An adaptive improvement method and device for reading ability assessment

Abstract

The application discloses a self-adaptive improvement method and device for reading ability evaluation, which comprises the following steps: terminating judgment is performed based on the ability preset condition of the IRT self-adaptive algorithm; a reading path, information marking behavior, draft reasoning process and final score of a tested person are recorded in real time to generate an ability and strategy report and determine a reading process analysis and evaluation result; based on a reading evaluation index, necessary parameters, optional parameters and unlocking logic are determined; meanwhile, a NLP model and voice emotion analysis are used to collect voice interaction data in the reading process of a user; a reading task library covering different school stages and different reading targets is built, and a matching operation guide is formulated; a standardization intervention script library corresponding to a fixed intervention prompt is constructed, and the level and timing of reading intervention are determined; and a front-end chart engine is used in combination with a back-end graph to generate a visual ability report for reading result interpretation, so that the reading level and literacy of a reader are improved.

Description

Technical Field

[0001] This invention belongs to the field of reading assessment technology, and in particular relates to an adaptive improvement method and apparatus for reading ability assessment. Background Technology

[0002] Currently, there are four main types of reading assessment methods: standardized achievement tests, performance-based assessments, dynamic assessments, and portfolio assessments. These methods all have significant drawbacks: A) Standardized reading assessments have limited sampling, with fixed question counts and text types, failing to cover all necessary reading skills (such as digital reading strategies and cross-text integration abilities), and are prone to incomplete assessments due to sampling bias; B) They prioritize results over process: focusing only on final scores, they fail to track students' reading thought processes (such as reasoning logic and information filtering strategies), making it difficult to determine whether "cannot do" a question stems from a lack of ability or inappropriate strategies; C) They are outdated: test content is updated frequently (usually every 1-3 years), making it difficult to adapt to the new forms of digital reading (such as social media texts and short video scripts), resulting in a disconnect between the assessment scenario and real-world reading scenarios. B. Performance-based reading assessment is highly subjective, lacking a unified scoring standard. Judgments on task outcomes (such as the depth of reading reflections and the logic of mind maps) are easily influenced by the experience and preferences of the raters, compromising the reliability of the results. Implementation costs are high; designing tasks that align with the objectives (such as interdisciplinary reading tasks) and grading a large number of non-standardized outcomes (such as debate videos) require significant time and manpower, making large-scale promotion (such as school-wide or district-wide assessments) difficult. Furthermore, it is susceptible to interference from non-reading abilities; task completion may depend on other skills (such as writing and oral expression). For example, a student may have excellent reading comprehension but receive a low report score due to poor writing, easily leading to misjudgments of reading level. C. Dynamic reading assessment has a low degree of standardization. Intervention methods (such as the type and timing of prompts) depend on the assessor's judgment, and the operation varies greatly among different assessors, making it difficult to compare assessment results horizontally (e.g., the potential levels of students in Class A and Class B cannot be directly compared). It places high demands on assessors, requiring them to receive professional training and have the ability to judge "student bottlenecks" and "appropriate prompt intensity," which is difficult for ordinary teachers or parents to conduct independently. The results are difficult to interpret, as the assessment results are mostly "potential development ranges" (e.g., "can master complex text reasoning after prompting") rather than explicit scores or grades, and schools and parents may overlook their value due to difficulty in understanding. D. Portfolio reading assessment lacks a unified comparative standard. The content of different students' portfolios may vary greatly (e.g., Student A focuses on science fiction reading, while Student B focuses on prose reading). It cannot quickly rank students horizontally like standardized tests, making it difficult to meet the "outcome-oriented" needs of school admissions and class placement. The management cost is high. Long-term collection, classification, and preservation of paper or electronic portfolios (such as reading materials from six years of primary school) requires a lot of space and time, and materials are easily lost or confused due to personnel changes or storage errors. The assessment focuses on the "past." Portfolios record reading behaviors and achievements that have already occurred, and have weak predictive power for students' future reading abilities, making it difficult to identify "imminent ability gaps" (such as gaps in the ability to analyze complex texts that will be learned later).

