An online adaptive reading and self-adaptive testing method, an online reading system, a device and a medium

Abstract

The application relates to the technical field of intelligent education, and provides an online adaptive reading and self-adaptive evaluation method, an online reading system, equipment and a medium. The method comprises the following steps: combining the multi-dimensional comparison of the same platform group performance in the system to construct a dynamic user ability portrait; obtaining a double-path candidate set based on the user ability portrait through a preset double-path recommendation framework, and obtaining an individualized article recommendation list based on the double-path candidate set; guiding a preset large language model to generate adaptive questions through a probability theory, performing real-time probabilistic semantic scoring on submitted answer contents to obtain a semantic scoring result, and outputting feedback information according to a hierarchical feedback strategy; adopting a target large language model matched from a preset model pool to analyze any touched word or sentence, and obtaining analysis contents; and generating a learning report based on the dynamic user ability portrait, the semantic scoring result, the analysis contents and the feedback information. The application can meet the individualized learning demand while improving the running efficiency.

Description

Technical Field

[0001] This application relates to the field of smart education technology, and in particular to an online adaptive reading and adaptive evaluation method, an online reading system, and an online reading system for devices and media. Background Technology

[0002] Currently, existing online English reading platforms still suffer from many common industry limitations in terms of technical architecture and functional implementation. Existing technical solutions have failed to form an effective solution path. For example, the recommendation logic of existing platforms mostly relies on shallow behavioral data such as user clicks, favorites, and page dwell time to construct the recommendation logic. Moreover, user reading level levels are mostly based on external general grading standards, the source of platform test questions is mostly limited to fixed question banks, and all tasks are processed through a single Large Language Model (LLM). This not only fails to dynamically generate appropriate questions based on the user's real-time reading ability level, but also results in a low degree of matching between recommended content and the user's actual learning needs. It is difficult to adapt to the personalized learning needs of users with different ability levels, the rating results are not well adapted to the actual learning scenario, the assessment results cannot truly reflect the user's real-time reading ability, and the efficiency of reading recommendation and learning assessment is low. Summary of the Invention

[0003] This application provides an online adaptive reading and evaluation method, online reading system, device, and medium, which can realize a closed-loop reading and learning strategy that includes personalized dual-path recommendation, probability theory-guided adaptive evaluation, asynchronous parallel weak point identification, multi-model computing power optimization, and learning report generation. This improves the targeting of learning and the efficiency of system operation, and solves common industry pain points such as low recommendation accuracy, lack of theoretical support for evaluation, low efficiency of sequential modules, and waste of computing resources in existing reading systems.

[0004] In one aspect, this application provides an online adaptive reading and adaptive evaluation method, applied to an online reading system, the method comprising:

[0005] Obtain the user's full-dimensional historical learning data, and construct a dynamic user capability profile based on the full-dimensional historical learning data and multi-dimensional comparison information of the performance of the same platform group within the system;

[0006] In response to a recommendation trigger command, a dual-path candidate set is obtained based on the dynamic user capability profile through a preset dual-path recommendation architecture. The dual-path candidate set is then fused and reordered to obtain a personalized article recommendation list. In the preset dual-path recommendation architecture, the main path recommendation branch performs matching recommendations based on the dynamic user capability profile, while the side path recommendation branch performs mining recommendations based on collaborative filtering and sequence patterns.

[0007] In response to the selection instruction of the target article in the personalized article recommendation list, the system uses probability theory to guide the preset large language model to generate an appropriate question, performs real-time probabilistic semantic scoring on the answers submitted for the appropriate question to obtain a semantic score result, and outputs feedback information according to a hierarchical feedback strategy.

[0008] Weak words and phrases are collected from the low-scoring question information indicated by the semantic scoring results. The collected weak words and phrases are annotated with multi-dimensional information, and the weak words and phrases with multi-dimensional information annotation are added to a preset word and phrase library. The low-scoring question information is pushed asynchronously through a message queue.

[0009] In response to a touch command on any word or phrase in the preset word and phrase library, a task parsing request is generated for the touched word or phrase. Based on the task type, task complexity, and real-time system load of the task parsing request, a target large language model is matched from the preset model pool. The target large language model instance is used to parse the touched word or phrase to obtain the parsed content.

[0010] A learning report is generated based on the dynamic user capability profile, the semantic scoring results, the parsed content, and the feedback information.

[0011] On the other hand, this application provides an online reading system, the system comprising:

[0012] The user profile and ability scoring module is used to obtain users' full-dimensional historical learning data and construct dynamic user ability profiles based on the full-dimensional historical learning data and multi-dimensional comparison information of the performance of the same platform group within the system.

[0013] The personalized article recommendation module is used to respond to recommendation trigger commands, obtain a dual-path candidate set based on the dynamic user capability profile through a preset dual-path recommendation architecture, and merge and reorder the dual-path candidate set to obtain a personalized article recommendation list; wherein, in the preset dual-path recommendation architecture, the main path recommendation branch performs matching recommendation based on the dynamic user capability profile, and the side path recommendation branch performs mining recommendation based on collaborative filtering and sequence pattern.

[0014] The adaptive reading assessment module is used to respond to the selection instruction of the target article in the personalized article recommendation list, guide the preset large language model to generate adaptive questions through probability theory, perform real-time probabilistic semantic scoring on the answers submitted for the adaptive questions to obtain semantic scoring results, and output feedback information according to the hierarchical feedback strategy.

[0015] The intelligent word and phrase collection and parsing module is used to collect weak words and phrases from the low-scoring question information indicated by the semantic scoring results, perform multi-dimensional information annotation on the collected weak words and phrases, and add the multi-dimensional information-annotated weak words and phrases to the preset word and phrase library; the low-scoring question information is pushed asynchronously through a message queue;

[0016] The multi-model intelligent routing module is used to respond to touch commands for any word or phrase in the preset word and phrase library, generate a task parsing request for the touched word or phrase, and match a target large language model from the preset model pool according to the task type, task complexity and real-time system load of the task parsing request. The target large language model instance is used to parse the touched word or phrase to obtain the parsed content.

[0017] The learning report generation module is used to generate a learning report based on the dynamic user capability profile, the semantic scoring results, the parsed content, and the feedback information.

[0018] In another aspect, this application also provides an electronic device, including: a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements any of the online adaptive reading and adaptive evaluation methods described above.

[0019] In another aspect, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements any of the online adaptive reading and adaptive evaluation methods described in the present invention.

[0020] In another aspect, this application also provides a computer program product containing instructions that, when run on a computer, cause the computer to perform the online adaptive reading and adaptive evaluation methods described in the above aspects.

[0021] The online adaptive reading and evaluation method, online reading system, device, and medium provided in this application construct a dynamic user ability profile based on the user's full-dimensional historical learning data and multi-dimensional comparison information of the performance of the same platform group within the system. Responding to recommendation trigger commands, a dual-path recommendation architecture is used to obtain a dual-path candidate set based on the dynamic user ability profile. The dual-path candidate sets are then fused and reordered to obtain a personalized article recommendation list. In this architecture, the main recommendation path performs matching recommendations based on the dynamic user ability profile, while the side recommendation path performs recommendations based on collaborative filtering and sequence pattern mining. Then, responding to the selection command for target articles in the personalized article recommendation list, a pre-defined large language model is guided by probability theory to generate adaptive questions, and the submitted answers are processed... The system performs real-time probabilistic semantic scoring to obtain semantic score results and outputs feedback information according to a hierarchical feedback strategy. Furthermore, it can collect weak words and phrases from low-scoring questions asynchronously pushed via a message queue, annotate these weak words and phrases with multi-dimensional information, and add them to a preset word and phrase library. In response to touch commands on any word or phrase in the preset library, it generates a task parsing request for that specific word or phrase. Based on the task type, complexity, and real-time system load of the parsing request, it matches a target large language model from a preset model pool and uses an instance of the target large language model to parse the touched word or phrase, obtaining the parsed content. Finally, it generates a learning report based on dynamic user ability profiles, semantic score results, parsed content, and feedback information. This application enables a closed-loop reading and learning strategy that includes personalized dual-path recommendation, probabilistically guided adaptive assessment, asynchronous parallel weak point accumulation, multi-model computational optimization, and learning report generation. This improves the targeting of learning and the efficiency of system operation, addressing common industry pain points such as low recommendation accuracy, lack of theoretical support for assessment, low efficiency due to sequential module operation, and waste of computational resources in existing reading systems. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the architecture of an online reading system provided in an embodiment of this application;

[0023] Figure 2 This is a flowchart illustrating the steps of an online adaptive reading and adaptive evaluation method provided in an embodiment of this application.

[0024] Figure 3 This is a schematic diagram illustrating the implementation process of personalized article recommendation provided in the embodiments of this application;

[0025] Figure 4 This is a schematic diagram illustrating the implementation process of the adaptive reading assessment provided in the embodiments of this application;

[0026] Figure 5 This is a schematic diagram illustrating the implementation process of the dynamic scheduling large language model instance provided in the embodiments of this application;

[0027] Figure 6 This is a structural block diagram of an electronic device provided in an embodiment of this application;

[0028] Figure 7 This is a structural block diagram of a computer-readable storage medium provided in an embodiment of this application. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] Currently, existing online English reading platforms still suffer from many common industry limitations in terms of technical architecture and functional implementation. Existing technical solutions have failed to provide effective solutions, specifically in the following aspects:

[0031] First, the accuracy of content recommendations is insufficient. Most existing platforms rely solely on superficial behavioral data such as user clicks, favorites, and page dwell time to build their recommendation logic. They fail to deeply integrate refined learning profiles, such as user answering abilities and mastery of various article types. Furthermore, they lack collaborative mining and analysis of group learning behavior data. The industry has not yet built a fusion recommendation system that combines personalized gap-filling with group learning patterns. This results in a low match between pushed content and users' actual learning needs, frequently leading to the pain points of recommended content being too difficult, discouraging learning, or too easy, failing to achieve the desired training effect.

[0032] Secondly, the learning assessment lacks flexibility and scientific basis. The platform's test questions are mostly limited to fixed question banks, failing to dynamically generate suitable questions based on users' real-time reading ability levels, and lacking a core mechanism for dynamic question generation and probabilistic scoring. Simultaneously, the industry has failed to build an intelligent, tiered, accurate scoring and comprehensive ability assessment system based on large language models, still relying on traditional experience-based question generation and mechanical right-or-wrong scoring models. This makes it difficult to adapt to the personalized learning needs of users at different ability levels, and the assessment results cannot objectively and truthfully reflect users' real-time English reading ability.

