Document recommendation problem generation method and device, equipment and storage medium

Abstract

The present disclosure provides a document recommendation question generation method, device and equipment and storage medium, the method comprises: obtaining a to-be-processed document and performing hierarchical division on the to-be-processed document to obtain at least two hierarchical text segments; obtaining and generating the prompt word of each hierarchical text segment based on the pre-configured question generation parameter for each hierarchy; based on the prompt word of each hierarchical text segment, using a large language model to generate a recommended question for each hierarchical text segment. In the present disclosure, by dividing the to-be-processed document into layers and generating a recommended question for each hierarchical text segment one by one, a multi-level coverage recommended question is formed, ensuring complete coverage of the recommended question for different content depth and breadth of the to-be-processed document.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and storage medium for generating document recommendation problems. Background Technology

[0002] In fields such as intelligent question answering, retrieval systems, and educational assessment, automatically generating relevant questions for documents is a key technology supporting efficient human-computer interaction. It is directly related to the system's depth of understanding and the efficiency of user information retrieval.

[0003] Existing technical solutions typically focus on generating questions at a single granularity, resulting in a question set that fails to comprehensively and hierarchically cover the core content and key details of a document. This makes it difficult for users to gain a progressively deeper understanding through the question sequence. Therefore, how to generate multi-layered recommendation questions for documents is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, this disclosure proposes a method, apparatus, device and storage medium for generating document recommendation questions, which can form multi-level covered recommendation questions, ensuring that the recommendation questions fully cover the depth and breadth of different content of the documents to be processed.

[0005] According to a first aspect of this disclosure, a method for generating document recommendation questions is provided, comprising:

[0006] Obtain the document to be processed and divide the document into hierarchical levels to obtain text fragments of at least two levels;

[0007] Obtain and generate prompt words for text fragments at each level based on pre-configured question generation parameters for each level;

[0008] Based on the prompt words of each level of text fragment, a large language model is used to generate recommendation questions for each level of text fragment.

[0009] In one possible implementation, the problem generation parameters include at least one of the following: problem type, problem difficulty, number of problems, and problem output format.

[0010] In one possible implementation, when generating prompts for text fragments at each level based on pre-configured question generation parameters, the following is included:

[0011] Get pre-configured prompt word templates;

[0012] The question generation parameters and the text fragments of each level are filled into the prompt word template to obtain the prompt words for each level of text fragment.

[0013] In one possible implementation, when generating a recommendation question for each level of text fragment based on the prompt words of each level of text fragment using a large language model, the process includes:

[0014] Based on the prompt words of each level of text fragment, a large language model is used to generate initial recommendation questions for each level of text fragment;

[0015] The initial recommendation questions for each level of text fragment are filtered to obtain the recommendation questions for each level of text fragment.

[0016] In one possible implementation, the process of filtering the initial recommendation questions for each level of text fragment to obtain the recommendation questions for each level of text fragment includes:

[0017] At least one filtering operation, including semantic similarity detection, format detection, and coverage evaluation, is performed on the initial recommendation question of each level of text fragment to obtain the recommendation question of each level of text fragment.

[0018] In one possible implementation, after obtaining the recommendation questions for each of the aforementioned hierarchical text fragments, the method further includes:

[0019] The operation involves sorting the recommendation questions for each of the aforementioned text fragments.

[0020] In one possible implementation, ranking the recommendation problem for each of the aforementioned hierarchical text fragments includes:

[0021] Calculate the historical recommendation value and title similarity value of the recommendation question for each of the aforementioned text fragments;

[0022] Based on the historical recommendation values ​​and title similarity values ​​of the recommendation questions for each level of text fragment, the recommendation questions for each level of text fragment are ranked.

[0023] According to a second aspect of this disclosure, an apparatus for generating document recommendation questions is provided, comprising:

[0024] The hierarchical division module is used to acquire the document to be processed and perform hierarchical division on the document to be processed to obtain text fragments of at least two levels;

[0025] The prompt word generation module is used to obtain and generate prompt words for text fragments at each level based on the pre-configured question generation parameters for each level;

[0026] The recommendation question generation module is used to generate recommendation questions for each level of text fragment based on the prompt words of each level of text fragment and using a large language model.

[0027] According to a third aspect of this disclosure, an apparatus for generating document recommendation questions is provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to perform the method described in the first aspect of this disclosure.

[0028] According to a fourth aspect of this disclosure, a non-volatile computer-readable storage medium is provided that stores computer program instructions thereon, wherein the computer program instructions, when executed by a processor, implement the method described in the first aspect of this disclosure.

[0029] This disclosure provides a method, apparatus, device, and storage medium for generating document recommendation questions. The method includes: acquiring a document to be processed and dividing the document into hierarchical levels to obtain text fragments at least two levels; acquiring and generating prompt words for each level of text fragment based on pre-configured question generation parameters for each level; and generating recommendation questions for each level of text fragment based on the prompt words of each level of text fragment using a large language model. This disclosure, by dividing the document to be processed into layers and generating recommendation questions for each level of text fragment one by one, forms multi-level covered recommendation questions, ensuring complete coverage of the depth and breadth of different content in the document to be processed.

