AI reading control method and system based on artificial intelligence

Abstract

The invention relates to the technical field of natural language processing, in particular to an AI reading management and control method and system based on artificial intelligence, and the method comprises the following steps: carrying out text word segmentation processing on an input original text, segmenting the text into independent sentences, recognizing basic word units in each sentence, analyzing semantic adjacency relationships and syntactic structure features among vocabularies, and carrying out word segmentation processing on the word units; screening and extracting potential phrases representing paragraph meanings, and establishing a candidate semantic unit set; according to the method, through word segmentation, syntactic structure recognition and semantic adjacency analysis of the original text, potential phrases capable of representing paragraph significance are extracted, the candidate semantic unit set is constructed, and modeling of the semantic structure in the text is achieved. And associating the set with the reading fixation duration and the playback action of the user sentence by sentence to obtain a reading behavior response of each semantic fragment, and executing semantic weighting and reading behavior cross analysis according to the reading behavior response. Through the linkage mode, the semantic focus actually focused by the user at present can be recognized.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and in particular to an AI reading control method and system based on artificial intelligence. Background Art

[0002] Natural language processing (NLP) is a key branch of artificial intelligence (AI), dedicated to enabling computers to understand, generate, and interact with human language. Its core goal is to enable machines to recognize, parse, generate, and respond to natural language, thereby enabling effective text-based or voice-based communication between humans and machines.

[0003] Natural language processing focuses on the structural processing and semantic modeling of text. While it can perform basic tasks such as part-of-speech tagging, dependency analysis, and context modeling, it lacks a behavioral perception mechanism when faced with reading scenarios in real applications. The processing flow only analyzes the static structure of the text, ignoring dynamic variables such as the user's rhythm control, attention shifts, and information reception intensity during the reading process. Without the intervention of user behavioral feedback, the selection of semantic focus cannot match the individual's understanding path, often leading to misjudgment of the text's key points. Therefore, improvement is needed. Summary of the Invention

[0004] The purpose of the present invention is to solve the shortcomings of the existing technology and propose an AI reading control method and system based on artificial intelligence.

[0005] To achieve the above objectives, the present invention adopts the following technical solution, an AI reading control method based on artificial intelligence, comprising the following steps: Perform word segmentation on the input raw text, segmenting the text into independent sentences and identifying the basic word units within each sentence. Analyze the semantic adjacency and syntactic structure features between words, screen and extract potential phrases representing the meaning of the paragraph, and establish a set of candidate semantic units. Presenting text content associated with the candidate semantic unit set, tracking user gaze duration and rewinding actions, obtaining user interaction behavior records, performing semantic weighting based on the user interaction behavior records and the candidate semantic unit set, verifying and determining the current highlighted semantic focus and corresponding reading behavior features, and generating verified semantic focus and behavior features; Calculate a reading fluency index based on the verified semantic focus and user reading behavior characteristics in the behavioral characteristics, and based on the reading speed change rate and the frequency of review. Based on the reading fluency index and with reference to the verified semantic focus in the behavioral characteristics, select an adjustment category of simplification or information supplementation, and formulate a personalized adjustment strategy plan. According to the adjustment category in the personalized adjustment strategy, word replacement is performed on the target text segment to generate the text content to be presented. Based on the text content to be presented, reading path planning is performed, the current focus and reading suggestions are highlighted, and an adaptive guided reading interface is constructed.

[0006] Preferably, the steps of obtaining the candidate semantic unit set are: The original text is segmented sentence by sentence according to punctuation. For each sentence, continuous word group segments are extracted in sequence. The part-of-speech tag, starting character position in the sentence, ending character position, and syntactic dependency distance with adjacent words of each segment are recorded to form a basic semantic segment attribute table. Calculating the intra-sentence aggregation value of each semantic segment according to the basic semantic segment attribute table; According to the intra-sentence aggregation value, the fragments with intra-sentence aggregation value higher than the median in each sentence are selected, and the fragments are sorted and merged according to the order of the fragments in the text and the start and end character positions. The fragment combinations in the scoring set in continuous or adjacent sentences are extracted across sentences to generate a set of candidate semantic units.

[0007] Preferably, the steps for obtaining the user interaction behavior record are: Presenting text content associated with the candidate semantic unit set, locating the corresponding paragraph of each semantic segment in the candidate semantic unit set in the original text, highlighting or marking the corresponding paragraph in the reading interface, and synchronously projecting the text content into the user visual interface area to generate user-visible candidate semantic unit-associated text content; Based on the text content associated with the candidate semantic units visible to the user, the user's facial gaze is monitored in real time, the duration of the user's gaze at the text position corresponding to each semantic segment is recorded, and whether the gaze moves back to the read area is detected again. Each pause and replay event is marked with a timestamp and paragraph number to generate raw observation data of user behavior; Based on the original observation data of user behavior, the gaze event sequence corresponding to each paragraph of text is analyzed, and the average gaze time, first look-back interval, number of looks-backs and total stay period of each semantic segment are extracted. A behavior comparison table is established in chronological order and paragraph number to generate a user interaction behavior record.

[0008] Preferably, the steps of obtaining the verified semantic focus and behavioral features are: Based on the user interaction behavior record, each gaze event in the user interaction behavior record is matched with the semantic segments in the candidate semantic unit set, and the total gaze duration, the actual start time of each replay, the duration of each replay, the total number of replays, and the total character length of the paragraph in which the segment is located are extracted for each semantic segment to generate a semantic segment gaze behavior feature set; Calculating the reading focus intensity of each semantic segment according to the semantic segment gaze behavior feature set; Based on the reading focus intensity, the reading focus intensity of all semantic segments in a set of continuous paragraphs consisting of three adjacent paragraphs of text are sorted, the 75th percentile of the reading focus intensity is set as the judgment threshold, and the semantic segments with reading focus intensity not lower than the judgment threshold are screened. The gaze duration, first review interval and number of reviews are extracted to generate verified semantic focus and behavioral characteristics.

[0009] Preferably, the steps for obtaining the reading fluency index are: Based on the verified semantic focus and user reading behavior features in the behavior features, the sentence number, actual reading time of each sentence, the triggering number and the first occurrence time of each sentence replay event in each semantic focus are extracted in sequence, and a sentence-level reading time series and a replay number series are constructed in the order of sentence numbers to generate a sentence-level reading behavior dataset within the semantic focus; Based on the sentence-level reading behavior dataset within the semantic focus, the reading time of each sentence, the corresponding number of replays, and the time difference between the previous and next sentences are calculated, and the pause intervals between adjacent sentences are recorded as auxiliary features to obtain a combined sequence of reading speed changes and replay frequency changes; Based on the combined sequence of reading speed changes and review frequency changes, a reading fluency index of semantic focus is calculated.

[0010] Preferably, the steps for obtaining the personalized adjustment strategy are: Based on the reading fluency index, extract the corresponding reading fluency index value from each semantic focus, number the semantic focuses according to the order of the original text, calculate the median and 75th percentile of the reading fluency index values of all semantic focuses in sequence, set the values below the median as "stable reading focus", between the median and the 75th percentile as "mild non-fluent focus", and above the 75th percentile as "significant non-fluent focus", classify and label each semantic focus, and generate a semantic focus reading fluency graded label set; Based on the semantic focus reading fluency classification label set, for each semantic focus labeled as "mild non-fluent focus" and "significant non-fluent focus", the total fixation duration, first re-look interval, and number of re-looks in the semantic focus are extracted. The average values of the three behavioral characteristics of all semantic focuses in the stable reading focus are compared to determine whether the semantic focus simultaneously meets the following conditions: the total fixation duration exceeds the mean by more than 10%, the first re-look interval is less than 5 seconds, and the number of re-looks is greater than 1.3 times the mean. If two or more conditions are met, the semantic focus is marked as a segment to be intervened, forming a set of target semantic focuses for intervention recommendations; Based on the set of semantic focuses targeted by the intervention recommendations, the sentences corresponding to the original text positions in each semantic focus are extracted. The number of clauses, the number of long compound word structures, the number of introduced definition terms, and the number of references to external information in the sentences are analyzed. If the number of clauses is greater than 2 or includes 3 or more long compound word structures, the semantic simplification category is selected. If the number of introduced definition terms is 0 and the number of references to external information is 1 or more, the supplementary information category is selected to generate a personalized adjustment strategy plan.

[0011] Preferably, the steps of obtaining the text content to be presented are: Based on the adjustment categories in the personalized adjustment strategy, the original text segments corresponding to the semantic focus marked as semantic simplification or supplementary information are retrieved, the syntactic structure of each original text segment is parsed, the clause nesting relationship, term position distribution and compound word boundary index are extracted, and a set of original text structure features of the semantic focus is generated; Based on the semantic focus original text structure feature set, for text segments marked as semantic simplification categories, structural splitting is performed to convert multiple clauses into parallel short sentences, and word meaning decomposition is performed on compound word structures. After extracting core word units, a simplified substitution vocabulary is called to replace the core words. For text segments marked as supplementary information categories, equivalent phrase descriptions are inserted after the term position to generate semantic focus processing results. Based on the semantic focus processing result, the modified sentence is re-joined to the original text position, all rewritten areas are subjected to boundary verification and semantic consistency review, the replacement segment position is marked, and the text content to be presented is generated.

