Reading behavior analysis system based on artificial intelligence

Abstract

The invention relates to the technical field of reading behavior recognition, in particular to an artificial intelligence-based reading behavior analysis system, which comprises a behavior acquisition module, a skip reading recognition module, a path construction module, a label generation module and a cognitive modeling module. According to the invention, through high-frequency collection of rolling and clicking events, the integrity of a reading operation chain is improved, and the analysis granularity of student reading behavior data is enhanced; by utilizing a chapter structure threshold value and non-rolling area analysis, the jump reading recognition dimension is expanded, and the accuracy of behavior judgment is improved; the residence time change in the return path is collected, a comparison system with the first access is established, and the recognition capability of the deep reading behavior of the student is enhanced; constructing a learning label system with distinct hierarchy according to an increasing trend of residence time; and fusing double indexes of behavior frequency and time fluctuation, mapping to a knowledge graph based on a clustering result, strengthening dynamic comparison and difference analysis of chapter reading behaviors, and realizing accurate positioning and analysis of student reading behaviors.

Description

Technical Field

[0001] This invention relates to the field of reading behavior recognition technology, and in particular to a reading behavior analysis system based on artificial intelligence. Background Technology

[0002] The field of reading behavior recognition technology encompasses multiple aspects, including the collection, modeling, and analysis of user behavior characteristics during the reading process. It primarily involves the digital identification and processing of readers' behaviors in receiving, understanding, and interacting with text content. The core content includes collecting reading data and combining it with statistical modeling, behavioral feature matching, and data analysis methods to identify and classify user reading behavior patterns. This aims to achieve structured perception, personalized modeling, and dynamic monitoring of reading behavior, providing a data foundation and analytical support for scenarios such as educational evaluation, intelligent recommendation, and personalized teaching. It is one of the key application directions in the integrated development of artificial intelligence and education.

[0003] Among them, the AI-based reading behavior analysis system refers to using artificial intelligence to extract and classify features from collected user reading behavior data to identify and analyze behavioral characteristics such as attention distribution, information processing strategies, and comprehension methods of primary and secondary school students during text learning. By accessing reading interaction record data (e-readers, etc.), including page browsing trajectories, reading time distribution, and interaction feedback click sequences, this data is converted into multi-dimensional vector sequences using feature encoding methods. Based on supervised learning model training, reading behavior label classification is performed to determine user reading patterns and attribute behavioral processes. This system is primarily applied to primary and secondary school classroom teaching and learning data analysis scenarios, representing an extension of AI technology in basic education.

[0004] The current process of identifying reading behavior among primary and secondary school students suffers from several structural shortcomings. The granularity of behavioral data is too coarse, often relying on single-page browsing time as the measurement basis, neglecting the precise relationship between the start and end points of operations. This results in the loss of the temporal dimension of reading activities, affecting the judgment of the complete reading process. Skipping identification lacks a structural reference mechanism; behavioral path judgment depends on the order of operation sequences, failing to consider the impact of textbook chapter arrangement on the meaning of behavior, easily overlooking behavioral anomalies during cross-section skipping. Deep processing identification relies on inferences based on single dwell time, ignoring the cognitive input signals implied in repeated visits, making it difficult to identify the trend of deepening understanding in multiple behaviors. Tag generation methods are mostly based on single-point behavioral features, failing to reveal the evolution and superposition of behavior over time, leading to a one-sided expression of learning characteristics. Behavioral clustering relies solely on single indicators such as time average and total amount, failing to capture the interaction between temporal fluctuations and operation frequency, limiting the effectiveness of behavioral classification under complex teaching content, making it difficult to effectively reveal individual and group differences, and weakening the accuracy and explanatory power of data-driven teaching feedback. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an artificial intelligence-based reading behavior analysis system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an artificial intelligence-based reading behavior analysis system includes:

[0007] The behavior collection module acquires scrolling event data of primary and secondary school students in e-readers, records the page index and operation time of all click operations, and generates a dataset of page dwell time.

[0008] The skip reading recognition module calculates the index difference between the scroll termination page index and the click trigger page index in the page dwell time dataset and determines whether the click position skips across segments in advance. It also detects whether there are areas without scroll records in the continuous skip path and obtains the skip reading behavior trajectory.

[0009] The path construction module monitors whether students scroll back from the current page to the knowledge point page in the textbook based on the skipping behavior trajectory, marks the deep reading behavior, and generates a backtracking reading path record;

[0010] The tag generation module analyzes whether the knowledge point page number is the overlapping position of the jump segment endpoint in the skipping behavior trajectory based on the back-reading path record, marks the corresponding anchor page number, back-reading number, dwell time increment and corresponding behavior paragraph, and generates a set of learning behavior feature tags.

[0011] Based on the learning behavior feature label set, the cognitive modeling module statistically analyzes the frequency distribution and standard deviation of students' behavior in the target chapter, performs cluster analysis, and generates a heat map of primary and secondary school students' reading behavior.

[0012] As a further embodiment of the present invention, the page dwell time dataset includes page scroll range, scroll duration, click position index, click operation time, and single-page dwell time; the skipping behavior trajectory includes cross-segment page number pairs, skipping time periods, interactive event density classification, and continuous non-scrolling area identifiers; the backtracking reading path record includes backtracking start and end page numbers, backtracking scroll duration, backtracking dwell time, and comparison results between the first and backtracking dwell times; the learning behavior feature label set includes anchor page numbers, backtracking times, dwell time increment, and corresponding behavior paragraphs; and the reading behavior heatmap includes chapter behavior frequency distribution, chapter dwell time fluctuation values, student behavior pattern categories, and knowledge point association mapping positions.

[0013] As a further aspect of the present invention, the behavior acquisition module includes:

[0014] The scrolling behavior acquisition submodule obtains the start page index, end page index, scrolling start time, and scrolling stop time for each scrolling event of the e-reader, determines the validity of the scrolling process, records and numbers them in chronological order, and generates scrolling interval sequence data.

