A Method for Evaluating the Quality of Mathematical Modeling Competition Papers Based on Multi-Feature Fusion

Abstract

This invention discloses a method for evaluating the quality of mathematical modeling competition papers based on multi-feature fusion, relating to the field of natural language processing. The method includes: performing full-element analysis and structured segmentation on the paper to obtain the abstract, main text, appendix, and competition question text; calculating the necessary chapter coverage and supporting material completeness based on the main text and appendix to obtain structural standardization feature vectors; implementing a "competition task decomposition - problem-solving evidence extraction - coverage verification" thought chain process using a large language model based on the competition question text and main text to obtain competition task alignment feature vectors; performing multi-granular semantic interaction calculations using deep retrieval, cross-encoders, and natural language inference techniques based on the abstract and main text to obtain semantic consistency feature vectors; and fusing multi-dimensional features through a multi-layer perceptron model to output the award probability. This invention comprehensively simulates expert review decisions, effectively filling the gap in evaluation technology in this field and possessing significant application value.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing, data mining, and intelligent education assessment, specifically to a method for automated initial screening and award potential assessment of mathematical modeling competition papers based on multi-feature fusion, utilizing large language models and deep learning technology. Background Technology

[0002] With the modernization of educational assessment systems, automated essay scoring technology has emerged. Automated essay scoring (AES) refers to a technology that assesses the quality of student essays by automatically identifying technical errors and understanding elements such as narration, coherence, and other aspects. AES can effectively reduce the burden of essay scoring for teachers, while also providing students with objective and timely feedback. It is an important application of Natural Language Processing (NLP) in the field of education.

[0003] With the deepening application of artificial intelligence technology in educational assessment and the improvement of complex document processing capabilities, Automated Academic Paper Rating (AAPR) has become a key technology to assist peer review and reduce the burden on reviewers. Since 2018, research in this field has mainly focused on predicting the probability of paper acceptance. Existing technologies have recognized the unique structure of academic papers and proposed a variety of deep learning-based methods: for example, early researchers proposed a modular structure based on hierarchical convolutional neural networks to capture hierarchical representations of different modules such as titles and abstracts; combined hierarchical attention networks and structural labels to attempt to preserve important structural information in the text; and integrated semantic and structural information from knowledge graphs to enhance text representation efficiency.

[0004] However, despite the significant progress made by the aforementioned deep learning methods in general academic evaluation, in the specific field of mathematical modeling competition paper review, mainstream technologies remain at the stage of traditional multi-attribute decision models such as AHP and TOPSIS. These methods only focus on weighted calculations and ranking of predetermined numerical indicators, rely on manually pre-set scoring matrices, lack the ability to automatically analyze the semantics of long texts, struggle to verify the logical closure between the paper's answers and the competition questions, and are unable to identify the illusion that the abstract content deviates from the facts of the main text. Therefore, there is an urgent need for an automated review method that can comprehensively simulate the expert review decision-making process, integrating structural standardization, competition task alignment, and semantic consistency features. Summary of the Invention

[0005] Objective: The purpose of this invention is to construct a multi-feature evaluation architecture that comprehensively simulates the expert review and decision-making process. This architecture specifically encompasses rule-based structural normative features for applying to mathematical modeling competition papers, task alignment features using a large language model to achieve deep logical alignment between the competition questions and the paper's answers, and factual and semantic consistency features based on deep retrieval and inference techniques for the abstract and main text. Based on this architecture, this invention maps these three heterogeneous features to a unified feature vector space, performs nonlinear feature fusion using a multilayer perceptron model, and finally outputs binary classification probability values ​​using a sigmoid activation function, thereby achieving automated initial screening of whether a paper has award-winning potential.

[0006] Technical Solution: To achieve the above objectives, this invention provides a method for evaluating the quality of mathematical modeling competition papers based on multi-feature fusion, comprising the following steps:

[0007] S1. Perform full-element analysis and structured segmentation on mathematical modeling competition papers to obtain a standardized dataset containing abstracts, main text, appendices, and competition questions.