[0033] Furthermore, the depth of learning feedback is relatively limited. Existing assessments only use a simple binary right-or-wrong scoring method, which cannot conduct in-depth semantic understanding assessment and error attribution analysis of user responses. This makes it difficult to accurately identify learning problems and effectively guide users to carry out targeted skills improvement training.

[0034] Fourth, the platform suffers from inefficient accumulation of knowledge gaps. It still relies on users manually saving and organizing weak vocabulary and phrases, failing to automatically identify and intelligently analyze weak knowledge points based on user error data. This creates a gap in the learning loop, making it difficult for users to systematically review and fill knowledge gaps in their English reading.

[0035] Fifth, it is difficult to balance computing resources and user experience. The industry generally uses a single large language model to carry out all business tasks. Complex reading assessment and semantic analysis tasks suffer from high response latency and high computing costs, while simple basic tasks result in idle and wasted computing resources. Furthermore, there is no hierarchical scheduling mechanism and dynamic load balancing strategy based on task complexity, and no dedicated intelligent routing and dedicated model-specific technical architecture optimization has been implemented for English reading scenarios.

[0036] Sixth, the overall system's business processing efficiency is low. All functional modules of the platform adopt a serial working mode, lacking a cross-module asynchronous parallel processing architecture. The industry has failed to achieve collaborative parallel computing for core business functions such as LLM intelligent assessment, automatic collection of weak terms and phrases, and intelligent model routing. Single-task queuing results in slow overall system response speed, severely impacting user experience.

[0037] Seventh, the presentation of learning reports suffers from severe homogenization. Existing learning reports merely provide a simple statistical listing of basic learning data, lacking intelligent and in-depth analysis based on the entire learning process. They also fail to generate actionable suggestions for improving skills, making it difficult for users to develop personalized learning plans tailored to their individual circumstances.

[0038] Finally, the user ability rating system is poorly adapted to different scenarios. The platform's classification of user reading levels often directly applies externally accepted grading standards without conducting multi-dimensional horizontal comparisons based on the actual learning performance of users within the same user group. As a result, the rating results do not align well with the platform's actual learning scenarios and cannot provide accurate basic rating data support for core businesses such as content recommendation and intelligent assessment.

[0039] This application embodiment constructs a multi-dimensional dynamic user profile by combining the group performance within the system, introduces collaborative filtering and sequence pattern mining to optimize recommendation strategies, integrates the core theories of probability theory into the entire LLM assessment process to achieve scientific assessment, designs a multi-model intelligent routing strategy to achieve optimal allocation of computing resources, and adopts an asynchronous parallel mechanism to achieve efficient collaborative work of each module, ultimately forming a complete closed-loop learning path. While improving the targeting and efficiency of learning, it also takes into account the economy and smoothness of system operation. Specifically, by integrating users' full-dimensional historical learning data and combining it with multi-dimensional comparisons of performance among users on the same platform, a dynamic user ability profile is constructed, achieving precise recommendations that combine personalization and group patterns. Through a pre-designed dual-path recommendation architecture, collaborative filtering and sequence pattern mining methods for group learning data mining are integrated with personalized gap-filling methods that match three-dimensional profiles, further achieving precise recommendations that balance personalized gap-filling and group learning patterns. Throughout the reading process, probability theory guides LLM to automatically generate suitable questions, provide real-time probabilistic semantic scoring, intelligent hierarchical feedback, and automatically collect weak words and phrases. Multi-model intelligent routing and asynchronous parallel mechanisms optimize computing resource allocation and improve operational efficiency. Finally, a learning report is generated based on multi-dimensional learning data.

[0040] Reference Figure 1 The diagram illustrates the architecture of an online reading system provided in an embodiment of this application. Specifically, the system may include a user profiling and ability scoring module 101, a personalized article recommendation module 102, an adaptive reading assessment module 103, a word and sentence intelligent acquisition and parsing module 104, a multi-model intelligent routing module 105, a learning report generation module 106, and a photo reading OCR (Optical Character Recognition) extension module 107.

[0041] The user profiling and competency scoring module 101 can be used to obtain a user's comprehensive historical learning data. Based on this comprehensive historical learning data and multi-dimensional comparisons of scores from other users within the same platform, a dynamic user competency profile can be constructed. Specifically, it can integrate a user's historical question type scores and article type scores with multi-dimensional comparisons of scores from other users within the system to build a dynamic user profile.

[0042] The personalized article recommendation module 102 can be used to obtain a dual-path candidate set using a preset dual-path recommendation architecture, and then merge and reorder the dual-path candidate sets to output a personalized article recommendation list. Specifically, it can respond to recommendation trigger commands, obtain a dual-path candidate set based on dynamic user capability profiles through the preset dual-path recommendation architecture, and then merge and reorder the dual-path candidate sets to obtain a personalized article recommendation list. In the preset dual-path recommendation architecture, the main path recommendation branch can perform matching recommendations based on dynamic user capability profiles, while the side path recommendation branch can perform mining recommendations based on collaborative filtering and sequence patterns.

[0043] The Adaptive Reading Assessment Module 103 is primarily an LLM-driven end-to-end adaptive reading assessment module. It leverages probability theory to guide a large language model to automatically generate appropriate questions, perform real-time probabilistic semantic scoring, provide intelligent tiered feedback, and automatically collect weak words and phrases throughout the reading process. Specifically, it responds to selection commands from a personalized article recommendation list, uses probability theory to guide a pre-defined large language model to generate appropriate questions, performs real-time probabilistic semantic scoring on submitted answers to these questions to obtain semantic scores, and outputs feedback information according to a tiered feedback strategy.

[0044] The intelligent word and sentence acquisition and parsing module 104 can be used to annotate the acquired words and sentences with multi-dimensional information, enabling the multi-model intelligent routing module 105 to match the corresponding LLM model when the user clicks on any sentence / word. Specifically, the adaptive reading assessment module 103 can asynchronously push low-scoring question information indicating semantic scoring results through a message queue. The intelligent word and sentence acquisition and parsing module 104 can collect weak words and sentences from the asynchronously pushed low-scoring question information, annotate the collected weak words and sentences with multi-dimensional information, and add the annotated weak words and sentences to a preset word and sentence library. The preset word and sentence library can be considered as a word and sentence collection library.

[0045] The multi-model intelligent routing module 105 can dynamically schedule large language model instances with different parameter amounts and training accuracies based on task type, task complexity, and real-time system load. Specifically, it can respond to touch commands on any word or phrase in a preset word / phrase library, generate a task parsing request for the touched word / phrase, match a target large language model from a preset model pool based on the task type, task complexity, and real-time system load of the parsing request, and then use the target large language model instance to perform customized and contextualized parsing on the touched word / phrase to obtain the parsed content. Optionally, the parsed content can also be displayed in a pop-up window.

[0046] The learning report generation module 106 can collect multi-dimensional learning data from the user profiling and ability scoring module, the adaptive reading assessment module, and the intelligent word and sentence collection and parsing module in an asynchronous, non-blocking manner, and then generate a learning report based on the multi-dimensional learning data using a large language model. Specifically, it can generate structured and personalized learning reports based on dynamic user ability profiles, semantic scoring results, parsed content, and feedback information.

[0047] The image-based OCR extension module 107 can preprocess, perform high-precision recognition, and text structuring of uploaded English articles in paper / image format. After recognition and processing, it asynchronously pushes the processed text to the LLM-driven adaptive reading assessment process via a message queue. Specifically, it responds to image upload commands, preprocesses and recognizes the uploaded image, performs grammatical verification and semantic correction on the recognition results, performs structured annotation on the corrected text and matches it to the system's article tagging system, and finally asynchronously pushes the recognized and processed text through a message queue, generating a selection command for the target article. Responding to the selection command initiates the LLM-driven adaptive reading assessment process.

[0048] In this embodiment, a multi-dimensional dynamic user profile is constructed by combining the group performance within the system. Collaborative filtering and sequence pattern mining are introduced to optimize the recommendation strategy. The core theory of probability is integrated into the entire LLM assessment process to achieve scientific assessment. A multi-model intelligent routing strategy is designed to achieve optimal allocation of computing resources. An asynchronous parallel mechanism is adopted to achieve efficient collaborative work of each module, and finally a complete closed-loop learning path is formed. While improving the targeting and efficiency of learning, the economy and smoothness of system operation are also taken into account.

[0049] It should be noted that the online reading system provided in this application embodiment can be adapted to any subject, such as online adaptive reading and adaptive evaluation of English, Chinese, etc., and this application embodiment does not impose any limitations on this.

[0050] In practical applications, after entering the online reading system, users can actively trigger reading instructions or receive intelligent passive recommendations from the system based on the user's learning interval and profile updates.

[0051] Reference Figure 2 This document illustrates a flowchart of the steps involved in an online adaptive reading and adaptive evaluation method provided in an embodiment of this application, applicable to, for example... Figure 1 The online reading system shown may specifically include the following steps:

[0052] Step S201: Obtain the user's full-dimensional historical learning data. Based on the full-dimensional historical learning data and multi-dimensional comparison information of the performance of the same platform group within the system, construct a dynamic user capability profile.

[0053] In this embodiment, a dynamic user capability profile can be constructed by integrating the user's full-dimensional historical learning data and combining it with multi-dimensional comparison information of the performance of the same platform group within the system, thereby achieving accurate recommendations that integrate personalization and group patterns.

[0054] In some embodiments of this application, the user's full-dimensional historical learning data can be recorded and updated through the user profile and ability scoring module.

[0055] Optionally, the data dimensions of the comprehensive historical learning data may include the user's historical question type scores, article type scores, performance on articles of different difficulty levels, and the learning and mastery of weak historical vocabulary and phrases. The user's historical question type scores may include scores and accuracy trends for each question type (such as detail comprehension, reasoning, main idea, vocabulary guessing, and inference extension); the article type scores may include scores and mastery levels for each article type (such as narrative, expository, argumentative, popular science, and practical writing). This application embodiment does not impose any limitations.

[0056] In some embodiments of this application, a dynamic user competency profile can be constructed by integrating a user's historical question type scores and article type scores with a competency scoring module, and combining these with multi-dimensional comparisons of group scores within the system. Specifically, a dynamic, real-time updatable user competency profile can be formed based on the aforementioned data using a weighted algorithm. This can be achieved through percentile ranking calculations and weighted comprehensive evaluations.

[0057] Optionally, the core dimensions of a dynamic user capability profile may include capability strength and weakness tags, current reading level, dynamically updated weak knowledge points and vocabulary database, learning progress and improvement trend curve, etc.

[0058] Among them, the labels for strengths and weaknesses can be, for example, strong comprehension of details, weak reasoning and judgment, and low mastery of popular science articles. Specifically, these labels can be based on the scores of detailed comprehension questions, reasoning and judgment questions, and popular science articles, and the higher of a fixed passing score (e.g., 60 points) and the average score can be used as the threshold for judgment.

