Cryptography intelligent question and answer method based on maximum marginal correlation retrieval

Abstract

The invention discloses a cryptographic intelligent question and answer method based on maximum marginal correlation retrieval, and relates to the technical field of natural language process.The method comprises the steps that firstly, a cryptographic field document is subjected to analysis and semantic partitioning, and the integrity of mathematical formulas and algorithm logic is ensured by dynamically calculating text similarity; and information splitting caused by traditional fixed partitioning is avoided. And then, obtaining knowledge fragments which are related to the question and have various contents from the vector database by utilizing a maximum marginal correlation retrieval algorithm, and providing multi-angle reliable contexts for answer generation. And finally, through field adaptive prompt engineering and dynamic parameter adjustment, guiding the large language model to play a role of a cryptography expert based on the retrieved context, and generating an accurate personalized answer meeting the user demand.

Description

A cryptographic intelligent question answering method based on maximum marginal relevance retrieval Technical Field

[0001] This invention relates to the field of natural language processing technology, and more specifically to a cryptographic intelligent question answering method based on maximum marginal relevance retrieval. Background Technology

[0002] In the rapid digital wave of information technology development, cryptography, as the cornerstone of information security, has become increasingly strategically important, profoundly impacting the overall landscape of e-commerce, data privacy protection, and even national security. However, the inherent theoretical depth and abstractness of cryptography, especially the computational complexity theories such as large integer factorization and discrete logarithms upon which public-key cryptosystems rely, constitute significant learning barriers, leading to difficulties for students in further study and Q&A. While current traditional cryptography teaching models can systematically impart basic knowledge face-to-face, they struggle to meet students' growing needs for personalized and interactive learning, particularly in providing immediate feedback, resolving practical difficulties, and handling massive amounts of cryptographic knowledge.

[0003] In recent years, artificial intelligence technologies, represented by Large Language Models (LLM), have made groundbreaking progress, providing a new path for reshaping the education paradigm. With its superior natural language understanding and generation capabilities, LLM has demonstrated enormous potential in personalized tutoring, intelligent question answering, and the automated construction of educational resources, significantly improving teaching efficiency and the learning experience. However, general-purpose LLM still faces key bottlenecks when dealing with highly specialized disciplines like cryptography, which heavily rely on accuracy and factual consistency. Its knowledge may lag behind the latest developments and is prone to generating "illusionary" content that contradicts the facts of the discipline, severely limiting its reliable application in serious educational scenarios. To overcome this limitation, Retrieval-Augmented Generation (RAG) has emerged and quickly become a key pathway to enhance the specialized capabilities of LLM.

[0004] Therefore, how to provide cryptography course Q&A services based on RAG to achieve accurate transfer of cryptography knowledge and efficient learning is a problem that urgently needs to be solved by those in the field. Summary of the Invention

[0005] In view of the above problems, this invention proposes a cryptographic intelligent question-answering method based on maximum marginal relevance retrieval to overcome or at least partially solve these problems. It addresses the limitations of traditional teaching tools, such as their lack of personalized interactive capabilities; the tendency of general-purpose large-scale language models to produce factual errors and lag in knowledge updates within specialized fields; and the inability of existing retrieval systems to effectively handle the mathematical and structured knowledge system unique to cryptography. By constructing an intelligent question-answering method based on Retrieval Augmentation Generation (RAG) technology, this invention overcomes these technical bottlenecks, achieving accurate delivery and efficient learning of cryptographic knowledge. It employs innovative technologies such as semantic segmentation, maximum marginal relevance retrieval, and domain-adaptive generation, specifically optimized for the characteristics of cryptographic data, fundamentally improving the technical service level in the field of professional education. This invention solves the problem of traditional RAG retrieval relying solely on vector similarity, leading to high content redundancy and a single dimension, trapping learners in an "information cocoon" of homogeneous information.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, embodiments of the present invention provide a cryptographic intelligent question answering method based on maximum marginal relevance retrieval, comprising the following steps: Step 1: Collecting original documents related to cryptographic knowledge, and generating cryptographic text after parsing and preprocessing; Step 2: Using a semantic text segmenter to divide the cryptographic text into blocks to obtain several text blocks, and constructing a retrieval knowledge base; Step 3: Obtaining user questions, and using the maximum marginal relevance algorithm to perform similarity retrieval in the retrieval knowledge base to obtain context documents; Step 4: Generating answers based on user questions and context documents using a pre-trained large language model.