[0008] S2. Based on the main text and appendices in the standardized dataset, calculate the necessary chapter coverage and supporting material integrity to obtain the structural normative feature vector;

[0009] S3. Based on the competition question text and body text in the standardized dataset, use a large language model to decompose the competition question task, extract the problem-solving evidence, and verify the coverage to obtain the competition question task alignment feature vector.

[0010] S4. Based on the summary and text in the standardized dataset, multi-granularity semantic interaction calculation is performed using deep retrieval, cross encoder and natural language reasoning techniques to obtain semantic consistency feature vectors.

[0011] S5. Based on the structural normative feature vector, the competition task alignment feature vector, and the semantic consistency feature vector, feature fusion and prediction are performed using a multilayer perceptron model to obtain the paper's award probability, thereby achieving automated initial screening of papers.

[0012] Further, in step S1, the full-element analysis and structured segmentation of the mathematical modeling competition paper specifically includes: converting the original PDF document into Markdown format using the MinerU tool; employing a regular expression positioning strategy to extract the abstract area by identifying the "Abstract" and "Keywords" heading lines; extracting the main text area by identifying the heading lines from "Problem Restatement" to "References," while preserving the chapter hierarchy; and extracting the appendix area by identifying the "Appendix" or "Appendix" heading lines. Simultaneously, the text content in the competition document is extracted. This results in a standardized dataset containing the abstract, main text, appendix, and competition text.

[0013] Furthermore, in step S2, the specific calculation method for obtaining the structural normative feature vector based on the main text and appendix is ​​as follows:

[0014] Define the set of necessary chapter keywords This includes "problem restatement", "symbol explanation", "model assumptions", "model building", "model solution", "results analysis", and "model verification".

[0015] Calculate the necessary chapter coverage: Traverse the main text title lines and count the included sets. The number of unique keywords in Chinese The calculation formula is: ;

[0016] Calculate the completeness of supporting materials: Check if an appendix title exists at the end of the document; if it does, then... ,otherwise ;

[0017] By concatenating the coverage rate of the necessary chapters with the integrity index of the supporting materials, a structural normativity feature vector is obtained. .

[0018] Furthermore, in step S3, obtaining the task alignment feature vector based on the competition question text and the main body includes question decomposition, evidence extraction, and coverage calculation. The specific calculation method is as follows:

[0019] (1) Task breakdown: Extracting from the competition text using a large language model Each independent subproblem is solved, forming a set. ;

[0020] (2) Extraction of evidence for problem-solving: The main text is divided into a sequence of text blocks. A sliding window strategy is adopted, and a large language model is used to determine text blocks. Does it include sub-problems? The solution aggregates all local evidence and rewrites it to generate a summary of the evidence. ;

[0021] (3) Coverage determination: The evaluation pair Input a large language model and output a satisfaction confidence score based on the dimensions of existence, relevance, and completeness. ;

[0022] (4) Calculation of overall coverage: The arithmetic mean of the scores of all sub-problems is calculated as the feature index, and the formula is as follows:

[0023] ;

[0024] Finally, the alignment feature vector of the competition task is obtained. .

[0025] Furthermore, in step S4, the semantic consistency feature vector obtained from the abstract and the main text is used to verify whether the abstract is supported by the main text. The specific calculation steps and formulas are as follows:

[0026] (1) The summaries in the standardized dataset are segmented into sentence sets. The text in the standardized dataset is segmented into overlapping fragment sets using a sliding window mechanism. Using Chinese word segmentation tools to analyze the set Perform word segmentation and construct a BM25 inverted index. For each summary sentence... After word segmentation, calculate its relationship with the fragment. The BM25 score is calculated using the following formula:

[0027] ;

[0028] in, For word segmentation terms in the summary sentence, The frequency of a term in a segment. The length of the segment. The average length of all segments, This is the word frequency saturation parameter, used to adjust the impact of word frequency on the score. This is a length normalization parameter used to eliminate document length deviations;

[0029] Inverse document frequency (IVF) is used to measure the rarity of terms. Its specific calculation formula is as follows:

[0030] ;

[0031] in, A collection of text fragments The total number of segments in For included terms The number of fragments, a constant Used for smoothing to prevent the denominator from being zero or producing negative values;

[0032] Based on BM25 scores The top-N segments are recalled to obtain a set of candidate text segments. This step quickly narrows down the search scope for subsequent deep semantic computation.