[0003] Therefore, the four mainstream reading assessment methods mentioned above, as well as other non-mainstream assessment methods, lack a systematic approach, are not integrated with computer technology, and lack dynamic assessment data. They are not suitable for widespread application in reading in the new era and require improvement in areas such as interactive communication and dynamic assessment. Summary of the Invention

[0004] This invention aims to overcome the shortcomings of existing technologies and provide an adaptive method and apparatus for improving reading ability assessment. This method allows users to dynamically assess their reading skills, eliminates the need for manual intervention, saves costs, and saves dynamic reading data, significantly improving readers' reading level and literacy. The difficulty and type of questions are dynamically adjusted based on real-time reading performance (such as accuracy and answering time), automatically covering multiple skills and avoiding sampling bias from a fixed number of questions.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: An adaptive improvement method for reading ability assessment includes: The IRT adaptive algorithm is used to determine termination based on preset conditions. Logs or backups record the test taker’s reading path, information marking behavior, and draft reasoning process in real time. Combined with the final score, they generate ability reports and strategy reports to determine the reading process analysis and evaluation results. Based on reading assessment metrics, the required parameters, optional parameters, and unlocking logic are determined; at the same time, NLP models and voice emotion analysis are used to collect voice interaction data during the user's reading process. Establish a reading task library covering different educational stages and reading goals, and develop corresponding operation guidelines; Build a standardized intervention script library corresponding to setting intervention prompts, and determine the level and timing of reading interventions; By combining a front-end chart engine with back-end graphics, a visual report explaining the reading results is generated.

[0006] The present invention also provides an adaptive improvement device for reading ability assessment, comprising: The preset module is used to make a termination judgment based on the preset conditions of the IRT adaptive algorithm. The analysis module is used to determine the required parameters, optional assessment parameters, and unlocking logic based on reading assessment indicators; at the same time, it uses NLP models and voice emotion analysis to collect voice interaction data during the user's reading process. Supporting modules are used to build a reading task library covering different learning stages and reading goals, and to develop supporting operation guidelines; The building module is used to build a standardized intervention script library corresponding to setting intervention prompts and to determine the level and timing of reading interventions; The visualization module is used to generate visual reports that explain reading results by combining a front-end chart engine with back-end graphics.

[0007] The reporting module is used to record the test taker's reading path, information marking behavior, and draft reasoning process in real time during the reading process, and combine them with the final score to generate ability reports and strategy reports, and determine the reading process analysis and evaluation results. Compared with the prior art, the present invention has the following beneficial effects: This invention is low-cost, requiring only one software program to store thousands of public domain books and copyrighted books owned by the company, creating free resources; and the reading and evaluation process no longer requires human intervention, greatly saving labor costs, time costs, and evaluation space costs.

[0008] The system combines real-time dynamic evaluation (process-oriented evaluation) with summative evaluation. Each reading session collects data on the reader's reading behavior, including the time of reading, duration of each reading session, preferred book types, reading speed, reading assessment results, and a reading ability growth curve, automatically generating process-oriented evaluations. Summative evaluations are provided for each book read, each semester's reading, and each academic year's reading, with periodic summaries forming a perfect blend of process and summative evaluation.

[0009] During the reading process, large-scale artificial intelligence models such as DeepSeek can be used to assist in reading and evaluation. These models can tailor personalized assessment tools based on the reader's evaluation results, making the assessment more individualized.

[0010] The assessment data is more objective and has a higher degree of credibility. Compared with manual reading assessment, reading assessment methods and systems based on large models are more intelligent, avoiding the emotional factors introduced by human assessment, and the assessment results and data are more objective, fair and credible. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the adaptive improvement method for reading ability assessment according to the present invention. Figure 2 This is a schematic diagram of the hardware layer of the adaptive improvement system for reading ability assessment of the present invention; Figure 3 This is an operation diagram of the adaptive improvement method for reading ability assessment of the present invention; Figure 4 This is a schematic diagram of the adaptive improvement device for reading ability assessment according to the present invention. Detailed Implementation

[0012] To better illustrate the present invention, a detailed description is provided below in conjunction with the accompanying drawings.

[0013] Example 1

[0014] like Figure 1 As shown, this invention provides an adaptive improvement method for reading ability assessment, comprising: S1. Termination judgment is made based on the preset conditions of the IRT adaptive algorithm; S2. During the reading process, logs or backups record the test taker's reading path, information marking behavior, and draft reasoning in real time. Combined with the final score, an ability report and a strategy report are generated to determine the reading process analysis and evaluation results. S3. Based on reading assessment indicators, determine the required parameters, optional assessment parameters, and unlocking logic (preconditions); at the same time, enable NLP models and voice emotion analysis to collect voice interaction data during the user's reading process; S4. Establish a reading task library covering different learning stages and reading goals, and develop corresponding operation guidelines. Classify reading assessments by type and provide task templates (such as mind map frameworks and debate scripts). The system can adjust the reading assessment type based on the student's selected grade (such as changing the text topic), reducing time and manpower costs. Reading assessment types include basic reading (such as summarizing paragraph main ideas), in-depth reading (such as comparing viewpoints in two texts), and applied reading (such as designing solutions based on texts).