[0007] Preferably, original documents related to cryptographic knowledge in different formats are collected and parsed into document text. The document text is then preprocessed to obtain cryptographic text. The preprocessing includes removing irrelevant metadata, duplicate content, and special characters.

[0008] Preferably, the process of segmenting cryptographic text using a semantic text segmenter is as follows: Step 21: Use a sliding window to traverse the cryptographic text sequentially, and use a pre-trained vector embedding model to calculate the embedding vector of the text segment within the sliding window; Step 22: Calculate the cosine similarity value between the embedding vectors of the current text segment within the sliding window and the adjacent text segments to obtain the text similarity; Step 23: If the text similarity is lower than a preset similarity threshold, then segment the current text segment within the sliding window and the adjacent text segments to obtain a new text block; otherwise, do not segment, slide the sliding window to the next text segment, and return to step 21.

[0009] Preferably, the vector embedding model adopts the BAAI / bge-large-zh-v1.5 model, which is trained using relevant technical literature in cryptography education, thus reducing semantic distortion of technical terms.

[0010] Preferably, in step 2, a number of candidate text blocks are retrieved using the FAISS vector search engine based on all text blocks to construct a retrieval knowledge base. FAISS supports fast approximate nearest neighbor search, reducing retrieval latency and making it suitable for real-time question answering scenarios. Simultaneously, FAISS supports incremental index updates, facilitating dynamic expansion of the knowledge base (such as adding new cryptographic standards) and adapting to the evolution of domain knowledge.

[0011] Preferably, the specific process of step 3 is as follows: Step 31: Use the maximum marginal relevance algorithm (MMR) to select several target text blocks that are both relevant and diverse from the candidate text blocks in the retrieved knowledge base; Step 32: Clean, deduplicate and concatenate the target text blocks, and integrate them into a structured context document under the premise that the total length does not exceed the limit of the pre-trained large language model window.

[0012] Preferably, by adjusting the balance parameter in the maximum marginal correlation algorithm Balancing correlation and diversity; balancing parameters The default setting is 0.6.

[0013] Preferably, in step 4, a domain-specific prompt template is preset. The user's question and context document are dynamically inserted into the prompt template to generate prompt words. The Large Language Model (LLM) generates the answer based on the prompt words. The prompt template includes role setting, task description, context injection, question presentation, and output requirements. By using prompt words as instructions, illusion is reduced and fidelity is improved.

[0014] Preferably, the large language model uses a dynamic parameter mapping mechanism to generate answers based on prompts. This mechanism includes: using a question classifier to identify the type of user question in the prompts; if it is determined to be a concept definition type, the temperature parameter of the large language model is set to a low threshold, forcing the model to output deterministic facts, suppressing illusions, and minimizing randomness to ensure the rigor and uniqueness of cryptographic terminology interpretation; if it is determined to be a comparative analysis or logical reasoning type, the temperature parameter is set to a medium threshold, allowing the model to perform limited logical reasoning to enhance the flexibility of the model's thought process while ensuring factual accuracy; otherwise, the temperature parameter is set to a default balanced parameter.

[0015] Preferably, the low threshold can be 0.1, the medium threshold can be 0.3, and the default balance parameter is 0.5.

[0016] Preferably, the ChatGPT-4o-mini model is used for large language models.

[0017] As can be seen from the above technical solutions, compared with the existing technology, the present invention discloses a cryptographic intelligent question answering method based on maximum marginal relevance retrieval. Targeting the cryptographic education scenario, it utilizes the adaptability of semantic segmentation to mathematical symbols and formulas, the adaptation of the MMR retrieval strategy to the cryptographic knowledge structure, and the constraint generation of professional terms by the hint engineering. Combined with an authoritative cryptographic knowledge base, it improves the accuracy of model question answering and reduces the generation of illusionary content, providing learners with accurate, reliable, and personalized question answering services. This solves the problems of highly structured cryptographic data and strong concept dependence, significantly improving the accuracy of question answering. The specific beneficial effects include: (1) At the technical level, the semantic segmentation strategy maintains the integrity of the cryptographic algorithm description, the maximum marginal relevance retrieval achieves multi-angle knowledge coverage, and the domain-adaptive hint engineering ensures the accuracy of the generated content, ultimately forming a complete closed loop of retrieval enhancement generation technology. This achieves a low-illusion, highly timely, and verifiable professional education support effect, effectively improving the overall effect of cryptographic teaching and learning; (2) At the application effect level, it realizes the intelligent transformation of the cryptographic education model. Addressing the characteristics of cryptography theory—its depth, rapid knowledge iteration, and the need for Chinese-English translation—this approach solves key problems in LLM such as illusory information, outdated knowledge, and unverifiable sources. It provides accurate, timely, and reliable intelligent support for teaching and research, offering personalized and precise learning support for learners at different levels, effectively solving the "one-size-fits-all" problem in traditional teaching.

