Cross-platform mathematical application question automatic solving system based on large model and subject knowledge enhancement

Abstract

The invention relates to the field of education, and discloses a cross-platform mathematical application question automatic solving system based on a large model and subject knowledge enhancement, and the system comprises the steps: firstly constructing a subject knowledge recognition model, and achieving the joint recognition of knowledge points and situation knowledge of mathematical application questions; proposing a prompt design framework, and designing multiple types of prompts to enable the large language model to meet the needs of subject knowledge enhancement; then, a server is built, a Qwen2.5-72B-Math and GOT-OCR2.0 large model is deployed and operated, and automatic problem solving and character recognition services are provided; finally, a subject knowledge recognition model is deployed, integration is completed, and construction of a cross-platform mathematical application question automatic solving system based on a large model and subject knowledge enhancement is completed. Through the integrated photographing recognition question solving function and the subject knowledge recognition function, the mathematical application question problem is solved for a learner, through recognition of knowledge points and situational knowledge, the mathematical application question solving efficiency is improved, and the mathematical application question solving efficiency is improved. Therefore, the learner can clearly know the knowledge points and the situation knowledge of the mathematical application question, and the learner can be promoted to deeply understand the subject knowledge and deeply learn the subject knowledge.

Description

Technical Field

[0001] The present invention relates to the field of education, and in particular to a cross-platform automatic solution system for mathematical word problems based on a large model and enhanced subject knowledge. Background Art

[0002] When solving mathematical word problems, big models and subject knowledge enhancement can effectively promote mathematical subject learning. The big model significantly improves the ability to understand and generate mathematical language through deep learning and training with a large number of mathematical text data sets, thereby realizing the automatic solution of mathematical word problems. Subject knowledge enhancement uses graph convolutional neural networks to generate corresponding knowledge points and contextual knowledge labels for corresponding questions. In mathematics learning, knowledge points and contextual knowledge labels are crucial to improving the ability to solve mathematical word problems. By extracting knowledge points, learners can deepen their understanding of knowledge, quickly classify problem types and required mathematical concepts, and thus choose appropriate problem-solving strategies. By generating contextual knowledge labels, learners can determine key elements based on contextual knowledge and propose problem solutions. This system drives the big language model to create by designing specific prompts, producing a large number of similar question types, so that learners can strengthen the practice of similar question types. It can be concluded that the automatic solution system for mathematical word problems based on big models and subject knowledge enhancement can promote mathematical subject learning.

[0003] In the field of education, knowledge points are the basis for understanding more complex concepts and solving related problems. Learners need to relate knowledge points to each other in order to build a complete knowledge structure of a subject. The integration of contextual knowledge can significantly enhance the depth and breadth of problem description, thereby promoting students to closely link mathematical knowledge points with real-life situations. This integration helps students not only stay at the level of abstract symbol operations when solving mathematical problems, but also apply mathematical knowledge to the process of solving practical problems.

[0004] For example, the invention disclosed in the publication (announcement) number: CN112580653A discloses a method and device for automatically solving problems. This patent belongs to the field of education and discloses a method and device for automatically solving problems. By pre-setting a trained convolutional neural network classifier, a character sequence that can be recognized, analyzed, and calculated by a computer is obtained, and the problem is automatically solved through a question parsing tree of preset different models.

[0005] Another example is the invention with the publication (announcement) number: CN118657140B, which discloses a method and system for labeling test knowledge points based on artificial intelligence. This patent belongs to the technical field of test knowledge point labeling. The patent content is as follows: A method and system for labeling test knowledge points based on artificial intelligence is disclosed, and the method includes: collecting test data to be labeled, the test data includes a question text and multiple answer texts; performing knowledge point contribution analysis on multiple answer texts in turn to obtain multiple knowledge point contribution scores; based on the question text, integrating the multiple knowledge point contribution scores corresponding to each answer text to obtain multi-solution knowledge point labeling data.