[0030] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0031] The accompanying drawings, which are included in and form part of this specification, illustrate exemplary embodiments, features, and aspects of this disclosure together with the specification and serve to explain the principles of this disclosure.

[0032] Figure 1 A flowchart illustrating a method for generating document recommendation questions according to an embodiment of the present disclosure is shown.

[0033] Figure 2 A schematic block diagram of an apparatus for generating document recommendation questions according to an embodiment of the present disclosure is shown.

[0034] Figure 3 A schematic block diagram of a device for generating document recommendation questions according to an embodiment of the present disclosure is shown. Detailed Implementation

[0035] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0036] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0037] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0038] <Method Implementation>

[0039] Figure 1 A flowchart illustrating a method for generating document recommendation questions according to an embodiment of this disclosure is shown. Figure 1 As shown, the method includes steps S1100-S1300.

[0040] S1100: Obtain the document to be processed and divide the document (hereinafter referred to as the document) into levels to obtain text fragments of at least two levels. The level division can be set according to the user's specific needs.

[0041] In one possible implementation, when classifying documents, the classification can be divided into two levels: document level and fragment level.

[0042] When generating document-level text fragments, the process includes: generating a summary of the document using a large language model, and using the generated summary as a document-level text fragment.

[0043] When generating fragment-level text fragments, the following steps may be included:

[0044] First, the document is initially divided based on its inherent natural breakpoints (such as periods, paragraph marks, line breaks, etc.) to obtain an initial sequence of text fragments. This initial sequence of text fragments includes multiple divided text fragments, which are arranged in the order they appear in the document.

[0045] Second, combining semantic coherence analysis and pre-defined length constraints, the initial text segment sequence is merged or refined to obtain text segments with independent themes and knowledge points, which are then used as segment-level text segments. The specific steps are as follows:

[0046] First, semantic coherence analysis is performed on each text segment in the initial sequence: adjacent text segments are checked to see if they discuss the same topic or closely related knowledge points. If so, they are preferentially merged into a larger text segment to maintain the semantic integrity of the text segments.

[0047] Next, the initial text segment sequence undergoes length analysis: if a text segment's length exceeds the upper limit threshold (>1000 characters), it is split into smaller, semantically more complete text segments based on its internal natural breaks or logical structure (such as clauses or enumeration items). If adjacent text segments are all short (<500 characters) and semantically highly related, they are merged into a text segment that meets the minimum length requirement. If the length of a text segment is within the target length range (500-1000 characters) and semantically independent, no further processing is performed.

[0048] Finally, the text fragments after length analysis are checked and fine-tuned according to the preset semantic independence criteria (such as topic concentration and concept uniformity) to obtain multiple final text fragments with relatively independent topics and knowledge points. These text fragments are the fragment-level text content.

[0049] It should be noted here that, while ensuring semantic independence, length constraints should be followed as much as possible. If semantic integrity conflicts with strict length constraints, semantic independence should be prioritized.

[0050] In one possible implementation, particularly for technical documents containing program code, mathematical formulas, or structured tables, the generation of fragment-level text fragments also includes:

[0051] Identify special content units: Utilize a syntax parser or predefined pattern recognition rules to detect embedded program code blocks (such as code segments enclosed in \`` tags), mathematical formulas (such as LaTeX or MathML formats), or structured tables in text fragments.

[0052] Create auxiliary semantic units: Treat each identified block of program code, mathematical formula, or structured table as an independent auxiliary semantic unit.

[0053] Semantic unit fusion:

[0054] If an auxiliary semantic unit is semantically closely related to its surrounding natural language text (e.g., a code block is described in natural language to indicate its function, a formula is explained to indicate its meaning, and tabular data is summarized in text), then the auxiliary semantic unit and its directly associated natural language text are merged into a single fragment-level text fragment. During merging, it must be ensured that the total length of the merged fragment does not exceed a preset upper limit threshold (e.g., 1200 characters).

[0055] If the auxiliary semantic unit is relatively independent (e.g., only a code example, a complete list of formulas in the appendix, or an independent data table), and its own length already meets the minimum fragment length requirement (e.g., >200 characters), then it is directly treated as an independent fragment-level text fragment.

[0056] This step aims to ensure that elements crucial to understanding the technical details of the document, such as program logic, core formulas, or key data tables, can be used as independent semantic units or closely integrated with related descriptions to generate more targeted recommendation questions (such as code function explanations, formula derivations, and data interpretation questions).

[0057] In another possible implementation, when classifying documents, the documents can be parsed into four levels: document level, chapter level, section level, and fragment level.

[0058] The method for generating document-level text fragments is described above and will not be repeated here.