[0018] (3) A RAG retrieval knowledge base for cryptography courses was constructed. By deeply integrating authoritative knowledge sources such as classic textbooks, standards and academic papers, and introducing a dynamic update mechanism that is constantly evolving, the problem of "illusion" information and slow knowledge base updates caused by insufficient professional knowledge in cryptography teaching of general large language models is effectively solved. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 is a schematic diagram of the cryptographic intelligent question answering process based on maximum marginal relevance retrieval provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] This invention discloses a cryptographic intelligent question answering method based on maximum marginal relevance retrieval, as shown in Figure 1, comprising the following steps: S1: Collecting original documents related to cryptographic knowledge, and generating cryptographic text after parsing and preprocessing; S2: Using a semantic text segmenter to divide the cryptographic text into blocks to obtain several text blocks, and constructing a retrieval knowledge base; S3: Obtaining user questions, and using the maximum marginal relevance algorithm to perform similarity retrieval in the retrieval knowledge base to obtain context documents; S4: Generating answers based on user questions and context documents using a pre-trained large language model.

[0023] Furthermore, original documents related to cryptography knowledge in different formats are collected and parsed into document text. The document text is then preprocessed to obtain cryptographic text. The preprocessing includes removing irrelevant metadata, duplicate content, and special characters.

[0024] The technical effect of the above solution is that it solves the problem of logical coherence of cryptographic text and reduces information fragmentation through the above parsing and preprocessing process.

[0025] Furthermore, the process of segmenting cryptographic text using a semantic text segmenter is as follows: S21: The cryptographic text is traversed sequentially using a sliding window, and the embedding vector of the text segment within the sliding window is calculated using a pre-trained vector embedding model; S22: The cosine similarity value between the embedding vectors of the current text segment within the sliding window and the adjacent text segments is calculated to obtain the text similarity; S23: If the text similarity is lower than a preset similarity threshold, the text segment within the current sliding window and the adjacent text segments are segmented to obtain a new text block; otherwise, no segmentation is performed, the sliding window slides to the next text segment, and returns to S21.

[0026] The technical effect of the above solution is that by employing a semantic text segmenter and dynamically dividing text into blocks based on an embedding similarity threshold, the integrity of mathematical derivations and algorithm descriptions is ensured. This achieves the logic of segmenting Document objects into smaller text blocks, particularly semantic blocks. In this process, a `SemanticChunker` class is defined that receives a list of Documents (cryptographic text) and an embedding model as input. The segmentation point is determined by calculating the similarity of the embedding vectors of sentences or paragraphs within adjacent sliding windows. When the similarity is lower than a preset threshold, segmentation is performed. For example, the segmentation of public-key cryptographic documents maintains the continuity of the "key generation-encryption-decryption" process, rather than mechanically cutting it, prioritizing the preservation of semantic boundaries. Through this semantic segmentation of cryptographic text, the challenge of segmenting mathematical formulas and algorithm descriptions is solved, improving the quality of the retrieval context.

[0027] Furthermore, the vector embedding model adopts the BAAI / bge-large-zh-v1.5 model, which is trained using relevant technical literature in cryptography education, thus reducing semantic distortion of technical terms.

[0028] The technical effects of the above-mentioned technical solution are as follows: the BAAI / bge-large-zh-v1.5 model is good at handling technical content, can better understand the logical relationships of cryptographic concepts, and supports multilingual embedding to ensure the consistency of terminology vectorization. After converting cryptographic text into vectors, the model makes semantically similar concepts closer in the vector space, which makes it easier for the retrieval module to quickly locate relevant contexts, accurately capture the semantics of cryptographic terms, and improve retrieval relevance.

[0029] Furthermore, in S2, the FAISS vector search engine is used to recall several candidate text blocks based on all text blocks to construct a retrieval knowledge base. FAISS supports fast approximate nearest neighbor search, reducing retrieval latency and making it suitable for real-time question answering scenarios. At the same time, FAISS supports incremental index updates, facilitating dynamic expansion of the knowledge base (such as adding new cryptographic standards) and adapting to the evolution of domain knowledge.

