Intelligent learning companion method, system, and storage medium

Abstract

The application provides a kind of intelligent learning accompanying method, system and storage medium, the method includes: obtaining the reading behavior event of user in learning reading, and the feature analysis of the reading behavior event, obtain learning state feature;According to the learning state feature determines the posterior probability of card point, and when the posterior probability of card point meets preset trigger condition, the card point type of current card point is identified;Current learning accompanying mode is obtained, and according to the posterior probability of card point, the current learning accompanying mode, the current card point and the card point type are actively prompted.This embodiment of the application can automatically actively prompt the user based on the posterior probability of card point, current learning accompanying mode, current card point and card point type, improve the user experience.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent learning companion method, system, and storage medium. Background Technology

[0002] As the digitization of printed materials continues to advance, e-readers, with their advantages of eye-friendly displays, resource integration, and portable storage, have become the mainstream medium for after-school independent reading and academic learning. They are widely used in primary and secondary schools for in-class and extracurricular self-study, text analysis, and knowledge consolidation. Current learning models rely on e-readers to carry textbooks, supplementary materials, and explanatory texts for various subjects, allowing learners to independently browse texts, complete reading, and engage in self-study, thus meeting the needs of regular, fragmented, and independent learning.

[0003] Existing e-readers typically offer basic functions such as table of contents browsing, highlighting, note-taking, searching, and bookmarking, but they cannot proactively provide learning prompts to users, thus reducing the user experience. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent learning companion method, system, and storage medium to solve the problem that existing technologies cannot proactively provide learning prompts to users.

[0005] The present invention is implemented as follows: an intelligent learning companion method, the method comprising:

[0006] The system acquires reading behavior events of users during their learning and reading process, and performs feature analysis on these reading behavior events to obtain learning state features.

[0007] The posterior probability of the checkpoint is determined based on the learning state characteristics, and the checkpoint type of the current checkpoint is identified when the posterior probability of the checkpoint meets the preset triggering condition.

[0008] Obtain the current learning support mode and provide proactive learning prompts based on the posterior probability of the stuck point, the current learning support mode, the current stuck point, and the stuck point type.

[0009] Preferably, feature analysis is performed on the reading behavior events to obtain learning state features, including:

[0010] The system retrieves the viewing duration, highlighted text length, note length, question length, number of clicks on the explanation, exercise error rate, and current reading content from the reading behavior event.

[0011] The viewing duration, highlighted text length, note length, question length, number of analysis clicks, and exercise error rate are normalized to obtain the viewing duration vector, highlighted text vector, note vector, question vector, analysis click vector, and exercise error vector.

[0012] Feature aggregation is performed on the current page, current paragraph, current knowledge unit, current chapter, and current learning session in the current reading content to obtain a reading vector;

[0013] The learning state features are obtained by combining the viewing dwell vector, the highlighted text vector, the note vector, the question vector, the parsing click vector, the exercise error vector, and the reading vector.

[0014] Preferably, determining the posterior probability of the checkpoint based on the learning state features includes:

[0015] The expected dwell time is determined based on the text length of the current knowledge unit, the user's historical reading speed, and the difficulty of the unit knowledge. The dwell anomaly is determined based on the viewing dwell time and the expected dwell time. The dwell evidence strength is determined based on the dwell anomaly.

[0016] Obtain the number of times a user rewatches a video, and determine the rewatch intensity based on the number of rewatches;

[0017] The system obtains the user's knowledge question in the current knowledge unit, calculates the semantic similarity between the knowledge question and the current knowledge unit, and determines the intensity of repeated questions based on the semantic similarity.

[0018] The strength of the explanatory dependency is determined based on the number of clicks, the user's historical mastery of the current knowledge unit is obtained, and the mastery gap is determined based on the historical mastery.

[0019] The knowledge difficulty and conceptual confusion of the current knowledge unit are obtained, and the observation likelihood is determined based on the strength of evidence of dwell time, the intensity of revisiting, the intensity of repeated questioning, the intensity of explanation dependence, the mastery gap, the knowledge difficulty, and the conceptual confusion.

[0020] The prior probability is determined based on the probability of being stuck in the previous time window, the difficulty of knowledge, the gap in mastery, and the degree of conceptual confusion. The posterior probability of being stuck is then determined based on the prior probability and the observation likelihood.

[0021] Preferably, the formulas used to determine the prior probability based on the probability of getting stuck in the previous time window, the difficulty of the knowledge, the mastery gap, and the degree of conceptual confusion include:

[0022] π_t =P_stuck(t1)(1-μ)+(1-p_stuck(t1))•[1-(1-β0)(1-βg g_t)(1-βm m_t)(1-βc c_t)(1-βi i_t)]

[0023] Wherein, π_t represents the prior probability, β0, βg, βm, βc, and βi represent preset prior triggering parameters, p_stuck(t1) represents the probability of being stuck in the previous time window t1, μ represents the state retention parameter, g_t represents the knowledge difficulty, m_t represents the mastery gap, c_t represents the concept confusion degree, and i_t represents the knowledge importance.

[0024] Preferably, determining the posterior probability of the checkpoint based on the prior probability and the observed likelihood includes:

[0025] p_stuck(t)=π_t•L1_t / [π_t•L1_t+(1-π_t)•L0_t]

[0026] L1_t=ε+(1-ε)[1-∏_{s∈S_t}(1-ρ_s•s)]

[0027] L0_t=ε+(1-ε)∏_{s∈S_t}(1-ρ_s·s)

[0028] Where p_stuck(t) represents the posterior probability of the checkpoint, L1_t represents the strong likelihood, L0_t represents the weak likelihood, ε is the smoothing term, ρ_s is the confidence weight of different observation evidence, S_t represents the set of evidence in the set of evidence, and s represents the evidence in the set of evidence in the set of evidence.

[0029] Preferably, the checkpoint type for identifying the current checkpoint includes:

[0030] The intent of the knowledge question is classified to obtain the question intent, and the unit type of the current knowledge unit is obtained;

[0031] When the degree of abnormality of the stay is greater than the degree of abnormality threshold, and the intention of the question is the first preset intention, then the type of the stuck point is determined to be conceptual incomprehension.

[0032] When the review intensity is greater than the intensity threshold and / or the mastery gap is greater than the gap threshold, the type of the bottleneck is determined to be the prior knowledge deficiency type.

[0033] When the unit type is a preset type and the knowledge question contains a first preset word, the obstacle type is determined to be either the derivation process difficulty type or the formula symbol difficulty type.