[0015] The click behavior recording submodule monitors the trigger page index and operation time of all click operations on the page based on the corresponding start page index in the scroll interval sequence data, filters operation records whose operation time is between the scroll stop time and the next scroll start time, and obtains effective click frequency data.

[0016] The page dwell time calculation submodule calculates the difference between the scroll stop time and the next scroll start time in the scroll interval sequence data, combines it with the effective click frequency data to perform interval filtering and classification, combines the time difference of each segment with the operation frequency, analyzes the dwell performance of each page content in the e-reader, and generates a page dwell time dataset.

[0017] As a further aspect of the present invention, the skip-reading identification module includes:

[0018] The index difference calculation submodule obtains the scrolling termination page index and the click trigger page index from the page dwell time dataset, calculates the page number difference, and compares it with the chapter span threshold. If the difference is greater than or equal to the chapter span threshold and the click position is located in the previous chapter area, it is determined to be an early cross-segment skipping behavior, and a cross-segment skipping identification record is generated.

[0019] The jump path detection submodule retrieves all page numbers between the jump start point and the jump landing point based on the confirmed jump operation in the cross-segment jump recognition record, and judges the missing pages by combining the scroll record mark of the corresponding page in the page dwell time dataset, and obtains the missing page chain.

[0020] The behavior trajectory classification submodule performs event frequency analysis and dwell duration division on each segment of the page chain based on the dwell time and click frequency of each page segment in the missing page chain, calculates the skipping behavior intensity value, performs segmented interval clustering on all intensity values, and generates the skipping behavior trajectory.

[0021] As a further aspect of the present invention, the path construction module includes:

[0022] The knowledge point location comparison submodule obtains the target page number in the page jump chain segment recorded in the skipping behavior trajectory, and matches the target page of each jump segment with the page where the knowledge point is located based on the knowledge point location mark index in the textbook, and generates knowledge point jump comparison results.

[0023] The return path identification submodule obtains the scrolling time and dwell time of the corresponding starting page and target knowledge point page in the page dwell time dataset based on the return behavior path confirmed in the knowledge point jump comparison results, extracts the scroll stop time of the jump segment starting page and the scroll start time of the jump segment target page, calculates the time span of the page chain, and obtains the return path dwell parameter data.

[0024] The deep reading judgment submodule compares the dwell time of all paths in the return path dwell parameter data to determine whether the dwell time of the target page after jumping back is greater than the initial dwell time. It also performs a comprehensive scoring analysis by combining the frequency of interaction events and the average page dwell length of the corresponding page during the return process, and calculates the deep reading judgment value under the return path. If it is greater than the preset threshold, the corresponding path is marked as a deep reading chain, and a backtracking reading path record is established.

[0025] As a further aspect of the present invention, the label generation module includes:

[0026] Based on the backtracking reading path record, the knowledge point anchoring submodule extracts the page numbers of all knowledge points that are looped back, extracts the corresponding dwell time series, determines the candidate knowledge point anchors, and extracts three key parameters: the knowledge point identifier number, the number of loops, and the change in dwell time for the corresponding page, and establishes an anchor point identification parameter record.

[0027] The behavior path overlay submodule records the anchor point identification parameters, compares the end positions of all jump segments recorded in the jump reading behavior trajectory one by one, determines whether the knowledge point page number is also the end page in the jump reading trajectory, constructs the binding relationship between behavior segment number and anchor point page and stores the structure, and generates path overlay matching results.

[0028] The tag set generation submodule comprehensively encodes the page behavior features based on the path overlay matching results, associates each anchor page with the behavior paragraph number, constructs structural tag entries, and outputs them in sorted order according to the anchor number to obtain the learning behavior feature tag set.

[0029] As a further aspect of the present invention, the cognitive modeling module includes:

[0030] The partition behavior statistics submodule obtains the dwell time increment and backtracking number from the learning behavior feature tag set, classifies and statistically analyzes all dwell time data and backtracking number data according to chapter number, counts the number of valid behaviors recorded for each student in each chapter and adds them up to generate chapter behavior statistics results.

[0031] The chapter behavior clustering submodule clusters the behavior distribution of all students in the same chapter based on the fluctuation data of behavior frequency and dwell time of different students in each chapter recorded in the chapter behavior statistics results, records the classification results, and obtains chapter behavior pattern records.

[0032] The heatmap generation submodule maps chapter numbers to knowledge point numbers one-to-one based on the recorded chapter behavior patterns. Each node is assigned a color label based on the clustering results of student behavior, and the color depth is encoded according to the frequency of behavior to establish a heatmap of reading behavior for primary and secondary school students.

[0033] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0034] This invention improves the completeness of the reading operation chain and enhances the granularity of student reading behavior data analysis by frequently collecting scrolling and click events; expands the dimension of skipping recognition and improves the accuracy of behavior judgment by utilizing chapter structure thresholds and analysis of non-scrolling areas; collects changes in dwell time in the return path and establishes a comparison system with the first visit to enhance the ability to identify students' deep reading behavior; extracts the progressive trajectory of knowledge processing based on the increasing trend of dwell time and constructs a hierarchical learning tag system; integrates the dual indicators of behavior frequency and time fluctuation, maps the clustering results to the knowledge graph, generates a heat response structure, strengthens the dynamic comparison and difference analysis of chapter reading behavior, and achieves precise positioning and analysis of student reading behavior. Attached Figure Description

[0035] Figure 1 This is a system flowchart of the present invention;

[0036] Figure 2 This is a flowchart of the behavior acquisition module of the present invention;

[0037] Figure 3 This is a flowchart of the skip-read recognition module of the present invention;

[0038] Figure 4 This is a flowchart of the path construction module of the present invention;

[0039] Figure 5 This is a flowchart of the label generation module of the present invention;

[0040] Figure 6 This is a flowchart of the cognitive modeling module of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0042] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0043] Please see Figure 1 Artificial intelligence-based reading behavior analysis systems include:

[0044] The behavior collection module obtains the starting page index, ending page index, scroll start time, and scroll stop time of each scrolling event by primary and secondary school students in the e-reader, and synchronously records the triggering page index and operation time of all click operations to generate a page dwell time dataset.