[0006] However, the above invention has some defects:

[0007] 1. Large models are rarely used in specific subject education fields, especially in science subjects such as mathematics, physics, and chemistry. Since large models face customization and adaptation challenges when applied in specific fields, they are relatively rarely used in vertical fields. These models often need to be deeply optimized according to the characteristics of specific fields, which requires a lot of domain knowledge and resource investment. This limitation has led to a slow application and promotion of large models in vertical fields. There are few existing systems applied to the field of mathematics education, which makes it difficult to meet the diverse needs of the field of mathematics education for intelligent solutions;

[0008] 2. Existing Chinese question-answering systems based on large models are difficult to effectively support problem solving in the field of mathematics. At present, there are a large number of Chinese question-answering systems developed based on large models, such as Wenxin Yiyan, Zhipu Qingyan, ChatGPT and Qianwen, which can help solve many problems. However, although they are highly versatile, they are not specifically developed for the field of mathematics and lack deep integration of mathematical knowledge, so they are not fully applicable to professional problems in the field of mathematics;

[0009] 3. The application scenarios of educational big models are limited, and it is difficult to deeply integrate them with subject teaching. Although there are systems developed based on big models and applied in the field of mathematics, these systems lack the embedding of subject knowledge and are mainly used to solve mathematical problems. The application scenarios are relatively limited, and their practicality and applicability are relatively limited for mathematics educators.

[0010] 4. The application of big models in the field of education often ignores cross-platform support. At present, the systems developed based on big models and applied in the field of mathematics are mainly limited to use on the Web and lack compatibility support for cross-platform operations. This affects the popularity and user-friendliness of the system and limits the promotion and application of big models in education.

[0011] To this end, the present invention proposes a cross-platform automatic solution system for mathematical word problems based on a large model and enhanced subject knowledge to solve the above problems. Summary of the invention

[0012] The purpose of the present invention is to provide a cross-platform automatic solution system for mathematical word problems based on a large model and enhanced subject knowledge to solve the problems raised in the above-mentioned background technology.

[0013] To achieve the above object, the present invention provides the following technical solution: a cross-platform automatic mathematical word problem solving system based on a large model and enhanced subject knowledge, the construction of which includes the following steps:

[0014] Step S1, constructing a subject knowledge recognition model to realize the joint recognition of knowledge points and contextual knowledge of mathematical word problems;

[0015] Step S2, propose a prompt design framework and design multiple types of prompts to enable the large language model to meet the needs of subject knowledge enhancement;

[0016] Step S3, build a server, deploy and run Qwen2.5-72B-Math and GOT-OCR2.0 large models, and provide automatic problem solving and text recognition services;

[0017] Step S4, deploy the subject knowledge recognition model, complete the integration, and complete the construction of a cross-platform automatic solution system for mathematical word problems based on a large model and subject knowledge enhancement.

[0018] As a preferred embodiment: the subject knowledge recognition model construction process of step S1 is as follows: the obtained math word problem corpus is cleaned, and the knowledge points and contextual knowledge are annotated; then, the jieba word segmentation is used to build a vocabulary and stop words are removed; secondly, the math word problem text is vectorized using Sentence2Vec, and the primary school math knowledge graph is mapped to a low-dimensional space using the Node2Vec algorithm, and the features of the two are fused and extracted through a 1D-expansion network, and the feature matrix is ​​output. Subsequently, an adjacency matrix is ​​generated based on the primary school math word problem data set, and it is fused with the feature matrix to generate a weighted feature matrix. Next, a graph convolutional neural network GCN model is constructed, the weighted feature matrix is ​​input into the model through forward propagation, and back propagation is used for optimization. Finally, the model performance is evaluated using a test set, and the tuned model is downloaded.

[0019] As a preferred embodiment: the subject knowledge recognition model construction process in step S1 uses the Node2Vec algorithm to map the entities and relationships in the elementary school mathematics knowledge graph to a low-dimensional vector space. Node2Vec generates node sequences by performing random walks on the graph. These sequences are regarded as "sentences". The transition probability of the random walk is controlled by the following formula to determine whether the walk tends to be a breadth-first search or a depth-first search.

[0020]

[0021] in, πvx is the adjacency matrix from node v to x, α is a preference parameter used to control the random walk, and Z is a normalization factor.

[0022] As a preferred embodiment: the subject knowledge recognition model construction process in step S1 uses one-dimensional dilated convolution to further fuse the elementary school mathematics knowledge graph feature vector and the elementary school mathematics word problem data set feature vector, and the output formula of the dilated convolution is as follows:

[0023]

[0024] Among them, K is the convolution kernel, X is the fused feature signal, d is the dilation rate, and n is the length of the convolution kernel.

[0025] Preferably, step S2 designs a "COR" framework based on prompt words that integrate subject knowledge and situational knowledge, and designs multiple types of prompts, wherein the "COR" framework consists of three elements: "C" stands for context, which is used to help the large language model grasp the user background information related to the task to generate content that is appropriate to the user background; "O" stands for goal, which is used to clarify personal intentions and express the tasks that the large language wants to perform in the form of propositions, including asking questions, clarifying needs, etc.; "R" stands for output requirements, indicating the requirements to be output, and constraining the content of the large model output.