[0015] S5. Construct a standardized intervention script library corresponding to the intervention prompts, and determine the level and timing of reading interventions. Specific operation: Pre-set common reading obstacles (such as "unfamiliar with keywords," "unable to understand cause-and-effect relationships," "unable to understand long and complex sentences"), and provide 3-4 levels of intervention prompts for each reading obstacle (e.g., for "keyword obstacle": Level 1 prompt "find context synonyms," Level 2 prompt "view text annotations," Level 3 prompt "directly explain the meaning of the word"). The system executes according to the standardized intervention scripts (keyword obstacle, cause-and-effect obstacle, Level 3 prompt), avoiding subjective adjustments to the prompt intensity and ensuring consistent operation across different assessment scenarios.

[0016] S6. Using a front-end charting engine combined with back-end graphics, generate a visual report explaining reading results, including transforming potential developmental ranges into visual charts (such as radar charts and step charts) showing the current level and improvement path. Specific operation: After each student's assessment, the system automatically generates a report, using a radar chart to show the difference between "independent reading ability" and "ability after prompting," and using a step chart to indicate "skills that can be improved next" (e.g., "Currently can understand single texts, can understand cross-text with prompting, next step is to train independent cross-text integration"). Parents and teachers can directly understand "what the student can do and to what extent they can improve."

[0017] Preferably, S1, the termination judgment based on the capability preset conditions of the IRT adaptive algorithm includes: pre-completing parameter calibration through the IRT model to determine the required parameters for each question; Initial ability estimation: At the start of the assessment, an initial ability estimate is set based on prior information; if there is prior information about the test takers, the initial ability value or ability distribution is set based on the prior information; if there is no prior information, the initial ability value is set to the group average level (e.g., θ=0) to provide an initial basis for subsequent topic selection. Dynamic question selection: The project selection algorithm is invoked, and the next question that best meets the selection criteria is selected from the project library and pushed to the test taker based on the current ability estimate; real-time dynamic evaluation (process evaluation) and summative evaluation are combined; Ability parameter update: After the test taker completes the test, the system collects the test results and updates the original ability value or ability distribution through the ability estimation model to obtain a new ability estimate that matches the current test performance; Termination Condition Determination: After each ability value update, a determination is made based on the preset conditions in the termination rules module. If the termination conditions are met, the assessment ends and the final ability evaluation result is output; if not, the process returns to the dynamic question selection stage and repeats the above process until the termination conditions are met. The preset conditions aim to balance assessment accuracy and assessment efficiency.

[0018] Preferably, in step S2, the log or backup records the test subject's reading path, information marking behavior, and draft reasoning in real time, combined with the final score. This includes: combining the reading path with the final score through attribution analysis, where the reading path indicates the reasons for achieving the score or existing problems, and the final score displays the reading ability level. The reading path includes a combination of tasks such as collecting the test subject's input via touch / electromagnetic pen coordinates, constructing a text block mapping map, and analyzing the temporal sequence of reading behavior. The system records the test subject's reading path, information marking behavior, and draft reasoning process in real time, achieving digital tracking and standardized recording of the reading process.

[0019] Specific steps: Verify the authenticity of the results: By comparing the "complexity of the draft reasoning" with the "correctness of the final answer", determine whether the final score is based on genuine understanding or guesswork.

[0020] Diagnostic Capability Deficiencies: During the reading process, the system records the test-taker's reading path (e.g., whether they repeatedly reread a certain section of text), information marking behaviors (marking specific sentences, such as underlining key sentences, making annotations, comments, likes, and replies), and draft reasoning processes (e.g., writing down the logical chain "because A, therefore B" online), and correlates these with the questions where points were ultimately lost. The system pinpoints whether the loss was due to an information retrieval error or an interruption in logical reasoning. It generates a capability report and strategy report, clarifying whether "cannot do" is due to "not understanding the text" or "not finding the right method." Information marking behaviors include multimodal interactive data collection, structured content association, and sequential reading behavior serialization, integrating touch / electromagnetic pen input, text semantic localization, and handwriting trajectory recognition modules.