[0059] The current reading level can be obtained based on multi-dimensional comparisons of scores among users on the same platform within the system. Specifically, it can be obtained by comparing the average scores of other users on the same platform. Optionally, the comparison dimensions correspond one-to-one with the user's learning data dimensions, including score comparisons for each question type (such as reasoning, main idea, and vocabulary comprehension), each article type, and different difficulty levels. The level of a dimension can be determined by calculating the user's percentile ranking in each dimension, and then the overall level can be obtained through a weighted summation. Reading level levels can be divided into beginner (0-20%), basic (21-40%), intermediate (41-60%), excellent (61-80%), and proficient (81-100%), achieving accurate determination of the user's level from dimensions such as question type, article type, and difficulty level. This application embodiment does not limit this.

[0060] The dynamically updated database of weak knowledge points and vocabulary can be linked to incorrect question types, article types, and reasons for errors. The incorrect question types, article types, and reasons for errors can be extracted from a user's performance on a particular article or type of article; that is, the user's submitted answers, their correctness, and the reasons for errors and feedback are all recorded in the database.

[0061] The learning progress and improvement trend curve can be obtained by comparing it with the average improvement speed of the system group.

[0062] In some embodiments of this application, after each reading assessment, users can update their data through the user profile and ability scoring module. Specifically, their historical learning data can be updated in real time, and multi-dimensional percentile rankings and skill levels can be recalculated to ensure the real-time nature and accuracy of the user profile.

[0063] Step S202: In response to the recommendation trigger command, a dual-path candidate set is obtained based on the dynamic user capability profile through a preset dual-path recommendation architecture. The dual-path candidate sets are then merged and reordered to obtain a personalized article recommendation list.

[0064] In this embodiment, a pre-defined dual-path recommendation architecture can be used to integrate collaborative filtering and sequence pattern mining methods for group learning data mining with three-dimensional profile matching for personalized gap filling, thereby achieving more accurate recommendations that take into account both personalized gap filling and group learning patterns.

[0065] In some embodiments of this application, a personalized article recommendation module can use a preset dual-path recommendation architecture to obtain a dual-path candidate set based on dynamic user capability profiles. After obtaining the dual-path candidate set, it can be fused and reordered to obtain a personalized article recommendation list, breaking through the limitations of traditional single recommendation logic and achieving a dual improvement in recommendation accuracy and group pattern adaptability.

[0066] In the pre-defined dual-path recommendation architecture, the main recommendation path can perform matching recommendations based on dynamic user capability profiles; the side recommendation path can perform mining recommendations based on collaborative filtering and sequence patterns.

[0067] Optionally, the main path recommendation branch implements three-dimensional profile matching recommendation, which may include generating a three-dimensional profile matching candidate set, initial screening of advantageous question types and suitable article types, weighting of weak knowledge points, and the zone of proximal development strategy.

[0068] The bypass recommendation implements collaborative filtering and sequence pattern mining recommendation, which can include generating a candidate set for collaborative filtering and sequence pattern mining, calculating cosine similarity to select the top N similar users, including articles with similar user accuracy ≥ a preset percentage, optimizing the ranking using the Apriori algorithm, and proposing articles that do not match the user's level / gap-filling needs through secondary screening.

[0069] Reference Figure 3 The diagram illustrates the implementation process of personalized article recommendation provided in this embodiment of the application.

[0070] In practical applications, online reading systems can support users actively triggering commands to enter the article selection page, or the system can provide passive intelligent recommendations based on user learning intervals and profile updates. In other words, recommendation trigger commands can be either actively or passively triggered.

[0071] The core input data may include full-dimensional data of dynamic user ability profiles, structured tags of the article resource library, and big data on the learning behavior and answer performance of all users within the system. Specifically, the full-dimensional data of dynamic user ability profiles can be used to indicate user strengths and weaknesses, current reading level, weak knowledge points and vocabulary, learning progress, and improvement trend curves; the structured tags of the article resource library may include subject matter, type, difficulty, core knowledge points, question types, vocabulary / syntactic complexity, etc.; the big data on the learning behavior and answer performance of all users within the system may include reading article sequences, answer accuracy, and collection / like behavior, etc. This application embodiment does not impose any limitations on these aspects.

[0072] like Figure 3 As shown, the main route recommendation branch can generate a candidate set based on dynamic user capability profiles in layers, including basic matching layer initial screening, precision matching layer weighted adjustment, and difficulty adaptation layer dynamic adjustment. After the above three layers of matching, the candidate set of the main route recommendation branch is obtained, that is, the personalized gap filling candidate set.

[0073] Specifically, a basic matching layer screening can be performed first. This involves initial screening of article resources based on user strengths and weaknesses tags and structured tags from the article resource library. Optionally, initial screening can be conducted based on user's strengths in question types and suitable article types, eliminating extremely difficult articles beyond the user's capabilities and easily understood articles that the user has already mastered. In practical applications, the initial screening pass rate can be controlled to within 30%.

[0074] For example, all articles in the article resource library can be denoted as the complete set Q. If a user's strength is reasoning and judgment questions, and their weakness is detail comprehension questions, assuming that more than 80% of the articles are of these two types, then articles that are not of these two types can be denoted as A, articles of the weaker type as B, and articles of the stronger type as C. The article resources can be initially screened using QA∩B. If, after the initial screening, the proportion of articles of these two types still exceeds the set initial screening pass rate control rate, such as 30%, then the article resources can be initially screened using QAC∩B. If, after the initial screening, the proportion of articles of these two types still exceeds 30%, then the excess portion can be randomly removed. If, after the initial screening, the proportion of articles of these two types is less than 30%, then the initial screening of articles is successful, and no further action is taken.

[0075] Secondly, precise matching layer weight adjustment can be implemented. Specifically, a knowledge gap-filling weight mechanism is introduced, assigning recommendation weights to articles containing weak knowledge points and relevant vocabulary after initial screening. Optionally, the recommendation weight coefficient is inversely correlated with the user's mastery of weak knowledge points; that is, it can be dynamically adjusted in reverse based on the user's mastery of weak knowledge points. The lower the mastery of weak knowledge points, the higher the recommendation weight of the corresponding article. In this case, the recommendation weight of articles containing the user's weak knowledge points can be increased.

[0076] Finally, difficulty adaptation layer matching can be performed. Specifically, a recent development adaptation strategy can be adopted, matching articles based on the current reading level. This involves matching the basic difficulty of recommended articles with the current reading level based on learning progress and improvement trend curves, configuring a preset proportion (e.g., 10%-20%) of advanced knowledge points, and iterating the difficulty adaptation coefficient in real time based on the answer accuracy rate indicated by big data analysis of all users' learning behavior and answer performance within the system. A personalized gap-filling candidate set is then output based on the difficulty adaptation coefficient. For example, when the answer accuracy rate is greater than 80%, the difficulty adaptation coefficient can be increased, i.e., the recommended difficulty can be increased; when the answer accuracy rate is less than 40%, the difficulty adaptation coefficient can be decreased, i.e., the recommended difficulty can be decreased. This application embodiment does not limit this.

[0077] like Figure 3As shown, the alternative recommendation branch can mainly generate candidate sets based on collaborative filtering and sequence pattern mining. Specifically, it can generate candidate sets of alternative recommendation branches, i.e., group pattern candidate sets, based on big data of user learning behavior and answer performance across the entire system, combined with collaborative filtering and sequence pattern mining.

[0078] Specifically, the first step is to generate a collaborative filtering candidate set. This involves using question type scores, article type preferences, and reading level as feature dimensions, and calculating the similarity of the target user's learned features with other users in the system using cosine similarity. Then, the top N (e.g., 50) similar users with a similarity score no lower than a preset similarity threshold (e.g., 0.7) are selected. Articles whose answer accuracy rate of the top N similar users is no lower than a preset percentage (e.g., 70%) are included in the initial candidate set, thereby excluding articles whose answer accuracy rate of similar users is lower than 50%, thus avoiding ineffective recommended content.

[0079] For example, the formula for calculating the learned feature similarity in collaborative filtering based on user similarity can be as follows:

[0080]

[0081] In the formula, n is the total number of feature dimensions; X 1i This refers to User 1, i.e., the value of the target user on the i-th feature dimension, such as the target user's accuracy rate on detailed comprehension questions, score on popular science articles, and score rate on advanced articles. It is usually a normalized value (0~1 or standardized to the same dimension); X 2i This refers to user 2, i.e., the value of other users in the system on the i-th feature dimension, and X. 1i They are feature values ​​of the same dimension and unit, which facilitates cross-user comparison.

[0082] Secondly, the Apriori algorithm can be used to mine sequence patterns in the initial candidate set (i.e., the collaborative filtering candidate set). Specifically, this involves setting a support threshold (not less than 0.5) and a confidence threshold (not less than 0.8) for frequent itemsets, mining frequent association patterns in the sequences of articles read by system users, and then ranking the initial candidate set based on these frequent association patterns.

[0083] Optionally, the Apriori algorithm can be used to mine frequent association patterns in the article sequences read by users within the system. A support threshold of ≥0.5 and a confidence threshold of ≥0.8 for frequent itemsets can be set. Then, the candidate set obtained from collaborative filtering can be initially ranked based on these association patterns, retaining articles that conform to the user's potential reading sequence patterns. Here, the association pattern refers to the frequency of co-occurrence of different set types, such as the frequency of co-occurrence between groups with low average scores and groups with low detail comprehension ability in scientific articles.

[0084] Finally, the ranked candidate set can be further filtered by combining dynamic user capability profiles. Specifically, articles that do not match the current reading level and the need to fill in knowledge gaps are removed from the ranked candidate set, and the size of the filtered side candidate set is controlled to a preset second percentage (such as 50%-80%) of the personalized gap-filling candidate set, thus obtaining the group pattern candidate set.

[0085] In some embodiments of this application, the dual-path candidate set fusion performance merges the personalized gap-filling candidate set and the group pattern candidate set, and removes duplicate articles to form a fused candidate set.

[0086] The reordering algorithm employs a weighted score ranking. Specifically, it obtains pre-defined main path recommendation weights and side path recommendation weights, then uses these weights to calculate the main path matching score and side path mining score for each article in the fusion candidate set. Finally, it uses these scores to calculate the comprehensive score for each article in the fusion candidate set, calculates the comprehensive matching degree based on the comprehensive score, and finally sorts each article in the fusion candidate set from highest to lowest according to the comprehensive matching degree, outputting a personalized article recommendation list.