[0045] The skip-reading recognition module calculates the index difference between the scroll termination page index and the click trigger page index in the page dwell time dataset and compares it with the chapter span threshold (the minimum page number difference for chapter switching determined according to the textbook layout) to determine whether the click position skips ahead of the segment, detects whether there are areas without scroll records in the continuous skip path, records the page chains with skip-reading characteristics using page number pairs and time periods, and classifies the dwell time and interaction event frequency corresponding to the page chain to obtain the skip-reading behavior trajectory;

[0046] The path construction module, based on the skipping behavior trajectory, compares the knowledge point location markers set in the textbook (knowledge point location index built based on the textbook's table of contents structure) to monitor whether students scroll back from the current page to the knowledge point page, and records the start and end page numbers, scrolling time, and dwell time in the return path. Paths with dwell time higher than the first dwell time are marked as deep reading behavior (a deep information processing behavior pattern that conforms to the definition of cognitive science), and a backtracking reading path record is generated.

[0047] The tag generation module analyzes whether the corresponding knowledge point page number is also the overlapping position of the jump segment endpoint in the skipping behavior trajectory based on the knowledge point page number that is repeatedly returned and the dwell time increases in the backtracking reading path record. It marks the corresponding anchor page number, backtracking number, dwell time increment and corresponding behavior segment to generate a set of learning behavior feature tags.

[0048] The cognitive modeling module, based on the dwell time increment (time difference analysis of multiple visits to the same content area) and backtracking number in the learning behavior feature label set, statistically analyzes the frequency distribution of student behaviors in the target chapters by chapter (distribution of the number of behaviors) and the standard deviation of dwell time (statistical indicator reflecting the volatility of reading time). It then performs cluster analysis on the behavior distribution structure of different students in the same chapter (using the K-means algorithm for behavior pattern classification) and maps it to a knowledge graph structure (a knowledge point association network built based on the curriculum standards) to generate a heat map of primary and secondary school students' reading behavior.

[0049] The page dwell time dataset includes page scroll range, scroll duration, click location index, click operation time, and single-page dwell time. Skipping behavior trajectory includes cross-segment page number pairs, skipping time periods, interaction event density classification, and continuous non-scrolling area identifiers. Backtracking reading path records include the start and end page numbers of the backtracking, backtracking scroll duration, backtracking dwell time, and the comparison results between the first and backtracking dwell times. The learning behavior feature label set includes anchor page numbers, backtracking times, dwell time increments, and corresponding behavior paragraphs. The reading behavior heatmap includes chapter behavior frequency distribution, chapter dwell time fluctuation values, student behavior pattern categories, and knowledge point association mapping locations.

[0050] Please see Figure 2 The behavior collection module includes:

[0051] The scrolling behavior acquisition submodule obtains the start page index, end page index, scrolling start time, and scrolling stop time for each scrolling event of the e-reader, determines the validity of the scrolling process, records and numbers them in chronological order, and generates scrolling interval sequence data.

[0052] The e-reader's embedded scrolling event monitoring component automatically activates when a student opens a textbook. This component collects events through the operating system's underlying touch event handling mechanism, monitoring the user's screen-swiping actions in real time. It records the starting and ending page numbers of each swipe as the starting and ending page indices, respectively. The start and end times of the swipe are recorded as the system's current timestamp in milliseconds. For example, if a fifth-grade student opens their e-textbook during morning reading and swipes from page 5 to page 9, with a starting page index of 5 and an ending page index of 9, the system records the start time as 1695001200000ms. The timeout was 1695001200320ms, and the scrolling duration was 320ms. The system's criteria for valid scrolling are a page span of at least one page and a scrolling time of at least 200ms. If both conditions are met, the scrolling record is considered valid. If the scrolling is less than one page or the time is below the threshold, the action is discarded. For example, if a junior high school student scrolls from page 12 to page 12 while browsing a history textbook, and the scrolling time is only 120ms, this data is filtered out. All valid records are stored in chronological order for subsequent behavior analysis and dwell time calculation. In one actual data collection, the following scrolling sequence was recorded from the reading behavior of eighth-grade students in their geography e-textbook:

[0053] Table 1 Scrolling Event Record Table

[0054] serial number Start Page Termination Page Startup time (ms) Stop time (ms) Scrolling duration (ms) 1 5 9 1695001200000 1695001200320 320 2 9 10 1695001200550 1695001200790 240 3 10 12 1695001201000 1695001201265 265

[0055] As shown in Table 1, the table clearly records the details of each student's swipe operation. All records that meet the conditions are retained to form a scroll interval sequence value, which is used by the subsequent click behavior determination and page dwell estimation modules to obtain the scroll interval sequence data.

[0056] The click behavior recording submodule monitors the trigger page index and operation time of all click operations on the page based on the corresponding starting page index in the scroll interval sequence data, filters operation records whose operation time is between the scroll stop time and the next scroll start time, and obtains effective click frequency data.

[0057] Based on the start and end page indices and scrolling time period information of each record in the scrolling interval sequence data, the system activates a click event monitoring mechanism to monitor and record the click operations performed by students on each page. The records include the page number that triggered the click and the time the click occurred. The system determines whether the click action occurs within the interval between two scrolling operations, specifically whether the click time is greater than the stop time of the previous scrolling record and less than the start time of the next scrolling record. For example, if the stop time of scrolling interval 1 is 1695001200320ms and the start time of interval 2 is 1695001200550ms, then if a student clicks the "Annotate" button on page 9 at 1695001200400ms, this action is considered a valid click; if the click time is less than the start time of the next scrolling record, the action is considered a valid click. A click occurring at 001200300ms or 1695001200600ms indicates that the page is scrolling or scrolling to the next interval, respectively, and is not counted in the valid click statistics for the current interval. In the above example, if three valid click events are collected in interval 1 (1695001200400ms, 1695001200425ms, and 1695001200500ms), and the clicked page is page 9, then the valid click frequency value for that page is 3. If no valid clicks occur in interval 2, then the click frequency for that page is recorded as 0. The system uses this method to judge and classify click events for each page, ensuring that the recorded click frequency can accurately reflect the interactive participation behavior of primary and secondary school students in each reading segment and obtain valid click frequency data.

