An intelligent question setting method

Abstract

The application provides an intelligent question setting method and device, and relates to the technical field of artificial intelligence. The method comprises the following steps: obtaining a question setting request information input by a user; determining a target knowledge slice from a plurality of knowledge slices included in a knowledge base according to the question setting request information, and generating an initial question according to the target knowledge slice; generating retrieval information respectively according to a stem, a candidate correct answer and each candidate incorrect answer of the initial question, retrieving the knowledge base based on the retrieval information, and obtaining support information corresponding to each candidate answer; verifying the initial question according to the support information, and obtaining a final question according to a verification result. The application aims to improve the problem that the consistency between the generated question and the content of the knowledge base is low due to the fact that the related art does not perform source review on the generated initial question.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an intelligent question generation method. Background Technology

[0002] With the development of information technology and artificial intelligence, corporate training assessments, job skills certifications, and online education tests are increasingly adopting intelligent methods to generate assessment questions. Intelligent question generation systems can determine relevant knowledge content from business documents or pre-set question banks based on user-inputted keywords, question types, and the number of questions. They then use question generation models to generate question stems, candidate answers, and answer explanations, thereby reducing the workload of manually writing questions.

[0003] Current methods typically involve filtering or rewriting questions based on a pre-set question bank, or directly generating questions by inputting business documents into a question generation model, followed by model evaluation or manual review. The former is limited by the scope of the question bank, while the latter makes it difficult to independently verify the source support of each candidate answer, which can easily lead to inconsistencies between the generated questions and the knowledge base content. Furthermore, manual review increases processing volume and affects the efficiency of generating large batches of questions. Summary of the Invention

[0004] The embodiments disclosed in this application provide an intelligent question generation method, which aims to improve the problem that the generated questions have low consistency with the knowledge base content because the related technologies do not perform source verification on the generated initial questions.

[0005] The embodiments of this application adopt the following technical solutions: Firstly, it provides an intelligent question-generating method, including: Obtain the question request information input by the user; Based on the question request information, the target knowledge slice is determined from multiple knowledge slices included in the knowledge base, and an initial question is generated based on the target knowledge slice. The initial question includes a question stem, candidate correct answers, and at least one candidate incorrect answer. First retrieval information is generated based on the question stem and candidate correct answers, and second retrieval information is generated based on the question stem and each candidate incorrect answer. The knowledge base is then searched based on the first retrieval information and each second retrieval information to obtain the first supporting information corresponding to the candidate correct answer and the second supporting information corresponding to each candidate incorrect answer. The initial questions are verified based on the first supporting information and each of the second supporting information, and the target questions are obtained based on the verification results.

[0006] In one possible implementation of the first aspect, before obtaining the question request information input by the user, the method further includes: Retrieve the business documents to be processed; The business document is sliced ​​according to preset slicing rules to obtain multiple knowledge slices; Semantic anchor points are extracted from each knowledge slice, and each knowledge slice is vectorized to obtain the corresponding slice vector. A knowledge base is constructed based on multiple knowledge slices, the semantic anchors corresponding to each knowledge slice, and the slice vector.

[0007] In one possible implementation of the first aspect, determining a target knowledge slice from multiple knowledge slices included in the knowledge base based on question request information includes: Based on the question request information, keyword matching is performed on multiple knowledge slices to obtain the first candidate knowledge slice set; Based on the question request information, semantic matching processing is performed on multiple knowledge slices to obtain a second candidate knowledge slice set; The first and second candidate knowledge slice sets are merged and deduplicated to obtain the candidate knowledge slice set.

[0008] In one possible implementation of the first aspect, determining the target knowledge slice from multiple knowledge slices included in the knowledge base based on the question request information further includes: Determine the correlation between each candidate knowledge slice in the candidate knowledge slice set and the question request information; The candidate knowledge slices are sorted according to their relevance, and the target knowledge slice is determined from the candidate knowledge slice set based on the sorting results. Based on the target knowledge slices and question request information, generate question generation prompts and input the question generation prompts into the question generation model to obtain the initial questions.

[0009] In one possible implementation of the first aspect, first retrieval information is generated based on the question stem and candidate correct answers, and second retrieval information is generated based on the question stem and each candidate incorrect answer, including: The question stem and candidate correct answers are combined to obtain the first combined text, and the first combined text is vectorized to obtain the first retrieval information; The question stem and each candidate incorrect answer are combined to obtain multiple second combined texts. Each second combined text is then vectorized to obtain the second retrieval information corresponding to each candidate incorrect answer.

[0010] In one possible implementation of the first aspect, the knowledge base is searched based on the first search information and each of the second search information, including: The first similarity between the first retrieval information and each knowledge slice is determined, and the first supporting knowledge slice and the first support degree are determined based on each first similarity to obtain the first supporting information; For each second retrieval information, the second similarity between the corresponding second retrieval information and each knowledge slice is determined, and the second supporting knowledge slice and the second support degree are determined based on each second similarity to obtain the corresponding second supporting information.

[0011] In one possible implementation of the first aspect, the initial question is verified based on the first supporting information and each of the second supporting information, including: The first support level is compared with the first threshold, and the difference between the first support level and each second support level is compared with the second threshold. Under the condition that the second support degree meets the preset interference conditions, the slice association relationship between the corresponding second support knowledge slice and the first support knowledge slice is determined; Based on the comparison results of the first support level and the first threshold, the comparison results of each difference and the second threshold, and the correlation of the slices, the verification results of the initial question are determined.

[0012] In one possible implementation of the first aspect, updating the initial questions based on the verification results includes: Based on the verification results, identify the items in the initial question that failed verification; Based on the items that failed verification and the target knowledge slices, generate question adjustment information; Based on the question adjustment information, the question content corresponding to the items that failed verification in the initial question was updated.

[0013] In one possible implementation of the first aspect, when a flowchart is associated with the target knowledge slice, generating an initial question based on the target knowledge slice includes: Perform topology analysis on the flowchart to obtain flowchart topology data. Flowchart topology data includes node information and edge information representing the connection relationship between nodes. Node information includes node content and the position information of the node content in the flowchart. Based on the question request information and flowchart topology data, determine the target assessment content and the target image region corresponding to the target assessment content; The target image region is occluded or its content is replaced to obtain the target question image; Initial questions are generated based on the target question image and the target assessment content.