[0059] When generating chapter-level text fragments, the document can be divided into multiple chapter-level text fragments based on the highest-level heading structure (such as H1, Chapter, Part, etc., which clearly identify chapters).

[0060] When generating section-level text fragments, the process includes: for each chapter-level text fragment, dividing the chapter-level text fragment according to its internal subheading structure (such as H2, Section, Subpart, etc., which identify sections) will yield multiple section-level text fragments corresponding to each chapter-level text fragment.

[0061] When generating fragment-level text segments, the process includes: dividing each section-level text segment according to its inherent natural breakpoints (such as periods, paragraph marks, line breaks, etc.), and then merging and refining the segments based on semantic coherence analysis and preset length constraints to obtain the corresponding fragment-level text segments. The specific calculation process is detailed above and will not be repeated here.

[0062] In one possible implementation, the system executing the method of this disclosure is provided with a hierarchy configuration interface, which offers the two hierarchy division methods mentioned above. The first method divides the system into two levels: document level and segment level. The second method divides the system into four levels: document level, chapter level, section level, and segment level. Users can select the hierarchy division method according to their specific needs. After the user selects the hierarchy division method, the text segments at each level will be divided according to the user's selected hierarchy division method.

[0063] Furthermore, when a user selects the second level division method, they can further choose at least one of document-level, chapter-level, section-level, and segment-level, allowing the system to divide the document into levels according to the user's specific selection. For example, if the user only selects document-level and chapter-level, the document will only be divided into document-level and chapter-level text segments. Similarly, if the user only selects chapter-level, section-level, and segment-level, the document will only be divided into text segments at these three levels. This increases the flexibility of document classification and meets the specific needs of different users.

[0064] In one possible implementation, the system also provides an option to generate cross-level related questions. When the user enables this option in the configuration interface:

[0065] Hierarchical relationship mapping: When generating prompts for text fragments at each level, the system automatically identifies and associates the text fragments or their core summaries / keywords with the text fragments of the current level (such as section level to chapter level, chapter level to document level) in addition to including the text fragments of the current level and the question generation parameters.

[0066] Related prompt injection: Dynamically add related generation instructions to the prompt template, such as: "At the same time, please generate 1-2 questions to explore the specific relationship between the above fragment content and its parent level name, such as: chapter's core theme '[parent level core theme / keyword]' or how it supports the theme." or "Based on the details of this fragment and its parent level name's overall goal '[parent level goal summary]', propose a question that reflects the local support for the global."

[0067] Result Identification and Integration: For recommendation questions generated by large language models, those belonging to this type of cross-level association question are specially identified in the output results (e.g., by adding the "[association]" prefix). In the subsequent ranking process, such questions can participate in the ranking based on their association strength (e.g., semantic similarity with parent-level keywords) or user-configured priority rules.

[0068] This option aims to address users' need to understand how document content supports each other at the macro and micro levels, and how local details serve the overall goal, by generating recommendation questions with vertical depth.

[0069] After the document is divided, step S1200 can be executed to obtain and generate prompts for text fragments at each level based on the pre-configured question generation parameters for each level. The question generation parameters for each level include at least one of the following: question type, question difficulty, number of questions, and question output format.

[0070] It should be noted that different levels of text fragments have different functions in the document. In order to enable the generated recommendation questions to help the text fragments achieve their corresponding effects, the system pre-matches different question types to text fragments at different levels.

[0071] Specifically, for document-level and chapter-level text fragments, the purpose of these fragments is to guide users to grasp the overall value of the document. Therefore, the matching question types can include at least one of the following: framing, intention-inference, and value-judgment. Framing questions assess the user's understanding of the core structure, components, or overall context of a document or chapter. For example, when the question type is framing, the generated recommended questions could be: What is the core research objective of the document? What are the components of the overall framework of the solution proposed in Chapter 3? etc. Intention-inference questions assess the user's ability to infer the fundamental purpose, motivation, or core problem that the document or specific chapter attempts to solve. For example, when the question type is intention-inference, the generated recommended questions could be: What is the author's fundamental intention in writing this section? What pain points in the industry does this chapter attempt to solve? etc. Value-judgment questions assess the advantages, significance, impact, or relative value of the solutions, methods, or viewpoints proposed in the document or chapter. For example, when the question type is value-judgment, the generated recommended questions could be: Compared to traditional methods, what are the main advantages of the solution proposed in this chapter? Where does the core value of this technical approach lie for industry development? etc.