[0021] Generate capability and strategy reports: The capability report displays a radar chart (results), while the strategy report provides recommendations based on path analysis.

[0022] The preferred approach uses a pre-defined scoring framework (front-end interaction + data modeling + back-end storage + real-time / non-real-time interaction), while simultaneously defining the logical connections between various reading functions and the context within the business scenario. Through front-end interaction via a multimodal data acquisition module, it enables touchscreen trajectory recording, online annotation, comment replies, "likes," and voice ASR (Automatic Speech Recognition). The overall architecture is as follows: Frontend / Tools: Native JS / TS (touchscreen trajectories), Canvas / SVG (annotation rendering), Vue / React (interactive components), WebSocket (real-time annotation / likes); Data storage / tools: MySQL (structured data: users, comments, likes), MongoDB (unstructured data: tracks, annotation coordinates), Redis (like count cache); Backend / Tools: Node.js / Java / Python (API development), JWT (user authentication), RESTful API (basic interaction), WebSocket service (real-time push); Auxiliary tools: Anti-shake / throttling (performance optimization for trajectory acquisition), CRDT algorithm (resolving conflicts in real-time annotation by multiple users), CDN (static resources).

[0023] Preferably, the S3 business unlocking logic engine correlates the NLP model (Natural Language Processing), voice sentiment analysis results (such as confusion and joy), and the stylistic characteristics of the current reading text and the assessment question type in real time through business processes. This is used to determine the timing of reading intervention or to revise the ability assessment model. The business process correlation logic refers to transforming unstructured text / voice sentiment interaction data into structured assessment indicators through computational layer processing, and then embedding them into the entire reading ability assessment process, driving a closed-loop operation of data collection → analysis and judgment → feedback intervention → iterative optimization.

[0024] The original process was: user comments / replies (text / voice) → data storage → display / interaction (likes / replies); after integrating NLP models with voice sentiment analysis, the voice sentiment analysis process became: speech-to-text → text preprocessing → sentiment / dimensional scoring → score storage → score application. Whether it's multi-dimensional text analysis or NLP model-based voice sentiment analysis, it's crucial to determine essential parameters (to ensure analysis feasibility), optional evaluation parameters (to improve accuracy), and preconditions (to ensure data quality). General voice sentiment analysis only outputs "positive / negative / neutral," while business scenarios require detailed dimensional scoring (quantification + labeling) to ensure feasibility.

[0025] Prerequisites for speech sentiment analysis include: acquiring user comments / replies as natural language input for ASR speech-to-text recognition; preprocessing the recognized text data; performing sentiment analysis and multi-dimensional text analysis using a multi-task NLP model; storing and applying scores; outputting the NLP model analysis and sentiment analysis structure; and outputting the sentiment analysis results. Multi-dimensional AI model analysis includes basic information extraction, complex reasoning, and digital text interpretation. Taking a whole-book reading assessment method as an example, the specific implementation is as follows: Step 1: Set reading levels according to reading ability classification and determine the basic question type model; Basic Question Type Model Library: Contains the definitions of all question types, and each question type comes with the three parameters of IRT (Item Response Theory) (discrimination a, difficulty b, guessing coefficient c).

[0026] Question type classification tags: such as "text decoding", "information retrieval", "induction and summarization" and other 6 types of ability tags.

[0027] Standardized question templates: Question structure (e.g., multiple choice answer format, scoring rules).

[0028] By selecting question types from the basic question type model library that match the proportion of age groups and determining the initial difficulty range based on IRT parameters, a personalized monitoring question type combination is formed.

[0029] Step 2: Obtain user identity information, allocate question type proportions, and form personalized monitoring question type combinations; Input: User identity information (age / educational stage); Output: Basic question type model library.

[0030] Output parameters: Personalized monitoring question type combination: A configuration table containing the weight of each question type (e.g., lower elementary grades: pronunciation and character shape 80% + simple search 20%).

[0031] Quantitative selection rules: The number of questions drawn from the question bank is allocated (e.g., a total of 20 questions, a certain number of questions are drawn from each question type).

[0032] Initial difficulty range: The difficulty range is preset according to the age group.

[0033] When incorporating stylistic features, determine the weight of the current user's question type combination and prioritize integrating stylistic features into question types with higher weights (for example, if the summarization question type accounts for a high proportion, prioritize embedding stylistic features in this type of question).