[0026] As a preferred embodiment: Step S3 downloads the Qwen2.5-72B-Math and GOT-OCR2.0 large models on the built server, deploys and runs them using ollama, and provides automatic problem solving and text recognition services through ollama's own REST API.

[0027] As a preferred embodiment: Step S4 performs UI design and API document writing, builds front-end and back-end platforms, performs page and API development, deploys subject knowledge recognition models, and constructs a cross-platform automatic mathematical word problem solving system based on large models and subject knowledge enhancement.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. The "COR" prompt word design framework was proposed, and two types of prompts were designed: knowledge point / contextual knowledge analysis and type question generation;

[0030] 2. Provide a photo-taking automatic problem-solving tool based on a large model, which can efficiently identify questions and help learners solve difficult math word problems;

[0031] 3. Provide a tool to enhance subject knowledge and promote learners’ in-depth understanding and learning of subject knowledge;

[0032] 4. Provide educators with a tool to efficiently generate exercises related to subject knowledge, thereby enriching the means of resource acquisition and realizing the implementation of large models in vertical fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a design flow chart of the present invention;

[0034] Figure 2 A method for jointly identifying knowledge points and situational knowledge of mathematical word problems enhanced by subject knowledge of the present invention;

[0035] Figure 3 This is a functional structure diagram of the present invention;

[0036] Figure 4 Functional logic relationship diagram of the present invention;

[0037] Figure 5 This is a knowledge graph of elementary school mathematics taking natural numbers as an example in an embodiment of the present invention;

[0038] Figure 6 This is a description of the COR framework of the present invention;

[0039] Figure 7-Figure 11 This is the front-end page of an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] Example

[0042] See also Figure 1-Figure 2 , a cross-platform automatic mathematical word problem solving system based on a large model and subject knowledge enhancement is shown in the figure. The construction of the system includes the following steps:

[0043] Step S1, constructing a subject knowledge recognition model to realize the joint recognition of knowledge points and contextual knowledge of mathematical word problems;

[0044] Step S2, propose a prompt design framework and design multiple types of prompts to enable the large language model to meet the needs of subject knowledge enhancement;

[0045] Step S3, build a server, deploy and run Qwen2.5-72B-Math and GOT-OCR2.0 large models, and provide automatic problem solving and text recognition services;

[0046] Step S4, deploy the subject knowledge recognition model, complete the integration, and complete the construction of a cross-platform automatic solution system for mathematical word problems based on a large model and subject knowledge enhancement.

[0047] See also Figure 2 In this embodiment, step S1 obtains public math word problem corpus from the Internet, performs text cleaning, further annotates knowledge points and contextual knowledge for valid corpus, performs word segmentation on all questions, constructs the corpus vocabulary, and removes stop words to reduce noise. Finally, Sentence2Vec is used to convert sentences in the text of the math word problem dataset into numerical vectors.

[0048] Then use the Node2Vec algorithm to map the elementary school mathematics knowledge, such as Figure 5 The entities and relations shown in the figure are mapped to a low-dimensional vector space. Node2Vec generates node sequences by performing random walks on the graph. These sequences are regarded as "sentences". The transition probability of the random walk is controlled by the following formula to determine whether the walk tends to be a breadth-first search or a depth-first search.

[0049]

[0050] in, πvx is the adjacency matrix from node v to x, α is a parameter used to control the preference of the random walk, and Z is the normalization factor.

[0051] Furthermore, a one-dimensional dilated convolution is used to further fuse the feature vector of the elementary school mathematics knowledge graph with the feature vector of the elementary school mathematics word problem dataset. The dilated convolution increases its receptive field by inserting spaces in the convolution kernel without adding additional parameters, which allows the model to capture a wider range of contextual information while retaining local features. The output formula of the dilated convolution is as follows:

[0052]

[0053] Among them, K is the convolution kernel, X is the fused feature signal, d is the dilation rate, and n is the length of the convolution kernel.

[0054] Then, an adjacency matrix is ​​generated based on the elementary school mathematics word problem dataset, and it is fused with the feature matrix extracted based on the 1D-expansion network to generate a weighted feature matrix. Then, a graph convolutional neural network (GCN) model framework is constructed. Through the forward propagation process, the weighted feature matrix is ​​input into the model, and the ReLU activation function is applied to each graph convolution layer to enhance the expressiveness of the model. During this period, the loss function is used to quantify the difference between the model prediction and the true label, and the parameter gradient is calculated through the back-propagation algorithm. The gradient descent method is used to update the model parameters to minimize the loss. In order to avoid overfitting, hyperparameter adjustment and early stopping are used, and the performance of the validation set is used to optimize the generalization ability of the model.