[0030] Furthermore, the specific process of S3 is as follows: S31: Use the maximum marginal relevance algorithm (MMR) to select several target text blocks that are both relevant and diverse from the candidate text blocks in the retrieved knowledge base; S32: Clean, deduplicate and concatenate the target text blocks, and integrate them into a structured context document under the premise of controlling the total length to not exceed the window limit of the pre-trained large language model.

[0031] Furthermore, by adjusting the balance parameters in the maximum marginal correlation algorithm... Balancing correlation and diversity; balancing parameters The default setting is 0.6. The technical effect of the above solution is that, during the retrieval process, by calculating vector similarity and combining iteratively with the maximum marginal relevance algorithm, text blocks with high semantic similarity to the user's question but significant differences from each other are selected, ensuring query relevance while maximizing result diversity. This retrieval process is deeply adapted to the special needs of cryptographic scenarios: on the one hand, cryptographic knowledge is highly structured, and traditional retrieval easily returns redundant information, while MMR balances parameters... Based on semantic segmentation, the algorithm leverages the FAISS vector database to rapidly recall candidate sets and then filters out target text blocks covering multiple perspectives, including algorithm principles and security applications, effectively addressing the issues of content redundancy and single-dimensionality. Furthermore, the diverse contexts generated by MMR provide a comprehensive basis for LLM, reducing the risk of generating illusory answers and thus meeting the stringent requirements for accuracy, timeliness, and personalized interpretations in cryptography education. MMR addresses the multi-faceted needs of concepts and avoids redundancy; FAISS enables fast approximate nearest neighbor search, reducing retrieval latency.

[0032] Furthermore, S4 pre-sets domain-specific prompt templates, dynamically inserting user questions and contextual documents into these templates to generate prompt words. The Large Language Model (LLM) then generates answers based on these prompt words. The prompt templates include role settings, task descriptions, context injection, question presentation, and output requirements. By using prompt words as instructions, illusions are reduced and fidelity is improved.

[0033] The prompt template is loaded using `use_optimized_prompts=True`. The prompt template contains placeholders for dynamically inserting user questions and retrieved context documents. An optimized prompt template for cryptography question-and-answer scenarios is as follows: 1. Role Setting: The LLM will act as a professional cryptography course teaching assistant; 2. Task Description: Explain that the task is to answer the user's question based on the provided context information; 3. Context Injection: Clearly present the K retrieved text blocks to the LLM; 4. Question Presentation: Includes the user's original question; 5. Output Requirements: May include various requirements regarding answer style (e.g., concise, detailed), format, and whether sources need to be cited.

[0034] Furthermore, the large language model employs a dynamic parameter mapping mechanism to generate answers based on prompts. This mechanism includes: using a question classifier to identify the type of user question in the prompts; if it is determined to be a concept definition type, the temperature parameter of the large language model is set to a low threshold, forcing the model to output definitive facts, suppressing illusions, and minimizing generation randomness to ensure the rigor and uniqueness of cryptographic terminology interpretations; if it is determined to be a comparative analysis or logical reasoning type, the temperature parameter is set to a medium threshold, allowing the large language model to perform limited logical reasoning to enhance the flexibility of the model's thought process while ensuring factual accuracy; otherwise, the temperature parameter is set to a default balanced parameter. Adaptive adjustment of the temperature parameter avoids errors caused by excessive model creativity, dynamic adaptation through the question classifier improves answer reliability, and the dynamic parameter mapping mechanism balances accuracy and creativity in real time during the generation phase, effectively avoiding professional knowledge illusions caused by excessive model divergence.

[0035] Furthermore, the low threshold can be set to 0.1, the medium threshold to 0.3, and the default balance parameter to 0.5.

[0036] Furthermore, the large language model adopts the ChatGPT-4o-mini model. OpenAI's ChatGPT-4o-mini model, which excels across multiple key dimensions, achieves an 82% score on the MMLU benchmark, capable of handling complex cryptographic reasoning (such as protocol analysis). It performs even better in answer relevance and contextual accuracy, ensuring the quality of professional content generation. This model not only has the ability to process text, images, and audio simultaneously, but can also directly parse mathematical formulas and protocol flowcharts in cryptography, helping students visualize abstract concepts. It performs excellently in adherence, reasoning, and other aspects, and can dynamically adjust the depth of responses based on student user levels, achieving personalized adaptive teaching adjustments. It supports 128K context windows, accommodating long documents (such as cryptography textbooks), and meets the deployment requirements of high-frequency interaction scenarios.