[0014] In one possible implementation of the first aspect, the method further includes: Associate and store the target question with the corresponding slice of target knowledge. In response to business document updates, the updated business document is sliced ​​to obtain multiple updated knowledge slices; The first node set and the second node set are constructed based on multiple updated knowledge slices and multiple original knowledge slices in the knowledge base, respectively. The corresponding node connection weights are determined based on the semantic similarity and semantic anchor point similarity between each updated knowledge slice and each original knowledge slice, and a bipartite graph of knowledge slices is obtained. Based on the knowledge slice bipartite diagram, determine the slice mapping relationship between each updated knowledge slice and each original knowledge slice, and based on the slice mapping relationship and the association relationship between the target question and the target knowledge slice, determine the historical questions associated with the business document update; When the slice mapping relationship is one-to-one, the historical questions are updated according to the corresponding updated knowledge slices; When the slice mapping relationship is not a one-to-one mapping relationship, the historical questions are regenerated based on the corresponding updated knowledge slices; If there is no corresponding updated knowledge slice for the original knowledge slice, the historical questions associated with the original knowledge slice will be set to invalid status.

[0015] The intelligent question generation method provided in this application constructs a knowledge base by slicing and vectorizing business documents. Based on keyword matching, semantic matching, and reordering, it determines target knowledge slices that match the question generation request information and generates initial questions based on these slices. Furthermore, it reconstructs retrieval information using the question stem and candidate answers, reverse-searches the knowledge base to obtain supporting information for each candidate answer, independently reviews the initial questions, and updates and re-verifies questions that fail the review, thereby improving the consistency between the generated questions and the source knowledge content, as well as the effectiveness of the candidate answers. In addition, by performing topological analysis and image processing on flowcharts, it can generate graphical questions containing process logic. By establishing a mapping relationship between new and old knowledge slices and updating, regenerating, or invalidating historical questions accordingly, the question bank can evolve synchronously with business document updates, improving the automation and reliability of intelligent question generation and question bank maintenance.

[0016] Secondly, an intelligent question-generating system is provided, the device comprising: The request retrieval module is used to retrieve the question request information input by the user; The question generation module is used to determine the target knowledge slice from multiple knowledge slices included in the knowledge base based on the question request information, and generate an initial question based on the target knowledge slice. The initial question includes a question stem, candidate correct answers, and at least one candidate incorrect answer. The supporting information determination module is used to generate first retrieval information based on the question stem and candidate correct answers, and generate second retrieval information based on the question stem and each candidate incorrect answer. Based on the first retrieval information and each second retrieval information, the knowledge base is searched to obtain the first supporting information corresponding to the candidate correct answer and the second supporting information corresponding to each candidate incorrect answer. The question verification module is used to verify the initial question based on the first supporting information and each of the second supporting information. If the initial question fails the verification, the initial question is updated based on the verification result, and the updated initial question is re-verified until the updated initial question passes the verification, thus obtaining the target question.

[0017] Thirdly, an electronic device is provided, comprising: a memory and at least one processor. The memory is communicatively connected to the processor. The memory is used to store computer program code, which includes computer instructions. When the processor executes the computer instructions, it causes the electronic device to perform the method as described in the first aspect.

[0018] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer instructions. When executed by a processor, these computer instructions are used to implement, for example, causing an electronic device to perform the method described in the first aspect.

[0019] Understandably, the technical effects of the second to fourth aspects refer to the technical effects of the first aspect and any of its embodiments, and will not be repeated here. Attached Figure Description

[0020] Figure 1 A flowchart illustrating the steps of an intelligent question generation method provided in this application embodiment; Figure 2 This is a block diagram of an intelligent question generation system provided in an embodiment of this application. Detailed Implementation

[0021] The technical solutions in some embodiments of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application are within the scope of protection of this application.

[0022] Unless the context otherwise requires, throughout the specification and claims, the term "comprising" is interpreted as open-ended and encompassing, meaning "including, but not limited to." In the description of the specification, terms such as "one embodiment," "some embodiments," "exemplary embodiment," "exemplary," or "some examples," etc., are intended to indicate that a particular feature, structure, material, or characteristic associated with that embodiment or example is included in at least one embodiment or example of this application. The illustrative representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, a particular feature, structure, material, or characteristic may be included in any suitable manner in any one or more embodiments or examples.

[0023] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this application, unless otherwise stated, "a plurality of" means two or more.

[0024] With the development of information technology, corporate training and employee assessment are gradually shifting from offline, manual methods to networked, digitalized, and intelligent approaches. Companies can use training and assessment systems to provide training content to employees in different positions and generate corresponding test questions based on job requirements, training progress, and assessment objectives to evaluate their mastery of business knowledge and job skills. In this process, the efficiency of test question generation and the consistency between the test content and the training knowledge directly affect the operational efficiency of the training and assessment system and the accuracy of the assessment results.

[0025] The knowledge content used in corporate training and assessment typically originates from business documents such as regulations, job operation manuals, business specifications, product descriptions, technical documents, and business process diagrams. These business documents are generally characterized by a large number of documents, complex content structures, dense technical terminology, and high update frequency. A single business document may also include multiple formats such as text, tables, and images. Intelligent question generation systems need to identify the knowledge content matching the question request from the relevant business documents and generate question data, including the question stem, answer, options, and explanations, based on the identified knowledge content. Therefore, improving the matching degree between knowledge content retrieval results and question requests, and ensuring the consistency between generated questions and source knowledge content, are problems that need to be solved in the intelligent question generation process.

[0026] In one feasible implementation of the relevant technology, a question bank can be pre-established, and existing questions in the question bank can be labeled with knowledge points, question types, and difficulty levels. Upon receiving a question generation request, matching existing questions are retrieved from the question bank according to the request. These retrieved existing questions can be directly selected for test paper generation, or the retrieved existing questions can be used as samples to input into the question generation model to generate new questions. This method can improve test paper generation efficiency by utilizing existing question data, but its operation depends on the pre-built question bank. For newly released regulations, training manuals, or new business content for which historical questions have not yet been established, corresponding basic questions need to be pre-written and organized, resulting in the question generation process being limited by the size and coverage of the question bank data. Furthermore, when rewriting or expanding based on existing questions, the generated questions may use the assessment content and question structure of the existing questions, easily leading to problems such as limited knowledge coverage or repetitive question content.