[0058] The page dwell time calculation submodule calculates the difference between the scroll stop time and the next scroll start time in the scroll interval sequence data, combines it with the effective click frequency data to perform interval filtering and classification, combines the time difference of each segment with the operation frequency, analyzes the dwell performance of each page content in the e-reader, and generates a page dwell time dataset.

[0059] Based on the scrolling stop time and the start time of the next scrolling segment recorded in the scrolling interval sequence data, the system calculates the static period between adjacent scrolling segments as the basis for determining the page dwell time. Simultaneously, it uses the effective click frequency value as a parameter for determining behavioral activity. For example, if the stop time of scrolling interval 1 is 1695001200320ms and the start time of interval 2 is 1695001200550ms, then the page dwell time is 230ms, and the effective click frequency within this time period is 3. The system then evaluates the dwell state of the reading behavior on this page, setting the initial dwell time baseline value to 300ms. `s` is used to identify whether there is stable reading behavior. If the dwell time is insufficient but the click frequency is high, it can be judged as short-term but active reading behavior. If both are below the threshold, the page dwell record is ignored. For example, a junior high school student dwells on page 10 of a history e-textbook for 400ms without any click events. This page is also recorded as low-interaction static reading. Finally, the system generates page dwell record entries with page number, dwell time and click frequency as fields, forming a complete page dwell time dataset. This dataset is used by subsequent learning behavior analysis systems or platform teaching feedback mechanisms to generate page dwell time datasets.

[0060] Please see Figure 3 The skip reading recognition module includes:

[0061] The index difference calculation submodule obtains the scroll termination page index and the click trigger page index from the page dwell time dataset, calculates the page number difference, and compares it with the chapter span threshold. If the difference is greater than or equal to the chapter span threshold and the click position is located in the previous chapter area, it is determined to be an early cross-segment skipping behavior, and a cross-segment skipping recognition record is generated.

[0062] To obtain the scroll-ending page index and click-triggered page index from the page dwell time dataset, firstly, extract the scroll-ending page index corresponding to each page dwell record. This index represents the page number where the user pauses after completing the scrolling action. For example, if a sixth-grade student scrolls from page 5 to page 8 in an e-textbook and then clicks to trigger page 3, the click-triggered page index is 3. The system calculates the difference between the current click-triggered page index and the scroll-ending page index, i.e., 8 minus 3 equals 5. This is the page number difference for skipping between paragraphs. Next, extract the chapter span threshold from the textbook structure table. The chapter span threshold is defined as the minimum number of pages that need to be spanned between adjacent chapters or paragraphs. In junior high school geography textbooks, the average chapter span is 3 pages, so this value is set to 3. If the difference is greater than or equal to the chapter span threshold, and the click position is in the forward direction relative to the scroll end position (i.e., the clicked page number is less than the scroll end page number), the system determines that the current operation is a forward jump. If the clicked page number is 10 and the scroll end page is 8, then although the difference is 2, the direction is backward, so it is not counted as a jump. In one collection, if operations such as "scrolling from page 12 to page 15 and then clicking page 9" or "scrolling from page 15 to page 19 and then clicking page 14" occur consecutively, the system calculates the page number difference and determines its direction in turn, and compares it with the chapter threshold of 3 pages. If the page number difference is 5 and the direction is forward, it is a valid jump. Finally, all jump operations that meet the difference and direction judgment are constructed into a structural record to obtain the cross-segment jump recognition record.

[0063] The jump path detection submodule retrieves all page numbers between the jump start point and the jump landing point based on the confirmed jump operation in the cross-segment jump recognition record. It then combines the scroll record markers of the corresponding pages in the page dwell time dataset to make a judgment, marks pages with missing paths, and obtains the chain of pages with missing paths.

[0064] Based on the confirmed skipping behavior paths in the cross-segment skipping identification records, the system sequentially extracts all page numbers between the starting and ending pages of the skipping segment and compares each page with the page dwell time dataset to see if there are corresponding page scrolling records. If a page has no scrolling event records within the time period and is not a click-triggered page (i.e., there is no actual user operation trace), then the page is marked as a missing path page. For example, if a skipping behavior jumps from page 18 to page 10, with intermediate page numbers 17, 16, 15, 14, 13, 12, and 11, the system checks the page dwell data and finds that pages 13 and 15 have no scrolling records and do not appear in the click events. If the system determines that these two pages are missing pages, it then retrieves the corresponding timestamp data from the page dwell records and matches the scroll stop time of the starting page with the click time of the target page of the jump segment. For example, if the scroll stop time of the starting page 18 is 1695003120000ms and the click time of the jump segment page 10 is 1695003125600ms, then the path time interval of the entire jump segment behavior is 6000ms. Combining the starting and ending page numbers with this time interval, a page chain structure is formed, and the missing page sequence is marked. The system constructs this type of jump segment path into a chain structure for subsequent calling and classification, thus obtaining the missing page chain.

[0065] The behavior trajectory classification submodule performs event frequency analysis and dwell duration classification for each segment in the page chain based on the dwell time and click frequency of each page segment in the missing path page chain, using the following formula:

[0066]

[0067] Calculate the skipping behavior intensity value H of the i-th page chain segment. i Segmented interval clustering is performed on all intensity values ​​to generate skip reading behavior trajectories, where R i T represents the total dwell time of the link segment on page i. i F represents the standard dwell time baseline value for page links. i D represents the click frequency. ij This represents the dwell time on page j in the i-th chain. E represents the average dwell time of the i-th chain. ik E represents the frequency of interaction events in the k-th segment of the i-th page chain. ref This is a reference value for the frequency of interactive events, where n is the number of pages in the page chain and m is the number of event frequency collection items;

[0068] Based on the dwell time and click frequency information of each page segment in the missing page chain, the system sequentially calls the page number in each page chain and extracts the dwell time of the corresponding page from the page dwell time dataset. Taking the page chain numbered P7→P5→P3 as an example, its dwell time is D. 31=340ms, D 32 =400ms, D 33 =370ms, the page link number is 3, the system first calculates the average dwell time of the page link. The total dwell time of the page link is R3 = 340 + 400 + 370 = 1110 ms. The system sets the standard dwell time baseline value for the page link to T3 = 1000 ms. The page click frequency is F3 = 4 times, and the corresponding interaction event frequencies are E... 31 =3, E 32 =1, E 33 =0, the interaction event reference value is set to E ref =2 times, then calculate the squared term of the dwell time deviation, and get The sum is 1800. Substituting this into the formula:

[0069]

[0070] The calculation results show that the skipping behavior intensity value of page chain 3 is 30.524. This value will be used for behavior trajectory classification processing to form a skipping behavior trajectory.