[0055] Finally, use the test set to evaluate the model performance, tune the model according to indicators such as accuracy, recall, and F1 score, and download the tuned model.

[0056] See also Figure 6 In step S2 of this embodiment, reference is made to the key elements of the design of the “ICIO” prompt word, and a “COR” framework suitable for the prompt word design of the present invention is proposed, which consists of three elements, wherein “C” represents context, which helps the large language model grasp the user background information related to the task, so as to generate content appropriate to the user background; “O” represents the goal, which clarifies personal intentions and expresses the tasks that the large language is expected to perform in the form of propositions, including asking questions, clarifying requirements, etc. The clarity and specificity of the goal directly affect the relevance and accuracy of the generated content; “R” represents output requirements, which indicates the requirements to be output and constrains the content output by the large model.

[0057] Then the “COR” framework is designed by integrating the prompt words of subject knowledge and situational knowledge, such as Figure 6 As shown, multiple types of prompts are designed to make the large language model meet the needs of subject knowledge enhancement in the present invention. In specific applications, {{knowledge points}} and {{contextual knowledge}} will be automatically replaced according to the recognition results of the subject knowledge recognition model.

[0058] The specific process of step S3 of this embodiment is to first build a server with a running memory of 80G, select a Linux environment, install ollama on the server, ollama is an open source tool that aims to simplify the process of running and customizing large language models, and install OpenWebUI on the server, which provides a visual UI interface for dialogue testing.

[0059] Then we chose Qwen2.5-72B-Math and GOT-OCR2.0 to provide automatic problem-solving and text recognition services respectively. The reasons are as follows: Qwen2.5-72B-Math has been deeply optimized for mathematical content. Through the training of a large number of mathematical texts and problem data sets, it can accurately understand and generate mathematical problems, solutions and related mathematical concepts, thereby providing users with precise mathematical problem solutions and detailed concept explanations, effectively assisting users in mathematical learning and teaching; GOT-OCR2.0 has powerful multimodal recognition capabilities. It can not only recognize and convert regular text, but also accurately capture and convert formulas, notes, charts and other complex visual information. It performs well when processing mathematical formulas and tables containing small text. Since the Qwen2.5-72B-Math and GOT-OCR2.0 large models have just been released and have not been uploaded to https: / / ollama.com / library, they cannot be directly downloaded using the ollama command. First, download the model files of Qwen2.5-72B-Math and GOT-OCR2.0 from modelscope, and use the "Import model from Safetensors weights" method given by the ollama official to create and run these two models in ollama.

[0060] Then open the Open WebUI page, register and log in, enter the settings, load the model just deployed in ollama, and perform a dialogue test to ensure that the Qwen2.5-72B-Math and GOT-OCR2.0 models can be correctly accessed through the RESTAPI provided by ollama to provide automatic problem-solving and text recognition services.

[0061] See also Figure 3-Figure 4 The specific process of step S4 of this embodiment is to plan and design the system front-end page, determine the structure, layout, color scheme and font selection of the page, such as Figure 7-11 As shown in the figure, use Modao tool to design page architecture diagrams for functions such as system login and registration, photo recognition and problem solving, subject knowledge recognition, and prompt-based generation of type questions, complete the interaction design between modules and pages, write interface documents, and record the information of each interface in detail, including request parameters, response parameters, sample requests and responses, etc.

[0062] Next, use PyCharm to build the system backend based on Python's flask framework, deploy the subject knowledge recognition model in the backend, develop the interface according to the API documentation, and use Swagger to test and manage the API. After the interface test passes, deploy the backend interface to the server and perform external network mapping to ensure that the system front end can access the backend interface.