[0034] If the knowledge question contains a second preset word, then the type of obstacle is determined to be conceptual confusion.

[0035] Preferably, active learning prompts are provided based on the posterior probability of the checkpoint, the current learning support mode, the current checkpoint, and the checkpoint type, including:

[0036] The first probability threshold, the second probability threshold, and the third probability threshold are determined based on the current learning companion mode, and learning prompt information is determined based on the current checkpoint and the checkpoint type.

[0037] If the posterior probability of the checkpoint is less than the first probability threshold, then the prompting stops;

[0038] If the posterior probability of the checkpoint is greater than or equal to the first probability threshold and less than the second probability threshold, then an information prompt port is displayed at the current checkpoint.

[0039] If a confirmation instruction is received for the information prompt port, the learning prompt information is displayed;

[0040] If the posterior probability of the checkpoint is greater than or equal to the second probability threshold and less than the third probability threshold, then the checkpoint type and explanation button are displayed on the current checkpoint.

[0041] If a confirmation instruction is received for the explanation button, the learning prompt information is displayed;

[0042] If the posterior probability of the checkpoint is greater than or equal to the third probability threshold, the learning prompt information is displayed directly.

[0043] Another objective of this invention is to provide an intelligent learning companion system, the system comprising:

[0044] The feature analysis module is used to acquire reading behavior events of users when learning to read, and to perform feature analysis on the reading behavior events to obtain learning state features;

[0045] The checkpoint verification module is used to determine the posterior probability of the checkpoint based on the learning state features, and to identify the checkpoint type of the current checkpoint when the posterior probability of the checkpoint meets the preset triggering conditions.

[0046] The learning prompt module is used to obtain the current learning companion mode and provide proactive learning prompts based on the posterior probability of the stuck point, the current learning companion mode, the current stuck point, and the stuck point type.

[0047] In this embodiment of the invention, by performing feature analysis on reading behavior events, the learning state characteristics of the user's current learning state can be effectively obtained. Based on the learning state characteristics, the posterior probability of getting stuck can be effectively determined. Based on the posterior probability of getting stuck, it can be effectively determined whether the user is currently experiencing a reading stuck phenomenon. When the posterior probability of getting stuck meets the preset triggering conditions, based on the posterior probability of getting stuck, the current learning companion mode, the current stuck point, and the stuck point type, the user can be automatically given proactive learning prompts, thereby improving the user experience. Attached Figure Description

[0048] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0049] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0050] Figure 1 This is a flowchart of the intelligent learning companion method provided in the first embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of the intelligent learning companion system provided in the second embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the structure of the terminal device provided in the third embodiment of the present invention. Detailed Implementation

[0053] The following specific embodiments illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] Example 1

[0056] Please see Figure 1 This is a flowchart of the intelligent learning companion method provided in the first embodiment of the present invention. This embodiment provides a proactive AI learning companion method based on the integration of reading behavior and the semantic structure of teaching materials. By acquiring electronic learning materials imported by the user, the method analyzes chapters, pages, paragraphs, charts, formulas, and text anchors to further identify concepts, definitions, formulas, examples, conclusions, prerequisite knowledge, easily confused points, importance, and difficulty, forming knowledge units bound to the original text. During the user's reading process, behavioral events such as page dwell, page turning, rereading, scrolling, underlining, note-taking, concept clicking, AI questioning, follow-up questions, quiz answers, and error recording are continuously collected. These behavioral events are associated with knowledge units to generate learning status characteristics of the user on specific knowledge points. Based on the learning status characteristics, knowledge unit difficulty, historical mastery, and user preferences, the probability of getting stuck is calculated, the type of getting stuck is identified, and proactive learning prompts are provided when trigger conditions are met.

[0057] This intelligent learning companion method can be applied to any device or system, and includes the following steps:

[0058] Step S10: Obtain reading behavior events of the user during learning and reading, and perform feature analysis on the reading behavior events to obtain learning state features;

[0059] This involves pre-acquiring user-imported e-learning materials, including PDFs, EPUBs, Word documents, HTML files, scanned OCR text, or other parsable documents. The e-learning materials are then parsed in a structured manner to obtain chapter structure, page structure, paragraph text, heading levels, charts, formulas, footnotes, citations, and page coordinates.

[0060] The specific steps involved in the structured parsing of e-learning materials in this process include:

[0061] Format recognition: After receiving electronic learning materials, the client or server identifies the material type based on the file extension, MIME type, file header information, and content characteristics. The material types include PDF with text layer, scanned PDF, EPUB, Word, HTML, image, or mixed documents.

[0062] Content extraction: For PDFs containing text layers, read the text objects, font size, coordinate boxes, and reading order of each page; for scanned PDFs or images, call the OCR module to obtain text, line blocks, paragraph boxes, and confidence scores; for EPUB / HTML, parse DOM nodes and heading levels; for Word documents, read paragraphs, heading styles, tables, and image objects.

[0063] Page layout reconstruction: The extracted results are uniformly converted into page-level page blocks. Each page block includes at least the following fields: page_id, block_id, block_type, text, bbox, font_size, reading_order, and confidence. The block_type includes at least one of the following: title, body text, charts, formulas, footnotes, headers / footers, and table of contents. The bbox is a bounding box, representing the coordinates of a section of text, formula, chart, or knowledge point within the original ebook page.

[0064] Table of Contents and Chapter Recognition: Identify chapter titles based on the table of contents page, title number, font level, indentation, bold style, and regular expression rules, and generate a mapping relationship between chapter_id, section_id, and page_id; when table of contents hyperlinks exist, they are given priority.

[0065] Paragraph and text anchor generation: Merge text blocks with the same reading order, adjacent coordinates and consistent style into paragraphs; generate stable text anchors for each paragraph, sentence and key phrase, including document_id, page_id, block_id, char_start, char_end and bbox.

[0066] Chart and formula processing: Preserve page coordinates and chart / table title text for the chart area; extract formula images, LaTeX / OCR text, or symbol sequences for the formula area and establish association with the surrounding paragraphs.

[0067] Output structured document objects: Generates a document structure data object, including a chapter tree, page list, paragraph list, page block list, chart and formula list, footnote reference list, and text anchor index, for subsequent semantic analysis and original text location.

[0068] Semantic analysis is performed on the parsed text to identify concepts, definitions, formulas, examples, conclusions, causal relationships, prerequisite knowledge points, difficult knowledge points, and chapter learning objectives.