[0012] Preferably, the steps for obtaining the adaptive guided reading interface are: Based on the text content to be presented, the starting position, ending position, and focus identification status of each semantic focus in the paragraph are extracted in sequence, the position coordinates of all semantic focuses in the complete text are mapped to the interface rendering framework, and semantic simplification or supplementary description tags are bound according to the semantic focus type. The reading priority order and highlighting status are configured for each segment, and a reading segment position index and a visual indication parameter set are generated; Based on the reading segment position index and visual indication parameter set, a reading guidance control sequence is constructed, segments are inserted into the reading navigation sequence according to the distribution order of focus in the text structure, and visual jump trigger points are set. Operation prompt information, reading suggestion labels and dynamic highlight borders are embedded in all focus segments in sequence to generate a reading path navigation execution task set; Based on the reading path navigation execution task set, the complete text structure and task trigger configuration are loaded, guided jump rendering is performed for each reading focus, the currently focused segment is highlighted with a border mark, and a reading suggestion window pops up on the right side of the segment. Other non-focus paragraphs are controlled for light and dark processing and interface masking effects, guiding users to focus on the defined reading path and generating an adaptive guided reading interface.

[0013] The present invention provides a reading control system, comprising: Semantic Extraction Module: This module performs word segmentation on the input raw text, splits the text into independent sentences, identifies the basic word units within each sentence, analyzes the semantic adjacency and syntactic structure features between words, screens and extracts potential phrases representing the meaning of the paragraph, and establishes a set of candidate semantic units. Behavior perception module: presents text content associated with the candidate semantic unit set, tracks user gaze duration and replay actions, obtains user interaction behavior records, performs semantic weighting based on the user interaction behavior records and the candidate semantic unit set, verifies and determines the current highlighted semantic focus and corresponding reading behavior characteristics, and generates verified semantic focus and behavior characteristics; Strategy formulation module: Calculates a reading fluency index based on the verified semantic focus and user reading behavior characteristics in the behavioral characteristics, the reading speed change rate, and the frequency of review; selects a simplification or information supplement adjustment category based on the reading fluency index and reference to the verified semantic focus in the behavioral characteristics, and formulates a personalized adjustment strategy plan; Content adaptation module: Based on the adjustment category in the personalized adjustment strategy, perform word replacement on the target text segment to generate the text content to be presented, plan the reading path based on the text content to be presented, highlight the current focus and reading suggestions, and build an adaptive guided reading interface.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are: The present invention extracts potential phrases that can represent the meaning of the paragraph through word segmentation, syntactic structure recognition and semantic adjacency analysis of the original text, and constructs a set of candidate semantic units, thereby modeling the internal semantic structure of the text. By associating this set with the user's reading gaze duration and review action sentence by sentence, the reading behavior response of each semantic segment can be obtained, and semantic weighting and reading behavior cross-analysis can be performed accordingly. Through this linkage method, the user's current real semantic focus can be identified, and its corresponding behavioral characteristics can be extracted synchronously. On this basis, the reading speed fluctuation and review frequency can be evaluated to form a dynamic reading fluency index. This index is used in conjunction with the identified focus segment to determine the position where the user may encounter difficulty in understanding during the reading process, and to perform targeted simplification processing or information supplementation. The target segment is then adjusted by word replacement, combined with reading suggestions and text highlighting strategies to achieve path-guided content reconstruction. It effectively improves the matching degree between content and readers, enhances the efficiency of conveying text information and the accuracy of understanding, and solves the problem that readers' reading load cannot be dynamically identified in traditional static content presentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the steps of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present 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 only used to explain the present invention and are not intended to limit the present invention.

[0017] See also Figure 1 The present invention provides a technical solution, an AI reading control method based on artificial intelligence, comprising the following steps: Perform word segmentation on the input raw text, segmenting the text into independent sentences and identifying the basic word units within each sentence. Analyze the semantic adjacency and syntactic structure features between words, screen and extract potential phrases representing the meaning of the paragraph, and establish a set of candidate semantic units. Present the text content associated with the candidate semantic unit set, track the user's gaze duration and rewinding actions, obtain user interaction behavior records, perform semantic weighting based on the user interaction behavior records and the candidate semantic unit set, verify and determine the current highlighted semantic focus and corresponding reading behavior characteristics, and generate verified semantic focus and behavior characteristics; Based on the verified semantic focus and user reading behavior characteristics in the behavioral features, the reading fluency index is calculated based on the reading speed change rate and the frequency of rereading. Based on the reading fluency index and the verified semantic focus in the behavioral features, the adjustment category of simplification or information supplementation is selected to formulate a personalized adjustment strategy plan; According to the adjustment categories in the personalized adjustment strategy plan, word replacement is performed on the target text segment to generate the text content to be presented. Based on the text content to be presented, reading path planning is carried out, the current focus and reading suggestions are highlighted, and an adaptive guided reading interface is constructed.

[0018] The steps to obtain the candidate semantic unit set are: The original text is segmented sentence by sentence according to punctuation. For each sentence, continuous word group segments are extracted in sequence. The part-of-speech tag, starting character position in the sentence, ending character position, and syntactic dependency distance with adjacent words of each segment are recorded to form a basic semantic segment attribute table. According to the basic semantic segment attribute table, the sentence aggregation value of each semantic segment is calculated using the following formula: ; in, For the The intra-sentence aggregation value of the semantic fragments, and are the ending and starting character positions of the segment respectively, represents the fragment span, The first The word frequency of each segment, For the The relative dependency distance between the segment and the current segment in the sentence, is the regulating factor, is the number of fragments that have dependencies with the current fragment; According to the intra-sentence aggregation value, the fragments with intra-sentence aggregation value higher than the median in each sentence are selected, and the fragments are sorted and merged according to the order of the fragments in the text and the start and end character positions. The fragment combinations in the scoring set in continuous or adjacent sentences are extracted across sentences to generate a set of candidate semantic units.

[0019] Specifically, the original text is first carefully preprocessed and accurately divided into a series of independent sentence units using standardized punctuation marks such as period ".", question mark "?", exclamation mark "!" and semicolon ";" as clear segmentation basis to ensure that the semantic integrity of each sentence is preserved. This process uses rule-based Chinese word segmentation and sentence segmentation tools, such as using the Jieba word segmentation library combined with regular expressions to identify sentence end marks, and complete the conversion of text to sentence lists. Then, for each independent sentence, the system starts the continuous word group fragment extraction process. This process uses a two-way maximum matching algorithm combined with a pre-loaded domain dictionary to identify potential word combinations, and iteratively extracts all continuous word sequences with a length between 2 and 5 words from the sentence as initial candidate fragments. Subsequently, for each extracted continuous word group fragment, each word in the fragment is assigned its corresponding part of speech. Tags, such as nouns (n), verbs (v), and adjectives (a), are serialized and attached to the segment attributes. At the same time, the exact starting and ending character positions of the segment in the original sentence are accurately recorded. The character positions are based on the character index under UTF-8 encoding. For example, in the sentence "Natural language processing is important", the starting position of the "natural language processing" segment is 0 and the ending position is 5 (excluding the end). Finally, the dependency parser is used to parse the syntactic dependencies between words within each segment and between the segment and other adjacent words, and calculate the syntactic dependency distance, which is defined as the number of nodes on the shortest path connecting two core words. If there is no direct dependency, it is quantified based on the path length under the common parent node. This information, including the segment itself, part-of-speech tag sequence, start and end positions, and dependency distance, is integrated and structured to form a basic semantic segment attribute table.

[0020] formula: The benefit of the formula is that it can achieve a quantitative evaluation of the aggregation value of the fragment sentence by comprehensively considering the length of the semantic fragment itself, the word frequency of other fragments that have syntactic dependencies with the fragment, and the relative dependency distance between these fragments. The fragment span factor It ensures that when other conditions are similar, segments of moderate length (not too long or too short) can obtain relatively high basic weights, avoiding the distortion of aggregate values caused by excessive non-core information in segments that are too long or insufficient information in segments that are too short. The introduction of , makes the fragments related to common important concepts gain higher aggregation potential, while the dependence distance factor It emphasizes that the segments that are closely related in syntactic structure contribute more to the aggregation degree of the current segment, and the adjustment factor The setting provides the model with the flexibility to adjust the influence of each factor according to different text characteristics or application requirements. Through the fusion calculation of such multi-dimensional information, the formula can more accurately identify the core semantic units within a sentence that have both a certain independent semantic expression ability and are closely related to other components of the sentence, thus providing a more reliable basis for subsequent screening of the candidate semantic unit set.

[0021] represents the position of the termination character of the th semantic segment. This parameter is determined according to the specific position of each semantic segment in the sentence during the text preprocessing stage after tokenizing and segment extraction of the original text. Its value is the index of the last character of the segment in the sentence plus one. For example, for the sentence "Machine intelligence is the trend of the future" and the semantic segment "Machine intelligence" within it, if the sentence starts from index 0 and the character "neng" is the 3rd character (index 3), then the value of is 4. For example, when processing the text "AI drives innovation and development", if the character index of the character "dong" in the segment "AI drives" in the sentence is 3, then its .

[0022] represents the starting character position of the th semantic segment. This parameter is also determined during the text preprocessing and segment extraction processes. It refers to the index value of the first character of the semantic segment in its所属 sentence. For example, in the sentence "Machine intelligence is the trend of the future", if the character "ji" of the semantic segment "Machine intelligence" is the 0th character (index 0) of the sentence, then the value of is 0. The way to obtain this parameter is to locate the identified segments in the tokenized sentence and record the offset of its first character relative to the beginning of the sentence. This is a direct text attribute extraction process. For example, when processing the text "AI drives innovation and development", if the character index of the character "A" in the segment "AI drives" in the sentence is 0, then its .