[0087] In this system, the main route recommendation weight is typically set higher than the side route recommendation weight to ensure personalized supplementation of wanted cases, while the side route weight supplements the adaptability of the group's learning patterns. For example, the main route recommendation weight can be set to 0.6, and the side route recommendation weight can be set to 0.4. Then, the formula for calculating the comprehensive score can be expressed as: Comprehensive Score = Main Route Score × 0.6 + Side Route Score × 0.4 (maximum score 100 points).

[0088] Optionally, in the output personalized article recommendation list, each article may also include a tag indicating the reason for its suitability (such as matching the user's need for supplementary popular science articles, or high-accuracy recommendations from similar users) and a description of its difficulty / focus. This embodiment of the application does not impose any limitations on this.

[0089] In some embodiments of this application, the user's performance in answering questions about recommended articles can be fed back to the user's ability profile in real time. At the same time, the weight coefficients of the main path recommendation algorithm and the user similarity calculation and sequence pattern mining association rules of the side path collaborative filtering can be updated, thereby realizing dynamic iterative optimization of "recommendation-learning-feedback-re-recommendation".

[0090] In some embodiments of this application, article resources can also be managed. Specifically, articles in the resource library can be structured and updated in real time, supporting categorized searches by topic, difficulty, and knowledge point. An article quality assessment mechanism can also be established to remove low-quality articles based on user experience and accuracy distribution. This application does not impose any limitations on these aspects.

[0091] Step S203: In response to the selection instruction of the target article in the personalized article recommendation list, the preset large language model is guided by probability theory to generate the appropriate questions, and the submitted answers to the appropriate questions are given a real-time probabilistic semantic score to obtain the semantic score result. Feedback information is output according to the hierarchical feedback strategy.

[0092] In this embodiment, during the entire reading process, probability theory can guide LLM to automatically generate suitable questions, perform real-time probabilistic semantic scoring, provide intelligent hierarchical feedback, and automatically collect weak words and phrases. At the same time, it introduces an asynchronous parallel mechanism of multi-process, task queue, and non-blocking IO (Input / Output) to overcome the limitations of fixed question banks, mechanical scoring, and serial processing, and achieve one question per person, one question per article, real-time adaptation, and efficient processing, which is scientific, innovative, and efficient.

[0093] In practical applications, it can respond to the selection command of the target article in the personalized article recommendation list and enter the LLM-driven adaptive reading evaluation process.

[0094] In some embodiments of this application, the three core theories of probability theory (including Bayesian posterior probability, maximum likelihood estimation, and binomial distribution) can be integrated into the entire process of LLM-driven adaptive reading assessment, providing scientific mathematical theoretical support for question generation, scoring mechanism, and difficulty adaptation.

[0095] The probabilistic difficulty adaptation of question generation can be achieved based on Bayesian posterior probability. Optionally, in the question generation stage, the Bayesian posterior mastery probability of each knowledge point is calculated based on the dynamic user ability profile to select knowledge points in the easy-to-improve range. Then, LLM generates questions according to the question generation strategy (such as a combination of basic, advanced, and gap-filling questions) and customization, while verifying the quality of the generated questions, and finally outputs the candidate questions that pass the verification.

[0096] For example, suppose P(A|X) is the posterior probability that a user has mastered knowledge point A under feature X (historical answer performance, level), P(A) is the prior probability of mastering knowledge point A (average mastery level of users in the system), P(X|A) is the conditional probability of feature X appearing when knowledge point A is mastered, and P(X) is the marginal probability of feature X; then the mastery probability of each knowledge point can be calculated using Bayes' theorem: P(A|X) = P(X)P(X|A)P(A), and a mastery probability interval threshold can be set. For knowledge points with a mastery probability between 30% and 70% (the range where users are likely to improve), the probability of generating corresponding questions is increased (60%), and the probability of generating questions is decreased (20% each) for knowledge points with a mastery probability >90% or <10%, thus achieving probabilistic and precise adaptation of question generation.

[0097] The probabilistic quantification of semantic scoring can be achieved based on maximum likelihood estimation. Optionally, in the semantic scoring stage, LLM can quantify the scoring based on a three-dimensional scoring system (such as semantic similarity, depth of understanding, and completeness of answer) while outputting the scoring criteria.

[0098] For example, based on maximum likelihood estimation, the semantic similarity of user answers is probabilistically evaluated. The semantic matching problem between user answers and standard answers is transformed into a probability distribution fitting problem. The semantic feature vector of the answer is extracted by LLM, and the likelihood value of the probability distribution of the feature vector of the user answer fitting the feature vector of the standard answer is calculated. Then, the likelihood value can be divided into 5 quantification intervals (0-0.2 / 0.2-0.4 / 0.4-0.6 / 0.6-0.8 / 0.8-1.0), which correspond to semantic similarity scores of 1-4 respectively. The higher the likelihood value, the higher the semantic similarity. Combining the likelihood value realizes the quantification and scientification of the score, replacing the traditional subjective judgment.

[0099] Dynamic probability adjustment of assessment difficulty can be achieved based on a binomial distribution. For example, suppose a user answers n questions of a certain difficulty with a correct answer rate of p. The probability of this correct answer rate occurring can be calculated using a binomial distribution B(n,p). If a low-probability threshold is set (P < 0.05), and the probability is less than 0.05 (a low-probability event), it can be determined that the current difficulty level does not match the user's skill level. In this case, the difficulty level can be adjusted upwards or downwards immediately, thus achieving dynamic probabilistic optimization of the assessment difficulty.

[0100] This can be achieved by comparing the user's actual accuracy rate with the expected accuracy rate for questions of the current difficulty level, and then using a binomial distribution low-probability test to determine whether to adjust the difficulty upwards or downwards. For example, for medium-difficulty questions, the average score can be used as the expected accuracy rate. As an example, if a user's consecutive accuracy rate is significantly higher than the expected accuracy rate for questions of this difficulty level, and the probability of this occurring is less than 5%, it indicates that the questions are too easy for the user, and the difficulty will be adjusted upwards. Conversely, if a user's consecutive accuracy rate is significantly lower than the expected accuracy rate for questions of this difficulty level, and the probability of this occurring is less than 5%, it indicates that the questions are too difficult for the user, and the difficulty will be adjusted downwards. This application's embodiments do not impose any limitations on this.

[0101] Reference Figure 4 The diagram illustrates the implementation process of the adaptive reading assessment provided in the embodiments of this application.

[0102] In practical applications, its implementation process includes the following steps: pre-model and resource preparation, probabilistic fusion modeling and rule pre-setting, adaptive intelligent question generation, semantic multi-dimensional intelligent scoring, and hierarchical intelligent feedback and evaluation data accumulation.

[0103] The pre-model and resource preparation steps involve first fine-tuning the basic large language model for reading assessment application scenarios. Specifically, this can be achieved by using a dedicated training corpus consisting of English reading texts of different themes and difficulty levels, standardized reading test questions, and answer analysis samples. This enhances the large language model's professional adaptability in reading question generation, semantic scoring, and answer feedback analysis tasks. Simultaneously, a multi-scenario customized prompt vocabulary library is constructed, categorized by question type, scoring dimension, and feedback type, to provide standardized support for intelligent invocation in subsequent business processes.

[0104] The probabilistic modeling and rule pre-setting steps involve integrating three core probabilistic theories—Bayesian posterior probability, maximum likelihood estimation, and binomial distribution—into the adaptive assessment process modeling. Based on Bayesian posterior probability and combined with users' historical answer performance and reading level characteristics, the probability of a user mastering each knowledge point is calculated, and pre-set rules for matching question generation to different mastery probability intervals are established. Based on maximum likelihood estimation, a probability distribution fitting model of the semantic feature vectors of user answers and standard answers is constructed, and pre-set quantification rules corresponding to the likelihood value intervals and semantic scoring results are established. Based on the binomial distribution, a probability judgment model for the accuracy of assessment answers is constructed, and a low-probability judgment threshold is pre-set as the basis for dynamically adjusting the assessment difficulty.

[0105] The adaptive intelligent question generation process involves acquiring input user profile data (such as level, weak question types / knowledge points, and mastery probability of each knowledge point) and structured information of the selected article (such as core content, knowledge points, and syntax / vocabulary difficulty). It calculates the target user's Bayesian posterior mastery probability of each knowledge point and pre-sets question generation ratio rules corresponding to different mastery probability intervals. Then, based on the Bayesian posterior mastery probability, it can select target knowledge points that fall within the target user's easy-to-improve interval (such as 30%-70%) and generate corresponding question types using a combination of basic, advanced, and remedial questions according to a preset ratio.

[0106] The preset proportions can be obtained from the question generation rules corresponding to different Bayesian posterior mastery probability intervals. Basic questions (e.g., 50%) test the understanding of the core information of the article and match the target user's current reading level; advanced questions (e.g., 30%) test advanced reasoning / extension abilities and are above the target user's current reading level; remedial questions (e.g., 20%) precisely target the target user's weak knowledge points to achieve targeted training; the difficulty of each question type for the target user can be determined based on Bayesian posterior probability.

[0107] like Figure 3As shown, the finely tuned large language model can be called to generate test questions according to the corresponding question type by combining the question generation strategy and customized prompt words (i.e., customized prompts). The quality of the generated test questions is then checked, and finally, the questions that pass the quality check are output.

[0108] For example, a customized Prompt could be "Generate questions of type {C} with difficulty {N} based on the current article {article}". In this case, the LLM can be invoked to perform the aforementioned operations. For generating basic article questions, several questions of each type can be generated based on the article information using the LLM and stored in the database. For generating customized article questions, appropriate questions can be generated by inputting the article information, the average student data set in the system, and the user ability profile into the corresponding Prompt.

[0109] Specifically, the system can verify and filter the generated questions based on their rationality, rigor, and relevance to the article content, eliminating incorrect, off-topic, and unusual questions to ensure question quality. It supports the generation of a mix of objective questions (such as multiple choice, true / false, and fill-in-the-blank) and subjective questions (such as short answer, main idea summary, and sentence translation), and dynamically matches the number of questions based on the article's word count; for example, 5 questions for articles under 300 words, 8-10 questions for articles between 300 and 800 words, and 10-15 questions for articles over 800 words. This embodiment of the application does not impose any limitations on these aspects.

[0110] The semantic multidimensional intelligent scoring process involves receiving user responses, the original text, the core points of the questions, and customized scoring prompts (i.e., customized scoring prompts). It then constructs a three-dimensional scoring system encompassing semantic similarity, depth of understanding, and completeness of the answer. Maximum likelihood estimation is used to probabilistically quantify the semantic similarity of the user's answer. Scores are then assigned separately for each dimension, including depth of understanding and completeness of answer coverage, and a comprehensive quantitative evaluation is performed. Finally, a large language model outputs the corresponding scoring criteria, ensuring the objectivity and interpretability of the scoring results. The customized scoring prompt refers to a prompt based on LLM (Language Modeling) for scoring student responses.