[0027] In another feasible implementation of the related technology, business documents or exam syllabi generated from business documents can be input into a large language model, and questions can be generated according to user-defined parameters such as question type, number of questions, and difficulty. The generated questions can be evaluated for correctness by the large language model, or the user can manually modify the question stems, answers, and options, and the differences between the modified and unmodified questions can be used to train the large language model. This method reduces reliance on a pre-set question bank, but the generated results of the large language model have a certain degree of uncertainty; the generated answers, options, or explanations may be inconsistent with the knowledge content in the business documents.

[0028] Furthermore, enterprise business documents are updated as business rules, operating procedures, and technical specifications change. During document updates, chapter adjustments, paragraph splits, paragraph merging, content modifications, or clause deletions may occur, resulting in differences in content structure, paragraph placement, and the number of knowledge units before and after the update. If historical questions still correspond to the content of the business documents before the update, the answers, explanations, or applicable conditions of the historical questions may be inconsistent with those in the updated business documents. Manually verifying historical questions against the updated business documents one by one requires processing a large amount of document content and question data; simply rebuilding the search index for the updated business documents is also insufficient to synchronously adjust the already stored historical questions, affecting the accuracy and maintenance efficiency of the question bank data.

[0029] Therefore, the intelligent question generation solutions in related technologies still have problems such as low consistency between generated questions and source knowledge content, insufficient verification accuracy of candidate answers, low efficiency in maintaining historical questions after business document updates, and insufficient utilization of image-based knowledge content.

[0030] To address the aforementioned issues, this application provides an intelligent question generation method. This method involves determining a target knowledge slice from a knowledge base based on a question request, generating initial questions based on the target knowledge slice, constructing retrieval information based on the question stem, candidate correct answers, and candidate incorrect answers of the initial questions, verifying the initial questions using supporting information retrieved from the knowledge base, and updating and re-verifying initial questions that fail verification. This improves the consistency between the generated questions and the source knowledge content, as well as the effectiveness of the candidate answers.

[0031] The intelligent question generation method provided in this application can be applied to scenarios such as enterprise employee training and assessment, job skills certification, online education testing, professional knowledge assessment, and automatic question bank construction.

[0032] The execution entity of the intelligent question generation method provided in this application embodiment can be a server, server cluster, cloud computing platform, or other electronic device with data processing capabilities. In one feasible implementation, the execution entity is a server, which can communicate with the user terminal and the knowledge base used to store knowledge slices to receive question generation request information sent by the user terminal and perform processing such as knowledge slice retrieval, question generation, question verification, and question updating.

[0033] Figure 1 This is a flowchart illustrating an intelligent question generation method provided in an embodiment of this application, with the server as the executing entity. For example, please refer to [link to example]. Figure 1 As shown, the method may include S101 to S105.

[0034] S101: Obtain the question request information input by the user.

[0035] Users can input or select question-generating conditions through the question-generating interface provided on the user terminal. The user terminal then generates a question-generating request based on the conditions and sends the request to the server. The user terminal can be a computer, tablet, smartphone, or other electronic device capable of interacting with the server.

[0036] The question request information may include at least one of the following: target knowledge base, question keywords, number of questions, question type, question difficulty, and question creation rules. The target knowledge base is used to limit the search scope of knowledge slices; question keywords may include knowledge point names, chapter names, business terms, or natural language descriptions entered by the user; question types may include at least one of the following: true / false questions, multiple choice questions, fill-in-the-blank questions, and graphic reasoning questions; question creation rules may be used to limit the question content, answer format, number of options, or output format.

[0037] In some embodiments, the user terminal can encapsulate the question request information into a request message and send the request message to the server through a preset communication interface.

[0038] S102: Based on the question request information, determine the target knowledge slice from multiple knowledge slices included in the knowledge base, and generate initial questions based on the target knowledge slice.

[0039] Knowledge slices can be text data obtained by slicing business documents. Each knowledge slice can be associated with information such as source document identifier, chapter information, slice identifier, semantic anchors, and slice vectors. The server can determine the corresponding knowledge base and search content based on the question request information, and determine the target knowledge slice from multiple knowledge slices included in the knowledge base based on the search content.

[0040] In some embodiments, the server can perform keyword retrieval on multiple knowledge slices based on question request information to obtain a first candidate knowledge slice set; it can also perform vectorization processing on the question request information and perform semantic retrieval on multiple knowledge slices based on the vectorization processing results to obtain a second candidate knowledge slice set. The server merges, deduplicates, and filters the first and second candidate knowledge slice sets to obtain a candidate knowledge slice set.

[0041] Furthermore, the server can determine the correlation between each candidate knowledge slice in the candidate knowledge slice set and the question request information, reorder each candidate knowledge slice according to the correlation, and determine a preset number of target knowledge slices from the reordered candidate knowledge slices. Thus, the target knowledge slices can include knowledge content that directly matches the question keywords, or knowledge content that is semantically related to the question request information.

[0042] The server constructs question generation hints based on the target knowledge slice and question request information, and inputs these hints into the question generation model to obtain initial questions. The initial questions include a question stem, candidate correct answers, and at least one candidate incorrect answer. In some embodiments, the initial questions may also include question type, answer explanation, and source knowledge slice identifier. The server can generate one or more initial questions based on the number and type of questions indicated in the question request information.

[0043] S103: Generate first retrieval information based on the question stem and candidate correct answers, and generate second retrieval information based on the question stem and each candidate incorrect answer. Search the knowledge base based on the first retrieval information and each second retrieval information to obtain the first supporting information corresponding to the candidate correct answer and the second supporting information corresponding to each candidate incorrect answer.

[0044] Specifically, the server can combine the question stem with each candidate correct answer to obtain a first combined text, and then vectorize the first combined text to obtain first retrieval information. The server can also combine the question stem with each candidate incorrect answer separately to obtain a second combined text corresponding to each candidate incorrect answer, and then vectorize each second combined text separately to obtain multiple second retrieval information.

[0045] The server determines the first similarity between the first combined text and each knowledge slice in the knowledge base based on the first retrieval information, and identifies the first similarity that meets the first screening criteria and the corresponding knowledge slice as the first supporting information. The first supporting information may include a first support degree and a first supporting knowledge slice. The first support degree characterizes the degree of matching between the candidate correct answer and the corresponding knowledge content in the knowledge base, and the first supporting knowledge slice characterizes the knowledge source corresponding to the candidate correct answer.