[0033] (2) Calculate the summary sentence using a cross encoder With candidate fragments The semantic relevance is considered, and the maximum value is selected as the re-ranking score. And obtain the best evidence fragments. :

[0034] ;

[0035] (3) Calculate using natural language reasoning models Contains The probability is used to obtain the logical implication score:

[0036] ;

[0037] (4) Calculate using the BERTScore algorithm and Contextual semantic similarity Its calculation is based on the accuracy of maximum cosine similarity. and recall rate F1 value:

[0038] ;

[0039] ;

[0040] ;

[0041] in, and These are the word vector sequences for the summary sentence and the best evidence fragment, respectively. and Sequences and Length, The F1 score, which better represents the balanced evaluation, is selected as the final context similarity score. ;

[0042] (5) The average of the above three indicators for all summary sentences is taken to obtain the three-dimensional semantic consistency feature vector:

[0043] .

[0044] Furthermore, in step S5, the specific steps and formulas for fusing multi-dimensional features with the initial screening prediction are as follows:

[0045] (1) Feature concatenation and standardization: The feature vectors are concatenated to obtain And obtain Z-Score standardization. The standardized formula is:

[0046] ,

[0047] in, and For the training set The mean and standard deviation of the dimensional features;

[0048] (2) Model prediction: Construct a multilayer perceptron model and input... After processing by the hidden layer, the probability of winning is calculated by the Sigmoid function of the output layer. :

[0049] ,

[0050] in, This is the output vector of the hidden layer. and These are model parameters;

[0051] (3) Threshold determination: Set a classification threshold ,like Greater than the preset threshold If it is, it will be judged as "recommended for nomination / award", otherwise it will be recorded as "not awarded".

[0052] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0053] (1) This invention breaks through the limitations of traditional single text features. By integrating features of three heterogeneous dimensions—structural normativity, task alignment, and semantic consistency—it more comprehensively simulates the cognitive process of expert review.

[0054] (2) In the structural standardization stage, by quantifying the coverage of necessary chapters and the completeness of supporting materials, papers with missing structures or that do not meet the competition's hard requirements can be quickly identified. This not only makes up for the deficiency of pure semantic models in perceiving structural compliance, but also effectively filters low-quality samples through low-cost rule calculation, thereby improving the overall initial screening efficiency;

[0055] (3) In the task alignment stage of the competition, a strategy of “block extraction-global aggregation” was proposed, which effectively solved the context window limitation and “intermediate forgetting” problem in the processing of long scientific papers by large language models, and ensured the integrity of the solution evidence chain;

[0056] (4) A semantic consistency verification mechanism based on deep retrieval and reasoning was introduced. Through multi-granularity calculation of cross encoder and NLI model, it can accurately identify the "illusion" phenomenon that the summary content deviates from the facts of the main text, which significantly improves the reliability of the initial screening results. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0058] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0059] like Figure 1 As shown, this embodiment of the invention provides a method for evaluating the quality of mathematical modeling competition papers based on multi-feature fusion. This method aims to provide a quantitative assessment of the paper's award potential by analyzing multiple dimensions. The specific implementation steps are as follows.

[0060] Step S1: Perform full-element analysis and structured segmentation on the mathematical modeling competition papers to obtain a standardized dataset;

[0061] This step is fundamental to subsequent feature extraction. Because mathematical modeling competition papers typically contain numerous formulas, diagrams, and complex hierarchical structures, directly reading plain text would result in the loss of crucial structural information.

