A method for constructing a subject knowledge base to assist digital teacher teaching

Abstract

The present invention relates to the field of artificial intelligence technology, and discloses a method for constructing a subject knowledge base that assists digital teachers in teaching, including: using a large language model to characterize text-format subject teaching materials, and constructing a knowledge representation database based on the teaching material representation results; constructing an interactive interface between the teacher and the large language model interface, and designing prompts to construct two large model answering methods: using a knowledge base to retrieve answers and not using a knowledge base to retrieve answers, using a large language model to match knowledge points in a knowledge representation database for teachers' teaching needs, calculating the similarity between each knowledge point embedding in the knowledge representation database and the teacher's needs, sorting the knowledge points in the knowledge representation database based on the similarity, obtaining demand teaching resources based on the knowledge point sorting, and using a large language model to answer the teacher's needs based on the demand teaching resources combined with the large model's own knowledge. The present invention can more accurately assist digital teachers in carrying out teaching activities in specific subject areas.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for constructing a subject knowledge base for assisting digital teacher teaching. Background Art

[0002] In today's information age, large language model technology has made significant progress and has been widely used in many fields. However, with the continuous development and application of large model technology, especially in the field of education, it still faces a series of challenges. Although large language model technology has achieved the understanding, generation and processing of natural language through advanced methods such as deep learning, especially in general tasks such as text generation and sentiment analysis, its application effect in specific fields, especially in the field of education, still has significant shortcomings.

[0003] First, large language models have shown certain limitations when dealing with specific problems in the field of education. Although these models can process a wide range of text information and answer various questions, when faced with highly professional subject areas, due to the lack of in-depth understanding of the knowledge in this field, the answers they generate are often difficult to achieve the expected accuracy and quality. For example, when teachers need to rely on models to provide accurate subject knowledge during the teaching process, existing large language models may not be able to provide sufficiently accurate support due to the lack of specialized understanding of the content of the textbooks. This limitation makes it difficult for teachers to fully trust the model to assist in teaching, thus limiting the widespread application of large models in educational scenarios.

[0004] Secondly, the current large language model technology is also insufficient in processing multimodal information. Although modern large language models are capable of processing text data, in educational applications, multimodal information such as teaching materials, images, and charts are equally important. For example, in the teaching process of digital teachers, it is necessary not only to process text, but also to effectively utilize the image and chart information in the teaching materials. However, the existing large model technology is still limited in this regard, resulting in limited understanding and answering capabilities for complex queries.

[0005] Large models also have problems with knowledge acquisition and updating. Most current large language models rely on pre-training data, which is often general and lacks domain-specific knowledge. This means that the model may not be able to respond in a timely manner or give effective answers when faced with emerging knowledge or constantly updated teaching materials in the field of education. In addition, large models usually lack a mechanism to dynamically integrate external knowledge into their own reasoning and generation processes, resulting in their inability to fully realize their potential when faced with complex teaching situations.

[0006] In order to solve the above problems, a new research trend has emerged in recent years, which is to enhance the capabilities of large language models by building external knowledge bases. This method can significantly improve the performance of large models in specific fields by structuring the knowledge in professional fields into knowledge bases and combining them with the powerful semantic understanding and generation capabilities of large models. In the field of education, especially in the teaching scenarios of digital teachers, building a dedicated teaching material knowledge base can not only provide the necessary background information for the large model, but also achieve accurate matching of the specific teaching needs of teachers. This method can greatly improve the accuracy and practicality of the model's answers, ensuring that the intelligent support in the teaching process is closer to actual needs. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a method for constructing a subject knowledge base to assist digital teachers in teaching. By combining the powerful semantic understanding and generation capabilities of a large language model, it is possible to deeply analyze the teaching content, especially in specific subject areas, by constructing a special knowledge base to achieve accurate matching of teachers' teaching needs. In addition, the present invention also comprehensively utilizes multimodal information processing technology, which can effectively process various information forms such as text, images, and charts in teaching materials, and enhance the system's understanding and response capabilities to complex teaching scenarios. At the same time, the knowledge base construction method designed by the present invention also has the ability of dynamic updating and real-time learning, ensuring the timeliness and accuracy of teaching resources, thereby providing digital teachers with more intelligent and personalized teaching support.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0009] A method for constructing a subject knowledge base to assist digital teacher teaching includes the following steps:

[0010] Step 1: Search for textbook knowledge resources and construct a textbook knowledge dataset;

[0011] Step 2: Block the text content in the textbook knowledge dataset and split the text content into text paragraphs; use a large language model to represent the textbook knowledge dataset to obtain a textbook knowledge vector library;

[0012] Step 3: Build an interactive interface between the user and the large language model, and prompt the user with two question-answering methods;

[0013] Step 4, obtaining the user's question, matching the semantic vector corresponding to the user's question with the content of the textbook knowledge vector library; merging the matched text paragraph with the user's question to form an overall input content;

[0014] Step 5: construct prompt words. The large language model performs reasoning based on the prompt words and the overall input content to generate answers that are consistent with the textbook knowledge and meet the user's needs.