[0023] represents the word frequency of the th segment that has a dependency relationship with the current semantic segment . The acquisition of this parameter first depends on the result of dependency parsing to determine the set of other segments related to the segment , and then, for each of these related segments (i.e., the The total number of times a word appears in a predetermined background corpus is counted. This background corpus can be a large-scale general corpus (such as an Internet text summary) or a document collection in a specific field to ensure the representativeness of word frequency. Word frequency statistics are completed by building a vocabulary index and counting. For example, if the current segment is "deep learning" and one of its dependent segments is "neural network", the number of occurrences of "neural network" is retrieved in the background corpus containing 1 million words. If it appears 500 times, then .

[0024] Indicates the Dependent fragments and the current fragment Relative dependency distance in the sentence. This parameter is obtained through dependency syntactic analysis. Specifically, it refers to the distance between the current segment and the dependency syntactic tree. The core word and The number of edges contained in the shortest path between the core words of the dependency fragments. If the two fragments have a direct dependency relationship (father-child or brother relationship), the distance is closer. For example, in the sentence "Natural language processing technology is applied to machine translation", the fragments "Natural language processing technology" and "Machine translation" have an indirect dependency relationship and are connected by the common verb "apply to". If "Natural language processing technology" is the subject and "Machine translation" is the object, they are both directly connected to "apply to", then the distance between them can be defined as 2 (for example, the distance from subject to verb is 1, the distance from verb to object is 1, and the total path is 2). If one fragment directly modifies another fragment, the distance is 1. This distance is calculated by the dependency tree structure output by the syntactic analyzer.

[0025] The parameter is a hyperparameter used to balance the effect of segment span and the effect of dependent segment features. Its specific value is not obtained directly from the text, but is determined by experimental tuning on the validation set. The sentence aggregation value of the semantic fragment is calculated based on these values, and the subsequent semantic unit extraction task is performed using these aggregation values. The model with the best performance is selected by evaluating the accuracy, recall rate or F1 value of the final extraction result. For example, in a series of tests, if it is found that When it is set to 0.7, the semantic units extracted by the system are most consistent with the reference standard of manual annotation, and Determine it to be 0.7. For this example, set .

[0026] Indicates that the current fragment The number of segments with dependency relationships is directly obtained from the results of dependency syntactic analysis. After the sentence is parsed by dependency syntax, a dependency graph or tree is generated, which indicates the syntactic connections between the words or fragments in the sentence. The number of other identified semantic fragments directly or indirectly connected (with dependency arcs) to the word (or its core word) can be determined. For example, if the fragment "efficient algorithm" has a direct dependency relationship with the two fragments "data processing" and "speed optimization" in the syntax tree, then , this parameter reflects the fragment The breadth of connections in sentence structure.

[0027] Computational process: semantic fragments For example, set its text content to "core technology". Get the sample value of each parameter: From the text analysis, get: the starting character position of the segment , the end character position . Therefore, the fragment span characters. Through dependency syntactic analysis, it is determined that there are two fragments that have a dependency relationship with "core technology", so . The first dependent fragment ( ) is a “breakthrough”, and its word frequency in the background corpus is , the dependence distance from "core technology" The second dependent fragment ( ) is "independently developed", and its word frequency in the background corpus , the dependence distance from "core technology" . Based on experience tuning or preset, adjust the factor .

[0028] Calculation process: Calculate the denominator of the fragment span penalty term: ; Calculate the contribution of each dependency fragment in the sum: For the first dependency fragment ( ): ; For the second dependent fragment ( ): ; Compute the sum of the summations: ; Calculate the aggregate value within a sentence : ; The results show that the intra-sentence aggregation value of the semantic segment "core technology" in the sentence context of this example is 56.70. This value comprehensively reflects the moderate length of the segment itself and the strength of its association with two important and relatively close syntactically related dependent segments ("breakthrough" and "independent research and development"). The higher the intra-sentence aggregation value, the greater the possibility that the semantic segment is a core meaning-bearing unit in the sentence.

[0029] According to the intra-sentence aggregation values of each semantic fragment calculated in the previous step, first for each sentence in the text, collect all the calculated intra-sentence aggregation values in the sentence to form a numerical list, and then calculate the median of this list. The median is used as a dynamic threshold. For example, if a sentence contains fragments whose intra-sentence aggregation values are [10.2, 25.5, 18.0, 30.1, 22.3], and after sorting, they are [10.2, 18.0, 22.3, 25.5, 30.1], then the median is 22.3. All intra-sentence aggregation values are 22.3. The fragments with values strictly greater than this median (in this example, the fragments corresponding to 25.5 and 30.1) will be initially screened out, retaining their original text content, start and end character positions and other attributes. Next, these high-aggregation value fragments screened out from all sentences will be globally sorted strictly according to the order in which they appear in the original text (that is, first by sentence number, then by the starting character position of the fragment in the sentence) to ensure the orderliness of subsequent processing. After sorting, we will try to merge the high-scoring fragments screened out in the same sentence that are adjacent or partially overlapping and semantically coherent. For example, if "advanced manufacturing technology" and the following "core process" are both selected and the syntactic relationship allows, they can be considered to be merged into a longer segment "advanced manufacturing technology core process". However, the merging strategy should be cautious to avoid improper connection that may lead to a decrease in the quality of the semantic unit. Usually, only directly adjacent segments that together form a more complete semantic unit are merged. After completing the initial screening and sorting of intra-sentence segments, cross-sentence segment combination extraction is further performed. This process focuses on identifying those segment combinations that are distributed in consecutive or adjacent sentences but are closely related semantically and have high intra-sentence aggregation values. For example, if the high-scoring segment selected at the end of the first sentence is "artificial intelligence algorithm" and the high-scoring segment selected at the beginning of the second sentence is "model optimization", and the context indicates that the two are closely related, then these two segments are recorded as a segment combination. This combination is not a simple text splicing, but rather marks them as a semantic unit cluster that points to a common topic. Ultimately, all segments and segment combinations that have undergone the above screening, sorting, intra-sentence merging (if any), and cross-sentence combination constitute the candidate semantic unit set.

[0030] The steps for obtaining user interaction behavior records are as follows: Presenting the text content associated with the candidate semantic unit set, locating the corresponding paragraph of each semantic segment in the candidate semantic unit set in the original text, and highlighting or marking it in the reading interface, and synchronously projecting the text content into the user's visual interface area to generate user-visible candidate semantic unit-associated text content; Based on the text content associated with candidate semantic units visible to the user, the system monitors the user's facial gaze in real time, records the time the user's gaze stays at the text position corresponding to each semantic segment, and detects whether the gaze moves back to the read area. It adds a timestamp and paragraph number to each pause and review event to generate raw observation data of user behavior; Based on the original observation data of user behavior, the gaze event sequence corresponding to each paragraph of text is analyzed, and the average gaze time, first look back interval, number of looks back and total stay period of each semantic segment are extracted. A behavior comparison table is established in chronological order and paragraph number to generate user interaction behavior records.

[0031] Specifically, the system first extracts each independent semantic segment and its associated original text position information from the candidate semantic unit set obtained in the previous step. Specifically, for each semantic segment in the candidate semantic unit set, the starting character position and the ending character position recorded in the original text are used to reversely locate the specific paragraph where the segment is located in the complete original text. The definition of the paragraph is based on the standard text structure, such as two consecutive line breaks or specific indentation as the paragraph separation mark. Once the paragraph to which the semantic segment belongs is determined, the system immediately performs visual enhancement processing on these specific semantic segments on the user reading interface. This highlighting operation can be specifically configured to change the background color of the segment to Set it to a uniform light yellow color (for example, RGB values are 255, 255, 224), or adjust the font weight of the text in the segment to bold, while ensuring that the highlighted area corresponds exactly to the start and end boundaries of the segment and does not affect the normal display of the surrounding text. Then, the text content containing these highlighted semantic segments, together with their context paragraphs, are completely and in real time rendered and projected into the visual interface area of the computer screen or designated reading device that the user is currently looking at, ensuring that the user can clearly see the marked key content. This process involves the dynamic update of the graphical user interface, converting the text data stream into pixel information and displaying it to the user, and finally generating the user-visible candidate semantic unit-associated text content.

[0032] Based on the text content associated with the candidate semantic units visible to the user presented on the user interface, the system initiates high-frequency eye tracking, and continuously captures the two-dimensional coordinates of the user's gaze points on the screen through an external eye tracker or a built-in camera combined with a computer vision algorithm (for example, using the OpenCV library to detect facial feature points and adopting a gaze estimation model). The coordinate data stream is collected at a rate of 60 data points per second. Subsequently, the system converts the original screen gaze point coordinates into specific character or word positions on the displayed text content through a pre-calibrated mapping relationship. In order to accurately calculate the dwell time, the system adopts a dispersion threshold-based (I-DT) gaze point recognition algorithm to identify continuous, spatially (for example, the dispersion range is less than 1 degree of visual angle) and temporally (for example, the duration exceeds 150 milliseconds) sufficiently concentrated gaze points. The system identifies the original gaze point sequence as a valid fixation point or dwell event, and records the start and end timestamps of each effective dwell on the text area corresponding to a specific semantic segment, thereby calculating the duration of a single dwell. At the same time, the system marks the text area where a gaze event has occurred and the user's line of sight has moved forward beyond a predetermined range (for example, at least two lines of text or more than 10 words after the current fixation point) as a "read area", and monitors in real time whether there is a rapid jump back from the current gaze area to any "read area". Once such a jump occurs, it is identified as a look back action. For each identified dwell event and look back event, the system records in detail the timestamp of its occurrence and the paragraph number of the text content associated with the event in the original document. The paragraph number is marked incrementally from the beginning of the document in natural paragraph order to form the original observation data of user behavior.