[0071] Table 2 Classification Intervals of Skipping Behavior Intensity Values

[0072] Classification type Intensity range Behavioral Judgment Explanation Weak Behavior Chain <![CDATA[0≤H i <15]]> Short page dwell time, sparse clicks, weak user activity Medium Behavior Chain <![CDATA[15≤H i <30]]> Page dwell time is normal, and the interaction is moderate. Forced Behavior Chain <![CDATA[H i ≥30]]> Significant dwell time, frequent clicks, and strong behavior.

[0073] The skimming behavior intensity value is used to quantify the activity level and abnormal characteristics of students when skimming across sections in electronic textbooks. This value comprehensively reflects the dwell time of each page in the page chain and the time deviation of the overall chain segment, click frequency, and distribution of interactive events. By calculating dwell duration, operation density, and behavior stability, the skimming behavior intensity value can characterize the degree of concentration and amplitude of students' behavior on the skimming path. The higher the value, the more concentrated the behavior on that page chain, the more frequent the operation, and the clearer the skimming intention. It is often used to identify non-linear reading behaviors such as target-oriented rapid positioning, chapter skipping and backtracking, and key review. It is a key quantitative indicator for constructing student learning behavior maps and predicting behavioral motivations.

[0074] The formula's calculation logic aims to comprehensively measure the multi-dimensional performance of page chains in jump behavior, firstly through |R i -T i The difference between the total dwell time on the page chain and a preset baseline value is calculated to reflect the degree of deviation in the overall time spent by students on that chain. This value serves as the basic behavior intensity factor, which is then used to calculate the click frequency F. i Sum of squared deviations from page dwell time The summation and square root calculation are used to integrate interaction activity and page dwell time fluctuations. The square of click frequency emphasizes the strong impact of high-frequency clicks, while dwell time fluctuations reflect whether students show unusual attention on certain pages. This composite value is then added to the dwell time deviation term to obtain the numerator of the behavior intensity. Finally, the interaction event frequency term |E| is used. ik -E ref Using the absolute deviation of | as the denominator, a suppression factor for the normal distribution of behavior is constructed. That is, if the event distribution is not concentrated or deviates too much from the baseline interaction value, the intensity coefficient is suppressed. This structure realizes the dynamic integration of three factors: time deviation, interaction frequency and behavior consistency. A quantitative evaluation model for cross-segment skipping behavior is established using the structures of addition, square, summation, square root and ratio.

[0075] Please see Figure 4 The path construction module includes:

[0076] The knowledge point location comparison submodule obtains the target page number in the page jump chain segment recorded in the skipping behavior trajectory, and matches the target page of each jump segment with the page where the knowledge point is located based on the knowledge point location mark index in the textbook to generate knowledge point jump comparison results.

[0077] After obtaining the target page number of the page jump link segment in the skipping behavior trajectory, the system first extracts the page number of the final jump page in each trajectory and calls the textbook knowledge point positioning mark table. This mark table is generated based on the official table of contents of the textbook and consists of a two-dimensional array structure with the chapter page starting position and the teaching focus page index. In the structure, a unique page number positioning identifier is set for each page, such as "Second Unit Knowledge Guidance Page" and "Chapter Eight Extension Page". For example, in the "Elementary School Science" textbook, page 7 is set as the starting page of the second chapter knowledge point. The system matches the target page 7 of the jump segment with the knowledge point page index to determine whether the target page falls within the knowledge point. Within the scope of the knowledge point location page, if a match is found, the jump path is marked as a knowledge point pointing behavior path. If there is no match, the link is removed and not included in the path construction module. In practice, if there is a jump record of "jumping from page 21 to page 7", and page 7 is the teaching page of "light refraction experiment" marked in the table of contents, then the path is determined to be a return to knowledge point behavior. The system traverses and matches all target pages accordingly, and constructs a comparison structure array with the jump link record number, target page number and corresponding knowledge point identifier number of all matching paths, and generates knowledge point jump comparison results.

[0078] The return path identification submodule obtains the scrolling time and dwell time of the corresponding starting page and target knowledge point page in the page dwell time dataset based on the return behavior path confirmed in the knowledge point jump comparison results. It also extracts the scroll stop time of the jump segment starting page and the scroll start time of the jump segment target page, calculates the time span of the page chain, and obtains the return path dwell parameter data.

[0079] Based on the records identified as knowledge point return paths in the knowledge point jump comparison values, the system extracts the starting page number and target page number from the corresponding jump path, and retrieves the scrolling events for the corresponding two pages from the page dwell time dataset, extracting scroll stop time, scroll start time, and dwell time data. For example, in a certain jump action, scrolling from page 19 to page 7, the system extracts the stop time for page 19 as 1695011204000ms, the start time for page 7 as 1695011207400ms, and the dwell time as 560ms. Then, it retrieves the first time of page 7 from the history record. The dwell time at the time of appearance is recorded as the reading dwell time when the page is first viewed. For example, if the initial dwell time is 380ms, a dwell difference structure is constructed for the two time periods. At the same time, it is extracted whether there are short scrolling and click events in the intermediate pages of the jump path, and the scrolling duration, total number of clicks and path page statistics are recorded. If the path is a direct and uninterrupted chain, its scrolling time is equal to the time difference between the two ends, which is 3400ms. Finally, a structured record is established for the five items of the path number, start and end page numbers, scrolling time, current dwell time and initial dwell time to obtain the dwell parameter data of the return path.