[0046] For each second retrieval piece of information, the server determines the second similarity between the corresponding second combined text and each knowledge slice in the knowledge base, and identifies the second similarity that meets the second filtering criteria and the corresponding knowledge slice as the second supporting information for the corresponding candidate incorrect answer. The second supporting information may include a second support degree and a second supporting knowledge slice. The second support degree characterizes the degree of matching between the corresponding candidate incorrect answer and the knowledge content in the knowledge base, and the second supporting knowledge slice characterizes the knowledge content matched by the corresponding candidate incorrect answer.

[0047] As a feasible implementation, the first support score can be the highest similarity between the first combined text and each knowledge slice, and the first supporting knowledge slice can be the knowledge slice corresponding to the highest similarity score; the second support score can be the highest similarity between the corresponding second combined text and each knowledge slice, and the second supporting knowledge slice can be the knowledge slice corresponding to the highest similarity score. The similarity score can be determined using cosine similarity or other parameters that can characterize the semantic relevance of the text.

[0048] S104: Verify the initial question based on the first supporting information and each of the second supporting information, and obtain the target question based on the verification results.

[0049] In this step, the server can determine whether a candidate correct answer has matching source knowledge content based on the first supporting information, and determine the degree of correlation between the corresponding candidate incorrect answer and the source knowledge content based on each second supporting information. The server can also determine whether each candidate incorrect answer meets the preset option differentiation conditions based on the relationship between the first supporting knowledge slice and each second supporting knowledge slice.

[0050] In some embodiments, the server may compare a first support level with a first threshold, determine the support level difference between the first support level and each second support level, and compare each support level difference with a second threshold. If the first support level is not less than the first threshold, and the support level difference between the first support level and each second support level is not less than the second threshold, it indicates that the candidate correct answer has corresponding knowledge content support, and the candidate correct answer has a high degree of content matching relative to each candidate incorrect answer.

[0051] If the second support level corresponding to a candidate incorrect answer meets the preset interference conditions, the server can compare the corresponding second support knowledge slice with the first support knowledge slice. When the second support knowledge slice differs from the first support knowledge slice, the server can further determine whether the candidate incorrect answer references other real knowledge content based on the business object, applicable conditions, or knowledge content corresponding to the two knowledge slices, and determine the candidate incorrect answer that meets the option differentiation conditions as a valid candidate incorrect answer.

[0052] The server determines the verification result of the initial question based on the comparison results of the first support level, the comparison results of the differences between each support level, and the differentiation results of the options corresponding to each candidate incorrect answer. If both the candidate correct answer and each candidate incorrect answer meet the corresponding verification conditions, the initial question is determined to have passed verification, and the verified initial question is designated as the target question.

[0053] If the initial question fails validation, the server determines the reason for the validation failure based on the unmet validation conditions, and generates question adjustment information based on the reason for the failure and the target knowledge slice. The server updates the corresponding question stem, candidate correct answers, or candidate incorrect answers in the initial question according to the question adjustment information, and re-executes the retrieval and validation process on the updated initial question until the updated initial question passes validation and the target question is obtained.

[0054] The intelligent question generation method provided in this application generates initial questions based on target knowledge slices, and reconstructs retrieval information using the question stem and each candidate answer. It then reverse-searches the knowledge base to obtain supporting information for each candidate answer, thereby achieving independent verification of the initial questions. For initial questions that fail the verification, the method updates and re-verifies them. This improves upon the problem in related technologies where the generated initial questions are not verified for their source, resulting in a lack of knowledge content support for the candidate answers. This enhances the consistency between the generated questions and the knowledge base content, as well as the reliability of the question generation results.

[0055] Before obtaining the user's input request information, the server can perform knowledge base preparation processing. In one feasible implementation, the server obtains the business document to be processed, which may include company rules and regulations, job operation manuals, business specifications, product descriptions, or technical documents. The business document can be a text file or a composite document containing text, tables, and flowcharts.

[0056] The server performs format recognition and content extraction on business documents, and then segments the extracted document content according to preset slicing rules to obtain multiple knowledge slices. These preset slicing rules can be used to determine the slice boundaries of knowledge slices based on at least one of the following: chapter title, paragraph level, clause number, table markers, and text length. As an example, the server can prioritize dividing the document content according to chapter titles and table markers, controlling the text length of each knowledge slice to within 512 characters, and maintaining a 50-character overlap between adjacent knowledge slices to reduce the loss of contextual information caused by slice boundaries.

[0057] The server assigns a corresponding slice identifier to each knowledge slice and records the source document identifier, document version, chapter position, and text content of each knowledge slice. The server can also extract chapter paths, hierarchical numbers, and core business entities from each knowledge slice to obtain the semantic anchor points of the corresponding knowledge slices. Semantic anchor points are used to represent the structural position of a knowledge slice within a business document and its corresponding core knowledge content.

[0058] The server inputs each knowledge slice into the embedding model for vectorization, obtaining the corresponding slice vector. The slice vector is used to represent the semantic features of the corresponding knowledge slice, enabling knowledge slices with different expressions but similar knowledge meanings to be matched through vector similarity.

[0059] Furthermore, the server associates and stores each knowledge slice with its corresponding slice identifier, source document information, semantic anchors, and slice vectors. It also builds a keyword index based on the text content of multiple knowledge slices and a vector index based on the slice vectors of multiple knowledge slices to construct a knowledge base. In some embodiments, the slice vectors and vector indexes can be stored in a vector database, while the text content, source information, and semantic anchors of the knowledge slices can be stored in a relational database. A correspondence between vector data and relational data is established through the slice identifiers.

[0060] For example, the business document to be processed is the "Job Safety Operation Specifications," which includes chapters such as "Work Preparation," "Equipment Startup," "Anomaly Handling," and "Shutdown Inspection." For the knowledge slices in the "Equipment Startup" chapter, the server can extract the corresponding chapter path, clause number, and equipment name as semantic anchors, and store the text content, slice vector, semantic anchors, and source chapters of the knowledge slice in association. After completing the above processing, the server enters the question generation stage.