[0033] Based on the original observation data of user behavior collected in the previous step, the system first performs a structured analysis on these data, and classifies the gaze points and review events with timestamps and paragraph numbers recorded in the original data stream according to the paragraph numbers to which they belong, forming a gaze event sequence arranged in chronological order within each paragraph of text. Then, for each semantic segment in the candidate semantic unit set, all gaze events whose spatial positions fall within the display area of the segment are screened out from the gaze event sequence of the paragraph in which it is located, and a series of reading behavior indicators are calculated based on these events: First, the average gaze time of the semantic segment is calculated, that is, the sum of the duration of all valid gaze events falling on the segment divided by the number of occurrences of these gaze events. For example, if a segment is gazed 3 times, with durations of 210 milliseconds, 180 milliseconds, and 240 milliseconds respectively, the average gaze time is (210+180+240) / 3=210 milliseconds. Secondly, the first review interval is extracted. It refers to the time difference between the user's first reading of the semantic segment (i.e., the time when the gaze point last left the segment) and the first time the user looked back at the segment from the subsequent text area (i.e., the starting time when the gaze point fell on the segment). If the segment is not looked back, this value can be recorded as a preset maximum value or marked as empty. Then, the total number of times the semantic segment was looked back is counted, that is, the number of all lookback events targeting the segment. Finally, the total dwell period of the user on the semantic segment is calculated, that is, the sum of the duration of all gaze events that fell on the segment (including the first reading and all lookbacks). After completing the calculation of these indicators, the system integrates each semantic segment and its corresponding average gaze time, first lookback interval time, number of looks back, total dwell period, together with its paragraph number and original order information within the paragraph, into a structured behavior comparison table. The table is organized according to the paragraph number and the order of appearance of the segment in the text to generate a record of user interaction behavior.

[0034] The steps for obtaining verified semantic focus and behavioral features are: Based on the user interaction behavior records, each gaze event in the user interaction behavior record is matched with the semantic segments in the candidate semantic unit set. The total gaze duration, the actual start time of each replay, the duration of each replay, the total number of replays, and the total character length of the paragraph in which the segment is located are extracted for each semantic segment to generate a semantic segment gaze behavior feature set; According to the semantic segment gaze behavior feature set, the reading focus intensity of each semantic segment is calculated using the following formula: ; in, For the The reading focus intensity of each semantic segment, is the total fixation duration of the segment, is the time interval between the first viewing and the first fixation of the segment, The total number of times the segment has been replayed. is the average number of times the user has reviewed the current paragraph set. The total length of the paragraph in which the fragment is located; Based on the reading focus intensity, the reading focus intensity of all semantic segments in a set of continuous paragraphs consisting of three adjacent paragraphs is sorted, and the 75th percentile of the reading focus intensity is set as the judgment threshold. The semantic segments with reading focus intensity not lower than the judgment threshold are screened, and the gaze duration, first replay interval and number of replays are extracted to generate verified semantic focus and behavioral characteristics.

[0035] Specifically, based on the user interaction behavior record and the candidate semantic unit set generated in the previous step, the system first traverses each semantic segment in the candidate semantic unit set, and uses the reading behavior data pre-stored for each semantic segment in the user interaction behavior record. For example, the user interaction behavior record has stored information such as the total stay period, the time of the first replay, the end time of the first gaze, and the list of replay events for each semantic segment. For each current semantic segment, the system directly extracts its total gaze time from the user interaction behavior record. The total gaze time is the cumulative gaze time of the segment during the entire reading process, including the gaze during the first reading and all subsequent replays. Then, for each event of replaying to the current semantic segment, the system extracts the actual start timestamp of the replay event from the user interaction behavior record. If a segment is replayed multiple times, a timestamp list will be extracted. At the same time, for each For a review event, the system further analyzes the total time the user's gaze remains in the current semantic segment during the review process. This is achieved by summarizing the duration of all continuous gaze points on the segment during the review. Then, the system extracts the previously counted total number of reviews of the current semantic segment. Finally, based on the paragraph number to which the current semantic segment belongs, the system calculates and extracts the total number of characters in the complete paragraph from the original text data. This character length does not include any formatting tags, only plain text characters are counted. The extracted information, including the total gaze duration of each semantic segment, a list containing the actual start timestamps of each review, a list containing the duration of each review on the segment, the total number of reviews, and the total character length of the paragraph to which the segment belongs, are organized together to form a record for each semantic segment. All these records are aggregated to generate a semantic segment gaze behavior feature set.

[0036] formula: The usefulness of the formula is that it constructs a comprehensive reading focus intensity index , which can quantify the user's attention to a specific semantic segment and the difficulty of understanding it. The formula combines the time dimension (total gaze duration) with the numerator. Interval from first review ratio) and frequency dimension (number of replays Average number of replays The ratio of the two squares (similar to the Euclidean distance) can amplify significant reading behavior characteristics, such as a shorter first re-reading interval ( Small) or higher relative viewing frequency ( The numerator value is significantly increased, indicating that the segment may be difficult or of high importance. The denominator is based on the total length of the paragraph in which the segment is located. , through the logarithmic form Normalization effectively adjusts the natural differences in reading time caused by different paragraph lengths, and avoids the fragments in long paragraphs from obtaining artificially high focus intensity simply because the paragraphs themselves have more content. This design makes It can more accurately identify those snippets that are particularly attractive to users or are read repeatedly, regardless of the context of their length.

[0037] For the The total fixation time of each semantic segment. This parameter indicates the time that the user's gaze stays on the first segment during reading. The cumulative time sum on each semantic segment is directly obtained from the semantic segment gaze behavior feature set generated in the previous step. The feature set has recorded this data for each semantic segment. For example, the original gaze data is collected by the eye tracking device. After processing, it is determined that the total duration of all gaze points of the user on the semantic segment "adaptive learning path" is 3500 milliseconds. .

[0038] For the The time interval between the first review and the first fixation of a semantic segment refers to the time interval between the first time the user completes the The length of time between the moment when a user looks at a segment (i.e., the gaze leaves the segment and moves to the subsequent content) and the moment when the gaze first moves back from the subsequent content and looks at the segment again, in milliseconds (ms). This data is also extracted from the semantic segment gaze behavior feature set, which calculates and stores this interval based on the user interaction behavior record. If a segment has not been reviewed after the first reading (i.e., ),but Theoretically, it is infinite. For the convenience of calculation, in this case, the ratio The value of 0 can be set by This can be achieved by setting it to a maximum value or directly treating the ratio as 0 in the calculation. For example, if a user watches the segment "Personalized Recommendation Algorithm" for the first time 5800 milliseconds after reading it, then .

[0039] For the The total number of times the user looks back at a semantic segment indicates that during the entire reading process, the user's gaze returns from other areas to the first segment. The total number of semantic segments, which is a non-negative integer, is directly obtained from the semantic segment gaze behavior feature set. This value reflects the frequency of the user's re-reading or reviewing of the information in the segment. For example, if the user moves his gaze back to the segment "Key Performance Indicator" three times to re-read it, then .

[0040] The average number of times a user looks back in the current paragraph set. Here, the "current paragraph set" specifically refers to the paragraph set containing The set of all semantic segments contained in a semantic segment and its adjacent paragraphs (usually a set of three consecutive text segments) is calculated as follows: First, obtain the total number of replays of all semantic segments in this three-segment set (i.e., the total number of times each segment is reviewed). value), then add up the total number of replays and divide it by the total number of semantic segments in the three-segment set, that is, ,in is the total number of segments in the three-segment set. To avoid the denominator being zero in subsequent calculations, if the calculated If the value is 0, adjust it to a small positive number, such as 0.1, when used in the formula. For example, in a set containing 3 paragraphs, there are 15 semantic segments, and the total number of their replays is 18. The average number of replays for this set is .

[0041] For the The total length of characters in the paragraph containing the semantic fragment is The total number of characters in the natural paragraph of the semantic segment (excluding spaces and special formatting characters, only text content characters) is obtained from the semantic segment gaze behavior feature set, which reflects the amount of text in the local context of the segment. For example, if the semantic segment "deep neural network" is located in a paragraph containing 450 characters, then .

[0042] Calculation process: For example, The reading focus intensity of the semantic segment "AI ethical dilemma" The following parameter values are obtained from the semantic segment gaze behavior feature set: total gaze duration milliseconds. The time interval between the first look back and the first fixation Milliseconds. Total number of replays The average number of times the clip is viewed within the current three text segments. times. Since this value is non-zero, it is used directly. The total character length of the paragraph in which this fragment is located characters.

[0043] Calculation process: Calculating ratios : ; Calculate ratios : ; Calculate the squares of the two ratios above: ; ; Calculate the square root of the sum of squares (numerator): ; Calculate the logarithm of the denominator (using avoid Problems caused by being too small): ; To calculate , you can use the base-changing formula ,For example . ; ; Calculating reading focus strength : ; The results show that the reading focus intensity of the semantic segment "AI Ethical Dilemma" is approximately 0.1992. This value is obtained after comprehensively considering the user's gaze duration, the immediacy and frequency of the review behavior, and adjusting it relative to the average reading behavior and text length of the paragraph. A higher reading focus intensity value (for example, significantly higher than the average value of other segments in the region) suggests that the segment may pose a reading difficulty for users, or it may be core content that users are particularly interested in.