[0034] Step 3: Based on the text type and stylistic features of the entire book, these are interspersed throughout the question type matching process; Input: The text type of the entire book (e.g., "Journey to the West" is a novel, containing characteristics such as mythology and chapters), and the combination of question types output by S2.

[0035] Output parameters: Genre-based question mapping rules: A correlation table indicates the question type that matches each genre characteristic (e.g., "character analysis" in a novel corresponds to the "evaluation and appreciation" question type; "plot summary" corresponds to the "induction and generalization" question type).

[0036] Text content semantic tags: NLP analyzes key information in the text (e.g., if a chapter contains "environmental description", it will be automatically tagged with "environmental description" for easy reference when creating questions later).

[0037] Based on the question type mapping rules, the characteristics of the text are embedded into the corresponding question type template to form assessment questions that are closely integrated with the text content (for example, after the first chapter of "Journey to the West", a "summarization" question is generated based on the "environment description" tag).

[0038] Step 4: Match standardized assessment question types with age-appropriate monitoring question types to combine genre characteristics with assessment. Input: Question type model library, question type combination weights, text type mapping rules, and the text content currently being read.

[0039] Output parameters: Final assessment question set: Questions generated for the current reading chapter (or the entire book) (including question stem, options, answers, IRT parameters, etc.).

[0040] Formative assessment data recording template: specifies the reading behavior data to be collected (such as touch coordinates, answer time, annotation content, etc.).

[0041] During the user's quiz process, the answer to each question (correct / incorrect, time taken, reading path data) is fed back to the IRT adaptive algorithm in real time to update the user's ability estimate and dynamically select the next question from the question bank accordingly. At the same time, all behavioral data enters the formative assessment system to generate periodic ability reports.

[0042] A quality threshold parameter is used as the criterion for process jumps / interruptions, ultimately forming a structured sentiment analysis result. The speech recognition system (ASR), also known as automatic speech recognition, is a system that converts human speech into text or commands through techniques such as feature extraction, pattern matching, and model training.

[0043] Specific Operation: Standardized scoring dimensions are preset (such as "accuracy of information extraction," "logical coherence," and "number of critical viewpoints"). Based on student comments, replies, and likes, the AI ​​model quantifies and scores each dimension, employing Natural Language Processing (NLP), machine learning algorithms, deep learning, and multimodal analysis. The quantification process includes data collection, feature extraction, multi-dimensional analysis, rule-constrained quantification, and a human-machine calibration closed loop, significantly reducing subjective bias. Quantitative scoring is performed across multiple dimensions, including information extraction accuracy, logical coherence, and the number of critical viewpoints. A multi-task NLP model combined with rule-constrained quantification algorithms and a human-machine calibration closed loop is used to quantify and score each dimension. This ensures objectivity while adapting to the specific characteristics of educational texts, greatly reducing subjective bias. It adopts a three-layer architecture of model layer, computation layer, and adaptation layer, transforming natural language text into structured dimensional scores while linking them to relevant business context (student comments, question / annotation IDs, etc.). When scoring, the AI ​​calls the data preprocessing module, the model inference module (basic model, rules), the score calculation and fusion module (weighted calculation engine, normalization algorithm, confidence evaluator), and the context association and storage module (business ID mapper, JSON serialization tool, database interface) to execute the relevant quantitative scoring tasks.

[0044] The data preprocessing module includes: data access: receiving student answer text, audio files, and reading behavior trajectory data collected from the front end; cleaning and denoising: correcting typos and converting the text to colloquial language; denoising and removing silence from the audio; removing abnormal coordinates from the behavior data; feature extraction: converting the text into word vectors using NLP technology; extracting MFCC acoustic features using speech processing technology; calculating features such as reading trajectory length and pause time through behavior analysis; and standardization and encapsulation: normalizing all features and encapsulating them into standardized data batches according to the input format of a multi-task NLP model for subsequent multi-dimensional quantitative scoring by the AI ​​model. The specific operations are as follows: Input student text / speech-to-text (with business ID: user ID, reading text ID), call the data preprocessing module: data preprocessing (word segmentation, text cleaning, custom stop word list) → output: standardized text + vocabulary sequence; Model Inference Module: Model Inference (Basic Model, Rules) → Output: Raw scores for each dimension + Inference Log + Confidence for each dimension; Score Calculation and Fusion Module: Raw score calculation and fusion (weighted calculation engine, normalization algorithm, confidence evaluator) → Output: Weighted score for each dimension + overall total score + global confidence score; Context Association and Storage Module: Context Association and Storage (Business ID Mapper, JSON Serialization Tool, Database Interface) → Output: Structured stored data + business-usable results; Reading assessment platform display, generating a visual ability report explaining students' reading results, and triggering personalized intervention actions.