[0069] This step involves semantic analysis of the parsed text, including:

[0070] Text preprocessing: Denoising, sentence segmentation, terminology normalization, formula placeholder replacement, and cross-page paragraph merging are performed on the paragraph text in the structured document object to obtain text fragments that can be processed by the model.

[0071] Candidate concept extraction: Candidate concepts are extracted using a rule dictionary, title weights, definition sentence matching, and sequence labeling models. Definition sentences may include "X is...", "X refers to...", "X is called...", "X is defined as...", etc.

[0072] Definition, formula, example, and conclusion identification: Use text classification models or rule classifiers to classify paragraphs into categories including concept definitions, formula derivations, examples, cases, summaries and conclusions, precautions, prerequisite knowledge, and common mistakes.

[0073] Relation extraction: Based on adjacent chapter relationships, citation relationships, syntactic dependency relationships, co-occurrence relationships, and semantic similarity, identify pre-relationships, inclusion relationships, causal relationships, contrast relationships, inference relationships, and easily confused relationships.

[0074] Importance score: The importance score is output by the graph attention personalized random walk model ImportanceGraphModel, with the output value importance_score∈[0,1], representing the centrality and propagation influence of a knowledge unit in the current document and the current learning objective. Specifically, a directed heterogeneous knowledge graph G=(V,E) is constructed with knowledge units as nodes. Edge types include preconditions, inclusions, causal / derivative relationships, chapter references, exercise references, graph / formula associations, and easily confused relationships. For node u, the node feature z_u=[level_u,def_u, objective_u, exercise_u, reuse_u, formula_u,Graph_degree_u] is constructed, where level_u represents the title level feature, def_u indicates whether a definition sentence exists, objective_u indicates whether it hits the chapter learning objective, exercise_u represents the normalized value of exercise reference count, and reuse_u represents the normalized value of cross-chapter reuse count. For edge (u,v), the edge feature is constructed:

[0075] r_uv=[edge_type, distance_uv, cooccur_uv, cite_uv]

[0076] The edge attention weights are calculated using a graph attention mechanism:

[0077] e_uv=LeakyReLU(aT[Wz_u || Wz_v || Rr_uv])

[0078] α_uv=exp(e_uv) / Σ_{k∈N(u)}exp(e_uk)

[0079] Where W, R, and a are parameters obtained through training or preset, N(u) is the set of neighboring nodes of node u, and || denotes vector concatenation. Then, a transition matrix P is constructed, and P_uv = α_uv is set. At the same time, a personalized prior vector q is generated according to the learning objective. If the knowledge unit hits the chapter objective, the exam objective, or the learning objective selected by the user, then q_u takes a larger value and satisfies Σ_uq_u = 1.

[0080] The importance vector π is obtained through iterative random walks with restart: π(t+1) = λq + (1-λ)PTπ(t), where 0 < λ < 1. The iteration stops when ||π(t+1) - π(t)||_1 < ε or the maximum number of iterations is reached, resulting in a stable vector π*.

[0081] importance_score(u)=(π*_u-min_vπ*_v) / (max_vπ*_v-min_vπ*_v+δ)

[0082] δ is a minimal constant to prevent the denominator from being zero. This model can express multi-hop propagation relationships. For example, if a knowledge point itself appears infrequently, but is depended on by multiple subsequent formulas, examples, and chapter objectives, random walks will transfer the importance of adjacent nodes to that knowledge point. Similarly, if a knowledge point is referenced by numerous practice questions and connected to multiple easily confused concepts, its edge attention weights and graph centrality will jointly increase its importance_score. Therefore, this importance score is the output of a nonlinear graph model based on heterogeneous knowledge graphs, multi-hop propagation, and attention normalization.

[0083] Difficulty Score: The difficulty score is output by the DifficultyModel, with the output value difficulty_score ∈ [0,1]. In one specific implementation, the DifficultyModel is implemented using a Gradient Boosting Regression Tree (GBRT), whose input includes both the historical user stuck rate and the features of the knowledge point itself. For knowledge unit u, the historical stuck rate h_u = (n_stuck(u) + α) / (n_read(u) + α + β) is calculated first, where n_stuck(u) is the number of users or sessions that have historically stuck at knowledge unit u, n_read(u) is the number of users or sessions that have read the knowledge unit, and α and β are smoothing parameters used to avoid distortion of the rate due to insufficient cold start samples. Then, the feature vector of the knowledge itself is constructed:

[0084] x_u=[h_u, len_u, formula_u,Prereq_u, example_u, term_u, symbol_u,abs_u]

[0085] Where len_u = log(1 + number of knowledge text tokens) / log(1 + L_max), formula_u = min(1, number of formulas / F_max), prereq_u = min(1, number of prerequisite knowledge points / P_max), example_u = exp(-number of examples / τ_e), term_u is the proportion of low-frequency terms, symbol_u is the symbol or variable density, and abs_u is the concept abstraction score. The GBRT model consists of M regression trees, initialized with F_0(x) = mean(y_i), where y_i is the proportion of historical true blocking or future window blocking in the training samples. During the m-th training round, the residual r_i(m) = y_i - F_(m-1)(x_i) is calculated, and a regression tree T_m(x) = Σ_j γ_mj•I(x∈R_mj) is fitted, where R_mj represents the j-th leaf region of the m-th tree, and γ_mj is the mean of the residuals within that leaf region or the leaf output obtained by minimizing the squared loss. The model recursion is F_m(x) = F_(m-1)(x) + ν•T_m(x), where ν is the learning rate. The final difficulty coefficient is difficulty_score(u) = min(1, max(0, F_M(x_u))). Since T_m(x) forms different leaf regions through multi-level conditional splitting, this model can express non-linear combinations such as "a large number of formulas and many prerequisite knowledge points", "long text but few examples", and "a high proportion of historical obstacles and many low-frequency terms", rather than simply performing linear weighting on each feature.

[0086] Original text evidence binding: Each concept, definition, formula, example, and relationship is bound to source_span and text_anchor to ensure that subsequent AI explanations, key annotations, and question-and-answer results can be traced back to the original text.

[0087] Establish a document knowledge unit model that associates concepts, paragraphs, chapters, page coordinates, text anchors, difficulty levels, importance levels, and original text evidence fragments.

[0088] In one specific implementation, the document knowledge unit model consists of a "structured data model + semantic encoding model + relation classification model".