[0044] Based on the reading focus strength of each semantic segment calculated previously, the system uses a sliding window method to analyze the text. Specifically, it takes three consecutive natural paragraphs as a processing unit or a "continuous paragraph set" and gradually moves this three-segment window in the entire document. For example, it first processes paragraphs 1, 2, and 3, then paragraphs 2, 3, and 4, and so on, until the end of the document. Within each current three-segment set, the system collects all the semantic segments contained in these paragraphs and extracts their respective reading focus strength values. These reading focus strength values are arranged in ascending order to determine the distribution of reading focus strength within the set. Subsequently, the system calculates the 75th percentile of this sorted strength value list and sets this percentile as the dynamic judgment threshold of the current three-segment set. The 75th percentile is calculated as follows: if there are N segments in the set, its index position is 0.75*N. If the result is a non-integer, the nearest neighbor ordinal interpolation is used. The exact value is determined by the value method or linear interpolation method. For example, there are 40 semantic segments in a three-segment set. After the reading focus intensity is sorted, the reading focus intensity value of the 30th segment (i.e., 0.75*40=30) is 0.28, and the 0.28 is set as the judgment threshold. Then, the system screens out all semantic segments in the current three-segment set whose reading focus intensity is greater than or equal to the 0.28 judgment threshold. These screened out segments are considered to be the parts with higher user attention in the current context. For each semantic segment with high focus intensity screened out in this way, the system extracts and associates the three specific reading behavior data, namely, the total gaze duration, the time interval between the first look back and the first gaze, and the total number of looks back, from the previously generated semantic segment gaze behavior feature set. Finally, all the semantic segments with high focus intensity screened out through this process and their associated three core behavior feature data are brought together to jointly generate the verified semantic focus and behavior features.

[0045] The steps to obtain the reading fluency index are: Based on the verified semantic focus and user reading behavior features, the sentence number, actual reading time of each sentence, triggering times and first occurrence time of each sentence replay event in each semantic focus are extracted in sequence. Sentence-level reading time series and replay count series are constructed in the order of sentence numbers to generate a sentence-level reading behavior dataset within the semantic focus. Based on a sentence-level reading behavior dataset within a semantic focus, we calculated the reading time of each sentence, the corresponding number of replays, and the time difference between the preceding and following sentences. We also recorded the pause intervals between adjacent sentences as auxiliary features, and obtained a combined sequence of changes in reading speed and replay frequency. Based on the combined sequence of reading speed changes and review frequency changes, the reading fluency index of semantic focus is calculated using the following formula: ; in, is the reading fluency index of the current semantic focus, For the The actual reading time of the sentence, For the The total number of times the sentence is reviewed, is the total number of sentences in the semantic focus, Indicates the change in the frequency of reviewing consecutive sentences.

[0046] Specifically, based on the verified semantic focus and behavioral features obtained in the previous step, especially the user reading behavior characteristics of each semantic focus contained therein (total fixation duration, first replay interval, total number of replays), the system further conducts a more detailed sentence-level reading behavior analysis for each verified semantic focus (i.e., a semantic segment with high attention and its context area). First, the system determines the text area that constitutes the current semantic focus and divides it into independent sentence units according to standard punctuation marks (such as period, question mark, exclamation mark), and assigns a consecutive number starting from 1 to each sentence in the semantic focus. Then, in order to obtain the actual reading time of each sentence, the system retrieves the original eye tracking data (derived from user interaction behavior records) with detailed timestamps and spatial coordinates, filters out all fixation points that fall within the text range of the current sentence, and accumulates the duration of these fixation points to obtain the actual reading time of the sentence ( ), at the same time, the system also records the timestamp when the user's gaze first enters the sentence ( ) and the timestamp before the last sentence is left and the next sentence is moved to, or a long jump occurs ( ), in order to calculate the pauses between sentences later. Then, for the current sentence, the system counts the total number of times all the replay events targeting the sentence are triggered ( ), and record the exact time when the sentence was first reviewed, including the sentence number, actual reading time , number of replays , first look back time, sentence first fixation start time , the time point of the last fixation of the sentence ) are organized in the order of sentence numbers, and a sentence-level reading time series and a review frequency series are constructed for each semantic focus, jointly generating a structured sentence-level reading behavior dataset within the semantic focus.

[0047] According to the sentence-level reading behavior dataset generated in the previous step, this dataset already contains the actual reading time of each sentence in each semantic focus ( ), the corresponding number of replays ( ), and the first fixation start time of the sentence ( ) and the last fixation end time point ( ), the system then computes a series of features used to characterize the dynamic changes in reading behavior. For each pair of consecutive sentences within the semantic focus (e.g., Sentence and sentences), the system calculates the difference between their actual reading times, i.e. ,This difference reflects the relative change in the user's speed when reading adjacent sentences. A positive value indicates that the time taken to read the next sentence increases (it may become slower or the sentence is more complex), and a negative value indicates that the time taken to read the next sentence decreases. In addition, the system also calculates the pause interval between adjacent sentences. The specific method is, Sentence and The pause between sentences is The first fixation start time of the sentence ( ) minus the The last fixation ending time of the sentence ( ) is obtained, that is , the pause interval is recorded as an important auxiliary feature, which reflects the cognitive processing load during sentence conversion. All these calculated reading time difference sequences, the number of replays (obtained directly from the input dataset) The sequence itself represents the sentence-by-sentence performance of the rereading frequency) and the pause interval sequence between adjacent sentences are combined to form a comprehensive combined sequence of reading speed changes and rereading frequency changes, providing a data basis for the subsequent calculation of reading fluency indicators.

[0048] formula: The benefit of the formula is that it provides a method to quantitatively evaluate the reading fluency of users when reading a specific semantic focus. This indicator takes into account the fluctuation of reading speed and the stability of rereading behavior. The term highlights the variation in reading time between consecutive sentences. A sharp change in time (whether it is a sudden increase or decrease) will cause this value to increase, suggesting that reading fluency may be disturbed. The change in the number of replays is taken into consideration. If the number of replays of consecutive sentences also changes significantly, it means that the user's reading strategy or text difficulty may be undergoing a transformation. This change explains the fluctuation of reading time to a certain extent. Therefore, the increase in the denominator will correspondingly reduce the contribution of this item to the overall degree of disfluency; on the contrary, if the number of replays is stable (the denominator is close to 1), the fluctuation of reading time will be more significantly amplified. The formula is weighted (squared term) and normalized (divided by the replay change plus one) by summing up this relative volatility of all adjacent sentence pairs within the semantic focus, and finally averaged to each sentence pair (divided by ), thus obtaining an index that can reflect the stability of the overall reading process, Larger values indicate less fluent reading and greater fluctuations.

[0049] For the The actual reading time of a sentence refers to the time when the user's sight is on the first sentence that constitutes the current semantic focus. The total time spent on a sentence, in milliseconds (ms). This data comes directly from the "semantic focus sentence-level reading behavior dataset" generated in the previous step. For example, when analyzing a semantic focus containing three sentences, if the user reads the first sentence ( ) took 4500 milliseconds, then .

[0050] For the The total number of times the user's sight jumps back from the subsequent text area or other areas to the first sentence when reading the current semantic focus. The number of times the sentence is reread is a non-negative integer, which is also directly obtained from the "semantic focus sentence-level reading behavior dataset". For example, if the first sentence of the semantic focus ( ) is viewed by the user twice, then .

[0051] is the number of consecutive sentence pairs used to calculate fluency in the current semantic focus. If a semantic focus contains sentences, then the number of consecutive sentence pairs is , so in the formula, , which represents the number of items in the summation operation and the divisor used in the final average. For example, if a semantic focus consists of 4 sentences, there are 3 pairs of consecutive sentences (sentence 1-sentence 2, sentence 2-sentence 3, sentence 3-sentence 4). .

[0052] Indicates a continuous sentence ( Sentence and It quantifies the change in the frequency of review behavior when users read adjacent sentences. This value is obtained by taking the Number of times the sentence is reviewed With the Number of times the sentence is reviewed The absolute value of the difference between the two is obtained. Both of these review times are obtained from the "Semantic Focus Sentence-Level Reading Behavior Dataset". For example, if Reread 2 times ( ), No. Reread 1 sentence ( ),but .

[0053] Calculation process: Consider a semantic focus, which contains 4 sentences ( ). Therefore, the number of sentence pairs used for calculation The reading time of each sentence was obtained from the "Semantic Focus Sentence-Level Reading Behavior Dataset" ( ) and number of replays ( ) is as follows: Sentence 1: ms, Second, sentence 2: ms, Next, sentence 3: ms, Second, sentence 4: ms, Second-rate; Calculate the value of each pair of consecutive sentences Item: For (Sentences 1 and 2): Item 1 ; for (Sentences 2 and 3): , item 2 ; for (Sentences 3 and 4): , item 3 ; Compute the sum of all terms: ; Calculating reading fluency index (in ): ; The results show that the current semantic focus reading fluency index It is approximately 694444.44. This value reflects the comprehensive fluctuation of the user's reading speed and review behavior when reading consecutive sentences within the semantic focus. The larger the value of , the more discontinuous the reading process is and the greater the volatility is, which means that the user encounters more obstacles or has a higher cognitive load when understanding the content. On the contrary, a smaller value of A value of 0 indicates that the reading process is relatively smooth and fluent.

