby improving the robustness and practicality of the classification; the mean thresholding method is simple and intuitive, and is suitable for preliminary or resource-constrained scenarios.

[0095] Example 4: Experimental Verification and Comparative Analysis This embodiment presents the experimental setup and comparative results: A dataset of 240 narrative essays (600-800 words each) from fourth and fifth grade elementary school students on a Chinese essay writing website was used as the benchmark, with manual annotation by five graduate students. The statistical differences in mean and standard deviation of fit between the semantic relation model based on domain-major words and the model based on non-domain-major words were compared (Table 7), as well as the performance comparison of the dynamic thresholding method and the mean thresholding method on the five-level classification task (Table 8). The experimental results show that the domain-major word model significantly outperforms the non-domain-major word model in mean fit (mean 0.51 vs 0.40, standard deviation 0.22 vs 0.40, t-test p-value close to 0), indicating that extracting the domain-major words of paragraphs significantly improves the effectiveness of topic aggregation determination. In the level classification task, the macro-average and weighted average F1 scores of the dynamic thresholding method are superior to those of the mean thresholding method, verifying the superiority of the dynamic thresholding method.

[0096]

[0097] Table 7. Experimental statistics of domain-major and non-domain-major models.

[0098] Table 8 compares the results of dynamic threshold method and mean threshold method in classifying topic aggregation degree levels. Understandably, this embodiment provides a data-driven basis for algorithm selection and threshold strategy through systematic experimental comparison, and demonstrates the feasibility and advantages of the domain subject method in elementary school narrative scenarios.

[0099] Example 5: Comparative Experiment with Large Model (ChatGPT-4) This embodiment constructs a zero-shot evaluation experiment to compare the evaluation of essay topic aggregation degree by GPT-4 with that of the domain name model. Instructions for GPT-4 were designed according to human scoring standards, and its accuracy, recall, and F1 score on a five-class classification task were statistically analyzed (Table 9). The results show that, at the macro-average and weighted average levels, the domain name model described in this embodiment outperforms GPT-4 on several key indicators, indicating that under this task and data setting, the corpus-driven and specially designed domain name method has advantages in judgment and stability; it also shows that the general large model still has room for improvement in determining aggregation degree at the text level and in a teaching-oriented manner.

[0100]

[0101] Table 9. Experimental results of ChatGPT-4 in five-class classification assessment. Understandably, this comparative experiment not only supports the effectiveness of this method, but also suggests that future exploration could explore ways to integrate domain subject signals with large model capabilities to further improve evaluation performance.

[0102] Please refer to the following: Figure 3 , Figure 3 This application provides an embodiment of an essay topic aggregation degree evaluation system 10 based on domain subject words, such as... Figure 3 The essay topic aggregation degree evaluation system 10 based on domain subject words shown includes at least the following parts: information acquisition module 11, domain subject word extraction module 12, vector space model construction module 13, and result output module 14.

[0103] In this embodiment, module 11, the information acquisition module, is used to acquire the essay to be evaluated. Please refer to the following for details. Figure 1 , 2 The details and their corresponding descriptions are not repeated here.

[0104] In this embodiment, the domain subject extraction module 12 is used to preprocess the text of the essay to be evaluated to obtain the paragraph boundaries and part-of-speech tagging results of the essay, and extract the domain subject from each paragraph according to the part-of-speech tagging results and the preset part-of-speech priority order. The domain subject includes pronoun domain subject and subject domain subject. Please refer to the following for details. Figure 1 , 2 The details and their corresponding descriptions are not repeated here.

[0105] In this embodiment, the vector space model construction module 13 is used to perform referential restoration of pronoun-type domain subjects in the sentence and surrounding paragraphs based on paragraph boundaries to determine the semantic content of the pronoun-type domain subjects. It then calculates weights based on the domain subject, its frequency in the text, and the word topic degree obtained through corpus training, thus constructing a vector space model for the domain subjects. In this vector space model, each paragraph is represented as a paragraph domain subject vector, and the essay to be evaluated is represented as the overall domain subject vector. Please refer to the following for details. Figure 1 , 2 The details and their corresponding descriptions are not repeated here.