[0045] Preferably, S4, it establishes a "ready-to-use" reading task library covering different grade levels and reading goals, and sets up accompanying operation guidelines. Specifically, it categorizes reading assessment types and provides task templates (such as mind map frameworks and debate scripts). The system can fine-tune tasks based on the student's selected grade level (such as changing the text topic), reducing time and manpower costs. The reading assessment types include basic reading (such as summarizing paragraph main ideas), in-depth reading (such as comparing viewpoints in two texts), and applied reading (such as designing solutions based on texts).

[0046] Preferably, in S5, standardized intervention scripts are constructed, a standardized intervention script library corresponding to intervention prompts is set, and the hierarchical structure and timing of interventions are determined to address the low degree of standardization. Specific operations include: pre-setting common reading obstacles (such as "unfamiliar with keywords," "unable to understand cause-and-effect relationships," and "unable to understand long and complex sentences"), with each reading obstacle accompanied by 3-4 levels of prompts (e.g., for "keyword obstacles": Level 1 prompt "find context synonyms," Level 2 prompt "view text annotations," and Level 3 prompt "directly explain the meaning of the word"). The system strictly follows the script execution, avoiding subjective adjustments to the prompt intensity and ensuring consistent operation across different assessment scenarios.

[0047] The intervention prompts include preset reading checkpoints, with each reading checkpoint accompanied by multiple levels of prompts.

[0048] Preferably, S6, the front-end chart engine includes using visualization libraries (such as ECharts, D3.js, Highcharts, etc.) to dynamically render multi-dimensional reading ability data (such as information extraction, logical analysis, critical viewpoints, etc.) in the form of radar charts and ladder charts to achieve an intuitive display of the ability map; the back-end graphics, when it is necessary to generate static reports or PDFs, use a server-side drawing library to pre-render charts and then embed them into the ability report page.

[0049] Preferably, the visualization report that generates reading results includes: visual charts (such as radar charts and step charts) that transform potential development ranges into current level + improvement path, addressing the issue of "difficulty in interpreting results". Specifically, after each student's assessment, the system automatically generates a report using a radar chart to show the difference between "independent reading ability" and "ability after prompting", and using a step chart to indicate "skills that can be improved next" (e.g., "Currently can understand single texts, can understand cross-text with prompting, next step is to train independent cross-text integration"). Parents and teachers can directly understand "what the student can do and to what extent they can improve".

[0050] In this invention, real-time dynamic evaluation and summative evaluation are combined to automatically generate process evaluations. Each book read, each semester's reading, and each academic year's reading, with each stage summary, constitutes a summative evaluation. During the reading process, large-scale artificial intelligence models such as DeepSeek can be invoked to assist reading and evaluation, making the assessment more personalized. The assessment data is more objective and has higher reliability. The reading assessment method and system based on large models are more intelligent, avoiding the emotional factors introduced by human evaluation, resulting in more objective, fair, and reliable assessment results and data.

[0051] Example 2

[0052] Based on the foregoing embodiments, this invention also proposes an adaptive improvement device for reading ability assessment, such as... Figure 4 As shown, it includes: a preset module, used to make termination judgments based on preset conditions of the IRT adaptive algorithm; The reporting module is used to record the test taker's reading path, information marking behavior, and draft reasoning process in real time during the reading process, and combine them with the final score to generate ability reports and strategy reports, and determine the reading process analysis and evaluation results. The analysis module is used to determine the required parameters, optional assessment parameters, and unlocking logic based on reading assessment indicators; at the same time, it uses NLP models and voice emotion analysis to collect voice interaction data during the user's reading process. Supporting modules are used to build a reading task library covering different learning stages and reading goals, and to develop supporting operation guidelines; The building module is used to build a standardized intervention script library corresponding to setting intervention prompts and to determine the level and timing of reading interventions; The visualization module is used to generate visual reports that explain reading results by combining a front-end chart engine with back-end graphics.

[0053] Preferably, this IRT-based adaptive algorithm comprises four modules: (1) Project Library Management Module: The foundation for algorithm operation. The questions in this module need to be pre-calibrated using the IRT model to clarify the core parameters of each question. The core parameters include the difficulty parameter, which is the level of difficulty of the question; the discrimination parameter, which refers to the ability of the question to differentiate between test takers of different ability levels; and the guessing parameter, which is mostly used for multiple-choice questions and reflects the probability of the test taker answering the question correctly by pure guessing.