[0072] For section-level text fragments, the purpose is to help users understand local logical chains and key knowledge points. Therefore, the matched question types can include at least one of the following: logical analysis, knowledge association, and method comparison. Logical analysis questions assess the user's understanding and analytical ability regarding the argumentation process, causal chain, or operational flow within a specific section. The aim is to guide users to identify local logical relationships or working mechanisms. For example, when the question type is logical analysis, recommended questions might include: How does the data processing flow presented in this section solve the data silo problem? What is the dependency between step A and step B? What is the complete execution path after the fault recovery mechanism is triggered? etc. Knowledge association questions assess the user's ability to identify the inherent connections between the knowledge points in the current section and other parts of the document (such as parent chapters, sibling sections, or key concepts). The aim is to guide users to establish cognitive connections across content modules. For example, when the question type is knowledge association, recommended questions might include: How is the security mechanism of this module related to the encryption scheme mentioned in Chapter 5? How does the optimization strategy here correspond to the core assumptions in the introduction? Does the experimental method described in this section overturn the theoretical model in Chapter 2? These are examples of questions. Method comparison questions examine the differences, advantages, disadvantages, or applicability of the methods / solutions proposed in this section compared to other alternatives. The aim of these questions is to guide users in critical thinking and decision support. For example, when the question type is method comparison, the generated recommendation questions might be: What are the differences in the applicable scenarios between solution X and solution Y? Why choose an iterative algorithm instead of a one-time solution? What is the basis for the throughput improvement of dynamic scheduling strategies compared to static strategies?

[0073] For fragment-level text snippets, the purpose of these snippets is to support users in quickly locating key details or performing specific operations. Therefore, the matching question types can include at least one of parameter verification, operation execution, and fact verification. Parameter verification questions assess the ability to accurately extract technical parameters, configuration items, or numerical indicators from a text snippet. The purpose is to guide users to locate key data or configuration details. For example, when the question type is parameter verification, the generated recommended questions might be: What temperature threshold parameter should be set to? What is the range of values ​​for the attenuation coefficient γ in Formula 7? etc. Operation execution questions assess the ability to reproduce specific operation steps or troubleshooting actions based on the text snippet description. The purpose is to guide users to complete atomic-level task execution. For example, when the question type is operation execution, the generated recommended questions might be: What are the three specific steps for configuring an SSL certificate? How do I check the ERROR_CODE field in the log? Which command line command is needed to reset the device password? etc. Fact verification questions assess the ability to verify objective facts, states, or conclusions explicitly stated in a text snippet. The purpose of these questions is to guide users to confirm undisputed details. For example, when the question type is a fact-checking question, the generated recommended questions could be: What is the final determined sample size of the experimental group? Does the protocol support encryption algorithms including AES-256?

[0074] It's important to note that, to meet the needs of different users, after selecting the hierarchical division method in the hierarchical settings configuration interface, users can further configure the question generation parameters for each level. Specifically, clicking on the question parameter configuration space for each level will bring up the corresponding question parameter configuration interface. This interface provides configuration options for question type, question difficulty, number of questions, and question output format. Users can configure the corresponding question generation parameters according to their needs. When configuring the question type, users can select the desired question type from the question types matching that level; when configuring the question difficulty, users can select the corresponding difficulty level; when configuring the number of questions, users can enter the required number of questions; and when configuring the question output format, users can choose the default output format or import a user-defined data output format.

[0075] After the user completes the question generation parameters for each level, step S1200 automatically reads the question generation parameters corresponding to each level. Then, by combining the question generation parameters for each level with the pre-constructed prompt word templates of the text fragment input values ​​for each level, the prompt words for each level of text fragment can be obtained. It should be noted that if there are multiple text fragments under a level, then for each text fragment, the prompt words for that text fragment and the question generation parameters of the level to which the text fragment belongs are combined with the prompt word templates of the input values ​​to obtain the prompt words for each text fragment.

[0076] After generating prompt words for text fragments at each level, step S1300 can be executed: based on the prompt words for each level of text fragments, a large language model is used to generate recommendation questions for each level of text fragments. Specifically, this may include the following steps: based on the prompt words for each level of text fragments, a large language model is used to generate initial recommendation questions for each level of text fragments. The initial recommendation questions for each level of text fragments are then filtered to obtain the recommendation questions for each level of text fragments. The output format of the initial recommendation questions is as follows:

[0077]

[0078]

[0079] In a specific example, the initial recommendation question for the text fragment "Water is one of the most common substances on Earth, existing in nature in three forms: liquid, solid, and gas" could be as follows:

[0080]

[0081] In other words, the recommendation questions for each text fragment include not only the initial recommendation questions generated for it, but also the specific content of the text fragment and the keywords it contains, making it convenient for users to see the correspondence between the questions and the text.

[0082] After outputting the recommended questions for each text segment according to the above format, the initial recommended questions in the recommended question information will be filtered, and then replaced with the final recommended questions, thus completing the update of the recommended information for each text segment. That is, the final recommended questions for each text segment are also output in this format.

[0083] It's important to note that the text fragments at each level are ordered according to their order of appearance in the document and the hierarchical relationship between them. Therefore, generating recommendation questions for each level of text fragment also follows this order. Specifically, the text fragments at each level are traversed in their order of appearance in the document. For the current text fragment, the prompt word is obtained and input into a pre-trained large language model, which generates an initial recommendation question for that fragment. After obtaining the initial recommendation question, the generated initial questions are filtered to obtain the final recommendation question for the current text fragment. Once the traversal is complete, the recommendation questions for each level of text fragment are obtained.