[0037] Based on the same inventive concept, this invention also provides a cryptographic intelligent question-answering system based on maximum marginal relevance retrieval, including a data acquisition module, a data processing and indexing module, a retrieval module, and a generation module. The data acquisition module collects original documents and user questions related to cryptographic knowledge. The data processing and indexing module parses and preprocesses the original documents to generate cryptographic text, and uses a semantic text segmenter to segment the cryptographic text into blocks to construct a retrieval knowledge base. The retrieval module uses the maximum marginal relevance algorithm to perform similarity searches in the retrieval knowledge base based on user questions to obtain context documents. The generation module deploys a pre-trained large language model to generate answers based on user questions and context documents.

[0038] Furthermore, the data processing and indexing module includes a document preprocessing unit and a text segmentation unit; the document preprocessing unit performs unified parsing and preprocessing of the original document to generate cryptographic text; the text segmentation unit deploys a vector embedding model to segment the cryptographic text into blocks, obtains several text blocks, and uses the FAISS vector search engine to construct a retrieval knowledge base from the text blocks.

[0039] Since the principle behind the problem solved by this system is similar to the aforementioned cryptographic intelligent question answering method based on maximum marginal relevance retrieval, the implementation of this system can refer to the implementation of the aforementioned method, and the repetitions will not be repeated.

[0040] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0041] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A cryptographic intelligent question answering method based on maximum marginal relevance retrieval, characterized in that, The process includes the following steps: Step 1: Collect original documents related to cryptographic knowledge, and generate cryptographic text after parsing and preprocessing; Step 2: Use a semantic text segmenter to divide the cryptographic text into blocks to obtain several text blocks, and build a retrieval knowledge base; Step 3: Obtain user questions, and use the maximum marginal relevance algorithm to perform similarity retrieval in the retrieval knowledge base to obtain context documents; Step 4: Generate answers based on user questions and context documents using a pre-trained large language model.

2. The cryptographic intelligent question answering method based on maximum marginal relevance retrieval as described in claim 1, characterized in that, Collect raw documents related to cryptography knowledge in different formats, parse them into document text, and preprocess the document text to obtain cryptographic text. The preprocessing includes removing irrelevant metadata, duplicate content, and special characters.

3. The cryptographic intelligent question answering method based on maximum marginal relevance retrieval as described in claim 1, characterized in that, The process of segmenting cryptographic text using a semantic text segmenter is as follows: Step 21: Use a sliding window to traverse the cryptographic text sequentially, and use a pre-trained vector embedding model to calculate the embedding vector of the text segment within the sliding window; Step 22: Calculate the cosine similarity value between the embedding vectors of the current text segment within the sliding window and the adjacent text segments to obtain the text similarity; Step 23: If the text similarity is lower than the preset similarity threshold, then segment the current text segment within the sliding window and the adjacent text segments to obtain a new text block; otherwise, do not segment, slide the sliding window to the next text segment, and return to step 21.

4. The cryptographic intelligent question answering method based on maximum marginal relevance retrieval as described in claim 1, characterized in that, In step 2, the FAISS vector search engine is used to recall several candidate text blocks based on all text blocks to build a retrieval knowledge base.

5. The cryptographic intelligent question answering method based on maximum marginal relevance retrieval as described in claim 4, characterized in that, The specific process of step 3 is as follows: Step 31: Use the maximum marginal relevance algorithm to select several target text blocks that are both relevant and diverse from the candidate text blocks in the retrieved knowledge base; Step 32: Clean, deduplicate and concatenate the target text blocks, and integrate them into a structured context document under the premise that the total length does not exceed the window limit of the pre-trained large language model.

6. The cryptographic intelligent question answering method based on maximum marginal relevance retrieval as described in claim 1, characterized in that, In step 4, a domain-specific prompt template is preset, and the user's question and context document are dynamically inserted into the prompt template to generate prompt words. The large language model generates the answer based on the prompt words.

7. The cryptographic intelligent question answering method based on maximum marginal relevance retrieval as described in claim 6, characterized in that, The large language model uses a dynamic parameter mapping mechanism to generate answers based on prompts. The dynamic parameter mapping mechanism includes: using a question classifier to identify the type of user question in the prompts; if it is determined to be a concept definition type, the temperature parameter of the large language model is set to a low threshold; if it is determined to be a comparative analysis or logical reasoning type, the temperature parameter is set to a medium threshold; otherwise, the temperature parameter is set to the default balanced parameter.

8. The cryptographic intelligent question answering method based on maximum marginal relevance retrieval as described in claim 1, characterized in that, The large language model adopts the ChatGPT-4o-mini model.

Images