[0111] Optionally, a probability distribution fitting model of the semantic feature vectors of user answers and standard answers constructed based on maximum likelihood estimation can be obtained, along with the corresponding quantification rules for the preset likelihood value intervals and semantic scoring results. Then, using the probability distribution fitting model and corresponding quantification rules, the answer content is scored separately from semantic similarity, comprehension depth, and answer completeness, and comprehensively quantified to obtain the semantic scoring result. The three-dimensional scoring system of semantic similarity, comprehension depth, and answer completeness can be out of 10 points. Semantic similarity (4 points) can be quantified based on the probability value of maximum likelihood estimation, comprehension depth (3 points) can be used to assess the target user's deep understanding of the question's testing points, and answer completeness (3 points) can be used to assess the accuracy of objective question options / the coverage of key points in subjective questions.

[0112] The tiered intelligent feedback and assessment data accumulation process is manifested in tiered feedback based on semantic scoring results. Specifically, a tiered feedback strategy is set according to the score range of the semantic score result. For users whose semantic score results are in the first score range, i.e., full marks / high scores, the feedback includes expanded knowledge points and problem-solving techniques for similar question types; for users whose semantic score results are in the second score range, i.e., average scores, the feedback includes explanations and guidance on problem-solving approaches; and for users whose semantic score results are in the third score range, i.e., low scores / incorrect scores, the feedback includes detailed explanations such as standard answers, error attribution, original text location, and breakdown of problem-solving steps.

[0113] Optionally, feedback output can take the form of single-question feedback synchronized with scoring results, reasons for errors, directions for improvement, and links to the original text. For subjective questions, excellent answer examples can be matched and pushed to help users understand the answering process. In addition, the test content, user answer information, semantic scoring results, and feedback information corresponding to this assessment can be stored in the user learning data center in real time, providing basic data support for subsequent weak word and phrase mining and personalized learning report generation.

[0114] In this embodiment, an asynchronous parallel mechanism of multi-process, task queue and non-blocking IO can also be used to realize the collaborative parallel work of each sub-module (such as question generation, scoring, feedback, data accumulation, etc.) in the adaptive reading assessment module and other modules of the system (such as word and sentence collection, multi-model routing, etc.). At the same time, multiple optimization strategies are used to ensure the stability and efficiency of parallel processing.

[0115] In practical implementation, the traditional serial processing mode can be abandoned, and an asynchronous parallel processing architecture based on multi-process, task queues, and non-blocking I / O can be constructed. A producer-consumer model is used to build a global task queue and dedicated task queues for each sub-module. Question generation, scoring, feedback, data accumulation, word and phrase collection, and model routing are encapsulated as independent task units, which are uniformly distributed and scheduled by a task scheduler. Independent core processes are allocated to high-complexity business sub-modules (such as question generation and in-depth scoring), while shared core processes are configured for low-complexity business sub-modules (such as basic verification, data accumulation, and feedback output). A shared process pool enables on-demand allocation of process resources; all model calls and data read / write operations adopt non-blocking I / O mode, thus avoiding module lag caused by single-task waiting (such as LLM model return results, database read / write, etc.); a four-level task priority scheduling rule is pre-configured with a timeout fallback mechanism for low-priority tasks; for large-batch processing tasks (such as multi-question scoring, batch data accumulation), a task sharding and batch parallel processing strategy is adopted, splitting a large task into multiple smaller tasks and allocating them to the shared process pool for parallel processing, thereby improving processing efficiency; for user profile basic data, High-frequency data such as article structure tags and prompt word templates are cached locally with an expiration update strategy to avoid frequent database queries and model calls, reducing IO wait time. Independent exception handling and fault-tolerant retry mechanisms are configured for each parallel task. A single task failure is retried independently without affecting the normal operation of other tasks. Failed tasks are re-added to the task queue for retry. Excessive retry limits are logged and reported to the user, ensuring system stability. Real-time monitoring of CPU and memory usage and task queue length for each process is implemented, with preset process resource usage thresholds (e.g., 80%). Overloaded processes are migrated across processes to idle processes for dynamic load balancing. A message queue enables asynchronous cross-module communication between the adaptive reading assessment module, the intelligent word and sentence acquisition module, and the multi-model intelligent routing module, triggering parallel execution of related tasks. For example, when the adaptive reading assessment module completes scoring and identifies low-scoring questions, it immediately sends a acquisition task message to the word and sentence acquisition module without waiting for the entire assessment process of its own module to finish, achieving cross-module parallel collaboration and improving overall system processing efficiency.

[0116] The reading assessment task is given a four-level priority rule, which can be expressed as: Question creation task (Level 1) > Scoring task (Level 2) > Feedback task (Level 3) > Data accumulation task (Level 4). This allows the task scheduler to prioritize the distribution of high-priority tasks, while a 10-second timeout fallback mechanism is set for low-priority tasks to prevent them from being blocked indefinitely. This application does not impose any limitations on this.

[0117] Step S204: Collect weak words and phrases from the low-scoring question information indicated by the semantic scoring results, perform multi-dimensional information annotation on the collected weak words and phrases, and add the weak words and phrases after multi-dimensional information annotation to the preset word and phrase library.

[0118] In this embodiment, the collected weak words and phrases can be annotated with multi-dimensional information through the intelligent word and phrase collection and parsing module, so that when the user clicks on any sentence / word, the multi-model intelligent routing module can be invoked to match the corresponding LLM model for customized and scenario-based parsing.

[0119] In some embodiments of this application, the adaptive reading assessment module can asynchronously push low-scoring question information, such as questions with an LLM three-dimensional score below a preset threshold of 4, to the intelligent word and sentence collection and analysis module via a message queue. The intelligent word and sentence collection and analysis module can automatically trigger a collection mechanism for the original sentences / core vocabulary corresponding to questions with an LLM three-dimensional score below the preset threshold of 4. It can annotate the collected words and sentences with multi-dimensional information, including the sentence / vocabulary source (such as article title, paragraph position, etc.), related question type, error type, score for this answer, and core knowledge points tested. Then, it adds the annotated weak words and sentences to a preset word and sentence library without manual user intervention. The preset word and sentence library can be a user-owned word and sentence collection library.

[0120] Step S205: In response to a touch command on any word or phrase in the preset word and phrase library, generate a task parsing request for the touched word or phrase. Based on the task type, task complexity, and real-time system load of the task parsing request, match a target large language model from the preset model pool, and use an instance of the target large language model to parse the touched word or phrase to obtain the parsed content.

[0121] In the embodiments of this application, the optimization of computing resource allocation and the improvement of operating efficiency can be achieved by relying on multi-model intelligent routing and asynchronous parallel mechanisms.

[0122] In some embodiments of this application, when a user clicks on any sentence / vocabulary in the preset word library, the system can respond to touch commands on any word in the preset word library and call the multi-model intelligent routing module to match the corresponding LLM model for customized and scenario-based parsing.

[0123] The parsing content can be dynamically adjusted according to the user's skill level. For example, beginner users focus on basic parsing, while advanced users focus on in-depth expansion.

[0124] Optional, specific analytical dimensions may include core vocabulary (such as phonetics, parts of speech, core meaning, contextual meaning, fixed collocations, synonyms / antonyms, and derivatives), sentence structure (such as syntactic component division, analysis of core grammar points, explanation of special sentence structures (inversion, emphasis, relative clauses, etc.), and rewriting of synonymous sentences), contextual function (i.e., the role of the sentence / vocabulary in the text, such as supporting the main idea, providing evidence, supplementing details, transitioning, and expressing emotions), and extended application (such as providing application examples of the word / phrase in exam scenarios and usage methods in similar contexts).

[0125] In practical applications, the analysis content can be displayed in a pop-up window, supporting interactive functions such as one-click bookmarking, pronunciation practice, and example sentence exercises. It also provides a text copying function for the analysis content.

[0126] In this embodiment, an intelligent routing strategy of task complexity grading, model capability matching, and dynamic load scheduling can be adopted. At the same time, an asynchronous parallel calling mechanism is used with other modules. The task complexity grading theory and dynamic load balancing theory are applied to the LLM model scheduling of the reading system, breaking through the computing power bottleneck of single model processing and the efficiency bottleneck of serial calling, so as to achieve dedicated model for dedicated use, optimal computing power, and smooth experience.

[0127] Reference Figure 5 The diagram illustrates the implementation process of the dynamic scheduling large language model instance provided in this application embodiment.

[0128] Optionally, LLM instances with different parameter values ​​and training precision can be dynamically scheduled based on task type, task complexity, and real-time system load. This minimizes computational costs and reduces response latency while ensuring feedback quality. At the same time, model switching is made completely transparent to users through a unified task scheduling interface, without affecting the user experience.

[0129] In some embodiments of this application, after receiving a task request, the task can be labeled with type and complexity to generate corresponding task type labels and complexity labels. Then, based on the task complexity labels, the corresponding candidate model can be matched from a preset model pool. A system load monitoring submodule can be introduced to monitor the current concurrency and resource utilization of each model instance in real time. If the resource utilization of a candidate model exceeds a set threshold (e.g., 80%), a model degradation / queuing strategy can be initiated.

[0130] The preset model pool is a pre-built multi-specification LLM model pool, which can include lightweight models (such as 0.5B parameters for simple tasks), medium-sized models (such as 3B parameters for medium-complexity tasks), and heavyweight models (such as 8B or more parameters for high-complexity tasks). All models are fine-tuned based on the reading scenario to ensure the processing capability of each model in the corresponding task scenario.

[0131] The task complexity classification standard can be represented by dividing tasks into three complexity levels—low, medium, and high—based on the task type of the reading system, and configuring corresponding task type and complexity labels for each level.

[0132] Optionally, task complexity can be categorized based on the computational characteristics of each stage of the process (such as semantic matching, multi-step reasoning, and multi-dimensional integration). Taking the semantic scoring stage as an example, matching options for objective questions is a low-complexity task that can be quickly completed using a lightweight model. However, scoring subjective questions across three dimensions requires the model to simultaneously handle judgments on semantic similarity, depth of understanding, and completeness of answers, making it a high-complexity task. In this case, the multi-model intelligent routing module will route the data to a heavyweight model to ensure the accuracy and interpretability of the scoring.

[0133] For example, specific task complexity grading criteria can be shown in Table 1 below:

[0134] Table 1 Task Complexity Classification Standards

[0135]

[0136] It should be noted that each model specification has a set rated concurrency level. The rated concurrency level refers to the critical value at which resource utilization reaches 80%. As long as the concurrency level does not exceed the aforementioned critical value, the model's resource utilization will not exceed the threshold (i.e., 80%), thus meeting the response latency requirements. In other words, the upper limit is the maximum concurrency level for different models on different machines multiplied by 80%.