[0084] Furthermore, the large language model generates multiple initial recommendation questions for the current text segment. Therefore, it is necessary to further filter the multiple initial recommendation questions to obtain the final recommendation question. In one possible implementation, the process of filtering the initial recommendation questions for the current text segment to obtain the final recommendation question may include the following steps: performing at least one filtering operation among semantic similarity detection, format detection, and coverage evaluation on the initial recommendation questions for the current text segment to obtain the recommendation question for the current text segment.

[0085] In one possible implementation, when performing semantic similarity detection on the initial recommendation questions for the current text segment, the following steps are included: using a sentence vector model (such as Sentence-BERT) to calculate the semantic similarity between each initial recommendation question, and removing initial recommendation questions with repetitive or highly similar expressions by setting a similarity threshold (such as 0.9).

[0086] In one possible implementation, when performing format detection on the initial recommendation question for the current text fragment, the following steps are included: obtaining the text length of each initial question, and deleting initial questions whose text length exceeds the upper limit and the lower limit, so as to ensure that the length of the final recommendation question is within a reasonable range.

[0087] In one possible implementation, after semantic similarity detection and format detection, the coverage of the remaining initial recommendation questions for the current text segment is evaluated. This can include the following steps: calculating the keywords for each initial recommendation question corresponding to the current text segment (i.e., the keywords of all initial recommendation questions generated by the large model), and deduplicating the keywords to form a keyword library for the current text segment. The coverage of the remaining initial recommendation questions with the keywords in the keyword library is calculated (coverage = number of keywords covered by the remaining initial recommendation questions ÷ total number of keywords in the keyword library): if the coverage is greater than a coverage threshold, the remaining initial recommendation questions are used as the final recommendation questions for the current text segment; if the coverage is less than or equal to the coverage threshold, uncovered keywords in the keyword library are searched, and new initial recommendation questions including these uncovered keywords are generated as a supplement. All the supplemented initial recommendation questions corresponding to the current text segment are then used as the final recommendation questions for the current text segment, thus ensuring that the final recommendation questions for the current text segment have high knowledge point coverage and comprehensiveness. The coverage threshold can be set according to specific needs, for example, it can be set to 70%.

[0088] In one possible implementation, to improve the user experience and cognitive friendliness of the final recommendation question, after completing semantic similarity detection, format detection, and coverage evaluation filtering, a cognitive load evaluation filtering step is introduced for the remaining recommendation questions:

[0089] Problem complexity modeling: For each remaining recommendation question, calculate its comprehensive cognitive load score. This score is calculated based on a weighted average of the following factors:

[0090] Syntactic complexity: quantified by analyzing indicators such as sentence length, number of nested clauses, and special sentence structures (negation, interrogation, condition).

[0091] Concept density: The number of core domain terms, technical terms, or undefined concepts that appear in a statistical problem.

[0092] Reasoning depth: Fine-tuned based on the predefined basic weights of the question type (e.g., fact-confirming type = 1, logical analysis type = 3, intentional reasoning type = 4), and combined with the keywords of reasoning steps implied in the question text (e.g., "why", "how", "impact", "comparison").

[0093] Load Threshold and Dynamic Adaptation: The system sets a baseline cognitive load score threshold. In the user configuration interface, users can select or input a difficulty preference factor (e.g., concise, standard, in-depth). This factor is used to dynamically adjust the cognitive load score threshold for actual applications (e.g., selecting concise lowers the threshold, while selecting in-depth raises it).

[0094] Load screening: The combined cognitive load score of the remaining recommended questions is compared to a dynamically adjusted threshold. For questions with scores significantly higher than the threshold (e.g., exceeding the threshold by 20%):

[0095] Try to simplify the restatement: Input it as a new "question" into the large language model, with the instruction: "Please restate the following question in a more concise and understandable sentence, keeping the core query intent unchanged: [original question]".

[0096] Replace or supplement: Replace the original high-load problem with the simplified problem obtained, or keep it as a supplementary option along with the original problem (this needs to be noted in the results).

[0097] This step aims to automatically identify and optimize recommended questions that are syntactically complex, verbose, or have excessively long thought processes, making them more aligned with the target user's cognitive level and current difficulty preferences, thereby improving the readability and answerability of the questions.

[0098] In one possible implementation, after obtaining the recommendation questions for each level of text fragments, the method further includes: sorting the recommendation questions for each level of text fragments. Specifically, this may include the following steps: calculating the historical recommendation values ​​and title similarity values ​​for the recommendation questions of each level of text fragments; and sorting the recommendation questions for each level of text fragments based on the historical recommendation values ​​and title similarity values.