[0062] S11. Data Format Conversion: Obtain the original PDF document of the mathematical modeling competition paper to be evaluated; to preserve the logical structure of the document, use the MinerU document parsing tool to convert the PDF document to Markdown format. Markdown format can clearly mark heading levels by the number of "#" symbols, which is crucial for subsequent chapter identification;

[0063] S12, Full Element Parsing: Employs a text positioning strategy based on regular expressions;

[0064] (1) Extract the abstract: Match the heading lines containing the words "Abstract" and extract the text below them up to the first level heading;

[0065] (2) Extract the main text: Match all content from "Restatement of the Problem" to "References", retaining the heading structure at all levels for subsequent analysis of chapter coverage;

[0066] (3) Extract appendix: Match the words "Appendix" at the end of the document and extract the code or supplementary materials that follow it;

[0067] At the same time, extract the text content from the competition document;

[0068] Through the above processing, a standardized dataset containing the abstract, main text, appendix, and competition questions is obtained.

[0069] Step S2: Based on the standardized dataset obtained in Step S1, calculate and obtain the structural normalization feature vector;

[0070] This step is used to quickly screen papers that do not conform to the basic mathematical modeling competition paper structure specifications;

[0071] S21. Calculate the coverage of necessary chapters: Mathematical modeling papers have strict format requirements. A set of keywords for necessary chapters is pre-defined. In this embodiment The values ​​can be {"Problem Restatement", "Symbol Explanation", "Model Assumptions", "Model Building", "Model Solution", "Results Analysis", "Model Validation"}, that is... Iterate through the text titles parsed in step S1 and count the sets contained in the text. The number of unique keywords in Chinese For example, if the text only contains "model building" and "results analysis," then... The formula for calculating coverage is: ;

[0072] S22. Quantitative Supporting Material Completeness: Check whether the paper provides supporting code or data. Detect the presence of an appendix area in the document parsing results. If an "Appendix" or "Appendix" heading is detected, then... ,otherwise ;

[0073] S23. Feature Vector Construction: Concatenate the two indicators mentioned above to obtain the structural normalization feature vector. .

[0074] Step S3: Based on the standardized dataset obtained in Step S1, construct the task-aligned feature vector;

[0075] The Large Language Model (LLM) is used to implement the thought process of "task decomposition - problem-solving evidence extraction - coverage verification" to quantify the completeness of the paper's problem-solving logic and determine whether the paper truly answers each sub-problem in the competition.

[0076] S31. Problem Breakdown: Input the problem text from the standardized dataset into the LLM model. Utilize the model's engineering requirements prompt: "Ignore the background description of the problem and break it down into..." Each subproblem is a specific, independent subproblem to be solved. The output set of subproblems is then generated. ;

[0077] S32, Problem-solving evidence extraction: Since the main body of the paper is usually longer than the context window of the LLM, and long texts are prone to cause the model to "forget", this embodiment adopts a sliding window strategy;

[0078] First, the text in the standardized dataset is segmented into sequences of text blocks. 20% overlap is maintained between adjacent text blocks;

[0079] Secondly, for each sub-problem , and the sequence Each text block in Combined input LLM, the instruction is: "Determine if a text block contains a subproblem". The model will determine the solution evidence; if present, extract it; otherwise, return empty. Does it contain anything targeting The solution provides partial information; if it is contained, the relevant statements are extracted as "partial evidence"; otherwise, an empty value is returned.

[0080] Finally, a global aggregation is performed, concatenating all non-empty "local evidence" strictly according to their order of appearance in the original text. This is then input into the LLM library again for deduplication and summary rewriting, generating a final, refined summary of solution evidence for this subproblem. If no valid evidence is extracted after traversing all text blocks, then... Marked as "no solution";

[0081] S33. Coverage Determination and Scoring: Based on the extracted evidence summary, perform coverage determination steps using multi-dimensional evaluation criteria. (Sub-problems) Summary of Evidence for Solving the Problem Composition Evaluation Input a large language model. To prevent subjective arbitrariness in model scoring and ensure the robustness of the evaluation results, specific scoring criteria constraints are incorporated into the prompting project, requiring the model to be comprehensively evaluated from the following three dimensions:

[0082] First, existence, that is, judgment. Does it provide a specific algorithm flow and quantified numerical results?