[0061] After receiving the question request information input by the user, the server can extract the target knowledge base, question keywords, number of questions, question types, and question creation rules from the request information. The server determines the retrieval scope of knowledge slices based on the target knowledge base and generates target retrieval information based on the question keywords or query text expanded from the question keywords.

[0062] The server performs keyword matching on multiple knowledge slices in the knowledge base based on the target retrieval information. As a feasible implementation, the server can calculate the keyword relevance score between the target retrieval information and each knowledge slice using the BM25 algorithm based on a pre-established keyword index. The keyword relevance score can be determined using formula (1): in, Representing knowledge slices Relative to target retrieval information Keyword relevance score; Indicates keywords in the target search information; Keywords In knowledge slices Word frequency in; Keywords Inverse document frequency; Representing knowledge slices The length of the text; This represents the average text length of multiple knowledge slices in the knowledge base. and This is the smoothing parameter. As an example... It can be set to 1.5. It can be set to 0.75.

[0063] The server sorts multiple knowledge slices based on keyword relevance scores and selects a preset number of knowledge slices from the sorted list to obtain a first candidate knowledge slice set. Keyword matching processing can be used to match content with clear textual characteristics, such as clause numbers, technical terms, equipment names, and business abbreviations.

[0064] The server can also vectorize the target retrieval information to obtain a query vector, and determine the vector similarity between the query vector and each slice vector in the knowledge base. The server sorts multiple knowledge slices based on vector similarity and selects a predetermined number of knowledge slices from the sorted list to obtain a second candidate knowledge slice set. Semantic matching processing can be used to retrieve knowledge content that differs in textual expression from the target retrieval information but is semantically related.

[0065] As a possible implementation, the first candidate knowledge slice set and the second candidate knowledge slice set can each include the top-ranked ones. A slice of knowledge The value is a positive integer. The server merges the first and second candidate knowledge slice sets and performs deduplication based on the slice identifier or the original text of the knowledge slice. For knowledge slices with missing text content, text lengths that do not meet preset requirements, or invalid knowledge content, the server can remove them from the merged knowledge slice set to obtain a candidate knowledge slice set.

[0066] The server combines each candidate knowledge slice in the candidate knowledge slice set with the target retrieval information to form unranked data pairs. These unranked data pairs are then input into a re-ranking model to obtain the correlation between each candidate knowledge slice and the question request information. The re-ranking model can perform joint semantic analysis on the question request information and candidate knowledge slices to further filter the keyword matching results and semantic matching results.

[0067] The server reorders the candidate knowledge slices based on their relevance and selects the top-ranked slices from the reordered list. Each candidate knowledge slice is used as the target knowledge slice. Therefore, the target knowledge slice can include knowledge content that directly matches the keywords used in the question, or it can include knowledge content that is semantically related to the question request information.

[0068] The server generates question generation hints based on the target knowledge slice and the question request information. The question generation hints may include the text content and source information of the target knowledge slice, and may also include the number of questions, question type, question difficulty, number of candidate answers, and output format indicated in the question request information.

[0069] The server inputs question generation hints into the question generation model to obtain an initial question. The initial question includes a stem, candidate correct answers, and at least one candidate incorrect answer. It may also include the question type, answer explanation, and source knowledge slice identifier. The server can receive question data output from the question generation model according to a preset question data format and perform field integrity checks on the question data. When the question data lacks a stem, candidate correct answers, or candidate incorrect answers, the server can supplement the question generation hints and re-invoke the question generation model.

[0070] In some embodiments, the server may receive the data output by the question generation model in a streaming manner, store multiple data fragments output by the question generation model into a streaming output buffer, and perform field parsing and integrity verification after obtaining complete question data.

[0071] When a target knowledge slice is associated with a flowchart, or when a question request indicates the generation of a text-based reasoning question, the server can generate the corresponding initial question based on the flowchart. Specifically, the server obtains the flowchart associated with the target knowledge slice and inputs it into a multimodal visual model. The multimodal visual model then performs topological parsing on the flowchart to obtain the flowchart topology data.

[0072] Flowchart topology data can include a set of nodes and a set of edges. Each node in the node set can include a node identifier, node content, node type, and the location information of the node content in the flowchart. Each edge in the edge set can include a source node identifier, a target node identifier, and condition information, used to represent the connection relationship between the corresponding nodes and the flow jump conditions. The location information of the node content can be represented by a bounding box, which can include the x-coordinate, y-coordinate, width, and height of the target node in the flowchart.

[0073] As an example, flowchart topology data can be stored in JSON format. Node data can include node identifier, node text, bounding box coordinates, and node type. Edge data can include source node identifier, target node identifier, and condition text.

[0074] Based on the question request information and flowchart topology data, the server determines the target assessment content from node content, node connection relationships, and conditional branches, and determines the target image region based on the location information corresponding to the target assessment content. The target assessment content can be key processing nodes, decision nodes, connection paths between nodes, or conditional branches.

[0075] The server performs occlusion or content replacement processing on the target image region to obtain the target question image. For example, when the target assessment content is a decision node in a flowchart, the server can occlude the bounding box region corresponding to the decision node and use the occluded node content as a candidate correct answer; when the target assessment content is a conditional branch, the server can occlude or replace the conditional text with errors and use the correct conditional text or flowchart path as a candidate correct answer.

[0076] The server generates initial questions in image and text format based on the target question image and the target assessment content, and associates the storage path of the target question image, the original flowchart identifier, the target image region, and the source knowledge slice identifier with the initial questions. These initial questions in image and text format can also proceed to the subsequent question review process.

[0077] After obtaining the initial question, the server does not directly determine the initial question as the final question result. Instead, it reconstructs the retrieval information based on the question stem of the initial question and each candidate answer in order to reverse search the knowledge base and verify the initial question.

[0078] Specifically, let the initial problem stem be... The correct answer is , No. The number of incorrect answers is The knowledge slice set in the knowledge base is The embedding function is The similarity function is .

[0079] The server responds to the question. and candidate correct answers The system performs a combination process to obtain the first combined text, and then vectorizes this first combined text to obtain the first retrieval information. This combination process can be either text concatenation or a structured combination of the question stem and candidate correct answers according to a preset template.