[0099] It should be noted that the sorting process for the corresponding recommendation questions is the same for each text fragment at each level. The following is a detailed explanation of the sorting process for the corresponding recommendation questions, using the current text fragment as an example.

[0100] Specifically, the historical recommendation values ​​of each recommended question corresponding to the current text fragment and the title similarity between each recommended question and the document title are calculated. This yields the historical recommendation values ​​and title similarity for each recommended question corresponding to the current text fragment. If a recommended question has historical recommendation values, it is prioritized; the higher the historical recommendation value, the higher the ranking. If a recommended question has no historical recommendation values, it is ranked according to title similarity; the higher the similarity, the higher the ranking. In other words, it is first sorted by historical recommendation values, then by title similarity. Sorting by historical recommendation values ​​prioritizes displaying recommended questions that the user is interested in, prioritizing them over the user's interests. In the absence of historical recommendation values, recommendations closely related to the title are then presented to the user based on title similarity. In a specific example, the ranking results of the recommended questions corresponding to the current text fragment are shown below:

[0101]

[0102] In one possible implementation, calculating the historical recommendation value of the recommendation question corresponding to the current text fragment includes: obtaining the user's historical question record, which records multiple historical questions previously asked by the user and the number of times each historical question was asked; and for each recommendation question, calculating the historical recommendation value of the recommendation question based on the similarity between the recommendation question and each historical question, as well as the frequency of occurrence of each historical question. The formula for calculating the historical recommendation value is as follows:

[0103]

[0104] In the formula, M is the historical recommendation value of the recommended question, si is the similarity between the recommended question and the i-th historical question, wi is the number of times the i-th historical question appears, and i is the total number of historical questions in the historical question records.

[0105] In one possible implementation, after generating the final recommendation questions for each text segment at each level in sequence, the recommendation questions for each document segment at each level of the document structure are reorganized to construct a tree-indexed set of recommendation questions. This structure preserves both hierarchical logic and reading flow, enabling intuitive and efficient browsing from the user's perspective. The specific reorganization process is as follows:

[0106] First, hierarchical nesting: the questions are organized in a tree structure, with the top level being document-level recommendation questions, and the middle level nesting recommendation questions at the chapter-section level.

[0107] Second, chronological arrangement: Recommendation questions at the same level are arranged in strict accordance with the original order of appearance of the text fragments.

[0108] Third, spatial annotation: the source text location of each recommendation question is annotated (e.g., §3.2.1-P4), and a tree-indexed set of recommendation questions is output.

[0109] This disclosure provides a method for generating document recommendation questions, comprising: acquiring a document to be processed and dividing the document into hierarchical levels to obtain text fragments at least two levels; acquiring and generating prompt words for each level of text fragments based on pre-configured question generation parameters for each level; and generating recommendation questions for each level of text fragments based on the prompt words of each level of text fragments using a large language model. This disclosure, by dividing the document to be processed into layers and generating recommendation questions for each level of text fragments one by one, forms multi-level covered recommendation questions, ensuring that the recommendation questions fully cover the different depths and breadths of the content of the document to be processed.

[0110] <Device Embodiment>

[0111] Figure 2 A schematic block diagram of an apparatus for generating document recommendation questions according to an embodiment of the present disclosure is shown. Figure 2 As shown, the device 100 includes:

[0112] The hierarchical division module 110 is used to acquire the document to be processed and perform hierarchical division on the document to be processed to obtain text fragments of at least two levels.

[0113] The prompt word generation module 120 is used to obtain and generate prompt words for text fragments at each level based on the pre-configured question generation parameters for each level.

[0114] The recommendation question generation module 130 is used to generate recommendation questions for each level of text fragment based on the prompt words of each level of text fragment and using a large language model.

[0115] <Equipment Example>

[0116] Figure 3 A schematic block diagram of a device for generating document recommendation questions according to an embodiment of the present disclosure is shown. Figure 3 As shown, the document recommendation question generation device 200 includes a processor 210 and a memory 220 for storing executable instructions of the processor 210. The processor 210 is configured to implement any of the document recommendation question generation methods described above when executing the executable instructions.

[0117] It should be noted here that the number of processors 210 can be one or more. Furthermore, the document recommendation question generation device 200 in this embodiment may also include an input device 230 and an output device 240. The processors 210, memory 220, input device 230, and output device 240 can be connected via a bus or other means, without specific limitations here.

[0118] The memory 220, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and various modules, such as the program or module corresponding to the document recommendation question generation method of this disclosure embodiment. The processor 210 executes various functional applications and data processing of the document recommendation question generation device 200 by running the software program or module stored in the memory 220.

[0119] Input device 230 can be used to receive input digital numbers or signals. These signals may include key signals related to user settings and function control of the device / terminal / server. Output device 240 may include a display device such as a screen.