[0137] The model degradation / queuing strategy can manifest as the initiation of model degradation for high / medium complexity tasks (such as temporarily scheduling a 0.5B model for medium complexity tasks), while low complexity tasks are initiated in an orderly queue, and a user-friendly "processing" prompt is displayed to the user.

[0138] In some embodiments of this application, after the final candidate model (i.e. the target large language model) is obtained through matching, the scheduling interface can send a task request to the matching LLM instance. After the model is processed, the result can be returned to the front end through a unified interface, and the model call information of this task can be recorded at the same time.

[0139] It should be noted that the multi-model intelligent routing module and the LLM-driven adaptive reading assessment module work asynchronously and in parallel through a message queue. The scoring results of the assessment module are pushed to the multi-model intelligent routing module in real time, so the data collection task does not need to wait for the assessment process to end, which can improve the data collection efficiency.

[0140] In the specific implementation, all model calls can adopt an asynchronous non-blocking call mode. After initiating a model call request, the caller does not need to wait for the model to return a result and can continue to process other tasks. After the model finishes processing, the result is returned through a callback function, reducing task waiting time. For batch low-complexity tasks (such as batch vocabulary parsing and batch correct / incorrect judgment), an asynchronous batch scheduling strategy can be used to integrate multiple small tasks into a single batch task and send it to a lightweight model. After batch processing, the model returns a unified result, reducing the number of model calls and improving scheduling efficiency. The multi-model intelligent routing module can handle question generation and scoring for the assessment module. The task provides asynchronous model call support. After the evaluation module initiates a model request, it continues to execute other tasks. Subsequent processing is performed after the model returns the results, enabling deep parallel collaboration between the two modules. Furthermore, based on historical task model call data, the task complexity grading standard and model matching strategy can be continuously iterated, while the processing capacity of each model can be continuously evaluated, inefficient models can be eliminated, and models adapted to new task scenarios can be added. In addition, a computing power consumption statistics submodule can be established to collect real-time statistics on the computing power consumption and task processing volume of each model, providing data support for the resource allocation of the model pool and enabling refined management of computing power costs.

[0141] Step S206: Generate a learning report based on the dynamic user capability profile, semantic scoring results, parsed content, and feedback information.

[0142] In some embodiments of this application, structured and personalized learning reports can be generated based on asynchronously collected dynamic user capability profiles, semantic scoring results, parsed content, and feedback information.

[0143] Specifically, an asynchronous, non-blocking approach can be used to collect multi-dimensional learning data from the user profiling and ability scoring module, the adaptive assessment module, and the intelligent word and sentence collection and parsing module. This includes user performance (scores / accuracy rates for each question type, total score), mastery of article types, collection results of weak words and sentences, LLM scoring and feedback content, and user learning behavior data (reading time, time spent answering questions). This avoids report generation blocking due to data not being returned from a certain module. Then, the LLM generates a Prompt based on the multi-dimensional learning data and customized reports, automatically performing structured and personalized generation.

[0144] Optionally, the report content can include data statistics, problem analysis, and recommendations. The core content of the report can include: an overview of the learning session, including total score, accuracy rate, reading time, and completion rate, along with comparison curves with the user's historical level and the average level of the system group; an ability radar chart, which quantitatively displays four dimensions: mastery of question types, comprehension of article types, depth of semantic understanding, and answering speed, intuitively presenting strengths and weaknesses and comparisons with the system group; detailed analysis, including statistics on accuracy rates for each question type / article type, classification of incorrect questions (e.g., 60% of reasoning and judgment errors), and error attribution analysis (e.g., semantic comprehension bias, detail positioning errors, knowledge gaps, etc.); a summary of weak words and phrases, including a list of weak words and phrases collected during the session, along with core explanations and links to related incorrect questions; personalized improvement suggestions, specifically providing actionable, phased improvement suggestions based on weaknesses, including targeted article recommendations, question type training, and word and phrase learning plans, along with corresponding learning resources; and a follow-up learning plan, specifically generating a short-term (1-week) learning plan based on the user's current level and weaknesses, clearly defining daily learning tasks and goals.

[0145] In practical applications, the generated learning reports also support online viewing, image export, and one-click sharing. Furthermore, weak vocabulary and phrases in the report can be directly accessed through a preset phrase collection for further study, and recommended content in the improvement suggestions can also directly lead to a personalized article recommendation list. This application embodiment does not impose any limitations on these aspects.

[0146] In some embodiments of this application, in response to an image upload command, the uploaded image can be preprocessed and text recognized, the recognition result can be grammatically validated and semantically corrected, the corrected text can be structurally annotated and matched with the system's article tagging system, the recognized and processed text can be asynchronously pushed through a message queue, and a selection command for the target article can be generated.

[0147] Specifically, the uploaded English articles in paper / image format can be preprocessed, highly recognized, and structurally processed. After recognition and processing are completed, the text is asynchronously pushed to the LLM-driven adaptive reading assessment process via a message queue.

[0148] Optionally, preprocessing of user-uploaded paper / image English articles can include noise reduction, correction, and cropping to improve the accuracy of subsequent OCR recognition. Then, a high-precision OCR recognition engine can be called to convert the image into editable text. At the same time, a dual verification mechanism of grammatical verification and semantic correction can be used to remove incorrectly recognized words / sentences, improving the recognition accuracy to over 98%. Finally, the corrected text can be structurally annotated (such as dividing into paragraphs and annotating core sentences) and automatically matched with the system's article tagging system (such as subject matter, difficulty, and examination focus). After recognition and processing are completed, the text can be asynchronously pushed to the LLM-driven adaptive reading assessment process through a message queue, without requiring users to wait. This achieves the same assessment system for paper / image articles and digital articles, breaking down the learning barriers between digital and paper articles.

[0149] In this application embodiment, at the theoretical / method level, a probabilistic adaptive assessment method that deeply integrates probability theory and LLM is proposed. This method introduces the three core theories of probability theory—Bayesian posterior probability, maximum likelihood estimation, and binomial distribution—into the entire LLM reading assessment process. It constructs a technical solution for probabilistic question generation, probabilistic scoring, and dynamic probability adjustment of difficulty. It sets multiple thresholds and mapping rules, such as mastery probability intervals, likelihood value quantification, and low probability judgment, upgrading the assessment from experience-based to scientific. This method can overcome the industry bottleneck of existing assessments that lack mathematical theoretical support and rely solely on mechanical scoring and experience-based question generation.

[0150] At the theoretical / methodological level, a percentile ranking profile construction method based on multi-dimensional comparison of group performance within the system is proposed. This method is a multi-dimensional percentile ranking method based on the performance of users on the same platform within the system, replacing the traditional external general grading standard. It achieves accurate determination of user level from dimensions such as question type, article type, and difficulty level. It constructs a technical solution of "full-dimensional data collection - percentile calculation - dynamic level determination - real-time update", sets a five-level level division rule, and the rating results are more in line with the actual learning scenario, which can break through the industry bottleneck of the disconnect between the existing ability rating and actual learning needs.

[0151] At the technical architecture level, a dual-path intelligent recommendation architecture is proposed, which integrates personalization and group pattern analysis. This dual-path recommendation architecture for reading systems combines group learning data mining methods such as collaborative filtering and sequence pattern mining with personalized gap-filling methods such as three-dimensional profile matching. By setting core parameters such as weighted fusion weights for the main path and the side path, and user similarity thresholds, it can solve the industry contradiction that the existing single recommendation logic cannot take into account both personalized gap filling and group learning patterns, and achieve a substantial improvement in recommendation accuracy.

[0152] At the technical architecture level, a cross-module asynchronous parallel processing architecture based on the producer-consumer model is proposed. This architecture applies the producer-consumer model to the module collaboration of the reading system, constructs a multi-process, task queue, and non-blocking I / O technical solution, sets four-level priority rules, timeout fallback mechanism, and load balancing threshold for reading tasks, and realizes cross-module collaborative parallelism of modules such as evaluation, word and sentence collection, and model routing. This can overcome the industry bottleneck of low system efficiency and slow response caused by the existing full-module serial processing.

[0153] At the technical strategy level, a multi-model intelligent routing strategy that integrates task complexity and load balancing is proposed. This is manifested in the design of a three-level complexity classification standard for reading tasks (low / medium / high) and an adaptation scheme for multi-specification domain fine-tuning model pools. It proposes a dedicated model-specific and dynamically adapted LLM routing method, sets a model resource utilization threshold, and realizes asynchronous non-blocking model calls and batch task scheduling. This can solve the industry pain points of wasteful computing power and slow response caused by the existing single model processing all tasks, while taking into account both computing power cost and processing efficiency.

[0154] At the technical strategy level, a strategy for the automated collection and customized analysis of weak words and phrases based on scoring thresholds is proposed. This is manifested in the construction of a closed-loop technical solution of "low score trigger - automatic collection - multi-dimensional annotation - LLM customized analysis - asynchronous collaboration". Three-dimensional scoring trigger thresholds are set to realize the automated collection of weak words and phrases and the customized analysis based on user level. It works asynchronously and in parallel with the assessment module, which can break through the existing industry bottlenecks of relying on manual collection, lack of analysis closed loop, and low collection efficiency.

[0155] At the technical strategy level, an asynchronous adaptation and unified evaluation strategy for electronic and paper articles is proposed. This is manifested in the design of a technical solution of "OCR image preprocessing - syntactic and semantic dual verification - structured annotation - asynchronous evaluation access". This achieves asynchronous adaptation between digital articles and paper / image articles, and uses the same LLM adaptive evaluation system. The dual verification mechanism improves the OCR recognition accuracy to over 95%, which can break through the industry bottleneck of learning barriers between electronic and paper articles and low OCR recognition accuracy, and improve the system's versatility and practicality.

[0156] To facilitate a better understanding of the online adaptive reading and adaptive evaluation method provided in the embodiments of this application by those skilled in the art, the following description is provided in conjunction with specific embodiments:

[0157] First, when users enter the online reading system, they can actively trigger the "Smart Reading Companion" command or receive intelligent passive recommendations based on the user's learning intervals and profile updates. The system can also read full-dimensional data of the user profile (including multi-dimensional percentile rankings and level levels within the system, weak knowledge points, etc.). For example, the user's accuracy rate for detail comprehension questions is 60% (35th percentile within the system, basic level), mastery of popular science articles is 55% (30th percentile within the system, basic level), and the percentiles for reasoning and judgment and word and sentence comprehension are 45% and 50% respectively (advanced level). The overall reading level is basic, and the weak knowledge points are core vocabulary of popular science articles and detail positioning skills.