[0054] The steps to obtain a personalized adjustment strategy are as follows: Based on the reading fluency index, the corresponding reading fluency index value is extracted from each semantic focus. After the semantic focuses are numbered according to the original text order, the median and 75th percentile of the reading fluency index values of all semantic focuses are calculated in sequence. The values below the median are defined as "stable reading focus", the values between the median and the 75th percentile are defined as "mild non-fluent focus", and the values above the 75th percentile are defined as "significant non-fluent focus". Each semantic focus is classified and labeled to generate a set of semantic focus reading fluency classification labels. Based on the semantic focus reading fluency classification label set, for each semantic focus labeled as "mild non-fluent focus" or "significant non-fluent focus", the total fixation duration, first re-look interval, and number of re-looks in the semantic focus were extracted. The average values of these three behavioral characteristics of all semantic focuses in the stable reading focus were compared to determine whether the semantic focus simultaneously met the following conditions: the total fixation duration exceeded the mean by more than 10%, the first re-look interval was less than 5 seconds, and the number of re-looks was greater than 1.3 times the mean. If two or more conditions were met, the semantic focus was marked as a segment to be intervened, forming a set of target semantic focuses for intervention recommendations. Based on the set of semantic focuses targeted by intervention recommendations, the sentences corresponding to the original text positions in each semantic focus were extracted. The number of clauses, the number of long compound word structures, the number of introduced definition terms, and the number of external information referenced in the sentences were analyzed. If the number of clauses was greater than two or included three or more long compound word structures, the semantic simplification category was selected. If the number of introduced definition terms was zero and the number of external information referenced was one or more, the supplementary information category was selected to generate a personalized adjustment strategy plan.

[0055] Specifically, based on the reading fluency index of each semantic focus calculated in the previous stage ( The system first numbers all these semantic focuses according to their natural order of appearance in the original text for easy tracking and management. Then, the system collects the data of all numbered semantic focuses. values, forming a complete list of reading fluency index values, performing statistical analysis on this list of values, and calculating the median of the list ( ) and the 75th percentile ( ), when calculating specifically, first The values are arranged in ascending order. If there are N values, the median is the The value of the position (if N is an even number, it is the average of the two middle values), the 75th percentile is the The value of the position (if the index is not an integer, it is obtained by linear interpolation), for example, if there are 100 semantic focuses After sorting, the median is the average of the 50th and 51st values, and the 75th percentile is the 75th value. After obtaining these two key statistical thresholds, the system traverses each semantic focus according to its own The value is compared with these two thresholds and the classification annotation is performed: if the Value less than the median , it is marked as “stable reading focus”; if Value greater than or equal to the median and less than or equal to the 75th percentile , it is marked as “mild non-fluent focus”; if Value strictly greater than the 75th percentile , it is marked as "significant non-fluent focus". By assigning such a clear fluency level label to each semantic focus, the system finally generates a set of semantic focus reading fluency grading labels.

[0056] Based on the semantic focus reading fluency classification label set generated in the previous step, the system first screens out all semantic focuses marked as "stable reading focus" and extracts the specific reading behavior data of these stable focuses from the "verified semantic focus and behavior characteristics" data, including the total fixation duration of each stable focus ( ), first review interval ( , the unit is converted to seconds) and the total number of replays ( ), and then calculated the arithmetic mean of these three behavioral characteristics in all “stable reading focuses” to obtain the average total fixation time ( ), average first review interval ( ), and average number of replays ( ), these average values will serve as the baseline for subsequent judgments. Then, the system traverses all semantic focuses marked as "mildly non-fluent focus" or "significantly non-fluent focus". For each such unstable focus, the corresponding total fixation duration is extracted from the "verified semantic focus and behavioral features" ( ), first review interval ( , in seconds) and the total number of replays ( ), compare these extracted behavioral features with the previously calculated average behavioral features of stable focus, and judge whether the following three conditions are met: Condition 1: Whether the total fixation time exceeds 10% of the average total fixation time of stable focus, that is, whether the total fixation time exceeds 10% of the average total fixation time of stable focus. Is it established? The second condition is whether the first review interval is less than 5 seconds, that is, Whether it is established, this 5-second threshold is set based on the common view in reading behavior research, which believes that rapid review shorter than this time usually indicates immediate reading comprehension difficulties; the third condition is whether the number of review times is greater than 1.3 times the average number of review times for stable focus, that is, the judgment Whether it is true or not, if a "mild non-fluent focus" or "significant non-fluent focus" meets at least two of the above three conditions at the same time, then the focus is further marked as a "segment to be intervened". After all the semantic focuses marked in this way are aggregated, a set of semantic focuses of intervention recommendation targets is formed.

[0057] Based on the set of semantic focuses of intervention suggestion targets formed in the previous step, the system conducts an in-depth text content analysis on each semantic focus in the set marked as a "fragment to be intervened". First, the system accurately extracts the corresponding specific sentences or text paragraphs from the original document based on the original text position information of each fragment to be intervened. Then, the natural language processing toolkit (for example, the functions of the CoreNLP or spaCy library) is called on the extracted text content to perform syntactic structure and lexical feature analysis, specifically parsing and quantifying the following linguistic indicators: First, the number of clauses contained in the text fragment is counted. By analyzing the syntactic tree structure, the number of subordinate clauses in the main-subordinate compound sentence is identified and counted; second, the number of long compound word structures is counted, where "long compound word structure" is defined as a compound word composed of three or more independent morphemes or with a total character length of more than 7 Chinese characters. The system identifies and counts such structures by combining dictionary matching and morphological analysis; third, the number of newly introduced and defined terms in the text fragment is counted; fourth, the number of external uninterpreted information cited in the text fragment is counted. After obtaining the quantitative values of the above four linguistic features, the system applies preset rules to select the adjustment category: if the number of clauses in the current intervention segment is greater than 2, or the number of long compound word structures it contains reaches or exceeds 3, then the segment is classified as "semantically simplified category". These thresholds (greater than 2 clauses, 3 or more long compound words) refer to the empirical settings of sentence complexity and vocabulary difficulty in text readability research. It is generally believed that exceeding these thresholds will significantly increase the cognitive load of reading comprehension. On the other hand, if the number of terms introduced and defined in the current intervention segment is 0, and the number of external information referenced is 1 or more, then the segment is classified as "supplementary information category". This rule aims to identify text segments that may cause comprehension difficulties due to the lack of necessary background explanations or definitions. By applying these rules to all intervention segments and determining their adjustment categories, the system ultimately generates a personalized adjustment strategy.

[0058] The steps to obtain the text content to be presented are: Based on the adjustment categories in the personalized adjustment strategy, the original text segments corresponding to the semantic focus marked as semantic simplification or supplementary information are retrieved. The syntactic structure of each original text segment is parsed to extract the clause nesting relationship, term position distribution, and compound word boundary index, thus generating a set of structural features of the original text of the semantic focus. Based on the semantic focus original text structure feature set, for text segments marked as semantic simplification, structural splitting is performed, multiple clauses are converted into parallel short sentences, and word meaning decomposition is performed on compound word structures. After extracting core word units, the core words are replaced by using a simplified substitution vocabulary. For text segments marked as supplementary information, equivalent phrases are inserted after the term position to generate the semantic focus processing results. Based on the results of semantic focus processing, the modified sentences are re-spliced to the original text position, all rewritten areas are boundary checked and semantic consistency reviewed, the replacement segment position is marked, and the text content to be presented is generated.

[0059] Specifically, based on the adjustment category of each semantic focus determined in the personalized adjustment strategy scheme generated in the previous step (i.e., "semantic simplification category" or "supplementary information category"), the system first traverses the scheme and filters out all semantic focuses that are assigned these two adjustment categories. For each selected semantic focus, the system uses the start and end character position information recorded in the original text to accurately retrieve and extract the corresponding text fragment content from the original document. Next, for each extracted original text fragment, the system initiates a complete set of sophisticated syntactic and lexical structure parsing processes, which is implemented with the help of advanced natural language processing tool libraries such as spaCy or HanLP. The specific parsing content includes using a dependency parser or a component parser to identify all clauses in the fragment (such as adverbial clauses, attributive clauses, object clauses, etc.) and clarify the nesting hierarchy and master-slave modification relationship between them. For example, for For the sentence "Despite the bad weather, they decided to set off because the mission was very urgent", the system will identify the two clauses "Despite the bad weather" and "Because the mission was very urgent" and their logical relationship with the main clause. At the same time, the system will locate all the professional terms or domain keywords that appear in the segment and record their exact starting and ending character positions in the segment. This is usually achieved by matching a pre-built domain term dictionary. In addition, the system will also identify and index the boundaries of all compound words (especially multi-word compound words or compound words with more complex structures), clarify their components and specific positions in the segment. All of this parsed structured information, such as clause type, nesting depth, term list and its position, compound word list and its boundary index, is integrated to generate a detailed structural feature record for each semantic focus original text segment to be adjusted. These records together constitute the semantic focus original text structural feature set.