[0120] <Storage Medium Examples>

[0121] According to a fourth aspect of this disclosure, a non-volatile computer-readable storage medium is also provided, on which computer program instructions are stored, which, when executed by processor 210, implement the method for generating the document recommendation problem described above.

[0122] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical applications, or technical improvements to the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for generating document recommendation questions, characterized in that, include: Obtain the document to be processed and divide the document into levels to obtain text fragments at least two of the following levels: document level, chapter level, section level, and fragment level; Based on the pre-configured question generation parameters for each level, prompt words for text fragments at each level are generated. The question generation parameters include question types. For document-level and chapter-level questions, at least one of framework-type, intent-reasoning-type, and value-judgment-type questions are configured. For section-level questions, at least one of logical analysis-type, knowledge association-type, and method comparison-type questions are configured. For fragment-level questions, at least one of parameter verification-type, operation execution-type, and fact confirmation-type questions are configured. After obtaining the pre-configured question generation parameters for each level, the question generation parameters corresponding to each level and the text fragments at each level are input into the pre-built prompt word template to obtain the prompt words for the text fragments at each level. If there are multiple text fragments under a level, for each text fragment, the text fragment and the question generation parameters of the level to which the text fragment belongs are input into the prompt word template to obtain the prompt words for each text fragment. Based on the prompt words of each level of text fragment, a large language model is used to generate recommendation questions for each level of text fragment; When generating recommendation questions for each level of text fragment based on the prompt words of each level of text fragment using a large language model, the following steps are included: Based on the prompt words of each level of text fragment, a large language model is used to generate initial recommendation questions for each level of text fragment. The initial recommendation questions include: the content of the text fragment, the keywords in the text fragment, and at least one initial recommendation question generated for the text fragment. The initial recommendation questions for each level of text fragment are sequentially subjected to four screening operations: semantic similarity detection, format detection, coverage evaluation, and cognitive load, to obtain the recommendation questions for each level of text fragment. The initial recommendation process for each level of text fragments involves four screening operations: semantic similarity detection, format detection, coverage assessment, and cognitive load. These operations include the following steps: Iterate through each text segment, and for the currently iterated text segment, perform the following filtering operations in sequence: The semantic similarity between the initial recommendation questions of the current text segment is calculated using a sentence vector model. By setting a similarity threshold, initial recommendation questions with repetitive or highly similar expressions are eliminated. Get the text length of each initial question in the current text segment, and delete the initial questions whose text length is greater than the upper limit or less than the lower limit to ensure that the length of the final recommended questions is within a reasonable range; Coverage evaluation is performed on the remaining initial recommendation questions for the current text segment. Specifically, the keywords of each initial recommendation question for the current text segment are deduplicated to form a keyword library for the current text segment. The coverage of the remaining initial recommendation questions with respect to the keywords in the keyword library is calculated, where coverage = number of keywords covered by the remaining initial recommendation questions ÷ total number of keywords in the keyword library. When the coverage is greater than the coverage threshold, the remaining initial recommendation questions are used as the recommendation questions after coverage evaluation for the current text segment. When the coverage is less than or equal to the coverage threshold, the keywords that are not covered in the keyword library are searched, and initial recommendation questions including the uncovered keywords are regenerated as a supplement. All the initial recommendation questions corresponding to the supplemented current text segment are used as the recommendation questions after coverage evaluation for the current text segment. After assessing the coverage of the current text segment, the recommended questions are then filtered based on cognitive load, specifically including the following steps: 1) Problem Complexity Modeling: For each recommended question after coverage evaluation, calculate its comprehensive cognitive load score. The comprehensive cognitive load score is calculated based on the following weighted factors: Syntactic complexity: quantified by analyzing sentence length, number of nested clauses, and special sentence structure indicators; Concept density: The number of core domain terms, technical terms, or undefined concepts appearing in a statistical problem; Reasoning depth: Based on the predefined basic weights of the question type, and fine-tuned by combining the keywords of the reasoning steps implied in the question text; 2) Load threshold and dynamic adaptation: The system sets a baseline cognitive load score threshold. In the user configuration interface, users can select or input a difficulty preference factor, which is used to dynamically adjust the cognitive load score threshold of the actual application. 3) Load Filtering: Compare the comprehensive cognitive load score of the recommended questions after coverage assessment with the dynamically adjusted threshold. For questions with scores significantly higher than the threshold: Try to simplify and restate: Input the question as a new "question" into the large language model and add the instruction: "Please restate the following question in a more concise and easier-to-understand sentence, keeping the core query intent unchanged: [original question]". Replace the original high-load question with the simplified question, or keep it as a supplementary option along with the original question.

2. The method according to claim 1, characterized in that, After obtaining the recommendation questions for each of the aforementioned text fragment levels, the process also includes: The operation involves sorting the recommendation questions for each of the aforementioned text fragments.