[0015] Furthermore, step 2 specifically includes:

[0016] Text segmentation: The text content in the textbook knowledge data set is characterized by a segmentation strategy, which includes: using chunk technology and overlap technology to segment the text content into fixed-length text paragraphs, while ensuring that there are overlapping parts between paragraphs;

[0017] Semantic vectorization construction: After completing text segmentation, call the large language model to perform semantic vectorization on the segmented text paragraphs, represent the text paragraphs as low-dimensional dense vectors, and map them into the semantic space, so that texts with similar meanings are closer in the vector space.

[0018] Furthermore, step 3 specifically includes:

[0019] The Streamlit tool is used to build an interactive interface. Users can enter questions through the interactive interface and choose whether to use the textbook knowledge vector library retrieval function. When the user chooses not to use the textbook knowledge vector library retrieval, the large language model will directly answer the user's questions based on the knowledge it has learned.

[0020] Furthermore, step 4 specifically includes:

[0021] Understand the semantics of questions in natural language input by users and generate low-dimensional dense semantic vector representations through a large language model;

[0022] Matching the semantic vector corresponding to the user question with the textbook knowledge in the textbook knowledge vector library, specifically including: calculating the cosine similarity between the semantic vector of the user question and the vectors of each text paragraph in the textbook knowledge vector library, and returning the K text paragraphs with the highest cosine similarity; in order to ensure the integrity of the retrieved text paragraphs, a similar text overlap mechanism with a fixed neighborhood is set, and the content of a set number of words is expanded before and after the matched K text paragraphs;

[0023] The K matched text paragraphs are merged with the user's question to form a whole input content.

[0024] Compared with the prior art, the beneficial technical effects of the present invention are:

[0025] 1) High accuracy in teaching demand matching: The present invention achieves accurate understanding and response to the teaching needs proposed by teachers through the combination of a large language model and a teaching material knowledge base, and significantly improves the accuracy of teaching information retrieval.

[0026] 2) Knowledge base construction optimization: The semantic vectorization processing method is used to effectively capture the semantic information of the textbook content, enhance the system's storage and retrieval capabilities for large-scale corpus, and thus improve the accuracy of matching teaching needs.

[0027] 3) Strong flexibility and scalability: The present invention supports the import and processing of multiple document formats, is compatible with different types of large language models and semantic vector models, provides a highly customized user experience, and adapts to the teaching needs of different fields.

[0028] 4) High-quality and stable answers: By designing optimized prompt templates and leveraging the semantic understanding and generation capabilities of large language models, high-quality, context-relevant answers can be generated, effectively improving the digital teacher's auxiliary capabilities in the teaching process. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the overall flow of the method in an embodiment of the present invention;

[0030] Figure 2 A working principle diagram of the method in an embodiment of the present invention;

[0031] Figure 3 The figure is a schematic diagram of the process of constructing a teaching material knowledge vector library in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to more clearly illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose, the present invention is now described in detail in conjunction with the accompanying drawings and embodiments.

[0033] The accompanying drawings are exemplary illustrations, which may slightly exaggerate the width, size and shape of certain objects and are not drawn strictly to scale. It should be understood that the terms "including", "having" and "comprising" in this specification indicate the presence of the stated features, wholes, steps, operations, modules and / or components, but do not exclude the possibility of other features, wholes, steps, operations, modules, components and combinations thereof. In addition, when a statement such as "at least one of..." appears, this description modifies the entire list of features, rather than a single component.

[0034] When describing the embodiments of the present invention, the term “may” is used to mean “one or more embodiments of the present invention”, and the term “exemplary” means to illustrate an example.

[0035] The terms "substantially" and "approximately" used in the specification indicate approximate meanings rather than precise definitions, and are intended to reflect reasonable deviations in measurements or calculations that can be understood by ordinary technicians in the field. Unless otherwise specified, the terms (including technical terms and scientific terms) in the specification are used according to the common understanding of ordinary technicians in the field, and should not be interpreted in an idealized or overly formal manner.