[0063] Then, HBuilderX is used to build the system front end (APP and applet) based on the uni-app and vue3 frameworks. According to the above-mentioned UI design and API documentation, the front end pages are developed, mainly including (1) a personal information page; (2) a photo recognition and problem solving page. In this page, a page event response function is written, including event responses for the photo taking and problem solving submission buttons. The event response function of the photo taking button will call the GOT-OCR2.0 API to extract the text in the photo and update the result to the input box. The event response function of the problem solving submission button will call the Qwen2.5-72B-Math API to automatically solve mathematical word problems and save the returned answers. Since the answers returned by Qwen2.5-72B-Math contain paragraphs and formulas in Markdown and Latex formats, these contents need to be parsed before they are readable. The system of the present invention chooses to perform parsing when the page is rendered, and in order to improve rendering efficiency, the mp-html plug-in is used. The mp-html plug-in can parse the contents in Markdown and Latex formats in segments and render them on the page. However, the mp-html plug-in currently only supports Latex formulas marked with “$...$” at the line level and “$$...$$” at the block level, while the Latex formulas returned by Qwen2.5-72B-Math are marked with “\(…\)” and “\[…\]”. Therefore, a Latex formula marking mode conversion function is written to convert the Latex formula marking mode before passing the answer to the mp-html plug-in for parsing and rendering to ensure that the answer can be correctly parsed and displayed on the page; (3) Photo recognition problem-solving page, in which page event response functions are written, including taking a photo The event response function of the problem-solving submit button and the event response function of the photo-taking button will call the GOT-OCR2.0API to extract the text in the photo and update the result to the input box. The event response function of the problem-solving submit button will call the Qwen2.5-72B-MathAPI to automatically solve mathematical word problems and save the returned answers. Since the answers returned by Qwen2.5-72B-Math contain paragraphs and formulas in Markdown and Latex formats, these contents need to be parsed before they are readable. The system of the present invention chooses to parse when the page is rendered, and in order to improve rendering efficiency, the mp-html plug-in is used.The mp-html plugin can parse and render Markdown and Latex format content on the page in sections. However, the mp-html plugin currently only supports Latex formulas marked with "$...$" at the line level and "$$...$$" at the block level, while the Latex formulas returned by Qwen2.5-72B-Math are marked with "\(…\)" and "\[…\]". Therefore, a Latex formula marking mode conversion function is written. Before passing the answer to the mp-html plugin for parsing and rendering, the Latex formula marking mode is converted to ensure that the answer is correct. Parse and display on the page; (4) Subject knowledge recognition page, in which the API request function of the subject knowledge recognition model is written so that the function can be called when clicking the button of the subject knowledge recognition page or when performing photo recognition problem solving, and the knowledge points and contextual knowledge of the submitted mathematical word problems are recognized. A function is implemented in the page, which generates a corresponding prompt based on the recognized knowledge points and contextual knowledge. After the knowledge points and contextual knowledge are recognized, a prompt pops up above the input box to guide the user to ask questions. The event response function of the type question generation button is written to generate type questions based on the knowledge points and contextual knowledge and display them on the page.

[0064] Then, after implementing the functions of the mini program and APP based on the above steps, use HBuilder X to build a web framework based on uni-app and vue3, migrate the system front end, and modify the page based on the web UI design. Finally, configure the domain name resolution, package the web end and upload it to the server, completing the construction of a cross-platform automatic solution system for mathematical word problems based on large models and enhanced subject knowledge.

[0065] The present invention involves Qwen2.5-72B-Math: Qwen2.5-72B-Math is a large pre-trained language model designed for students to solve math problems. The model is based on the Qwen series, has 7.2 billion parameters (i.e. 72B), and is optimized for math-related content. The Qwen2.5-72B-Math model can understand and generate math problems, solutions, and related math concepts, and is designed to assist math learning and teaching in the field of education. It is trained through a large amount of math text and problem data sets to improve the ability to parse and generate mathematical language, thereby providing students with accurate math problem solutions and concept explanations.

[0066] GOT-OCR2.0: GOT OCR 2.0, or "General OCR Theory", is an innovative optical character recognition (OCR) model that combines the advantages of traditional OCR systems and large language models and adopts a unified end-to-end architecture. The model consists of an image encoder with about 80 million parameters and a speech decoder with 500 million parameters. It can compress 1024x1024 pixel images into tokens and convert them into text of up to 8000 characters. GOT OCR 2.0 has multimodal recognition capabilities and can recognize and convert a variety of visual information including text, formulas, notes, charts, etc. Its application advantages are reflected in the integration of multiple modules to improve recognition accuracy, support for multiple OCR types, and efficient reasoning and training. The emergence of GOT OCR 2.0 is a response to the limitations of OCR 1.0 and large-scale visual language models in OCR tasks, and aims to improve the performance and application scope of OCR technology through innovative architectures and methods.