[0158] Then, the personalized article recommendation module can start a dual-path recommendation architecture including the main path and the side path. The main path recommendation branch can generate a personalized gap-filling candidate set through three-dimensional profile matching, while the side path recommendation branch can generate a group pattern candidate set through collaborative filtering and sequence pattern mining. The personalized gap-filling candidate set and the group pattern candidate set can be weighted, fused and rearranged to output a personalized article recommendation list with adaptation reasons. For example, the main recommendation branch generates a personalized gap-filling candidate set of medium difficulty, containing core vocabulary and detailed comprehension questions from popular science articles, through three-dimensional profile matching. The side recommendation branch selects 50 similar users (basic level, weak in popular science articles / detail comprehension) with a similarity of ≥0.7 through cosine similarity calculation, and includes articles with an accuracy rate of 70% or higher in these users into the candidate set. Then, through sequence pattern mining (support ≥0.5, confidence ≥0.8), it discovers the association patterns of similar users reading similar advanced articles after reading articles with detailed comprehension questions in popular science articles, and sorts them initially. Finally, the candidate sets from both paths are weighted and merged according to the main path (0.6) and the side path (0.4) and then rearranged to output a recommendation list, along with recommendation tags that match the gap-filling needs of popular science articles and the high accuracy of similar users.

[0159] After selecting a target article, users can enter an LLM-driven adaptive reading assessment process, which can be guided by probability theory and processed asynchronously and in parallel.

[0160] Specifically, the adaptive reading assessment module can calculate the probability of a user's mastery of each knowledge point based on Bayesian posterior probability. This allows the LLM fine-tuning model to generate appropriate questions in a pattern of basic, advanced, and remedial questions, and then perform quality verification. For example, for LLM probabilistic adaptive question generation, if the adaptive reading assessment module calculates the user's mastery probability of core vocabulary and detailed knowledge points in a popular science article to be 40% and 35% respectively (assuming it falls within the 30%-70% easy-to-improve range) based on Bayesian posterior probability, then the LLM can generate 8 objective questions and 2 subjective questions focusing on these knowledge points, in a pattern of 50% basic questions, 30% advanced questions, and 20% remedial questions, and output the results after quality verification.

[0161] After a user answers and submits their answers, the adaptive reading assessment module can perform semantic probabilistic scoring based on maximum likelihood estimation. Simultaneously, it uses a binomial distribution P to determine the significance of the accuracy rate in real time; if P < 0.05, the assessment difficulty can be dynamically adjusted. For example, in asynchronous parallel assessment and data collection, if a user submits their answers and scores less than 6 points across the three dimensions of three detail comprehension questions, the adaptive reading assessment module can perform semantic probabilistic scoring based on maximum likelihood estimation (e.g., a likelihood value of 0.3 corresponds to a semantic similarity of 2 points), indicating the reasons for the lost points as errors in detail location and semantic comprehension. Simultaneously, it asynchronously pushes the original text information of these three incorrect questions to the intelligent word and sentence collection and parsing module via a message queue. This module can immediately and automatically collect the corresponding sentences into a preset sentence database, such as a "word and sentence treasure trove," and annotate them with multi-dimensional information, such as source, related question types, and error type, without waiting for the assessment process to finish.

[0162] When a user clicks on any word or phrase in a preset sentence library, such as the "Word and Phrase Treasure Trove," the system can asynchronously invoke the multi-model intelligent routing module to match the corresponding LLM model. The matched LLM model then returns customized parsing content in real time, displayed in a pop-up window, and supports interactive functions such as follow-up reading and practice. For example, for asynchronous model invocation and parsing, when a user clicks on a collected sentence, the system asynchronously invokes the multi-model intelligent routing module to match a 3B medium-sized model for customized parsing. This outputs the meaning / collocation of core vocabulary in the popular science text, the syntactic components of the sentence, and clues from the original text for locating details, all displayed in a pop-up window. Users can then perform follow-up reading, practice with example sentences, and other operations.

[0163] After reading and assessment, the learning report generation module can asynchronously collect comprehensive learning data from each module and automatically generate structured and personalized learning reports using LLM. Users can view these reports online, export them as PDFs, or share them with one click. For example, the learning report generation module asynchronously collects data from each module and generates personalized reports. For instance, a capability radar chart can visually display the basic level of detailed comprehension and understanding of popular science articles (compared to the system population). Error attribution analysis indicates that errors in detailed location are the main problem (accounting for 60%), and provides targeted improvement suggestions (such as recommending 5 medium-difficulty popular science articles for detailed location training and compiling a core vocabulary list for popular science articles). Simultaneously, a one-week short-term learning plan is generated, clearly defining daily learning tasks and goals.

[0164] Optionally, the system can also support users to select the photo reading function. Specifically, after uploading an article image, the photo reading OCR extension module can asynchronously complete image preprocessing, high-precision recognition, syntactic and semantic double verification, and text structuring processing. It is then pushed to the adaptive reading assessment module through a message queue, and repeats the LLM-driven adaptive reading assessment process and learning report generation steps to achieve adaptive learning of paper / image articles.

[0165] Furthermore, it can also perform full-process data feedback, which means that all the learning data of this room change is updated asynchronously to the user's data center in real time, recalculating the user's multi-dimensional percentile ranking and level, and updating the recommendation algorithm weights and multi-model routing strategy parameters. That is, recalculating the percentile ranking of the user's detailed understanding questions and popular science articles, and updating the weight coefficients of the dual-path recommendation algorithm and the parameters of the multi-model routing strategy, so as to realize the self-iterative optimization of the system.

[0166] In the embodiments of this application, the common problems existing in the industry of existing online reading systems can be fundamentally solved, and the learning effect and system efficiency can be improved in a dual way. Each beneficial effect and innovation corresponds one by one, and experimental verification shows that it has significant quantitative value.

[0167] Specifically, the percentile ranking profile construction method based on multi-dimensional comparison of group performance within the system makes profile construction more accurate. This is manifested in the dynamic profile based on multi-dimensional percentile ranking of group performance within the system, which improves the rating accuracy by more than 85% compared to existing external grading standards. The multi-dimensional level judgment is more in line with actual learning scenarios, providing more reliable basic data for subsequent recommendations and assessments. The probabilistic adaptive assessment method based on the deep integration of probability theory and LLM makes the assessment system more scientific. This is manifested in the LLM assessment guided by probability theory, achieving probabilistically accurate question generation for each student and each text, improving the objectivity of semantic scoring by 90% and the dynamic adaptability of assessment difficulty by 80%. This approach fundamentally addresses the problems of existing assessments being experience-based, subjective, and lacking dynamic adaptation, ensuring that assessment results accurately reflect users' real-time reading abilities. Its dual-path intelligent recommendation architecture, integrating personalization and group learning patterns, makes recommended content more relevant. This dual-path architecture balances personalized supplementation with group learning patterns, improving the matching degree between recommended content and user learning needs by over 75%. It effectively avoids the problems of overly difficult content discouraging users or overly easy content proving ineffective, ensuring that recommended content better meets users' actual improvement needs. Furthermore, its cross-module asynchronous parallel processing architecture based on a producer-consumer model makes the system more efficient. This cross-module asynchronous parallel processing mechanism improves the overall system processing efficiency. It improves performance by over 70%, reduces average single-task response latency by 80%, and completely solves the system lag and long waiting time problems caused by existing serial processing, significantly improving the user experience. Based on a multi-model intelligent routing strategy that integrates task complexity and load balancing, it optimizes computing resources through multi-model intelligent routing and asynchronous calling strategies. This enables on-demand allocation and dedicated use of computing resources, reducing computing costs by over 60%. High-complexity task response latency is <3s, and low-complexity task response latency is <500ms, balancing processing depth, computing cost, and system response efficiency, avoiding resource waste. Based on a scoring threshold-triggered automatic weak word / phrase accumulation and customized parsing strategy, it further enhances... The system achieves more efficient knowledge gap identification through a closed-loop process of automated identification and customized analysis triggered by scoring thresholds. This eliminates the need for manual user intervention, increasing the efficiency of knowledge gap identification by 100%. It seamlessly integrates error identification, knowledge accumulation, intelligent explanation, and interactive learning, helping users systematically fill knowledge gaps. Furthermore, the system's asynchronous adaptation to both electronic and printed texts, along with a unified assessment strategy, broadens its applicability. This strategy improves OCR accuracy to over 98%, breaking down learning barriers between digital and printed texts. It can adapt to various English learning scenarios, including offline printed reading and online digital reading, significantly enhancing the system's versatility and practicality.Based on the innovative integration of multiple modules, the learning reports are more instructive. These reports are structured and generated from asynchronously collected data across all dimensions, taking into account data statistics, error attribution, personalized suggestions, and follow-up planning. Compared to existing homogeneous data statistics reports, their instructiveness and executability are improved by 90%, enabling users to develop precise follow-up learning plans.

[0168] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of this application are not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of this application.

[0169] This application also provides an electronic device, see embodiments thereof. Figure 6 The provided electronic device 600 includes a memory 610, a processor 620, and a computer program 611 stored in the memory 610 and capable of running on the processor 620. When the computer program 611 is executed by the processor, it implements the various processes of the above-described online adaptive reading and adaptive evaluation method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0170] This application also provides a computer-readable storage medium, see embodiments thereof. Figure 7 The computer-readable storage medium 700 provides a computer program 611 stored on it. When the computer program 611 is executed by the processor, it implements the various processes of the above-described online adaptive reading and adaptive evaluation method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0171] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0172] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of the embodiments of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not necessarily limited to those steps or modules explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products, or devices. The division of modules in the embodiments of this application is merely a logical division; in actual applications, there may be other division methods. For example, multiple modules may be combined into or integrated into another system, or some features may be ignored or not performed. Additionally, the shown or discussed mutual coupling or direct coupling or communication connection may be through some interface, and the indirect coupling or communication connection between modules may be electrical or other similar forms, none of which are limited in the embodiments of this application. Furthermore, the modules or sub-modules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed among multiple circuit modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the embodiments of this application.

[0173] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0174] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0175] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or modules through some interfaces, and may be electrical, mechanical, or other forms.

[0176] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0177] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0178] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0179] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0180] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0181] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes; these computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0182] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0183] Finally, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0184] The technical solutions provided in the embodiments of this application have been described in detail above. Specific examples have been used in the embodiments of this application to illustrate the principles and implementation methods of the embodiments of this application. The description of the above embodiments is only for the purpose of helping to understand the methods and core ideas of the embodiments of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the embodiments of this application. Therefore, the content of this specification should not be construed as a limitation on the embodiments of this application.