[0054] (2) Ability Estimation Model Module: This module infers the potential ability value of the test taker based on their answer results and is the core logical carrier for the algorithm to achieve adaptation. Commonly used estimation methods include maximum likelihood estimation, which calculates the ability value that maximizes the probability of the test taker's current answer result as the estimated value. The calculation logic is simple, but the error is large in extreme cases such as the test taker getting all the answers right or all the answers wrong; Bayesian estimation, which introduces the test taker's prior ability information and updates the posterior probability by combining the likelihood function of the current answer to obtain the ability estimate value; it can effectively solve the estimation problem in extreme answer situations and make the results more stable.

[0055] (3) Project selection algorithm module: This module determines how to select the next optimal question based on the current ability estimate. The core principle is the principle of maximum information content. It is usually based on the Fisher information function, selecting the question that provides the most ability discrimination information (ability estimate) at the current ability level, and pushing the optimal question from the standardized question type library, thereby minimizing measurement error.

[0056] (4) Termination Rules Module: This module is used to balance assessment accuracy and efficiency, and to prevent assessments from being conducted without limits. There are three types of termination conditions: measurement error threshold, which means that the assessment results are reliable enough and can be terminated when the standard error of the ability estimate is lower than the preset value; information threshold, which means that the assessment of the test taker's ability is sufficient when the accumulated assessment information reaches the preset standard; and the maximum number of questions, which sets the maximum number of questions in the assessment to prevent the assessment time from being too long due to the pursuit of ultimate accuracy, and to ensure the practicality of the assessment.

[0057] Example 3

[0058] like Figure 2 As shown, the adaptive improvement system architecture for reading ability assessment includes: the backend system and the software architecture. Backend system: An embedded system based on Android, supporting SIM card networking, multiple application functions (audiobook / bookstore), and complex UI components (tabs / search / pagination); it reuses Android hardware drivers (such as wireless network, SIM card) and application frameworks to reduce development costs, while also allowing for customized trimming (removing redundant components and reducing e-ink screen power consumption).

[0059] The software architecture adopts a modular, layered architecture, which includes the following modules: Hardware layer: E-ink screen panel, low-power processor, SIM card module, WiFi module, touch screen, reading light; Driver layer: E-ink screen refresh driver (supports partial / global refresh, reducing ghosting), SIM card / network driver, touch driver (interaction optimization adapted to the low refresh rate of the e-ink screen).

[0060] Middleware Layer: Rendering Engine: Integrates MuPDF / EPUB parsing library, responsible for parsing and typesetting book formats (PDF / EPUB, etc.), supporting text reflow and font size / line spacing adjustment; Lightweight UI Component Framework: May be based on Android native View customization, or use e-ink screen-specific GUI libraries such as LVGL, adapted to low refresh rates (avoiding complex animations and simplifying controls); Network / Cloud Synchronization Module: Supports WiFi / SIM card network connection, enabling book download and reading progress synchronization (connecting to the bookstore's backend system).

[0061] Plugin extension layer: Access third-party services (speech synthesis, AI intelligent interface) through modular plugins, such as audiobook function and intelligent assistant (deepseek large model) entry.

[0062] The application layer (user interaction) adopts a modular functional division, corresponding to the core entry points in the interface; including the bookshelf / bookstore module, reading settings module, parent function module, and personal center module; Bookshelf / Bookstore module: Manage local books and connect to the cloud library (including categories of must-read books and optional books); Reading settings module: font size / line spacing adjustment, reading light control (hardware + software collaboration); Parental control features: access control and content filtering (typical features of educational readers); The Personal Center module includes: User Information, Reading History, and Settings (corresponding to the "My" interface).

[0063] like Figure 3 The specific operating steps are shown below: 1. User registration and information binding: Register in the software system and bind user information, including region, school, name, grade, and mobile phone number; 2. Establish personal reading bookshelves, where students can select required reading books, optional reading books, and books for free reading; 3. Reading process and interaction: During the reading process, you can comment on the content, reply to other people's comments, like the content, annotate the text, consult the dictionary, translate, save new words and famous quotes, use the intelligent assistant Deepseek to look up information, and complete the assessment after reading the chapter.