[0089] The structured data model KnowledgeUnit includes at least the following fields: unit_id, document_id, chapter_id, section_id, page_id, paragraph_id, text_anchor, bounding box, unit_type, concept_name, definition_text, source_text, source_span, importance_score, difficulty_score, prerequisite_unit_ids, related_unit_ids, confusable_unit_ids, embedding_vector, created_at, and updated_at.

[0090] The semantic encoding model can adopt a Transformer encoder structure. In one embodiment, a Chinese BERT or GPT-like encoder is used to vectorize paragraphs, concepts, and questions, and output a fixed-dimensional embedding_vector for semantic retrieval, concept deduplication, and similar knowledge point matching.

[0091] The concept recognition model can adopt a Transformer Encoder annotation structure. The input is a sequence of sentence tokens, and the output is a sequence of BIO tags, including B-CONCEPT, I-CONCEPT, B-DEFINITION, I-DEFINITION, B-FORMULA, I-FORMULA, and O.

[0092] The relation classification model can employ a dual-tower encoder or a cross-encoder structure. For two candidate knowledge units Ku and Kv, the model input consists of their text, chapter distance, co-occurrence frequency, and title-level features. The output is the probability that they belong to the categories of preceding, containing, causal, contrastive, derivative, easily confused, or unrelated.

[0093] Importance and difficulty scoring models can be implemented using gradient boosting trees, logistic regression, lightweight neural networks, or rule-weighted models. For ease of implementation, one embodiment uses rule-weighted initialization, followed by online calibration based on user behavior data.

[0094] During the user's reading process, reading behavior events are collected, including page dwell time, scrolling speed, page turning behavior, number of times the page is reread, underlined areas, note content, concept clicks, AI Q&A, quiz answers, error records, etc.

[0095] Based on reading behavior events and document knowledge unit models, user learning status characteristics are generated, including current reading position, number of times knowledge points are exposed, degree of abnormal dwell time, frequency of revisiting, question density, explanation click rate, test accuracy, and historical mastery.

[0096] Event collection: The reading client generates ReadingEvents when a user enters a page, leaves a page, scrolls, turns pages, looks back, underlines, creates notes, clicks on concept explanations, submits questions, or answers quiz questions.

[0097] Events collected from different platforms are standardized into fields such as user_id, document_id, session_id, event_type, timestamp, page_id, anchor_id, duration, selected_text, question_text, and answer_result.

[0098] Event Attribution: Match the reading behavior event with the KnowledgeUnit's text_anchor, bbox, or selected_text to attribute the event to one or more knowledge units.

[0099] Time window aggregation: Aggregates behavioral features at multiple granularities, including the current page, current paragraph, current knowledge unit, current chapter, and current learning session.

[0100] Baseline normalization: Based on users' historical reading speed, average dwell time on similar pages, knowledge unit text length and difficulty score, features such as dwell time, number of revisits, and number of questions are normalized.

[0101] Generate the learning state vector: This yields the LearningStateVector, which includes at least:

[0102] current_unit_id, normalized_dwell_time, revisit_count, scroll_pause_count, highlight_density, note_density, question_density, explanation_click_count, quiz_error_rate, historical_mastery_score, unit_importance_score, unit_difficulty_score and last_intervention_result.

[0103] Status Update: When a user accepts a prompt, ignores a prompt, continues to ask follow-up questions, or completes a quiz, the LearningStateVector is updated and written to the user's learning profile library and mastery record library.

[0104] Optionally, feature analysis is performed on the reading behavior events to obtain learning state features, including:

[0105] The system retrieves the viewing duration, highlighted text length, note length, question length, number of clicks on the explanation, exercise error rate, and current reading content from the reading behavior event.

[0106] The viewing duration, highlighted text length, note length, question length, number of analysis clicks, and exercise error rate are normalized to obtain the viewing duration vector, highlighted text vector, note vector, question vector, analysis click vector, and exercise error vector.

[0107] Feature aggregation is performed on the current page, current paragraph, current knowledge unit, current chapter, and current learning session in the current reading content to obtain a reading vector;

[0108] The learning state features are obtained by combining the viewing dwell vector, the highlighted text vector, the note vector, the question vector, the parsing click vector, the exercise error vector, and the reading vector.

[0109] Step S20: Determine the posterior probability of the checkpoint based on the learning state features, and identify the checkpoint type of the current checkpoint when the posterior probability of the checkpoint meets the preset triggering condition;

[0110] Among them, based on the temporal hidden state model and the Noisy-OR observation fusion model, the posterior probability p_stuck(t) of the user's block point in the current knowledge unit is calculated, and the block point type is further identified after the trigger threshold is reached.

[0111] Optionally, determining the posterior probability of the checkpoint based on the learning state features includes:

[0112] The expected dwell time is determined based on the text length of the current knowledge unit, the user's historical reading speed, and the difficulty of the unit knowledge. The dwell anomaly is determined based on the viewing dwell time and the expected dwell time. The dwell evidence strength is determined based on the dwell anomaly.

[0113] Obtain the number of times a user rewatches a video, and determine the rewatch intensity based on the number of rewatches;

[0114] The system obtains the user's knowledge question in the current knowledge unit, calculates the semantic similarity between the knowledge question and the current knowledge unit, and determines the intensity of repeated questions based on the semantic similarity.

[0115] The strength of the explanatory dependency is determined based on the number of clicks, the user's historical mastery of the current knowledge unit is obtained, and the mastery gap is determined based on the historical mastery.

[0116] The knowledge difficulty and conceptual confusion of the current knowledge unit are obtained, and the observation likelihood is determined based on the strength of evidence of dwell time, the intensity of revisiting, the intensity of repeated questioning, the intensity of explanation dependence, the mastery gap, the knowledge difficulty, and the conceptual confusion.

[0117] The prior probability is determined based on the probability of being stuck in the previous time window, the difficulty of knowledge, the gap in mastery, and the degree of conceptual confusion. The posterior probability of being stuck is then determined based on the prior probability and the observation likelihood.

[0118] Furthermore, the formulas used to determine the prior probability based on the probability of getting stuck in the previous time window, the difficulty of the knowledge, the gap in mastery, and the degree of conceptual confusion include:

[0119] π_t =P_stuck(t1)(1-μ)+(1-p_stuck(t1))•[1-(1-β0)(1-βg g_t)(1-βm m_t)(1-βc c_t)(1-βi i_t)]

[0120] Wherein, π_t represents the prior probability, β0, βg, βm, βc, and βi represent preset prior triggering parameters, p_stuck(t1) represents the probability of being stuck in the previous time window t1, μ represents the state retention parameter, g_t represents the knowledge difficulty, m_t represents the mastery gap, c_t represents the concept confusion degree, and i_t represents the knowledge importance.