Claims

1. An online adaptive reading and adaptive evaluation method, characterized in that, The method, applied to an online reading system, includes: Obtain the user's full-dimensional historical learning data, and construct a dynamic user capability profile based on the full-dimensional historical learning data and multi-dimensional comparison information of the performance of the same platform group within the system; In response to a recommendation trigger command, a dual-path candidate set is obtained based on the dynamic user capability profile through a preset dual-path recommendation architecture. The dual-path candidate set is then fused and reordered to obtain a personalized article recommendation list. In the preset dual-path recommendation architecture, the main path recommendation branch performs matching recommendations based on the dynamic user capability profile, while the side path recommendation branch performs mining recommendations based on collaborative filtering and sequence patterns. In response to the selection instruction of the target article in the personalized article recommendation list, the system uses probability theory to guide the preset large language model to generate an appropriate question, performs real-time probabilistic semantic scoring on the answers submitted for the appropriate question to obtain a semantic score result, and outputs feedback information according to a hierarchical feedback strategy. Weak words and phrases are collected from the low-scoring question information indicated by the semantic scoring results. The collected weak words and phrases are annotated with multi-dimensional information, and the weak words and phrases with multi-dimensional information annotation are added to a preset word and phrase library. The low-scoring question information is pushed asynchronously through a message queue. In response to a touch command on any word or phrase in the preset word and phrase library, a task parsing request is generated for the touched word or phrase. Based on the task type, task complexity, and real-time system load of the task parsing request, a target large language model is matched from the preset model pool. The target large language model instance is used to parse the touched word or phrase to obtain the parsed content. A learning report is generated based on the dynamic user capability profile, the semantic scoring results, the parsed content, and the feedback information; The dual-path candidate set includes a personalized gap-filling candidate set and a group pattern candidate set. The process of obtaining the dual-path candidate set based on the dynamic user capability profile using a preset dual-path recommendation architecture includes: The system acquires full-dimensional data of the dynamic user ability profile, structured tags of the article resource library, and big data on the learning behavior and answer performance of all users within the system. The full-dimensional data of the dynamic user ability profile is used to indicate the user's strengths and weaknesses, current reading level, weak knowledge points and vocabulary, and learning progress and improvement trend curve. Article resources are initially screened based on the user's strengths and weaknesses and the structured tags of the article resource library. A knowledge point gap-filling weight mechanism is introduced, assigning recommendation weights to articles containing the weak knowledge points and vocabulary after the initial screening. The weight coefficient of the recommendation weight is inversely correlated with the user's mastery of the weak points. Based on the learning progress and improvement trend curve, the basic difficulty of recommended articles is matched with the current reading level, and a preset proportion of advanced knowledge points are configured. Furthermore, the system considers the answer performance indicated by the big data on the learning behavior and answer performance of all users within the system. The accuracy rate iterates in real time to adjust the difficulty adaptation coefficient, and outputs a personalized gap-filling candidate set based on the difficulty adaptation coefficient. Using question type score, article type preference, and reading level as feature dimensions, the similarity of learning features between the target user and other users in the system is calculated using cosine similarity. The top N similar users with a similarity of no less than a preset similarity threshold are selected, and articles with an accuracy rate of no less than a preset percentage from the top N similar users are included in the initial candidate set. Frequent itemset support threshold and confidence threshold are set to mine frequent association patterns in the reading article sequences of system users, and the initial candidate set is sorted in the first round according to the frequent association patterns. Articles that do not match the current reading level and gap-filling needs are removed from the sorted candidate set, and the size of the filtered side candidate set is controlled to a preset second percentage of the size of the personalized gap-filling candidate set to obtain the group pattern candidate set.

2. The method according to claim 1, characterized in that, The process of fusing and reordering the dual-path candidate sets to obtain a personalized article recommendation list includes: The personalized gap-filling candidate set and the group pattern candidate set are merged to form a fusion candidate set; there are no duplicate articles in the fusion candidate set. Obtain the pre-set main path recommendation weight and side path recommendation weight, and use the main path recommendation weight and side path recommendation weight to calculate the main path matching score and side path mining score for each article in the fusion candidate set; The comprehensive score of each article in the fusion candidate set is calculated using the main path matching score and the side path mining score; the comprehensive score is used to indicate the comprehensive matching degree. The articles in the fusion candidate set are sorted from highest to lowest according to the overall matching degree, and a personalized article recommendation list is output; wherein, each article in the personalized article recommendation list is accompanied by a matching reason tag.

3. The method according to claim 1, characterized in that, The method of generating suitable questions by guiding a pre-set large language model through probability theory includes: Based on the dynamic user capability profile, calculate the Bayesian posterior mastery probability of the target user for each knowledge point; Based on the Bayesian posterior mastery probability, target knowledge points that fall within the target user's easy-to-improve range are selected, and a question-generating paradigm combining basic questions, improvement questions, and remedial questions is adopted, with corresponding question types generated according to a preset ratio; among them, basic questions are matched with the target user's current reading level, improvement questions are higher than the target user's current reading level, and remedial questions are targeted at the target user's weak knowledge points; The refined large language model is called in combination with customized prompt words to generate test questions according to the corresponding question types, and the quality of the generated test questions is verified. Output the questions that have passed the quality check.

4. The method according to claim 1 or 3, characterized in that, The step of performing real-time probabilistic semantic scoring on the submitted answers to the adapted questions to obtain semantic scoring results includes: Obtain the probability distribution fitting model of the semantic feature vectors of user answers and standard answers constructed based on maximum likelihood estimation, and obtain the corresponding quantization rules of the preset likelihood value interval and semantic scoring results; Using the probability distribution fitting model and the corresponding quantification rules, the answer content is scored separately from semantic similarity, comprehension depth, and answer completeness, and then comprehensively quantified to obtain a semantic score result. Among them, the semantic similarity is quantified based on the probability value of maximum likelihood estimation, the comprehension depth is used to evaluate the target user's deep understanding of the question's test points, and the answer completeness is used to evaluate the accuracy of the options for objective questions and the coverage of the key points for subjective questions.

5. The method according to claim 1 or 3, characterized in that, The hierarchical feedback strategy is set based on the score range of the semantic scoring result; the output of feedback information according to the hierarchical feedback strategy includes: For users whose semantic scoring results fall within the first score range, extended feedback such as knowledge point expansion and problem-solving techniques for similar question types will be provided. For users whose semantic scoring results fall within the second scoring range, targeted feedback including explanations and guidance on problem-solving strategies will be provided. For users whose semantic scores fall within the third score range, the system provides detailed feedback including standard answers, error attributions, original text location, and breakdown of problem-solving steps.

6. The method according to claim 1, characterized in that, The method further includes: In response to the image upload command, the system preprocesses and recognizes the uploaded image, performs grammatical verification and semantic correction on the recognition results, performs structured annotation on the corrected text and matches it with the system's article tagging system, asynchronously pushes the recognized and processed text through a message queue, and generates a selection command for the target article.

7. An online reading system, characterized in that, The system includes: The user profile and ability scoring module is used to obtain users' full-dimensional historical learning data and construct dynamic user ability profiles based on the full-dimensional historical learning data and multi-dimensional comparison information of the performance of the same platform group within the system. A personalized article recommendation module is used to respond to recommendation trigger commands. It obtains a dual-path candidate set based on the dynamic user capability profile using a preset dual-path recommendation architecture, and then merges and reorders the dual-path candidate sets to obtain a personalized article recommendation list. In the preset dual-path recommendation architecture, the main recommendation path performs matching recommendations based on the dynamic user capability profile, while the side recommendation path performs recommendations based on collaborative filtering and sequence pattern mining. The dual-path candidate set includes a personalized gap-filling candidate set and a group pattern candidate set. The selection process specifically includes: acquiring full-dimensional data of the dynamic user ability profile, structured tags of the article resource library, and big data on the learning behavior and answer performance of all users within the system; wherein, the full-dimensional data of the dynamic user ability profile is used to indicate the user's strengths and weaknesses tags, current reading level, weak knowledge points and vocabulary, and learning progress and improvement trend curve; the article resources are initially screened based on the user's strengths and weaknesses tags and the structured tags of the article resource library; a knowledge point gap filling weight mechanism is introduced to assign recommendation weights to articles containing the weak knowledge points and vocabulary after the initial screening, and the recommendation weights are... The weighting coefficient of the recommendation is inversely correlated with the user's mastery of weaknesses. Based on the learning progress and improvement trend curve, the basic difficulty of the recommended articles is matched with the current reading level, and advanced knowledge points are configured in a preset proportion. The difficulty adaptation coefficient is iterated in real time based on the answer accuracy rate indicated by big data on the learning behavior and answer performance of all users in the system, and a personalized gap-filling candidate set is output based on the difficulty adaptation coefficient. Using question type score, article type preference, and reading level as feature dimensions, the similarity of the learning features between the target user and other users in the system is calculated using cosine similarity, and users with a high similarity are selected. For the top N similar users with a preset similarity threshold, articles whose correct answer rate is not less than a preset percentage are included in the initial candidate set; a frequent itemset support threshold and a confidence threshold are set to mine frequent association patterns in the sequence of articles read by system users, and the initial candidate set is sorted in the first round according to the frequent association patterns; articles that do not match the current reading level and the need to fill in the gaps in weak knowledge points are removed from the sorted candidate set, and the size of the filtered side candidate set is controlled to a preset second percentage of the size of the personalized gap-filling candidate set to obtain the group pattern candidate set; The adaptive reading assessment module is used to respond to the selection instruction of the target article in the personalized article recommendation list, guide the preset large language model to generate adaptive questions through probability theory, perform real-time probabilistic semantic scoring on the answers submitted for the adaptive questions to obtain semantic scoring results, and output feedback information according to the hierarchical feedback strategy. The intelligent word and phrase collection and parsing module is used to collect weak words and phrases from the low-scoring question information indicated by the semantic scoring results, perform multi-dimensional information annotation on the collected weak words and phrases, and add the multi-dimensional information-annotated weak words and phrases to the preset word and phrase library; the low-scoring question information is pushed asynchronously through a message queue; The multi-model intelligent routing module is used to respond to touch commands for any word or phrase in the preset word and phrase library, generate a task parsing request for the touched word or phrase, and match a target large language model from the preset model pool according to the task type, task complexity and real-time system load of the task parsing request. The target large language model instance is used to parse the touched word or phrase to obtain the parsed content. The learning report generation module is used to generate a learning report based on the dynamic user capability profile, the semantic scoring results, the parsed content, and the feedback information.

8. An electronic device, characterized in that, include: A processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the online adaptive reading and adaptive evaluation method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, and when executed by a processor, the computer program implements the online adaptive reading and adaptive evaluation method as described in any one of claims 1 to 6.

Images