[0080] The deep reading judgment submodule compares the dwell time of all paths in the return path dwell parameter data to determine whether the dwell time of the target page after the jump is greater than the initial dwell time. It then combines this with a comprehensive score analysis based on the frequency of interaction events and the average page dwell time during the return process, using the following formula:

[0081]

[0082] Calculate the depth reading judgment value S under the g-th return path. g If the value exceeds a preset threshold, the corresponding path is marked as a deep reading chain, and a backtracking reading path record is established. This refers to the dwell time for the target page in path g. C is the initial dwell time of this page. gh Let h be the dwell time on page h in path g. E represents the average dwell time across all pages in path g. gh Let N be the frequency of interactive events on page h in path g, and N be the total number of pages contained in path g.

[0083] Based on the dwell time comparison information of each path in the return path dwell parameter data, the system determines whether the dwell time of the target page is higher than the dwell time when it first appears. If it is higher, the difference is recorded and combined with the dwell time series of all pages in the path for standard deviation analysis. The system sets the pages in the path to be pages 7, 8, and 9, with corresponding dwell times of 440ms, 360ms, and 400ms, respectively. The total number of pages is 3, and the average dwell time is [missing value]. The system calculates the squared difference between each page and the average value, obtaining (440-400)² = 1600, (360-400)² = 1600, and (400-400)² = 0. The sum is 3200. Then, the interaction frequency per page is extracted as 3 times, 1 time, and 2 times, and substituted into the formula:

[0084]

[0085] The results show that the depth reading judgment value of this path is 207.37, and the system's preset threshold value is 300ms. This value is lower than the threshold, so the path does not meet the conditions for deep reading and is therefore not marked. If other path values ​​exceed 300, they are marked as deep reading paths and added to the structure table to form a backtracking reading path record.

[0086] The Deep Reading Judgment Value is a comprehensive indicator used to measure whether students exhibit deep information processing characteristics during their return journey. This value takes into account the difference between the time students spend on a second visit to a target page and their first visit, as well as the reading continuity and interactive activity of each page within the path segment. By integrating the increase in dwell time and the level of page behavior fluctuation, it reflects whether students invest more attention and cognitive resources when returning to a specific knowledge point page. The higher the value, the more concentrated, continuous, and proactive the information processing behavior may have been in the path. Therefore, this judgment value can serve as an important basis for identifying deep reading behavior chains.

[0087] The formula's operational logic aims to comprehensively assess whether students exhibit deep processing characteristics during the return path by synthesizing the difference in dwell time and the degree of fluctuation in page behavior within the path. Its first part... The first part represents the absolute difference between the return dwell time and the initial dwell time, reflecting whether the page generates a significant increase in dwell time upon second access. The larger the difference, the stronger the cognitive engagement of the student upon revisiting the page. The second part is the page behavior fluctuation item within the path. in Calculate the deviation of the dwell time on each page from the mean to measure the stability of students' reading on each page, combined with the squared term of interaction frequency. This can enhance the response intensity to behavioral activity. The square root operation is used to uniformly handle the contribution of duration deviation and click intensity to the value, while mitigating the amplification effect of extreme values. Finally, the average value is taken to measure the degree of behavioral fluctuation throughout the entire path segment. This average value is added to the dwell time difference to form the overall evaluation value of the path depth processing score.

[0088] Please see Figure 5 The tag generation module includes:

[0089] The knowledge point anchoring submodule extracts the page numbers of all knowledge points that are repeatedly returned based on the backtracking reading path record, extracts the corresponding dwell time series, determines the candidate knowledge point anchors, and extracts three key parameters: the knowledge point identifier number corresponding to the page, the number of returns, and the change in dwell time, and establishes an anchor point identification parameter record.

[0090] Based on the frequently occurring target page numbers in the backtracking reading path records, the system first filters out pages that have been returned more than once, extracts the dwell time record value for each of these pages, and constructs a dwell time growth sequence. For example, in the elementary school math textbook, page 12 was detected to have an initial dwell time of 240ms, a second of 370ms, and a third of 455ms. The system compares the current dwell time with the previous dwell time to confirm that the dwell time increases with the number of returns. Subsequently, the system records the number of returns and the dwell time sequence for each target page number, and looks up the corresponding knowledge point number for that page from the teaching structure mapping table. For example, page 12 is marked as K005 in the knowledge point directory, which is the "fraction addition concept page". Combining the jump path to which the page first appears in the backtracking behavior belongs, the system extracts the path segment number information and establishes a relationship between the anchor page number and the path segment, knowledge point number, and dwell time growth information. Finally, the system obtains the anchor point recognition parameter record.

[0091] The behavior path overlay submodule records anchor point recognition parameters, compares the end positions of all jump segments recorded in the jump reading behavior trajectory one by one, determines whether the knowledge point page number is also the end page in the jump reading trajectory, constructs the binding relationship between behavior segment number and anchor point page and stores the structure, and generates path overlay matching results.

[0092] Based on the anchor point identification parameter record of all knowledge point page numbers, and the corresponding knowledge point number pointed to by each page, the system sequentially compares the endpoint pages of all jump paths in the jump behavior trajectory to determine whether any knowledge point page appears simultaneously in the behavior trajectory at the endpoint of a jump path. For example, in a set of trajectories, page 9 is recorded as the endpoint of jump segment 3, and appears twice in the return path record with dwell times of 310ms and 460ms respectively. This page meets the path overlap condition, and the system records its path number as T3 as the endpoint of the jump segment, and extracts the behavior segment number of the jump segment as S4. The behavior segment number, page dwell time growth rate, and knowledge point number are bound together with the current page to establish the combination relationship between path number and behavior segment. At the same time, the system organizes the path, page number, and dwell time structure data of each path that meets the condition, and outputs its knowledge point identifier number, behavior segment number, and jump segment index for the page that meets the condition to obtain the path overlap matching result.