[0080] The server uses the first retrieval information to search the knowledge base, determines the first similarity between the first combined text and each knowledge slice, and determines the first support degree corresponding to the candidate correct answer based on each first similarity. The first support level can be determined using formula (2): in, This indicates a text combination operation; This represents any knowledge slice in the knowledge base. The knowledge slice with the highest similarity to the first combined text can be identified as the first supporting knowledge slice. The first supporting information includes the first supporting degree and the first supporting knowledge slice, which are used to characterize the knowledge source of the candidate correct answer and the degree of matching between the candidate correct answer and the source knowledge content.

[0081] The server processed the question stem separately. The text is combined with each candidate incorrect answer to obtain multiple second combined texts. Each second combined text is then vectorized to obtain the second retrieval information corresponding to each candidate incorrect answer.

[0082] For each second retrieval piece of information, the server determines the second similarity between the corresponding second combined text and each knowledge slice in the knowledge base, and determines the second support degree corresponding to the corresponding candidate incorrect answer based on each second similarity. The second support level corresponding to each candidate incorrect answer It can be determined using formula (3): ; With the The knowledge slice with the highest similarity to the second combined text can be identified as the corresponding second supporting knowledge slice. The second support information includes the second support level. And the corresponding second supporting knowledge slice, used to characterize the first The degree of correlation between each candidate incorrect answer and the knowledge content in the knowledge base.

[0083] In some embodiments, the first and second supporting information may further include the source document identifier, chapter position, semantic anchor point, and text content of the corresponding knowledge slice, so that the server can further determine the relationship between the knowledge content matched by the candidate correct answer and the candidate incorrect answer. The similarity function can be a cosine similarity function, or other similarity functions that can characterize the degree of semantic relevance of the text.

[0084] The server verifies the degree of source support for the candidate correct answer and the degree of differentiation between the candidate correct answer and each candidate incorrect answer based on the first supporting information and each second supporting information.

[0085] Specifically, the server will provide the first level of support. With the first threshold Compare and determine the first support level for each. With each second support degree The support difference between them is compared with the second threshold. The comparison is then performed. The corresponding verification conditions can be expressed by formula (4): ; in, Indicates the number of candidate incorrect answers; first threshold Used to limit the minimum level of knowledge required for a candidate correct answer; second threshold Used to limit the minimum support difference required between candidate correct answers and candidate incorrect answers.

[0086] When the first support level is less than the first threshold, it indicates that the candidate correct answer lacks sufficient source knowledge content support; when the difference between the first support level and any second support level is less than the second threshold, it indicates that the corresponding candidate incorrect answer is not sufficiently distinguishable from the candidate correct answer, and the initial question may have an unclear answer.

[0087] Furthermore, the server can determine whether the corresponding candidate incorrect answer meets the preset interference conditions based on the second support level. The preset interference conditions can be used to limit the degree of correlation between the candidate incorrect answer and other real knowledge content in the knowledge base, while the second support level corresponding to the candidate incorrect answer is lower than the first support level corresponding to the candidate correct answer.

[0088] When the second support level meets the preset interference conditions, the server determines the slice association relationship between the corresponding second support knowledge slice and the first support knowledge slice. The slice association relationship can be determined based on at least one of the following: slice identifier, source chapter, semantic anchor, business object, and applicable conditions.

[0089] When the second supporting knowledge slice differs from the first supporting knowledge slice, and the second supporting knowledge slice corresponds to other real knowledge content in the knowledge base, it indicates that the corresponding candidate incorrect answer can use other knowledge content to create interference. This candidate incorrect answer can then be identified as a candidate incorrect answer that meets the option differentiation criteria. If a candidate incorrect answer lacks a corresponding second supporting knowledge slice, or the slice association between the second and first supporting knowledge slices does not meet the preset requirements, then the corresponding candidate incorrect answer can be identified as a failed verification item.

[0090] The server determines the verification result of the initial question based on the comparison results of the first support level and the first threshold, the comparison results of the differences in each support level and the second threshold, and the slice association relationship between each second support knowledge slice and the first support knowledge slice. When both the candidate correct answer and each candidate incorrect answer meet the corresponding verification conditions, the server determines that the initial question has passed the verification and designates the verified initial question as the target question.

[0091] When the initial question fails validation, the server determines the validation failure items of the initial question based on the unmet validation conditions. Validation failure items may include at least one of the following: insufficient support from candidate correct answers, insufficient discrimination of candidate answers, lack of effective knowledge connection between candidate incorrect answers, and mismatch between the question stem and candidate answers.

[0092] The server generates question adjustment information based on the failed verification items and the target knowledge slice. This adjustment information may include the question content corresponding to the failed verification item, the reason for the failure, the text content of the target knowledge slice, and the corresponding adjustment requirements.

[0093] For example, when the first support degree of a candidate correct answer is less than the first threshold, the question adjustment information can be used to instruct the regeneration of the candidate correct answer or the adjustment of the question stem based on the target knowledge slice; when the difference between the first support degree and any second support degree is less than the second threshold, the question adjustment information can be used to instruct the regeneration of the corresponding candidate incorrect answer; when the candidate incorrect answer does not meet the preset interference conditions, the question adjustment information can be used to instruct the regeneration of the corresponding candidate incorrect answer based on other knowledge slices that are different from the first supporting knowledge slice.

[0094] The server inputs the question adjustment information as feedback into the question generation model, and performs partial updates to the question content corresponding to the items that failed validation in the initial questions. Partial updates may include updating at least one of the following: the question stem, candidate correct answers, candidate incorrect answers, and answer explanations, without needing to regenerate the entire content of the initial questions.

[0095] The server regenerates the first and second search information for the updated initial question, and re-executes the knowledge base search and question verification process. If the updated initial question still fails verification, the server generates and updates the question based on the new failed verification items, until the updated initial question passes verification and the target question is obtained.

[0096] After obtaining the target question, the server associates and stores the target question with the corresponding target knowledge slice to establish a correspondence between the question data and the source knowledge content. For text and image questions, the server can also associate and store the storage path of the target question stem image, the original flowchart identifier, and the target image area with the target question.

[0097] In some embodiments, the server can set up a knowledge base table, a knowledge slice table, a question table, and a generation log table in a relational database. The knowledge base table stores the knowledge base identifier, knowledge base name, and knowledge base status; the knowledge slice table stores the knowledge slice identifier, knowledge slice original text, semantic anchors, and source document information; the question table stores the question type, question stem, candidate answers, answer explanations, source knowledge slice identifier, and question status; and the generation log table records question generation, question verification, and question update information.