[0083] Second is relevance, i.e., judgment. Does the provided result directly respond to the sub-problem? The core variables or key indicators in;

[0084] Third is completeness, which means judging whether the logical chain of the solution is complete, such as identifying whether there is an incomplete situation of "only assumptions without solutions";

[0085] Based on a comprehensive analysis of the above three dimensions, LLM ultimately outputs a summary of evidence. Pair problem Satisfaction confidence score This serves as the basis for calculating the overall problem-solving coverage.

[0086] S34. Overall Coverage Calculation: Calculate the arithmetic mean of the scores for all sub-problems.

[0087] ;

[0088] Finally, the alignment feature vector of the competition task is obtained. .

[0089] Step S4: Construct semantic consistency feature vectors based on the standardized dataset obtained in Step S1;

[0090] This step aims to combine the scores of three dimensions to deeply verify the factual and semantic consistency between the abstract and the main text, and to prevent "abstract illusion," that is, the abstract is well written, but the corresponding work is not actually done in the main text.

[0091] S41. Preprocessing and Initial Recall: Segment the summary text in the standardized dataset into sentence sets. Simultaneously, the text in the standardized dataset is segmented into a set of fragments with a certain overlap rate using a sliding window mechanism. Using Chinese word segmentation tools to analyze the set Perform word segmentation and construct a BM25 inverted index. For each summary sentence... After word segmentation, calculate its relationship with the fragment. The BM25 score is calculated using the following formula:

[0092] ,

[0093] in, For word segmentation terms in the summary sentence, The frequency of a term in a segment. The length of the segment. The average length of all segments, This is the word frequency saturation parameter, used to adjust the impact of word frequency on the score. This is a length normalization parameter used to eliminate document length deviations;

[0094] Inverse document frequency (IVF) is used to measure the rarity of terms. Its specific calculation formula is as follows:

[0095] ,

[0096] in, A collection of text fragments The total number of segments in For included terms The number of fragments, a constant Used for smoothing to prevent the denominator from being zero or producing negative values;

[0097] Based on BM25 scores The top-N segments are recalled to obtain a set of candidate text segments. This step quickly narrows down the search scope for subsequent deep semantic computation.

[0098] S42. Reordering and Best Evidence Matching: To filter out genuine semantic evidence from a coarse candidate set, a reordering operation is performed using a pre-trained cross-encoder BGE model. The summary sentence... With sets Each candidate fragment The input sequences are concatenated and fed into the model for deep interaction attention computation, outputting a scalar score of semantic relevance between the two. The maximum value is selected as the reordered relevance score. And match the fragment corresponding to the maximum value as a summary sentence. Best evidence fragment The calculation formula is as follows:

[0099] ;

[0100] S43. Logical Implication Reasoning: Utilizing a BERT model fine-tuned on a large-scale natural language reasoning dataset, specifically the StructBERT model, to perform logical consistency checks. The best evidence fragment is selected. As a premise, the summary sentence Assuming that the input model performs a three-class classification prediction of "entailment / neutral / contradictory", the Softmax probability value of the "entailment" category is extracted as the logical entailment score. This quantifies whether the abstract's content is supported by evidence in the main text. The calculation formula is as follows:

[0101] ;

[0102] S44. Contextual Semantic Similarity Calculation: The BERT Score algorithm architecture is used to calculate the semantic overlap between the summary sentence and the evidence fragment in the deep feature space. Specifically, a pre-trained BERT model is first used to calculate the semantic similarity between the summary sentence and the evidence fragment. and best evidence fragments Encode them and extract their respective context embedding vector sequences. Let... The embedding vector sequence is , The embedding vector sequence is Next, the cosine similarity matrix between the two sets of vector sequences is calculated, and the recall rate is calculated based on a greedy matching strategy. Accuracy and F1 value The F1 score, which better represents the balanced assessment, is selected as the final context similarity score. The calculation formula is as follows:

[0103] ;

[0104] ;

[0105] ;

[0106] S45. Index Aggregation: The average scores of all summary sentences across the three dimensions are used to obtain a three-dimensional semantic consistency feature vector.