[0093] The tag set generation submodule comprehensively encodes the page behavior features based on the path overlay matching results, associates each anchor page with the behavior paragraph number, constructs structural tag entries, and outputs them in sorted order according to the anchor number to obtain the learning behavior feature tag set;

[0094] Based on the page number, number of returns, dwell time change sequence, and path overlap segment number in the path overlap matching results, the system constructs structured tag entries for each knowledge point page. First, each page is used as the anchor page to record the basic index, and its dwell sequence, such as 250ms, 430ms, and 540ms, as well as the corresponding behavior segment numbers for each stage, such as S1, S3, and S5, are extracted. Combined with the page's click frequency in each behavior segment, such as 2 times, 3 times, and 1 time respectively, the system jointly encodes the anchor page with its click features and dwell time sequence, and adds knowledge point identifiers such as K008 and K010. Finally, a tag content field "Page Number - Knowledge Point Number - Number of Returns - Dwell Time Growth Sequence - Behavior Segment Combination - Click Frequency Sequence" is generated for each knowledge point page. After structural processing, all tag entries form a classification tag set, resulting in a learning behavior feature tag set.

[0095] Please see Figure 6 The cognitive modeling module includes:

[0096] The partition behavior statistics submodule obtains the dwell time increment and backtracking number from the learning behavior feature tag set, categorizes and statistically analyzes all dwell time data and backtracking number data according to chapter number, counts the number of valid behaviors recorded for each student in each chapter and sums them up to generate chapter behavior statistics results.

[0097] The system retrieves the number of backtracking attempts and the increment of dwell time from the learning behavior feature tag set. It groups and organizes all data according to the textbook chapter number. First, it extracts the student ID, knowledge point ID, number of backtracking attempts, and dwell time for each round from the tag set. Then, it maps the data to the corresponding chapters according to the knowledge point-to-chapter mapping table. Next, it counts the number of times each student's behavior occurs in each chapter, using this as a frequency count. Simultaneously, it uses the corresponding dwell time for each attempt to construct a time series and calculates the standard deviation to represent the degree of fluctuation in reading time. For example, a student might have 6 backtracking attempts in Chapter 3. The dwell time is 310ms, 285ms, 325ms, 280ms, 300ms, and 315ms, with a mean of 302.5ms and a standard deviation of approximately 15.6ms. Another student exhibits four behaviors with dwell times of 270ms, 265ms, 260ms, and 255ms, resulting in a standard deviation of only 6.5ms. The system iterates through the behavioral parameters of each student within each chapter and calculates the two indicators mentioned above. After summarizing, a dataset is formed with the chapter number as the index and the fields being the behavior frequency value and the dwell time fluctuation value, thus obtaining the chapter behavior statistics results.

[0098] The chapter behavior clustering submodule clusters the behavior distribution of all students in the same chapter based on the fluctuation data of behavior frequency and dwell time of different students in each chapter recorded in the chapter behavior statistics results, records the classification results, and obtains chapter behavior pattern records.

[0099] Based on the behavioral frequency and time fluctuation parameters of all students in each chapter from the chapter behavior statistics results, the system constructs two-dimensional behavioral feature points for each student. The horizontal axis is set as behavioral frequency, and the vertical axis is the standard deviation of dwell time. The system inputs the two-dimensional point set of all students in the chapter into the distribution structure and uses spatial partitioning logic to classify them into three categories, i.e., setting the number of categories to 3, and initializing the center value of each category. The classification boundary is iteratively optimized to finally form three student group categories, corresponding to behavioral characteristics such as high frequency and low fluctuation, medium frequency and high fluctuation, and low frequency and medium fluctuation, respectively. The system binds the number of each student in the classification results to its category and calculates the central behavioral feature point of each category to form a set of category statistical vectors. For example, the average frequency of students in one category is 8.1 times and the standard deviation is 12.3ms, while the average frequency of students in another category is 5.0 times and the standard deviation is 21.4ms. The system forms a structured mapping record from all cluster outputs to generate chapter behavior pattern results.

[0100] The heatmap generation submodule is based on chapter behavior pattern records. It maps chapter numbers to knowledge point numbers one by one, assigns each node a color mark based on the clustering results of student behavior, and encodes the color depth according to the frequency of behavior to create a heatmap of reading behavior for primary and secondary school students.

[0101] Based on the clustering information between students and chapters in the chapter behavior pattern results, the classified student group data is mapped to the curriculum standard knowledge graph. The system first extracts the mapping relationship between knowledge point numbers and chapter numbers, converts the chapter numbers into node numbers in the graph, and assigns color labels to the nodes. The colors are distinguished according to the student behavior category to which they belong, and the color depth is gradient encoded by the number of students of the same type appearing at that node. For example, the first type of student cluster node is represented by red, and the more people in the cluster, the darker the color. The system performs the above labeling on each knowledge point node in the graph, while keeping the original association edge structure between nodes unchanged to ensure the integrity of the graph structure. Finally, all student behavior clustering results are combined with the visual labels of graph nodes to form a structural graph with chapter cognitive classification features, and outputs a heat map of primary and secondary school students' reading behavior.

[0102] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An artificial intelligence-based reading behavior analysis system, characterized in that, The system includes: The behavior collection module acquires scrolling event data of primary and secondary school students in e-readers, records the page index and operation time of all click operations, and generates a dataset of page dwell time. The skip reading recognition module calculates the index difference between the scroll termination page index and the click trigger page index in the page dwell time dataset and determines whether the click position skips across segments in advance. It also detects whether there are areas without scroll records in the continuous skip path and obtains the skip reading behavior trajectory. The path construction module monitors whether students scroll back from the current page to the knowledge point page in the textbook based on the skipping behavior trajectory, marks the deep reading behavior, and generates a backtracking reading path record; The tag generation module analyzes whether the knowledge point page number is the overlapping position of the jump segment endpoint in the skipping behavior trajectory based on the back-reading path record, marks the corresponding anchor page number, back-reading number, dwell time increment and corresponding behavior paragraph, and generates a set of learning behavior feature tags. Based on the learning behavior feature label set, the cognitive modeling module statistically analyzes the frequency distribution and standard deviation of students' behavior in the target chapter, performs cluster analysis, and generates a heat map of primary and secondary school students' reading behavior.