[0121] Furthermore, determining the posterior probability of the checkpoint based on the prior probability and the observation likelihood includes:

[0122] p_stuck(t)=π_t•L1_t / [π_t•L1_t+(1-π_t)•L0_t]

[0123] L1_t=ε+(1-ε)[1-∏_{s∈S_t}(1-ρ_s•s)]

[0124] L0_t=ε+(1-ε)∏_{s∈S_t}(1-ρ_s·s)

[0125] Where p_stuck(t) represents the posterior probability of the checkpoint, L1_t represents the strong likelihood, L0_t represents the weak likelihood, ε is the smoothing term, ρ_s is the confidence weight of different observation evidence, S_t represents the set of evidence before the checkpoint, s represents the evidence in the set of evidence before the checkpoint, and ∏ represents the multiplication operation.

[0126] The probability of a user getting stuck is calculated using a temporal hidden state model and a Noisy-OR observation fusion model. For the state of users, documents, and knowledge units within a time window t, a hidden variable Z_t∈{0,1} is defined, where Z_t=1 indicates that a user has gotten stuck, and Z_t=0 indicates that no user has gotten stuck; the output is P_stuck(t)=P(Z_t=1|O_1:t).

[0127] The features are defined as follows: d_t represents the dwell anomaly score, d_t=max(0,T_real / T_exp-1), where T_real is the actual dwell time of the user in the current knowledge unit, and T_exp is the expected dwell time estimated based on text length, user's historical reading speed, and knowledge difficulty; s_d=1-exp(-d_t / τ_d) is the strength of evidence for dwell anomaly. r_t represents the revisit intensity, r_t=1-exp(-N_back / λ_r), where N_back is the number of times the user jumps back to the previous text or page from the current knowledge unit. q_t represents the repeated question intensity, q_t=1-exp(-N_simq / λ_q), where N_simq is the number of questions semantically similar to the current knowledge unit within the time window. e_t represents the explanation dependency intensity, e_t=1-exp(-N_explain / λ_e), where N_explain is the number of times the user clicks on concept explanations, formula explanations, or example explanations. h_t represents the density of highlighted notes, h_t=min(1,(highlight_chars+note_chars) / (unit_tokens+1)). a_t represents the test error rate, a_t=wrong_count / max(1,answer_count). m_t represents the mastery gap, m_t=1-mastery_score, where mastery_score is the user's historical mastery of the knowledge unit. g_t represents the knowledge difficulty, given by difficulty_score output by DifficultyModel. i_t represents the knowledge importance, given by importance_score output by ImportanceGraphModel. c_t represents the conceptual confusion level; when the user's question intent is to differentiate, compare, or distinguish similar concepts, c_t takes the maximum semantic similarity between the current knowledge unit and easily confused knowledge units; otherwise, it takes 0 or a smaller value. b_t represents the scrolling / jumping oscillation intensity, b_t=1-exp(-N_osc / λ_b), where N_osc is the number of scrolling back and forth or page navigations in a short period of time.

[0128] The state transition is calculated as follows:

[0129] π_t =P_stuck(t1)(1-μ)+(1-p_stuck(t1))•[1-(1-β0)(1-βg g_t)(1-βm m_t)(1-βc c_t)(1-βi i_t)]

[0130] Where μ represents the probability of the stuck state being resolved naturally, and β0, βg, βm, βc, and βi are prior triggering parameters. This formula indicates that when the difficulty, mastery gap, conceptual confusion, and importance are high, the probability of the user transitioning from a previously unstuck state to a stuck state increases.

[0131] The observed likelihood is calculated as follows:

[0132] Let the evidence set S_t={s_d,r_t,q_t,e_t,h_t,a_t,m_t,g_t,c_t,b_t}.

[0133] When Z_t=1, the strong likelihood of observing at least one type of strong checkpoint evidence is:

[0134] L1_t=ε+(1-ε)[1-∏_{s∈S_t}(1-ρ_s•s)]

[0135] When Z_t=0, the weak likelihood of the overall weak observational evidence is:

[0136] L0_t=ε+(1-ε)∏_{s∈S_t}(1-ρ_s·s)

[0137] Where ρ_s is the confidence weight of different observational evidence, and ε is the smoothing term.

[0138] The final posterior probability of the checkpoint is:

[0139] p_stuck(t)=π_t•L1_t / [π_t•L1_t+(1-π_t)•L0_t]

[0140] This model calculates the probability of checkpoints through hidden state recursion, Noisy-OR evidence fusion, and Bayesian updates. It can handle situations where multiple weak pieces of evidence coexist and can also maintain continuity by utilizing the state of the previous time window. It does not belong to a simple linear superposition model.

[0141] Preferably, the checkpoint type for identifying the current checkpoint includes:

[0142] The intent of the knowledge question is classified to obtain the question intent, and the unit type of the current knowledge unit is obtained;

[0143] When the degree of abnormality of the stay is greater than the degree of abnormality threshold, and the intention of the question is the first preset intention, then the type of the stuck point is determined to be conceptual incomprehension.

[0144] When the review intensity is greater than the intensity threshold and / or the mastery gap is greater than the gap threshold, the type of the bottleneck is determined to be the prior knowledge deficiency type.

[0145] When the unit type is a preset type and the knowledge question contains a first preset word, the obstacle type is determined to be either the derivation process difficulty type or the formula symbol difficulty type.

[0146] If the knowledge question contains a second preset word, then the type of obstacle is determined to be conceptual confusion.

[0147] Types of obstacles include, but are not limited to: lack of understanding of concepts, lack of prior knowledge, difficulty in derivation, difficulty with formula symbols, confusion of concepts, unclear application scenarios, and weak memory. Differentiated explanation methods can be selected based on the type of obstacle, such as concept comparison tables, variable symbol tables, step-by-step derivations, simple analogies, case demonstrations, or mini-quizzes.

[0148] Checkpoint trigger judgment:

[0149] The posterior probability of a user getting stuck is calculated over a time window t: P_stuck(t) = P(Z_t=1|O_1:t), where Z_t∈{0,1} represents the hidden learning state, Z_t=1 indicates that the user is stuck in the current knowledge unit, and O_t represents the observed reading behavior features within the current time window. A temporal hidden state model combined with a Noisy-OR observation fusion model is used for calculation: first, the prior probability π_t is calculated based on the probability of getting stuck in the previous time window, the knowledge difficulty, the mastery gap, and the degree of conceptual confusion; then, the observation likelihood is calculated based on observational evidence such as abnormal pauses, revisiting, repeated questions, explanation clicks, underlined notes, and quiz errors; finally, P_stuck(t) is obtained through Bayesian update. When P_stuck(t) is higher than a preset threshold, the user enters the stuck type identification process; if the threshold is not reached, only the state is recorded and observation continues.