[0060] Based on the previously generated set of original text structural features of the semantic focus and the adjustment category of each semantic focus in the personalized adjustment strategy, the system performs specific text rewriting operations for each semantic focus that needs to be adjusted. If a semantic focus is marked as a "semantic simplification category", the system first refers to the clause nesting relationship information in its structural feature record, and performs structural splitting on complex long sentences containing more than two clauses or with a deeper nesting level. For example, a long sentence containing multiple clauses connected by conjunctions such as "although...but..." and "because...so..." is rewritten into a set of parallel simple sentences with clear logical relationships and shorter length, or a sequence of short sentences with clearer master-slave relationships. At the same time, for the long compound word structure in the segment (for example, words composed of more than three morphemes or with more than seven characters), the system will perform word meaning decomposition, identify its core meaning unit, and then query a pre-built "simple meaning replacement vocabulary", which stores a large number of mappings of complex or professional vocabulary and their commonly used simple counterparts. relationship (for example, "implementation" is replaced by "conduct", "synergy" is replaced by "benefits of cooperation"), and the original core words are replaced by the concise synonyms or phrases found; if a semantic focus is marked as a "supplementary information category", the system will search for those terms in the segment that are determined to need additional explanation according to the strategic plan (usually terms that are not defined in the text but related to external reference information), and select corresponding explanatory phrases from a preset "equivalent phrase description library" (containing common terms and their concise explanations) after the first appearance of these terms based on their position distribution in the structural feature record, and insert them into the text in the form of bracket annotations or footnote link prompts, for example, "(a philosophical study on the nature of existence)" is inserted after the term "ontology". After these targeted simplifications or supplementary processing, the text content of each semantic focus has been modified, and these modified text segments together generate the semantic focus processing results.

[0061] Based on the semantic focus processing result generated in the previous step, which contains each modified semantic focus, the system begins to integrate these processed text fragments back into the complete document. First, the system creates a copy of the original text. Then, for each modified sentence or text fragment in the semantic focus processing result, the system accurately replaces the corresponding old text content in the original text copy based on its position identifier in the original text (such as paragraph number, sentence number, and precise offset of the starting and ending characters). During this replacement process, since the modification may cause the text length to change, the system will update the character position index of the subsequent text in the affected area in real time to ensure the accuracy of the replacement. After the replacement is completed, the system performs strict boundary checks on all areas where content has been rewritten, checks whether the modified fragment is naturally connected with the unmodified text before and after it, ensures that punctuation is used correctly, and does not produce incomplete sentences or paragraphs. For example, if a A long sentence is split into two short sentences. The system will check whether each new short sentence ends with an appropriate punctuation mark such as a period, and whether the spaces or line breaks between sentences comply with typesetting specifications. Then, the system will conduct a preliminary semantic consistency review by calling a text coherence assessment tool based on a large language model to calculate the semantic smoothness score of the rewritten area and its surrounding context, or through keyword co-occurrence analysis, topic model comparison and other methods to check whether the rewriting significantly deviates from the original theme and meaning. If potential semantic conflicts or logical incoherence are found, the system will mark these areas for subsequent manual review or further automatic correction. At the same time, the system will accurately record the new starting and ending positions of each replaced or modified text paragraph (or fragment) in the final document, forming a replacement area position index table. Finally, after all modification, splicing, verification and marking steps, the text content to be presented is formed with optimized content.

[0062] The steps to obtain the adaptive guided reading interface are: Based on the text content to be presented, the starting position, ending position, and focus identification status of each semantic focus in the paragraph are extracted in sequence. The position coordinates of all semantic focuses in the complete text are mapped to the interface rendering framework. Semantic simplification or supplementary description tags are bound according to the semantic focus type. The reading priority order and highlighting status are configured for each segment, and a reading segment position index and visual indication parameter set are generated. Based on the reading segment position index and visual indication parameter set, a reading guidance control sequence is constructed. Segments are inserted into the reading navigation sequence according to the distribution order of focus in the text structure, and visual jump trigger points are set. Operation prompt information, reading suggestion labels and dynamic highlight borders are embedded in all focus segments in sequence to generate a reading path navigation execution task set. Based on the reading path navigation, the task set is executed, the complete text structure and task trigger configuration are loaded, and guided jump rendering is performed for each reading focus. The currently focused segment is highlighted with a border mark, and a reading suggestion window pops up on the right side of the segment. Other non-focus paragraphs are controlled for light and dark processing and interface masking effects, guiding users to focus on the defined reading path and generating an adaptive guided reading interface.

[0063] Specifically, based on the text content to be presented finally generated in the previous step, the system starts the parameter preparation process before interface rendering. It will traverse all the semantic focuses that have been previously identified and processed. For each semantic focus, it first accurately extracts its latest starting character position and ending character position in the paragraph from the text content to be presented, and at the same time calls the focus identification status assigned to the focus in the "Semantic Focus Reading Fluency Grading Label Set", such as "Significantly Non-fluent Focus" or "Stable Reading Focus". Then, the system uses the coordinate conversion algorithm to accurately map these semantic focus position coordinates based on text characters to screen pixel coordinates that can be directly used by the interface rendering framework (such as HTML5Canvas or a specific GUI library) according to the configuration of the current display environment (such as font size, line spacing, screen resolution and text display area width), that is, the specific rectangular area (x, y, width, height) of each semantic focus on the screen, and according to the adjustment category determined in the "Personalized Adjustment Strategy Scheme" for the semantic focus (i.e., "Semantic Simplification Category" or "Supplementary Information Category" ), and a corresponding type tag is associated with it. For example, a simplified segment is associated with the "Simplified" tag data, and a segment with supplementary information is associated with the "With Supplementary Notes" tag data. In addition, the system also configures an initial reading priority order for each semantic focus segment. This order defaults to the natural appearance order of the segments in the text content to be presented, but can be fine-tuned based on their focus status. For example, a segment marked as "Significant Non-fluent Focus" and subjected to "Semantic Simplification" may have its priority appropriately advanced. At the same time, an initial highlight state parameter is also set for each segment. This parameter specifically defines the background color of the segment when it is not currently in reading focus (for example, all semantic focuses use a light gray background #F5F5F5 to distinguish it from ordinary text) or border style. All of this extracted and configured information, including each semantic focus's text position, screen position coordinates, focus state label, adjustment type label, reading priority, and initial visual highlight parameters, is integrated to generate a reading segment position index and visual indication parameter set.

[0064] Based on the reading segment position index and visual indication parameter set generated in the previous step, the system begins to construct a detailed control sequence to guide the user to read. First, it inserts these semantic focuses one by one into a main reading navigation sequence list in an orderly manner according to the configured reading priority order of each semantic focus in the set (usually their natural flow order in the text). The list constitutes the backbone of the reading path that the user will be guided to follow. For each semantic focus in the sequence, the system will set clear visual jump trigger points. These trigger points define when to automatically guide the user's reading focus from the current segment to the next segment in the sequence. The trigger conditions can be diverse. For example, the user's eye movement fixation time at the end area of the current focus segment exceeds a preset threshold (for example, it is empirically set to 1.5 seconds, indicating that the user may have finished reading), or the user clicks a clear "continue" or "next focus" button on the interface, or the system triggers the user according to the text of the current focus segment. The system estimates an expected reading time based on the length and average reading rate, and automatically triggers a jump after the time expires. Subsequently, the system embeds corresponding interactive elements and visual prompts for all semantic focus segments in the navigation sequence according to their characteristics (such as adjustment category, focus identification status), including embedding brief operation prompt information text (for example, for segments with supplementary explanations, a prompt of "click to view details" can be embedded), attaching easy-to-understand reading suggestion labels (for example, for a simplified complex concept segment, the label can be "The core concept has been simplified, please pay attention to understanding"), and pre-defines a dynamic highlight border style applied when the segment becomes the current active focus (for example, the border color becomes a striking blue, with a width of 2 pixels, and a slight breathing animation effect). All these navigation sequences, jump trigger logic, prompt information, suggestion labels and dynamic highlight parameters are integrated into the execution instructions of each focus to jointly generate a reading path navigation execution task set.

[0065] Based on the reading path navigation execution task set generated in the previous step, the system starts to render and activate the adaptive guided reading interface. First, it loads the complete text content to be presented into the user interface, and at the same time loads the specific location information, visual style parameters and preset task trigger configurations of all semantic focuses contained in the reading path navigation execution task set. The reading guidance process usually starts from the first semantic focus in the reading navigation sequence. The system performs a guided jump rendering, and displays the text area where the reading focus specified in the current sequence is located in the center or optimal reading position of the user's screen through smooth scrolling or instant positioning. Then, in order to highlight the current reading target, the system applies the dynamic highlight border mark defined in the task set to the semantic focus segment being focused. For example, a 2-pixel wide bright blue border will surround the segment, making it visually stand out from the surrounding text. At the same time, if the focus segment is associated If there are reading suggestions or operation prompts, a small, non-intrusive reading suggestion window will pop up in time on the right side or below the fragment to display relevant content, for example, "This paragraph contains important definitions, it is recommended to read carefully." In order to further strengthen the focus, the system will actively control other text paragraphs on the interface that are not currently in focus, and perform significant light and dark contrast processing on them. For example, the color of non-focused text is adjusted from the default black to a lighter gray (such as #888888), or its display transparency is reduced to 60%. For text areas farther away from the current focus, even a progressive interface masking effect can be applied, such as fading blur at the upper and lower edges. Through these combined visual guidance methods, including highlighting the current, weakening others, and providing instant suggestions, the system effectively guides the user's attention to flow along a pre-defined reading path, and ultimately generates and presents a highly optimized adaptive guided reading interface that can dynamically adapt to the user's reading needs.