[0036] like Figure 1 As shown, a method for constructing a subject knowledge base for assisting digital teacher teaching in the present invention comprises the following steps:

[0037] Step 1: Search for textbook knowledge resources to build a textbook knowledge dataset.

[0038] Step 2: Use the big model to characterize the textbook knowledge data set to obtain the textbook knowledge vector library.

[0039] Step 3: Build an interactive interface between the user and the large model, and prompt the user with two question-and-answer methods.

[0040] Step 4: Obtain the user's question and match the semantic vector corresponding to the user's question with the content of the teaching material knowledge vector library.

[0041] Step 5: construct a prompt to prompt the large model to retrieve the knowledge stored in the textbook knowledge vector library to answer the user's question.

[0042] according to Figure 2 to Figure 3 , the above-mentioned implementation steps of the present invention are described in detail.

[0043] In step 1, searching for teaching material knowledge resources to construct a teaching material knowledge data set specifically includes:

[0044] Document data collection: Search and collect textbooks for various subjects. Textbooks include various published textbooks, learning materials, reference books, etc., and duplicate or irrelevant textbook content is removed through manual review.

[0045] Text data import: supports importing multiple document file formats such as TXT, DOCX, MD, and supports importing image file formats such as PDF and JPG. Using optical character recognition (OCR) technology, the input image data can be converted into text data. After inputting the reserve files of the knowledge base, the system will automatically identify the document type and convert it into text data uniformly.

[0046] Through manual review, a textbook knowledge data set is established for the huge amount of subject textbook data. The textbook knowledge data set structure is constructed based on textbook content, knowledge points, themes, publishing institutions and other information. Through the indexing method, each chapter of the textbook is used as an entity node of the textbook knowledge data set, and finally the processed textbook source files are imported into the textbook knowledge data set.

[0047] In step 2, the use of a large model to characterize the textbook knowledge data set to obtain a textbook knowledge vector library specifically includes:

[0048] Semantic vector matching: Based on the teaching needs of teachers, semantic vector matching is directly performed on the knowledge points in the textbook knowledge data set to determine the knowledge points related to the needs.

[0049] Textbook knowledge data set retrieval: Based on the knowledge points obtained by semantic vector matching, the textbook knowledge data set is retrieved through similarity calculation to obtain the most relevant teaching resources.

[0050] Text segmentation: The text content in the textbook knowledge dataset is characterized by a segmentation strategy. The specific method is to use chunk and overlap techniques to segment the text into fixed-length paragraphs while ensuring that there is a certain amount of overlap between paragraphs. This processing method helps capture the contextual relevance of the text and improve the accuracy of semantic matching.

[0051] Semantic vectorization construction: After completing text segmentation, call the large language model to perform semantic vectorization on the segmented text, represent the text paragraphs as low-dimensional dense vectors, and map them into the semantic space, so that texts with similar meanings are closer in the vector space. This helps capture the semantic associations between texts and makes the representation of knowledge more accurate and rich. After vectorization is completed, the semantic vector is imported into the textbook knowledge vector library, which supports efficient similarity retrieval and clustering for dense vectors, providing accurate knowledge point matching and resource recommendations for teachers' teaching needs.

[0052] In step 3, the interactive interface between the user and the large model is constructed to prompt the user with two question-answering methods, specifically including:

[0053] In order to further improve the interaction efficiency between teachers and the large language model, the present invention also uses an interactive interface built by Streamlit. Teachers can input teaching needs through this interface and choose whether to use the textbook knowledge data set retrieval function. When teachers choose not to use the textbook knowledge vector library retrieval, the large language model will directly answer the teacher's questions based on the knowledge it has learned.

[0054] In step 4, the step of obtaining the user question and matching the semantic vector corresponding to the user question with the content of the teaching material knowledge vector library specifically includes:

[0055] The natural language questions input by users are semantically understood and a low-dimensional dense semantic vector representation is generated through a large language model.

[0056] Match the semantic vector corresponding to the user question with the textbook knowledge in the textbook knowledge vector library: Specifically, search and match the semantic vector of the user question with the text paragraph vector in the textbook knowledge vector library. Calculate the cosine similarity between the semantic vector of the user question and each text paragraph vector in the textbook knowledge vector library. The similarity value range is between [0,1]. The closer the value is to 1, the higher the semantic similarity.