[0150] Question Intent Recognition: If a user has already asked a question, the question_text is categorized by intent, including categories such as "seeking definition", "seeking distinction", "seeking reason", "seeking derivation", "seeking examples", "seeking explanation of formula symbols", and "seeking application scenarios".

[0151] Knowledge Unit Type Reading: Reads the unit_type of the current KnowledgeUnit, such as concept, definition, formula, example, diagram, conclusion, or point of confusion.

[0152] Preliminary rule assessment: When the dwell time is unusually high and the user's intention is to seek a definition, the initial assessment is that the user does not understand the concept; when the user repeatedly jumps to previous knowledge units or has a low grasp of the prerequisite knowledge, the initial assessment is that the user lacks the prerequisite knowledge; when the current knowledge unit is a formula or derivation and the user's question contains words such as "why," "how to calculate," or "derive," the initial assessment is that the user has difficulty with the derivation process or the formula symbols; when the question contains expressions such as "difference," "are they the same," or "compared to," the initial assessment is that the user is confused about the concept.

[0153] Verification: Input the LearningStateVector, KnowledgeUnit attributes, question intent, historical knowledge level, and contextual text into the checkpoint type classifier, and output the probability of each checkpoint type.

[0154] Fusion decision: The initial judgment result of the rule is fused with the classifier probability in a weighted manner, and the checkpoint type with the highest probability that exceeds the type threshold is selected; if the probability of each type is lower than the threshold, the general incomprehensible type is output.

[0155] Intervention forms of matching: matching for those who do not understand the concept (using colloquial explanations and restates of definitions); matching for those who lack prior knowledge (reviewing prior knowledge); matching for those who have difficulty in the derivation process (step-by-step derivation); matching for those who have difficulty with formula symbols (using symbol tables); matching for those who have confusion about the concept (using comparison tables); matching for those who have unclear application scenarios (using case demonstrations).

[0156] Step S30: Obtain the current learning companion mode, and provide active learning prompts based on the posterior probability of the stuck point, the current learning companion mode, the current stuck point, and the stuck point type;

[0157] Specifically, when the probability of getting stuck meets the triggering conditions, proactive learning prompts are generated based on the user's set companion mode, disturbance control rules, and current learning context.

[0158] Optionally, active learning prompts are provided based on the posterior probability of the checkpoint, the current learning support mode, the current checkpoint, and the checkpoint type, including:

[0159] The first probability threshold, the second probability threshold, and the third probability threshold are determined based on the current learning companion mode, and learning prompt information is determined based on the current checkpoint and the checkpoint type.

[0160] If the posterior probability of the checkpoint is less than the first probability threshold, then the prompting stops;

[0161] If the posterior probability of the checkpoint is greater than or equal to the first probability threshold and less than the second probability threshold, then an information prompt port is displayed at the current checkpoint.

[0162] If a confirmation instruction is received for the information prompt port, the learning prompt information is displayed;

[0163] If the posterior probability of the checkpoint is greater than or equal to the second probability threshold and less than the third probability threshold, then the checkpoint type and explanation button are displayed on the current checkpoint.

[0164] If a confirmation instruction is received for the explanation button, the learning prompt information is displayed;

[0165] If the posterior probability of the stuck point is greater than or equal to the third probability threshold, the learning prompt information is directly displayed.

[0166] For example, when p_stuck(t)<T1, no active prompt is provided; when T1≤p_stuck(t)<T2, a lightweight prompt entry is displayed; when T2≤p_stuck(t)<T3, a stuck point judgment statement and an explanation button are displayed; when P_stuck(t)≥T3, a step-by-step explanation, a review of prerequisite knowledge or a micro-quiz is displayed. In one embodiment, T1=0.45, T2=0.65, and T3=0.78, and the thresholds can be dynamically adjusted according to the silent mode, standard mode or active companionship mode selected by the user; if the user continuously ignores the prompts, T1, T2 and T3 are increased or ρ_s is decreased in subsequent windows.

[0167] In this embodiment, the document semantic structure is bound to the coordinates of the reading interface: chapters, paragraphs, sentences, formulas, charts, concepts and difficult points are bound to page coordinates or text anchors, so that the AI explanation can be directly superimposed on or traced back to the position of the original text.

[0168] Reading behavior event stream collection and learning state modeling: events such as stay, re-reading, underlining, clicking for explanation, questioning, and answering questions are collected to form knowledge point-level behavior features.

[0169] Stuck point recognition based on abnormal behavior and semantic difficulty: the stuck point probability is calculated by integrating stay abnormality, number of re-readings, question density, explanation click rate, test error rate, knowledge point difficulty and historical mastery degree.

[0170] Hierarchical triggering mechanism for active prompting: according to the stuck point probability, user companionship mode, cooldown time and historical feedback, a lightweight prompt, a one-sentence prompt, a step-by-step explanation, a review of prerequisite knowledge or a micro-quiz is selected.

[0171] Explanation generation based on original text evidence: AI explanations are bound to original text fragments, chapter positions and knowledge units, and original text basis and supplementary explanations are distinguished.

[0172] Post-reading mastery update and review closed loop: the mastery degree is updated according to reading behaviors, question and answer records, active prompt feedback and test results, and notes, wrong questions, knowledge cards and review plans are generated.

[0173] For example, a user reads the page where "marginal cost and average cost" is located in an economics textbook. "Marginal cost", "average cost" and "cost curve" have been recognized as knowledge units, and "marginal cost" is marked with an importance of 0.91 and a difficulty of 0.72.

[0174] The user stayed on this page for 240 seconds, while their average time spent on pages of similar length in the past was 90 seconds; the user looked back at the previous page 3 times, clicked on the explanation of "marginal cost" 2 times, and submitted the question "Why are marginal cost and average cost different?" These events can be attributed to the two knowledge units of "marginal cost" and "average cost".

[0175] Generate LearningStateVector, where normalized_dwell_time=2.67, revisit_count=3, question_density=1, explanation_click_count=2, historical_mastery_score=0.35, unit_difficulty_score=0.72, unit_importance_score=0.91.