[0064] 4. Data feedback and analysis: The system collects all the above data, such as reading time, reading speed, and assessment results, and provides real-time feedback through data visualization charts. It records the entire assessment process, assessment results, and analyzes reading behavior trends, combining process evaluation and summative evaluation in real time.

[0065] Although the present invention has been described in detail above with general descriptions, specific embodiments, and experiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. An adaptive improvement method for reading ability assessment, characterized in that, include: The IRT adaptive algorithm is used to determine termination based on preset conditions. Logs or backups record the test taker’s reading path, information marking behavior, and draft reasoning process in real time. Combined with the final score, an ability and strategy report is generated to determine the reading process analysis and evaluation results. Based on reading assessment metrics, the required parameters, optional parameters, and unlocking logic are determined; at the same time, NLP models and voice emotion analysis are used to collect voice interaction data during the user's reading process. Establish a reading task library covering different educational stages and reading goals, and develop corresponding operation guidelines; Build a standardized intervention script library corresponding to setting intervention prompts, and determine the level and timing of reading interventions; By combining a front-end chart engine with back-end graphics, a visual report explaining the reading results is generated.

2. The adaptive improvement method for reading ability assessment according to claim 1, characterized in that, The IRT-based adaptive algorithm includes: pre-calibrating parameters using the IRT model to determine the core parameters of each question; Initial ability estimation: At the start of the reading assessment, initial ability values ​​or ability distributions are set based on the test takers' prior information; if there is no prior information, the initial ability values ​​are set to the group average level, providing an initial basis for reading topic selection; Dynamic question selection: The project question selection algorithm is invoked, and combined with the current reading ability estimate, questions that meet the selection criteria are selected from the project library and pushed to the test taker; Ability parameter update: After the test taker completes the test, the test results are collected, and the original ability value or ability distribution is updated through the ability estimation model to obtain a new ability estimate that matches the current test performance; Termination condition determination: After each capability value update, a determination is made based on the preset conditions of the termination rules module.

3. The adaptive improvement method for reading ability assessment according to claim 1, characterized in that, The logs or backups record the test taker's reading path, information marking behavior, and draft reasoning in real time, combined with the final score, including: combining the reading path with the final score through attribution analysis.

4. The adaptive improvement method for reading ability assessment according to claim 1, characterized in that, The process of collecting voice interaction data during user reading by enabling NLP models and voice sentiment analysis includes: acquiring natural language input for ASR speech-to-text recognition, preprocessing the recognized text data, performing NLP model sentiment analysis / multi-dimensional text analysis, storing and applying scores, and outputting NLP model analysis and sentiment analysis structures.

5. The adaptive improvement method for reading ability assessment according to claim 1, characterized in that, The reading assessment indicators include the user's learning stage and reading goals.

6. The adaptive improvement method for reading ability assessment according to claim 1, characterized in that, The reading task library includes: categorized by reading assessment type, providing reading assessment task templates, and adjusting reading assessment tasks according to the grade level selected by the student.

7. The adaptive improvement method for reading ability assessment according to claim 1, characterized in that, The intervention prompts include preset reading checkpoints, with each reading checkpoint accompanied by multiple levels of prompts.

8. The adaptive improvement method for reading ability assessment according to claim 1, characterized in that, The front-end chart engine includes a visualization library that dynamically renders multi-dimensional reading ability data in the form of radar charts and ladder charts to display the ability map.

9. The adaptive improvement method for reading ability assessment according to claim 1, characterized in that, The ability to generate visualization reports that explain reading results includes: transforming potential development ranges into visual charts that show the current level and the path to improvement.

10. An adaptive improvement device for reading ability assessment, characterized in that, Includes a preset module for determining termination based on preset conditions of the IRT adaptive algorithm; The reporting module is used to log or back up the test taker's reading path, information marking behavior, and draft reasoning process in real time. Combined with the final score, it generates ability reports and strategy reports to determine the reading process analysis and evaluation results. The analysis module is used to determine the required parameters, optional assessment parameters, and unlocking logic based on reading assessment indicators; at the same time, it uses NLP models and voice emotion analysis to collect voice interaction data during the user's reading process. Supporting modules are used to build a reading task library covering different learning stages and reading goals, and to develop supporting operation guidelines; The building module is used to build a standardized intervention script library corresponding to setting intervention prompts and to determine the level and timing of reading interventions; The visualization module is used to generate visual reports that explain reading results by combining a front-end chart engine with back-end graphics.

Images