[0107] .

[0108] Step S5: Based on the feature vectors obtained in steps S2, S3, and S4, perform feature fusion and binary classification preliminary screening prediction;

[0109] S51, Feature Concatenation and Standardization: [This section appears to be incomplete and requires further context.] Concatenate them into a comprehensive feature vector Because different dimensions have different units of measurement, such as coverage being 0-1, the Cross-Encoder score might be a logarithmic value. Therefore, the Z-Score method is used for standardization to obtain the input vector. Among them, the first 3D features The standardized formula is:

[0110] ,

[0111] in and The mean and standard deviation are obtained based on statistical analysis of historical training data.

[0112] S52, Model Prediction: Construct a multilayer perceptron model, including an input layer, hidden layers, and an output layer. Introduce a Dropout layer in the hidden layer to prevent overfitting. Standardize the vectors... The input model, after undergoing linear transformation and activation function processing, uses the Sigmoid activation function in the output layer to map the output values ​​to probabilities between 0 and 1. This indicates the likelihood of the paper winning an award.

[0113] ,

[0114] in For hidden layer output, and For output layer weights and biases;

[0115] S53, Threshold Determination: Set the classification threshold. ,like If the paper is selected, it will be marked as "suggested for inclusion / award"; otherwise, it will be marked as "not awarded". This completes the initial screening of the competition papers.

[0116] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A method for evaluating the quality of mathematical modeling competition papers based on multi-feature fusion, characterized in that, Includes the following steps: S1. Perform full-element analysis and structured segmentation on mathematical modeling competition papers to obtain a standardized dataset containing abstracts, main text, appendices, and competition questions. S2. Based on the main text and appendices in the standardized dataset, calculate the necessary chapter coverage and supporting material integrity to obtain the structural normative feature vector; S3. Based on the competition question text and main body in the standardized dataset, implement the thought process of "competition question task decomposition - problem-solving evidence extraction - coverage verification" using a large language model to obtain the competition question task alignment feature vector, specifically including the following steps: S31. Task Breakdown: Input the competition text into a large language model and extract... Each specific, independently solvable subproblem forms a set. ; S32. Problem-solving evidence extraction: Using a sliding window strategy, the main text is segmented into a sequence of text blocks with overlapping ratios. For sets Each subproblem in , and the sequence Each text block in The input large language model is combined to extract local evidence; then, a global aggregation is performed to concatenate and rewrite all local evidence in the original text order, generating a summary of problem-solving evidence. ; S33. Coverage Judgment and Scoring: Sub-problems Summary of Evidence for Solving the Problem The evaluation component scores the input large language model based on three dimensions: existence, relevance, and completeness, and outputs a satisfaction confidence score. ; It also includes calculating the full-text solution coverage index based on the satisfaction confidence score, as shown in the following formula: ， in, For the number of subproblems, For the first Satisfaction confidence scores for each sub-problem ; Finally, the alignment feature vectors for the competition task are constructed. ; S4. Based on the summary and main text in the standardized dataset, perform multi-granularity semantic interaction calculation using deep retrieval, cross-encoder, and natural language inference techniques to obtain semantic consistency feature vectors. This includes the following steps: S41. Preprocessing and Initial Recall: The summary is segmented into a set of sentences. The main text is divided into a set of fragments. And construct a BM25 inverted index; for each summary sentence Using the keyword retrieval algorithm BM25, from the fragment set Recall the top-N candidate text fragments ; S42. Reordering and Best Evidence Matching: Calculate the summary sentence using a cross-encoder model. With candidate fragments The semantic relevance score is used to select the maximum value as the reordered relevance score. And the corresponding fragment is recorded as the best evidence fragment. The calculation formula is as follows: ； S43. Logical Implication Reasoning: Utilizing natural language reasoning models, with the best evidence fragments stated. Premise, summary sentence Assuming this, the probability value of the implied class is calculated as the logical implication score. The formula is as follows: ； S44. Contextual semantic similarity calculation: Using the BERTScore algorithm architecture, extract the summary sentences respectively. Embedded vector sequence and best evidence fragments Embedded vector sequence Calculate the contextual semantic similarity score between the two. The formula is as follows: ， ， ， in, and Sequences and Length, For recall rate, For accuracy, The F1 score, which better represents the balanced evaluation, is selected as the final context similarity score. ; S45. Index Aggregation: Based on the reordering relevance score, the logical implication score, and the contextual semantic similarity score, the average of the above three scores for all summary sentences is taken to obtain a three-dimensional semantic consistency feature vector. ; S5. Based on the structural normative feature vector, the competition task alignment feature vector, and the semantic consistency feature vector, feature fusion and prediction are performed using a multilayer perceptron model to obtain the paper's award probability, thereby achieving automated initial screening of papers.