[0176] Based on the feature definition, we can obtain:

[0177] d_t = max(0, 2.67 - 1) = 1.67;

[0178] s_d=1-exp(-1.67)≈0.81;

[0179] r_t=1-exp(-3 / 3)≈0.63;

[0180] q_t = 1 - exp(-1 / 1) ≈ 0.63;

[0181] e_t=1-exp(-2 / 2)≈0.63;

[0182] m_t = 1 - 0.35 = 0.65;

[0183] g_t=0.72, c_t=0.80 because the question includes "why are they different" and there is a confusing relationship between the two knowledge units.

[0184] Assuming the probability of being stuck in the previous time window is P_stuck(t1)=0.30, the state holding parameter is μ=0.20, and the prior triggering parameters are β0=0.05, βg=0.35, βm=0.30, and βc=0.25, then π_t =P_stuck(t1)(1-μ)+(1-p_stuck(t1))[1-(1-β0)(1-βg g_t)(1-βm m_t)(1-βc c_t)]≈0.62.

[0185] Then, inputting s_d, r_t, q_t, e_t, m_t, g_t, and c_t into the Noisy-OR observation model yields L1_t≈0.75 and L0_t≈0.30.

[0186] Finally, p_stuck(t) = π_tL1_t / (π_tL1_t + (1-π_t)L0_t) ≈ 0.81, which is higher than the level 3 prompt threshold T3 = 0.78, thus triggering the level 3 active prompt.

[0187] The user's question included "Why are they different?" and there were easily confused relationships between the current knowledge units. The checkpoint type classifier output: concept confusion type 0.76, concept incomprehension type 0.18, and prior knowledge deficiency type 0.06. Therefore, it was identified as a concept confusion type checkpoint.

[0188] The proactive intervention decision-making module reads that the user is in "Standard Companion Mode" and that no prompts have been triggered on this page within the past 5 minutes. Therefore, it displays a simple prompt on the right side of the reading interface: "Marginal cost and average cost are easily confused here. Would you like to see a comparative explanation?"

[0189] After the user clicks, the AI ​​explanation generation module retrieves the definition sentences, related charts, and examples of the two concepts from the textbook based on the anchor points in the original text, generating an explanation card that includes "a one-sentence distinction, a comparison table, numerical examples, a link to the original source, and a mini-quiz." Each conclusion in the explanation card is associated with a source_span, and the user can click to return to the original text.

[0190] If a user answers the quiz correctly, the mastery level of "marginal cost" and "average cost" will be updated from 0.35 to 0.58, and the frequency of similar prompts will be reduced; if the user answers incorrectly, a prior knowledge review card will be generated, and the relevant wrong questions will be added to the review plan.

[0191] Active learning tips include concept explanations, easy-to-understand analogies, example demonstrations, prior knowledge reviews, clues for locating information in the original text, step-by-step derivations, mini-quizzes, or suggestions for later review.

[0192] The system records user feedback, such as acceptance of proactive prompts, ignoring of prompts, follow-up questions, and answer results, back into the user's learning profile to update knowledge mastery and subsequent prompting strategies.

[0193] In this embodiment, by performing feature analysis on reading behavior events, the learning state characteristics of the user's current learning state can be effectively obtained. Based on the learning state characteristics, the posterior probability of getting stuck can be effectively determined. Based on the posterior probability of getting stuck, it can be effectively determined whether the user is currently experiencing a reading stuck phenomenon. When the posterior probability of getting stuck meets the preset triggering conditions, based on the posterior probability of getting stuck, the current learning companion mode, the current stuck point, and the stuck point type, the user can be automatically given proactive learning prompts, which improves the user experience. It can upgrade static document Q&A to process-aware learning companionship, discover potential learning stuck points when the user does not actively ask questions, improve the traceability of explanations through original text anchors, reduce disturbances through graded prompts, and form a closed loop of reading, Q&A, testing, and review through mastery.

[0194] Example 2

[0195] Please see Figure 2 This is a schematic diagram of the structure of the intelligent learning companion system 100 provided in the second embodiment of the present invention, including:

[0196] The feature analysis module 10 is used to acquire reading behavior events of users when learning to read, and to perform feature analysis on the reading behavior events to obtain learning state features.

[0197] Optionally, the feature analysis module 10 is also used to: obtain the viewing duration, highlighted text length, note length, question length, number of parsing clicks, exercise error rate, and current reading content in the reading behavior event;

[0198] The viewing duration, highlighted text length, note length, question length, number of analysis clicks, and exercise error rate are normalized to obtain the viewing duration vector, highlighted text vector, note vector, question vector, analysis click vector, and exercise error vector.

[0199] Feature aggregation is performed on the current page, current paragraph, current knowledge unit, current chapter, and current learning session in the current reading content to obtain a reading vector;

[0200] The learning state features are obtained by combining the viewing dwell vector, the highlighted text vector, the note vector, the question vector, the parsing click vector, the exercise error vector, and the reading vector.

[0201] The checkpoint verification module 11 is used to determine the posterior probability of the checkpoint based on the learning state features, and to identify the checkpoint type of the current checkpoint when the posterior probability of the checkpoint meets the preset triggering conditions.

[0202] Optionally, the checkpoint verification module 11 is further configured to: determine the expected dwell time based on the text length of the current knowledge unit, the user's historical reading speed, and the difficulty of the unit knowledge; determine the dwell anomaly degree based on the viewing dwell time and the expected dwell time; and determine the dwell evidence strength based on the dwell anomaly degree.

[0203] Obtain the number of times a user rewatches a video, and determine the rewatch intensity based on the number of rewatches;

[0204] The system obtains the user's knowledge question in the current knowledge unit, calculates the semantic similarity between the knowledge question and the current knowledge unit, and determines the intensity of repeated questions based on the semantic similarity.

[0205] The strength of the explanatory dependency is determined based on the number of clicks, the user's historical mastery of the current knowledge unit is obtained, and the mastery gap is determined based on the historical mastery.

[0206] The knowledge difficulty and conceptual confusion of the current knowledge unit are obtained, and the observation likelihood is determined based on the strength of evidence of dwell time, the intensity of revisiting, the intensity of repeated questioning, the intensity of explanation dependence, the mastery gap, the knowledge difficulty, and the conceptual confusion.

[0207] The prior probability is determined based on the probability of being stuck in the previous time window, the difficulty of knowledge, the gap in mastery, and the degree of conceptual confusion. The posterior probability of being stuck is then determined based on the prior probability and the observation likelihood.