[0106] In this embodiment, the result output module 14 is used to calculate the similarity between the main word vector of each paragraph and the overall main word vector of the essay to be evaluated, to obtain the topic fit of each paragraph with the essay to be evaluated, and to output the overall topic aggregation degree of the essay to be evaluated based on the topic fit of each paragraph. Please refer to the following for details. Figure 1 , 2 The details and their corresponding descriptions are not repeated here.

[0107] Figure 4 This is an electronic device 20 provided in one embodiment of this application. For example... Figure 4 As shown, the electronic device 20 includes at least the following components: a processor 21 and a memory 22.

[0108] In this embodiment, the memory 22 is used to store executable instructions of the processor 21, which, when configured to execute instructions, implement... Figure 1 The method for evaluating essay topic aggregation based on domain subject is shown.

[0109] In one embodiment of this application, the program operating in the electronic device 20 may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). The information processed by these devices is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (Flash ROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0110] It should be noted that a portion of the electronic device 20 described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0111] It should be noted that the term "computer" as used here refers to a computer built into electronic device 20, employing hardware including an operating system and peripheral devices. Furthermore, "computer-readable recording media" refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage devices such as hard drives built into the computer.

[0112] Furthermore, a "computer-readable recording medium" can include: a medium that dynamically stores a program for a short period of time, such as a communication line used when transmitting a program via a network such as the Internet or a communication line such as a telephone line; or a medium that stores a program for a fixed period of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining with programs already recorded in the computer.

[0113] Furthermore, the electronic device 20 in the above embodiments can also be implemented as an assembly (device group) composed of multiple devices. Each device constituting the device group can possess some or all of the functions or functional blocks of the electronic device 20 in the above embodiments. As a device group, it is sufficient to have all the functions or functional blocks of the electronic device 20.

[0114] It is understood that the essay topic aggregation degree evaluation method, system 10, and electronic device 20 based on domain subject words provided in this application extract domain subject words that can represent core semantic units from the essay text, restore their referential meaning to clarify semantic content, construct a semantic relationship model of domain subject words, and finally quantify the topic aggregation degree based on dynamic thresholds, thus realizing a deep semantic analysis of the essay content organization logic. This method effectively overcomes the problems of low efficiency and strong subjectivity in traditional manual review, and also makes up for the shortcomings of existing automatic grading technologies, which are mostly limited to grammatical error correction and have difficulty in evaluating content fit. In particular, considering the lack of morphological markers in the semantic language of Chinese, through the precise positioning and semantic mining of domain subject words, it achieves an objective and accurate measurement of the consistency between each paragraph of the essay and the overall theme from the content level. Compared with the unstable performance of general large language models on this task, the evaluation results of this solution are more reliable, and its dynamic threshold mechanism can adapt to the characteristics of different types of essays, significantly improving the accuracy of evaluation and teaching applicability, and providing an efficient and accurate intelligent auxiliary tool for writing teaching.

[0115] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for evaluating the aggregation degree of essay topics based on domain subject words, characterized in that, The method includes: Obtain the essays to be evaluated; The text of the essay to be evaluated is preprocessed to obtain the paragraph boundaries and part-of-speech tagging results of the essay to be evaluated; Based on the part-of-speech tagging results and the preset part-of-speech priority order, the main topic words are extracted from each paragraph. The main topic words include pronoun-type main topic words and subject-type main topic words. Based on the paragraph boundaries, the referential meaning of the pronoun domain subject is restored in the sentence and surrounding paragraphs to determine the semantic content of the pronoun domain subject. The vector space model of the domain term is constructed based on the domain term, the frequency of the domain term in the text, and the word topic degree obtained through corpus training. In the vector space model, each paragraph is represented as a paragraph domain main word vector, and the essay to be evaluated is represented as the overall domain main word vector; Calculate the similarity between the main word vector of each paragraph and the overall main word vector of the essay to be evaluated to obtain the thematic fit between each paragraph and the essay to be evaluated; The overall thematic convergence of the essay to be evaluated is output based on the thematic relevance of each paragraph.