2. The artificial intelligence-based reading behavior analysis system according to claim 1, characterized in that, The page dwell time dataset includes page scroll range, scroll duration, click location index, click operation time, and single-page dwell time. The skipping behavior trajectory includes cross-segment page number pairs, skipping time periods, interaction event density classification, and continuous non-scrolling area identifiers. The backtracking reading path record includes the start and end page numbers of the backtracking, the scroll duration of the backtracking, the dwell time of the backtracking, and the comparison results between the first and backtracking dwell times. The learning behavior feature label set includes anchor page numbers, backtracking times, dwell time increments, and corresponding behavior paragraphs. The reading behavior heatmap includes chapter behavior frequency distribution, chapter dwell time fluctuation values, student behavior pattern categories, and knowledge point association mapping positions.

3. The artificial intelligence-based reading behavior analysis system according to claim 1, characterized in that, The behavior acquisition module includes: The scrolling behavior acquisition submodule obtains the start page index, end page index, scrolling start time, and scrolling stop time for each scrolling event of the e-reader, determines the validity of the scrolling process, records and numbers them in chronological order, and generates scrolling interval sequence data. The click behavior recording submodule monitors the trigger page index and operation time of all click operations on the page based on the corresponding start page index in the scroll interval sequence data, filters operation records whose operation time is between the scroll stop time and the next scroll start time, and obtains effective click frequency data. The page dwell time calculation submodule calculates the difference between the scroll stop time and the next scroll start time in the scroll interval sequence data, combines it with the effective click frequency data to perform interval filtering and classification, combines the time difference of each segment with the operation frequency, analyzes the dwell performance of each page content in the e-reader, and generates a page dwell time dataset.

4. The artificial intelligence-based reading behavior analysis system according to claim 1, characterized in that, The skip-reading recognition module includes: The index difference calculation submodule obtains the scrolling termination page index and the click trigger page index from the page dwell time dataset, calculates the page number difference, and compares it with the chapter span threshold. If the difference is greater than or equal to the chapter span threshold and the click position is located in the previous chapter area, it is determined to be an early cross-segment skipping behavior, and a cross-segment skipping identification record is generated. The jump path detection submodule retrieves all page numbers between the jump start point and the jump landing point based on the confirmed jump operation in the cross-segment jump recognition record, and judges the missing pages by combining the scroll record mark of the corresponding page in the page dwell time dataset, and obtains the missing page chain. The behavior trajectory classification submodule performs event frequency analysis and dwell duration division on each segment of the page chain based on the dwell time and click frequency of each page segment in the missing page chain, calculates the skipping behavior intensity value, performs segmented interval clustering on all intensity values, and generates the skipping behavior trajectory.

5. The artificial intelligence-based reading behavior analysis system according to claim 1, characterized in that, The path construction module includes: The knowledge point location comparison submodule obtains the target page number in the page jump chain segment recorded in the skipping behavior trajectory, and matches the target page of each jump segment with the page where the knowledge point is located based on the knowledge point location mark index in the textbook, and generates knowledge point jump comparison results. The return path identification submodule obtains the scrolling time and dwell time of the corresponding starting page and target knowledge point page in the page dwell time dataset based on the return behavior path confirmed in the knowledge point jump comparison results, extracts the scroll stop time of the jump segment starting page and the scroll start time of the jump segment target page, calculates the time span of the page chain, and obtains the return path dwell parameter data. The deep reading judgment submodule compares the dwell time of all paths in the return path dwell parameter data to determine whether the dwell time of the target page after jumping back is greater than the initial dwell time. It also performs a comprehensive scoring analysis based on the frequency of interaction events and the average page dwell length of the corresponding page during the return process, calculates the deep reading judgment value under the return path, and if it is greater than the preset threshold, the corresponding path is marked as a deep reading chain, and a backtracking reading path record is established.

6. The artificial intelligence-based reading behavior analysis system according to claim 1, characterized in that, The label generation module includes: Based on the backtracking reading path record, the knowledge point anchoring submodule extracts the page numbers of all knowledge points that are looped back, extracts the corresponding dwell time series, determines the candidate knowledge point anchors, and extracts three key parameters: the knowledge point identifier number, the number of loops, and the change in dwell time for the corresponding page, and establishes an anchor point identification parameter record. The behavior path overlay submodule records the anchor point identification parameters, compares the end positions of all jump segments recorded in the jump reading behavior trajectory one by one, determines whether the knowledge point page number is also the end page in the jump reading trajectory, constructs the binding relationship between behavior segment number and anchor point page and stores the structure, and generates path overlay matching results. The tag set generation submodule comprehensively encodes the page behavior features based on the path overlay matching results, associates each anchor page with the behavior paragraph number, constructs structural tag entries, and outputs them in sorted order according to the anchor number to obtain the learning behavior feature tag set.

7. The artificial intelligence-based reading behavior analysis system according to claim 1, characterized in that, The cognitive modeling module includes: The partition behavior statistics submodule obtains the dwell time increment and backtracking number from the learning behavior feature tag set, classifies and statistically analyzes all dwell time data and backtracking number data according to chapter number, counts the number of valid behaviors recorded for each student in each chapter and adds them up to generate chapter behavior statistics results. The chapter behavior clustering submodule clusters the behavior distribution of all students in the same chapter based on the fluctuation data of behavior frequency and dwell time of different students in each chapter recorded in the chapter behavior statistics results, records the classification results, and obtains chapter behavior pattern records. The heatmap generation submodule maps chapter numbers to knowledge point numbers one-to-one based on the recorded chapter behavior patterns. Each node is assigned a color label based on the clustering results of student behavior, and the color depth is encoded according to the frequency of behavior to establish a heatmap of reading behavior for primary and secondary school students.

Images