2. The method for evaluating the quality of mathematical modeling competition papers based on multi-feature fusion as described in claim 1, characterized in that, In step S1, the mathematical modeling competition papers undergo full-element analysis and structured segmentation to obtain a standardized dataset containing abstracts, main text, appendices, and competition problem texts. This includes the following steps: S11. Obtain the original documents of the mathematical modeling competition papers to be evaluated and the corresponding competition questions, and use a document parsing tool to convert the original documents into Markdown format; S12. Using a regular expression positioning strategy, extract the abstract text area by identifying the "Abstract" and "Keywords" heading lines; extract the main text content area by identifying the heading lines from "Restatement of the Problem" to "References", while preserving the structure of chapter headings at all levels; extract the appendix area by identifying the "Appendix" or "Appendix" heading lines; at the same time, extract the text content from the competition document; and obtain a standardized dataset containing the abstract, main text, appendix, and competition text.

3. The method for evaluating the quality of mathematical modeling competition papers based on multi-feature fusion as described in claim 1, characterized in that, In step S2, based on the main text and appendices in the standardized dataset, the necessary chapter coverage and supporting material integrity are calculated to obtain the structural normative feature vector, including the following steps: S21. A pre-defined set containing keywords from essential chapters of standard mathematical modeling competition papers. This includes sections such as "Problem Restatement," "Symbol Explanation," "Model Assumptions," "Model Building," "Model Solving," "Results Analysis," and "Model Validation." It iterates through all header lines in the main text and counts the included sets. The number of unique keywords in Chinese The necessary chapter coverage is calculated using the following formula: ， in, For set The total number of elements; S22. Check if there is a title line containing the words "Appendix" at the end of the document. If so, mark it as part of the completeness of the supporting materials. Otherwise remember ; S23. Concatenate the necessary chapter coverage rate with the supporting material integrity index to construct a structural normative feature vector. .

4. The method for evaluating the quality of mathematical modeling competition papers based on multi-feature fusion according to claim 1, characterized in that, In step S5, the structural normativity feature vector, the competition task alignment feature vector, and the semantic consistency feature vector are fused to obtain the paper's award probability, including the following steps: S51. Feature Concatenation and Standardization: Concatenate the structural normalization feature vector, the competition task alignment feature vector, and the semantic consistency feature vector to obtain the original feature vector. The Z-Score method was used to evaluate the performance of the target audience. The input vector is obtained by standardization. ; S52, Model Prediction: Construct a multilayer perceptron model and input vector... After linear transformation and activation function processing, the winning probability is calculated using the Sigmoid activation function in the output layer. The formula is as follows: ， in, For hidden layer output, and For output layer weights and biases; S53, Threshold Determination: Set the classification threshold. If the winning probability If so, the paper is deemed to have the potential to win an award.

Images