3. The method according to claim 2, characterized in that, When ranking the recommendation questions for each of the aforementioned levels of text fragments, the following is included: Calculate the historical recommendation value and title similarity value of the recommendation question for each of the aforementioned text fragments; Based on the historical recommendation values ​​and title similarity values ​​of the recommendation questions for each level of text fragment, the recommendation questions for each level of text fragment are ranked.

4. A device for generating document recommendation questions, characterized in that, include: The hierarchical division module is used to acquire the document to be processed and perform hierarchical division on the document to be processed to obtain text fragments at least two of the following levels: document level, chapter level, section level, and fragment level. The prompt word generation module is used to obtain and generate prompt words for text fragments at each level based on pre-configured question generation parameters for each level. The question generation parameters include question types. For document-level and chapter-level questions, at least one of framework-type, intent-reasoning-type, and value-judgment-type questions are configured. For section-level questions, at least one of logical analysis-type, knowledge association-type, and method comparison-type questions are configured. For fragment-level questions, at least one of parameter verification-type, operation execution-type, and fact confirmation-type questions are configured. After obtaining the pre-configured question generation parameters for each level, the corresponding question generation parameters for each level and the text fragments at each level are input into the pre-built prompt word template to obtain the prompt words for the text fragments at each level. If there are multiple text fragments under a level, for each text fragment, the text fragment and the question generation parameters of the level to which the text fragment belongs are input into the prompt word template to obtain the prompt words for each text fragment. The recommendation question generation module is used to generate recommendation questions for each level of text fragment based on the prompt words of each level of text fragment and using a large language model. When generating recommendation questions for each level of text fragment based on the prompt words of each level of text fragment using a large language model, the following steps are included: Based on the prompt words of each level of text fragment, a large language model is used to generate initial recommendation questions for each level of text fragment. The initial recommendation questions include: the content of the text fragment, the keywords in the text fragment, and at least one initial recommendation question generated for the text fragment. The initial recommendation questions for each level of text fragment are sequentially subjected to four screening operations: semantic similarity detection, format detection, coverage evaluation, and cognitive load evaluation, to obtain the recommendation questions for each level of text fragment. The initial recommendation process for each level of text fragments involves four screening operations: semantic similarity detection, format detection, coverage assessment, and cognitive load assessment. These operations include the following steps: Iterate through each text segment, and for the currently iterated text segment, perform the following filtering operations in sequence: The semantic similarity between the initial recommendation questions of the current text segment is calculated using a sentence vector model. By setting a similarity threshold, initial recommendation questions with repetitive or highly similar expressions are eliminated. Get the text length of each initial question in the current text segment, and delete the initial questions whose text length is greater than the upper limit or less than the lower limit to ensure that the length of the final recommended questions is within a reasonable range; Coverage evaluation is performed on the remaining initial recommendation questions for the current text segment. Specifically, the keywords of each initial recommendation question for the current text segment are deduplicated to form a keyword library for the current text segment. The coverage of the remaining initial recommendation questions with respect to the keywords in the keyword library is calculated, where coverage = number of keywords covered by the remaining initial recommendation questions ÷ total number of keywords in the keyword library. When the coverage is greater than the coverage threshold, the remaining initial recommendation questions are used as the recommendation questions after coverage evaluation for the current text segment. When the coverage is less than or equal to the coverage threshold, the keywords that are not covered in the keyword library are searched, and initial recommendation questions including the uncovered keywords are regenerated as a supplement. All the initial recommendation questions corresponding to the supplemented current text segment are used as the recommendation questions after coverage evaluation for the current text segment. After assessing the coverage of the current text segment, the recommended questions are then filtered based on cognitive load, specifically including the following steps: 1) Problem Complexity Modeling: For each recommended question after coverage evaluation, calculate its comprehensive cognitive load score. The comprehensive cognitive load score is calculated based on the following weighted factors: Syntactic complexity: quantified by analyzing sentence length, number of nested clauses, and special sentence structure indicators; Concept density: The number of core domain terms, technical terms, or undefined concepts appearing in a statistical problem; Reasoning depth: Based on the predefined basic weights of the question type, and fine-tuned by combining the keywords of the reasoning steps implied in the question text; 2) Load threshold and dynamic adaptation: The system sets a baseline cognitive load score threshold. In the user configuration interface, users can select or input a difficulty preference factor, which is used to dynamically adjust the cognitive load score threshold of the actual application. 3) Load Filtering: Compare the comprehensive cognitive load score of the recommended questions after coverage assessment with the dynamically adjusted threshold. For questions with scores significantly higher than the threshold: Try to simplify and restate: Input the question as a new "question" into the large language model and add the instruction: "Please restate the following question in a more concise and easier-to-understand sentence, keeping the core query intent unchanged: [original question]". Replace the original high-load question with the simplified question, or keep it as a supplementary option along with the original question.

5. A device for generating document recommendation questions, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to implement the method of any one of claims 1 to 3 when executing the executable instructions.

6. A non-volatile computer-readable storage medium storing computer program instructions thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 3.

Images