[0066] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. The AI reading control method based on artificial intelligence is characterized by: The following steps are involved: Perform word segmentation on the input raw text, segmenting the text into independent sentences and identifying the basic word units within each sentence. Analyze the semantic adjacency and syntactic structure features between words, screen and extract potential phrases representing the meaning of the paragraph, and establish a set of candidate semantic units. Presenting text content associated with the candidate semantic unit set, tracking user gaze duration and rewinding actions, obtaining user interaction behavior records, performing semantic weighting based on the user interaction behavior records and the candidate semantic unit set, verifying and determining the current highlighted semantic focus and corresponding reading behavior features, and generating verified semantic focus and behavior features; Calculate a reading fluency index based on the verified semantic focus and user reading behavior characteristics in the behavioral characteristics, and based on the reading speed change rate and the frequency of review. Based on the reading fluency index and with reference to the verified semantic focus in the behavioral characteristics, select an adjustment category of simplification or information supplementation, and formulate a personalized adjustment strategy plan. According to the adjustment category in the personalized adjustment strategy, word replacement is performed on the target text segment to generate the text content to be presented. Based on the text content to be presented, reading path planning is performed, the current focus and reading suggestions are highlighted, and an adaptive guided reading interface is constructed.

2. The AI reading control method based on artificial intelligence according to claim 1 is characterized in that: The steps of obtaining the candidate semantic unit set are: The original text is segmented sentence by sentence according to punctuation. For each sentence, continuous word group segments are extracted in sequence. The part-of-speech tag, starting character position in the sentence, ending character position, and syntactic dependency distance with adjacent words of each segment are recorded to form a basic semantic segment attribute table. Calculating the intra-sentence aggregation value of each semantic segment according to the basic semantic segment attribute table; According to the intra-sentence aggregation value, the fragments with intra-sentence aggregation value higher than the median in each sentence are selected, and the fragments are sorted and merged according to the order of the fragments in the text and the start and end character positions. The fragment combinations in the scoring set in continuous or adjacent sentences are extracted across sentences to generate a set of candidate semantic units.

3. The AI reading control method based on artificial intelligence according to claim 1 is characterized in that: The steps for obtaining the user interaction behavior record are: Presenting text content associated with the candidate semantic unit set, locating the corresponding paragraph of each semantic segment in the candidate semantic unit set in the original text, highlighting or marking the corresponding paragraph in the reading interface, and synchronously projecting the text content into the user visual interface area to generate user-visible candidate semantic unit-associated text content; Based on the text content associated with the candidate semantic units visible to the user, the user's facial gaze is monitored in real time, the duration of the user's gaze at the text position corresponding to each semantic segment is recorded, and whether the gaze moves back to the read area is detected again. Each pause and replay event is marked with a timestamp and paragraph number to generate raw observation data of user behavior; Based on the original observation data of user behavior, the gaze event sequence corresponding to each paragraph of text is analyzed, and the average gaze time, first look-back interval, number of looks-backs and total stay period of each semantic segment are extracted. A behavior comparison table is established in chronological order and paragraph number to generate a user interaction behavior record.

4. The AI reading control method based on artificial intelligence according to claim 1 is characterized in that: The steps for obtaining the verified semantic focus and behavioral features are: Based on the user interaction behavior record, each gaze event in the user interaction behavior record is matched with the semantic segments in the candidate semantic unit set, and the total gaze duration, the actual start time of each replay, the duration of each replay, the total number of replays, and the total character length of the paragraph in which the segment is located are extracted for each semantic segment to generate a semantic segment gaze behavior feature set; Calculating the reading focus intensity of each semantic segment according to the semantic segment gaze behavior feature set; Based on the reading focus intensity, the reading focus intensity of all semantic segments in a set of continuous paragraphs consisting of three adjacent paragraphs of text are sorted, the 75th percentile of the reading focus intensity is set as the judgment threshold, and the semantic segments with reading focus intensity not lower than the judgment threshold are screened. The gaze duration, first review interval and number of reviews are extracted to generate verified semantic focus and behavioral characteristics.

5. The AI reading control method based on artificial intelligence according to claim 1 is characterized in that: The steps for obtaining the reading fluency index are as follows: Based on the verified semantic focus and user reading behavior features in the behavior features, the sentence number, actual reading time of each sentence, the triggering number and the first occurrence time of each sentence replay event in each semantic focus are extracted in sequence, and a sentence-level reading time series and a replay number series are constructed in the order of sentence numbers to generate a sentence-level reading behavior dataset within the semantic focus; Based on the sentence-level reading behavior dataset within the semantic focus, the reading time of each sentence, the corresponding number of replays, and the time difference between the previous and next sentences are calculated, and the pause intervals between adjacent sentences are recorded as auxiliary features to obtain a combined sequence of reading speed changes and replay frequency changes; Based on the combined sequence of reading speed changes and review frequency changes, a reading fluency index of semantic focus is calculated.

6. The AI reading control method based on artificial intelligence according to claim 1 is characterized in that: The steps for obtaining the personalized adjustment strategy are as follows: Based on the reading fluency index, the corresponding reading fluency index value is extracted from each semantic focus. After the semantic focuses are numbered according to the original text order, the median and 75th percentile of the reading fluency index values of all semantic focuses are calculated in sequence. The values below the median are defined as "stable reading focus", the values between the median and the 75th percentile are defined as "mild non-fluent focus", and the values above the 75th percentile are defined as "significant non-fluent focus". Each semantic focus is classified and labeled to generate a set of semantic focus reading fluency classification labels; Based on the semantic focus reading fluency classification label set, for each semantic focus labeled as "mild non-fluent focus" and "significant non-fluent focus", the total fixation duration, first re-look interval, and number of re-looks in the semantic focus are extracted. The average values of the three behavioral characteristics of all semantic focuses in the stable reading focus are compared to determine whether the semantic focus simultaneously meets the following conditions: the total fixation duration exceeds the mean by more than 10%, the first re-look interval is less than 5 seconds, and the number of re-looks is greater than 1.3 times the mean. If two or more conditions are met, the semantic focus is marked as a segment to be intervened, forming a set of target semantic focuses for intervention recommendations; Based on the set of semantic focuses targeted by the intervention recommendations, the sentences corresponding to the original text positions in each semantic focus are extracted. The number of clauses, the number of long compound word structures, the number of introduced definition terms, and the number of references to external information in the sentences are analyzed. If the number of clauses is greater than 2 or includes 3 or more long compound word structures, the semantic simplification category is selected. If the number of introduced definition terms is 0 and the number of references to external information is 1 or more, the supplementary information category is selected to generate a personalized adjustment strategy plan.

7. The AI reading control method based on artificial intelligence according to claim 1 is characterized in that: The steps for obtaining the text content to be presented are: Based on the adjustment categories in the personalized adjustment strategy, the original text segments corresponding to the semantic focus marked as semantic simplification or supplementary information are retrieved, the syntactic structure of each original text segment is parsed, the clause nesting relationship, term position distribution and compound word boundary index are extracted, and a set of original text structure features of the semantic focus is generated; Based on the semantic focus original text structure feature set, for text segments marked as semantic simplification categories, structural splitting is performed to convert multiple clauses into parallel short sentences, and word meaning decomposition is performed on compound word structures. After extracting core word units, a simplified substitution vocabulary is called to replace the core words. For text segments marked as supplementary information categories, equivalent phrase descriptions are inserted after the term position to generate semantic focus processing results. Based on the semantic focus processing result, the modified sentence is re-joined to the original text position, all rewritten areas are subjected to boundary verification and semantic consistency review, the replacement segment position is marked, and the text content to be presented is generated.

8. The AI reading control method based on artificial intelligence according to claim 1 is characterized in that: The steps for obtaining the adaptive guided reading interface are: Based on the text content to be presented, the starting position, ending position, and focus identification status of each semantic focus in the paragraph are extracted in sequence, the position coordinates of all semantic focuses in the complete text are mapped to the interface rendering framework, and semantic simplification or supplementary description tags are bound according to the semantic focus type. The reading priority order and highlighting status are configured for each segment, and a reading segment position index and a visual indication parameter set are generated; Based on the reading segment position index and visual indication parameter set, a reading guidance control sequence is constructed, segments are inserted into the reading navigation sequence according to the distribution order of focus in the text structure, and visual jump trigger points are set. Operation prompt information, reading suggestion labels and dynamic highlight borders are embedded in all focus segments in sequence to generate a reading path navigation execution task set; Based on the reading path navigation execution task set, the complete text structure and task trigger configuration are loaded, guided jump rendering is performed for each reading focus, the currently focused segment is highlighted with a border mark, and a reading suggestion window pops up on the right side of the segment. Other non-focus paragraphs are controlled for light and dark processing and interface masking effects, guiding users to focus on the defined reading path and generating an adaptive guided reading interface.

9. The reading control system according to any one of claims 1 to 8, characterized in that: include: Semantic Extraction Module: This module performs word segmentation on the input raw text, splits the text into independent sentences, identifies the basic word units within each sentence, analyzes the semantic adjacency and syntactic structure features between words, screens and extracts potential phrases representing the meaning of the paragraph, and establishes a set of candidate semantic units. Behavior perception module: presents text content associated with the candidate semantic unit set, tracks user gaze duration and replay actions, obtains user interaction behavior records, performs semantic weighting based on the user interaction behavior records and the candidate semantic unit set, verifies and determines the current highlighted semantic focus and corresponding reading behavior characteristics, and generates verified semantic focus and behavior characteristics; Strategy formulation module: Calculates a reading fluency index based on the verified semantic focus and user reading behavior characteristics in the behavioral characteristics, the reading speed change rate, and the frequency of review; selects a simplification or information supplement adjustment category based on the reading fluency index and reference to the verified semantic focus in the behavioral characteristics, and formulates a personalized adjustment strategy plan; Content adaptation module: Based on the adjustment category in the personalized adjustment strategy, perform word replacement on the target text segment to generate the text content to be presented, plan the reading path based on the text content to be presented, highlight the current focus and reading suggestions, and build an adaptive guided reading interface.

Images