[0098] Based on the association between the target question and the target knowledge slice, when the server receives the updated business document, it can trigger the update process of the knowledge base and related questions. The server performs slice processing on the updated business document according to the slicing rules corresponding to the original business document, obtains multiple updated knowledge slices, and extracts the semantic anchor points and slice vectors of each updated knowledge slice.

[0099] Let the set of original knowledge slices in the knowledge base be . Update the knowledge slice set to The server uses a collection of raw knowledge slices. and updating the knowledge slice set Construct a bipartite graph of knowledge slices, using each slice as a separate set of nodes. and any updated knowledge slice The server determines the corresponding node connection weights based on the vector similarity and semantic anchor similarity between the original knowledge slice and the updated knowledge slice. The node connection weights can be determined using formula (5): in, Represents a slice of original knowledge With updated knowledge slices The node connection weights between them; Indicates vector similarity; Indicates edit distance; Represents a slice of original knowledge semantic anchors; This indicates an update to the knowledge slice. semantic anchors; This represents the weight coefficients corresponding to vector similarity. As an example... It can be set to 0.6.

[0100] The server filters the connection edges in the bipartite graph of knowledge slices based on the connection weights of each node, and determines the slice mapping relationship between each original knowledge slice and each updated knowledge slice based on the filtering results. Slice mapping relationships can include one-to-one mapping relationships, non-one-to-one mapping relationships, and no mapping relationship.

[0101] Among them, a one-to-one mapping relationship is used to represent that an original knowledge slice corresponds to an updated knowledge slice; a non-one-to-one mapping relationship can include one original knowledge slice corresponding to multiple updated knowledge slices, multiple original knowledge slices corresponding to one updated knowledge slice, or multiple original knowledge slices corresponding to multiple updated knowledge slices; no mapping relationship is used to represent that an original knowledge slice does not have a corresponding updated knowledge slice, or an updated knowledge slice does not have a corresponding original knowledge slice.

[0102] After the server determines the slice mapping relationship, it queries the historical questions associated with each original knowledge slice based on the source knowledge slice identifier recorded in the question table, and processes the historical questions according to the corresponding slice mapping relationship.

[0103] When the slice mapping relationship is one-to-one, it indicates that the corresponding knowledge content has mainly undergone local modifications. The server can retain the question stems of historical questions, update the candidate answers and answer explanations according to the corresponding updated knowledge slices, and then re-generate and review the updated historical questions. After the updated historical questions pass verification, the server re-establishes the association between them and the corresponding updated knowledge slices.

[0104] When the slice mapping relationship is not one-to-one, it indicates that the corresponding knowledge content has been split, merged, or restructured. The server regenerates historical questions based on one or more updated knowledge slices that have a mapping relationship with the original knowledge slices, and performs post-generation verification on the regenerated historical questions. After the regenerated historical questions pass verification, the server stores them in association with the corresponding updated knowledge slices.

[0105] If no corresponding updated knowledge slice exists for the original knowledge slice, it indicates that the relevant knowledge content has been deleted from the updated business document. The server will set the historical questions associated with the original knowledge slice to an invalid state to prevent the corresponding historical questions from being used in subsequent assessments.

[0106] If the updated knowledge slice does not have a corresponding original knowledge slice, it indicates that the updated business document has added new knowledge content. The server can write the updated knowledge slice into the knowledge base, and when it subsequently receives question generation requests related to the updated knowledge slice, it can generate corresponding questions based on the updated knowledge slice.

[0107] After completing the business document update process, the server updates the corresponding keyword index and vector index, and records the content before and after the update, the slice mapping relationship, and the processing results for each historical question. When knowledge slice retrieval fails, question generation times out, or question data storage is abnormal, the server can record the corresponding exception information and the current processing status to support the re-execution of the corresponding processing procedure.

[0108] The intelligent question generation method provided in this application constructs a knowledge base by slicing and vectorizing business documents. Based on keyword matching, semantic matching, and reordering, it determines target knowledge slices that match the question generation request information and generates initial questions based on these slices. Furthermore, it reconstructs retrieval information using the question stem and candidate answers, reverse-searches the knowledge base to obtain supporting information for each candidate answer, independently reviews the initial questions, and updates and re-verifies questions that fail the review, thereby improving the consistency between the generated questions and the source knowledge content, as well as the effectiveness of the candidate answers. In addition, by performing topological analysis and image processing on flowcharts, it can generate graphical questions containing process logic. By establishing a mapping relationship between new and old knowledge slices and updating, regenerating, or invalidating historical questions accordingly, the question bank can evolve synchronously with business document updates, improving the automation and reliability of intelligent question generation and question bank maintenance.

[0109] Figure 2 This is a block diagram of an intelligent question-generating system provided in an embodiment of this application. Figure 2 As shown, the intelligent question-generating system includes: The request retrieval module 201 is used to retrieve the question request information input by the user; The question generation module 202 is used to determine the target knowledge slice from multiple knowledge slices included in the knowledge base according to the question request information, and generate an initial question based on the target knowledge slice. The initial question includes a question stem, candidate correct answers, and at least one candidate incorrect answer. The supporting information determination module 203 is used to generate first retrieval information based on the question stem and candidate correct answers, and generate second retrieval information based on the question stem and each candidate incorrect answer. Based on the first retrieval information and each second retrieval information, the knowledge base is searched to obtain the first supporting information corresponding to the candidate correct answer and the second supporting information corresponding to each candidate incorrect answer. The question verification module 204 is used to verify the initial question based on the first supporting information and each of the second supporting information. If the initial question fails the verification, the initial question is updated based on the verification result, and the updated initial question is re-verified until the updated initial question passes the verification, thus obtaining the target question.

[0110] The intelligent question-generating system provided in this application can execute the methods shown in the above method embodiments. Its implementation principle and beneficial effects can be found in the relevant descriptions in the method embodiments, and will not be repeated here. Furthermore, each module in the above intelligent question-generating system can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0111] In other embodiments, an electronic device is provided, which may be the intelligent question-generating system in the above method embodiments, for executing the method steps performed by the terminal in the above method flow.

[0112] This application also provides a computer-readable storage medium storing computer instructions. When these computer instructions are executed by a processor, they are used to implement the technical solution of the intelligent question generation method described in the above embodiments. Specifically, when the computer instructions are executed by a processor, the computer device can execute the technical solution of the intelligent question generation method provided in the above embodiments.