[0067] Graph Convolutional Neural Networks: Graph Convolutional Neural Networks, or GCNs for short, is a feature engineering method that condenses graph structure information. In graph structured data, each node can have multiple attributes, and the nodes are connected to each other through edges. Graph convolutional networks are specifically designed to capture the relationship between these nodes and the characteristics of the nodes themselves. By leveraging the advantages of convolutional neural networks in local structure modeling and the ubiquitous dependencies between nodes in graph data, graph convolutional neural networks have become one of the most active and influential research directions in the field of building deep learning models on graph structures.

[0068] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0069] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A cross-platform automatic mathematical word problem solving system based on large models and subject knowledge enhancement, characterized in that: The construction of the system includes the following steps: Step S1, constructing a subject knowledge recognition model to realize the joint recognition of knowledge points and contextual knowledge of mathematical word problems; Step S2, propose a prompt design framework and design multiple types of prompts to enable the large language model to meet the needs of subject knowledge enhancement; Step S3, build a server, deploy and run the Qwen2.5-72B-Math and GOT-OCR2.0 large models, and provide automatic problem solving and text recognition services; Step S4, deploy the subject knowledge recognition model, complete the integration, and complete the construction of a cross-platform automatic solution system for mathematical word problems based on a large model and subject knowledge enhancement.

2. A cross-platform automatic mathematical word problem solving system based on large models and subject knowledge enhancement according to claim 1, characterized in that: The process of constructing the subject knowledge recognition model in step S1 is as follows: cleaning the acquired mathematical word problem corpus, and labeling the knowledge points and contextual knowledge; Then, the jieba word segmentation is used to build a vocabulary and remove stop words; secondly, Sentence2Vec is used to vectorize the math word problem text, and the Node2Vec algorithm is used to map the elementary school mathematics knowledge graph to a low-dimensional space. The features of the two are fused and extracted through a 1D-expansion network, and the feature matrix is ​​output. Subsequently, an adjacency matrix is ​​generated based on the elementary school mathematics word problem data set, and it is fused with the feature matrix to generate a weighted feature matrix. Next, a graph convolutional neural network GCN model is constructed, the weighted feature matrix is ​​input into the model through forward propagation, and back propagation is used for optimization. Finally, the test set is used to evaluate the model performance, and the tuned model is downloaded.

3. A cross-platform automatic mathematical word problem solving system based on large models and subject knowledge enhancement according to claim 2, characterized in that: The subject knowledge recognition model construction process in step S1 uses the Node2Vec algorithm to map the entities and relationships in the elementary school mathematics knowledge graph to a low-dimensional vector space. Node2Vec generates node sequences by performing random walks on the graph. These sequences are regarded as "sentences". The transition probability of the random walk is controlled by the following formula to determine whether the walk tends to be a breadth-first search or a depth-first search. in, πvx is the adjacency matrix from node v to x, α is a preference parameter used to control the random walk, and Z is a normalization factor.

4. The cross-platform automatic mathematical word problem solving system based on large models and subject knowledge enhancement according to claim 3 is characterized by: The subject knowledge recognition model construction process in step S1 uses one-dimensional dilated convolution to further fuse the elementary school mathematics knowledge graph feature vector and the elementary school mathematics word problem dataset feature vector. The output formula of the dilated convolution is as follows: Among them, K is the convolution kernel, X is the fused feature signal, d is the dilation rate, and n is the length of the convolution kernel.

5. A cross-platform automatic mathematical word problem solving system based on large models and subject knowledge enhancement according to claim 4, characterized in that: The step S2 designs a "COR" framework based on prompt words that integrate subject knowledge and situational knowledge, and designs multiple types of prompts, wherein the "COR" framework consists of three elements, "C" represents context, which is used to help the large language model grasp the user background information related to the task, so as to generate content appropriate to the user background; "O" stands for goal, which is used to clarify personal intentions and express the tasks that you want the big language to perform in the form of propositions, including asking questions, clarifying needs, etc. "R" stands for output requirement, which indicates the desired output requirement and constrains the output content of the large model.

6. A cross-platform automatic mathematical word problem solving system based on large models and subject knowledge enhancement according to claim 5, characterized in that: The step S3 downloads the Qwen2.5-72B-Math and GOT-OCR2.0 large models on the built server, deploys and runs them using ollama, and provides automatic problem solving and text recognition services through the REST API provided by ollama.

7. A cross-platform automatic mathematical word problem solving system based on large models and subject knowledge enhancement according to claim 6, characterized in that: The step S4 performs UI design and API document writing, builds front-end and back-end platforms, performs page and API development, deploys subject knowledge recognition models, and constructs a cross-platform automatic mathematical word problem solving system based on a large model and subject knowledge enhancement.

Images