2. The essay topic aggregation degree evaluation method based on domain subject as described in claim 1, characterized in that, The step of extracting the main topic word from each paragraph based on the part-of-speech tagging results and a preset part-of-speech priority order includes: The candidate words in the paragraph are sorted according to the preset word class priority and the part-of-speech tagging results to obtain a candidate word sequence; The candidate word sequence is screened according to the representativeness evaluation rules to obtain the domain main word; The representativeness evaluation rule includes evaluating based on at least one of the frequency of occurrence of the candidate word in the paragraph and the whole text, and the semantic relevance of the candidate word to other words in the paragraph.

3. The essay topic aggregation degree evaluation method based on domain subject as described in claim 2, characterized in that, The method further includes: Obtain the topic degree of the words in the domain of discourse by training on a large-scale classification corpus; Obtain the frequency of the domain term in the essay to be evaluated; Obtain the positional importance factor of the domain subject within the paragraph; The vector space model of the domain subject is constructed based on the topic degree, frequency of occurrence, and positional importance factor of the domain subject.

4. The essay topic aggregation degree evaluation method based on domain subject as described in claim 1, characterized in that, The method further includes: Obtain the statistical distribution characteristics of the thematic fit between each paragraph and the essay to be evaluated; The overall thematic aggregation level of the essay to be evaluated is determined based on the statistical distribution characteristics.

5. The essay topic aggregation degree evaluation method based on domain subject as described in claim 4, characterized in that, The method further includes: Calculate the mean and standard deviation of the thematic relevance of each paragraph; The mean and standard deviation are compared with the grade threshold range determined by the dynamic threshold method; The overall thematic cohesion level of the essay to be evaluated is output based on the comparison results.

6. The essay topic aggregation degree evaluation method based on domain subject as described in claim 5, characterized in that, The dynamic threshold method includes: The dataset containing multiple essays with labeled aggregation levels was divided into several subsets and subjected to five-fold cross-validation. For each level, the combination of mean threshold and standard deviation threshold is iteratively adjusted within a reasonable range with a preset step size, and the corresponding F1 value is calculated. The threshold combination with the largest F1 value is selected as the threshold range for that level. In each cross-validation, the training subset is used to determine the threshold range for each level, and the test subset is used to evaluate the model performance. The final topic aggregation degree was determined by combining the results of multiple cross-validations.

7. The essay topic aggregation degree evaluation method based on domain subject as described in claim 4, characterized in that, The method further includes: Generate the topic relevance of each paragraph; Identify off-topic paragraphs whose topic relevance is below a preset threshold; Based on the analysis results of the domain subject, improvement suggestions including keyword supplementation and semantic enhancement are generated for the off-topic paragraph.

8. The method for evaluating essay topic aggregation based on domain subject as described in claim 1, characterized in that, The preprocessing of the text of the essay to be evaluated to obtain the paragraph boundaries and part-of-speech tagging results of the essay includes: Generate a structured representation of the text of the essay to be evaluated, wherein the structured representation includes at least the paragraph boundaries, terms, and the part-of-speech tags corresponding to the terms.

9. A system for evaluating the aggregation degree of essay topics based on domain subject words, applied to implement the essay topic aggregation degree evaluation method based on domain subject words as described in any one of claims 1 to 8, characterized in that, The system includes: The information acquisition module is used to acquire the essays to be evaluated; The domain subject extraction module is used to preprocess the text of the essay to be evaluated to obtain the paragraph boundaries and part-of-speech tagging results of the essay to be evaluated, and extract the domain subject from each paragraph according to the part-of-speech tagging results and the preset part-of-speech priority order. The domain subject includes pronoun domain subject and subject domain subject. The vector space model construction module is used to perform referential restoration of the pronoun-type domain subject in the sentence and surrounding paragraphs based on the paragraph boundaries to determine the semantic content of the pronoun-type domain subject. It then calculates weights based on the domain subject, its frequency in the text, and the word topic degree obtained through corpus training, and constructs a vector space model for the domain subject. In this model, each paragraph is represented as a paragraph domain subject vector, and the essay to be evaluated is represented as the overall domain subject vector. The result output module is used to calculate the similarity between the main word vector of each paragraph and the overall main word vector of the essay to be evaluated, to obtain the theme fit of each paragraph and the essay to be evaluated, and to output the overall theme aggregation degree of the essay to be evaluated based on the theme fit of each paragraph.

10. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to implement the essay topic aggregation degree evaluation method based on domain subject as described in any one of claims 1 to 8 when executing the instructions.

Images