[0113] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the technical solution of the intelligent question generation method provided in the above embodiments.

[0114] The technical parameters in the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical parameters in the above embodiments are described. However, as long as these combinations of technical parameters do not contradict each other, they should be considered within the scope of this specification. The above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application's patent. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical parameters in the formula. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. An intelligent question generation method, characterized in that, include: Obtain the question request information input by the user; Based on the question request information, a target knowledge slice is determined from multiple knowledge slices included in the knowledge base, and an initial question is generated based on the target knowledge slice. The initial question includes a question stem, candidate correct answers, and at least one candidate incorrect answer. First retrieval information is generated based on the question stem and the candidate correct answers, and second retrieval information is generated based on the question stem and each of the candidate incorrect answers. The knowledge base is then searched based on the first retrieval information and each of the second retrieval information to obtain first supporting information corresponding to the candidate correct answers and second supporting information corresponding to each of the candidate incorrect answers. The initial question is verified based on the first supporting information and each of the second supporting information, and the target question is obtained based on the verification results.

2. The intelligent question generation method according to claim 1, characterized in that, Before obtaining the question request information input by the user, the method further includes: Retrieve the business documents to be processed; The business document is sliced ​​according to preset slicing rules to obtain multiple knowledge slices; Semantic anchor points are extracted from each knowledge slice, and each knowledge slice is vectorized to obtain the corresponding slice vector. The knowledge base is constructed based on the multiple knowledge slices, the semantic anchors corresponding to each knowledge slice, and the slice vector.

3. The intelligent question generation method according to claim 1, characterized in that, The step of determining the target knowledge slice from multiple knowledge slices included in the knowledge base based on the question request information includes: Based on the question request information, keyword matching processing is performed on the multiple knowledge slices to obtain a first candidate knowledge slice set; Based on the question request information, semantic matching processing is performed on the multiple knowledge slices to obtain a second candidate knowledge slice set; The first candidate knowledge slice set and the second candidate knowledge slice set are merged and deduplicated to obtain a candidate knowledge slice set.

4. The intelligent question generation method according to claim 3, characterized in that, The step of determining the target knowledge slice from multiple knowledge slices included in the knowledge base based on the question request information further includes: Determine the correlation between each candidate knowledge slice in the candidate knowledge slice set and the question request information; The candidate knowledge slices are sorted according to the correlation degree, and the target knowledge slice is determined from the candidate knowledge slice set according to the sorting result. Based on the target knowledge slice and the question request information, a question generation prompt is generated, and the question generation prompt is input into the question generation model to obtain the initial question.

5. The intelligent question generation method according to claim 1, characterized in that, The step of generating first search information based on the question stem and the candidate correct answers, and generating second search information based on the question stem and each of the candidate incorrect answers, includes: The question stem and the candidate correct answers are combined to obtain a first combined text, and the first combined text is vectorized to obtain the first retrieval information; The question stem and each of the candidate incorrect answers are combined to obtain multiple second combined texts, and each of the second combined texts is vectorized to obtain second retrieval information corresponding to each of the candidate incorrect answers.

6. The intelligent question generation method according to claim 5, characterized in that, The step of searching the knowledge base based on the first search information and each of the second search information includes: The first similarity between the first retrieval information and each of the knowledge slices is determined respectively, and the first supporting knowledge slice and the first support degree are determined based on each of the first similarities to obtain the first supporting information; For each of the second retrieval information, a second similarity between the corresponding second retrieval information and each of the knowledge slices is determined, and a second supporting knowledge slice and a second support degree are determined based on each of the second similarities to obtain the corresponding second supporting information.

7. The intelligent question generation method according to claim 6, characterized in that, The step of verifying the initial question based on the first supporting information and each of the second supporting information includes: The first support level is compared with the first threshold, and the difference between the first support level and each of the second support levels is compared with the second threshold. When the second support degree meets the preset interference conditions, the slice association relationship between the corresponding second support knowledge slice and the first support knowledge slice is determined. The verification result of the initial question is determined based on the comparison results of the first support degree and the first threshold, the comparison results of each difference and the second threshold, and the correlation relationship of the slices.

8. The intelligent question generation method according to claim 7, characterized in that, The step of updating the initial question based on the verification results includes: Based on the verification results, identify the items in the initial question that failed verification; Based on the failed verification items and the target knowledge slices, generate question adjustment information; Based on the question adjustment information, the question content corresponding to the verification failure item in the initial question is updated.

9. The intelligent question generation method according to claim 1, characterized in that, When the target knowledge slice is associated with a flowchart, the step of generating initial questions based on the target knowledge slice includes: The flowchart is subjected to topological parsing to obtain flowchart topological data. The flowchart topological data includes node information and edge information representing the connection relationship between nodes. The node information includes node content and the position information of the node content in the flowchart. The target assessment content and the target image region corresponding to the target assessment content are determined based on the question request information and the flowchart topology data. The target image region is occluded or its content is replaced to obtain the target question image. The initial questions are generated based on the target question image and the target assessment content.

10. The intelligent question generation method according to any one of claims 1 to 9, characterized in that, The method further includes: The target question is associated and stored with the target knowledge slice corresponding to the target question; In response to business document updates, the updated business document is sliced ​​to obtain multiple updated knowledge slices; A first node set and a second node set are constructed based on the multiple updated knowledge slices and the multiple original knowledge slices in the knowledge base, respectively. The corresponding node connection weights are determined based on the semantic similarity and semantic anchor point similarity between each updated knowledge slice and each original knowledge slice, thereby obtaining a bipartite graph of knowledge slices. The slice mapping relationship between each updated knowledge slice and each original knowledge slice is determined based on the knowledge slice bipartite graph. Based on the slice mapping relationship and the association relationship between the target question and the target knowledge slice, the historical question associated with the business document update is determined. When the slice mapping relationship is a one-to-one mapping relationship, the historical questions are updated according to the corresponding updated knowledge slices; If the slice mapping relationship is not a one-to-one mapping relationship, the historical questions will be regenerated according to the corresponding updated knowledge slices; If there is no corresponding updated knowledge slice for the original knowledge slice, the historical questions associated with the original knowledge slice will be set to an invalid state.

Images