[0057] The content most similar to the user's question is retrieved based on the similarity, and the K text paragraphs with the highest matching degree are returned. In order to ensure the integrity of the retrieved text paragraphs, a similar text overlap mechanism with a certain neighborhood (for example, 50 words) is set up at the same time, and the K retrieved text paragraphs are expanded by 50 words before and after to prevent the complete sentence from being split and cut off. By using the embedding model for semantic vector matching, it has strong generalization ability and is suitable for knowledge bases in multiple fields, thereby improving the applicability of the system.

[0058] Merge the K text paragraphs matched in the knowledge base with the user's question. This step forms a holistic input content by integrating the relevant text paragraphs with the user's question, providing a basis for subsequent reasoning.

[0059] In step 5, a prompt is constructed to prompt the large model to retrieve the knowledge stored in the textbook knowledge vector library to answer the user's question, including:

[0060] Using the constructed prompt, the K text paragraphs and user questions matched in the knowledge base are used as input, and the large language model is called for reasoning to generate answers that are consistent with the textbook knowledge and meet the teacher's needs. This method can effectively link the retrieved knowledge with user needs and provide teachers with accurate and targeted teaching support.

[0061] In this embodiment, the semantic understanding and generation capabilities of the large language model are utilized to implement rapid calling of the large language model and optimization of prompt templates through an API to complete the answer to the user's natural language question.

[0062] This embodiment provides a calling interface for a large language model, supports a variety of open source large language models, and supports multiple rounds of historical conversations. The conversation model parameters currently used are Zhipu ChatGLM3, and the representation model adopts Zhipu's Embedding-2 model. The K text paragraphs matched in the above knowledge base are merged with the user's question, input into the defined prompt template, and generate the prompt information required by the large language model according to the prompt. Based on the prompt, the large language model uses its Chinese understanding and generation capabilities to extract available information and infer and generate answers. At the same time, the model returns the matching paragraph as a reference source.

[0063] In summary, the present invention can not only improve the response speed of digital teachers in complex teaching scenarios, but also provide more accurate and personalized teaching support, thereby significantly improving the quality and efficiency of digital teaching.

[0064] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0065] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A method for constructing a subject knowledge base to assist digital teacher teaching, characterized in that: The following steps are involved: Step 1: Search for textbook knowledge resources and construct a textbook knowledge dataset; Step 2: Block the text content in the textbook knowledge dataset and split the text content into text paragraphs; use a large language model to represent the textbook knowledge dataset to obtain a textbook knowledge vector library; Step 3: Build an interactive interface between the user and the large language model, and prompt the user with two question-answering methods; Step 4, obtaining the user's question, matching the semantic vector corresponding to the user's question with the content of the textbook knowledge vector library; merging the matched text paragraph with the user's question to form an overall input content; Step 5: construct prompt words. The large language model performs reasoning based on the prompt words and the overall input content to generate answers that are consistent with the textbook knowledge and meet the user's needs.

2. The method for constructing a subject knowledge base for assisting digital teacher teaching according to claim 1, characterized in that: Step 2 specifically includes: Text segmentation: The text content in the textbook knowledge data set is characterized by a segmentation strategy, which includes: using chunk technology and overlap technology to segment the text content into fixed-length text paragraphs, while ensuring that there are overlapping parts between paragraphs; Semantic vectorization construction: After completing text segmentation, call the large language model to perform semantic vectorization on the segmented text paragraphs, represent the text paragraphs as low-dimensional dense vectors, and map them into the semantic space, so that texts with similar meanings are closer in the vector space.

3. The method for constructing a subject knowledge base for assisting digital teacher teaching according to claim 1, characterized in that: Step 3 specifically includes: The Streamlit tool is used to build an interactive interface. Users can enter questions through the interactive interface and choose whether to use the textbook knowledge vector library retrieval function. When the user chooses not to use the textbook knowledge vector library retrieval, the large language model will directly answer the user's questions based on the knowledge it has learned.

4. The method for constructing a subject knowledge base for assisting digital teacher teaching according to claim 1, characterized in that: Step 4 specifically includes: Understand the semantics of questions in natural language input by users and generate low-dimensional dense semantic vector representations through a large language model; Match the semantic vector corresponding to the user question with the textbook knowledge in the textbook knowledge vector library, specifically including: calculating the cosine similarity between the semantic vector of the user question and the vectors of each text paragraph in the textbook knowledge vector library, and returning the K text paragraphs with the highest cosine similarity; in order to ensure the integrity of the retrieved text paragraphs, a similar text overlap mechanism with a fixed neighborhood is set, and the content of a set number of words is expanded before and after the matched K text paragraphs; The K matched text paragraphs are merged with the user's question to form a whole input content.

Images