[0208] Furthermore, the checkpoint verification module 11 is also used to: classify the intent of the knowledge question, obtain the question intent, and obtain the unit type of the current knowledge unit;

[0209] When the degree of abnormality of the stay is greater than the degree of abnormality threshold, and the intention of the question is the first preset intention, then the type of the stuck point is determined to be conceptual incomprehension.

[0210] When the review intensity is greater than the intensity threshold and / or the mastery gap is greater than the gap threshold, the type of the bottleneck is determined to be the prior knowledge deficiency type.

[0211] When the unit type is a preset type and the knowledge question contains a first preset word, the obstacle type is determined to be either the derivation process difficulty type or the formula symbol difficulty type.

[0212] If the knowledge question contains a second preset word, then the type of obstacle is determined to be conceptual confusion.

[0213] The learning prompt module 12 is used to obtain the current learning companion mode and provide active learning prompts based on the posterior probability of the stuck point, the current learning companion mode, the current stuck point, and the stuck point type.

[0214] Optionally, the learning prompt module 12 is further configured to: determine a first probability threshold, a second probability threshold, and a third probability threshold based on the current learning companion mode, and determine learning prompt information based on the current checkpoint and the checkpoint type;

[0215] If the posterior probability of the checkpoint is less than the first probability threshold, then the prompting stops;

[0216] If the posterior probability of the checkpoint is greater than or equal to the first probability threshold and less than the second probability threshold, then an information prompt port is displayed at the current checkpoint.

[0217] If a confirmation instruction is received for the information prompt port, the learning prompt information is displayed;

[0218] If the posterior probability of the checkpoint is greater than or equal to the second probability threshold and less than the third probability threshold, then the checkpoint type and explanation button are displayed on the current checkpoint.

[0219] If a confirmation instruction is received for the explanation button, the learning prompt information is displayed;

[0220] If the posterior probability of the checkpoint is greater than or equal to the third probability threshold, the learning prompt information is displayed directly.

[0221] Optionally, in another embodiment, the intelligent learning companion system 100 includes:

[0222] The document import module is used to receive e-books, textbooks, papers, lecture notes, or web page materials uploaded by users.

[0223] The document parsing module is used to parse the table of contents, chapters, page numbers, paragraphs, charts, formulas, heading levels, and page coordinates.

[0224] The semantic structure modeling module is used to identify concepts, definitions, examples, formulas, difficulties, key points, prerequisite knowledge, and relationships between knowledge points.

[0225] The original text anchor module is used to bind knowledge points, explanations, and AI outputs to original text fragments, page areas, or coordinate positions.

[0226] The reading behavior collection module is used to collect events such as pauses, page turning, scrolling, rewinding, underlining, note-taking, clicking, asking questions, and answering questions during the user's reading process.

[0227] The learning state modeling module is used to convert reading behavior events into learning state characteristics of users on specific knowledge points.

[0228] The obstacle recognition module is used to calculate the probability of a user having difficulty understanding a particular page, paragraph, or knowledge point, and to identify the type of obstacle.

[0229] The proactive intervention decision-making module is used to determine whether to proactively prompt users and the form of prompts based on factors such as the probability of being blocked, user preferences, learning objectives, disturbance thresholds, and historical feedback.

[0230] The AI ​​explanation generation module is used to generate explanations, analogies, examples, derivations, quizzes, and review suggestions based on original evidence and user context.

[0231] The mastery update module is used to update the user's mastery of knowledge points based on reading behavior, question-and-answer behavior, and test results.

[0232] The note and review generation module is used to generate chapter summaries, Cornell notes, knowledge cards, error collections, and review plans.

[0233] The user interface module is used to display key annotations, concept entries, AI explanation cards, proactive prompts, Q&A areas, and learning progress on the original e-book text.

[0234] In this embodiment, by performing feature analysis on reading behavior events, the learning state characteristics of the user's current learning state can be effectively obtained. Based on the learning state characteristics, the posterior probability of getting stuck can be effectively determined. Based on the posterior probability of getting stuck, it can be effectively determined whether the user is currently experiencing a reading stuck phenomenon. When the posterior probability of getting stuck meets the preset triggering conditions, based on the posterior probability of getting stuck, the current learning companion mode, the current stuck point, and the stuck point type, the user can be automatically given proactive learning prompts, thereby improving the user experience.

[0235] Example 3

[0236] Figure 3 This is a structural block diagram of a terminal device 2 provided in the third embodiment of this application. For example... Figure 3 As shown, the terminal device 2 in this embodiment includes: a processor 20, a memory 21, and a computer program 22 stored in the memory 21 and executable on the processor 20, such as a program for an intelligent learning companion method. When the processor 20 executes the computer program 22, it implements the steps in the various embodiments of the intelligent learning companion methods described above.

[0237] For example, the computer program 22 may be divided into one or more modules, which are stored in the memory 21 and executed by the processor 20 to complete this application. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 22 in the terminal device 2. The terminal device may include, but is not limited to, the processor 20 and the memory 21.

[0238] The processor 20 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0239] The memory 21 can be an internal storage unit of the terminal device 2, such as a hard drive or memory of the terminal device 2. The memory 21 can also be an external storage device of the terminal device 2, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the terminal device 2. Furthermore, the memory 21 can include both internal and external storage units of the terminal device 2. The memory 21 is used to store the computer program and other programs and data required by the terminal device. The memory 21 can also be used to temporarily store data that has been output or will be output.

[0240] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0241] If an integrated module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. This computer-readable storage medium can be non-volatile or volatile. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the contents of a computer-readable storage medium may be appropriately added to or subtracted from the contents as required by the legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, a computer-readable storage medium may not include electrical carrier signals and telecommunication signals.

[0242] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for intelligent learning companionship, characterized in that, The method includes: The system acquires reading behavior events of users during their learning and reading process, and performs feature analysis on these reading behavior events to obtain learning state features. The posterior probability of the checkpoint is determined based on the learning state characteristics, and the checkpoint type of the current checkpoint is identified when the posterior probability of the checkpoint meets the preset triggering condition. Obtain the current learning support mode and provide proactive learning prompts based on the posterior probability of the stuck point, the current learning support mode, the current stuck point, and the stuck point type.

2. The intelligent learning companion method as described in claim 1, characterized in that, Feature analysis is performed on the reading behavior events to obtain learning state features, including: The system retrieves the viewing duration, highlighted text length, note length, question length, number of clicks on the explanation, exercise error rate, and current reading content from the reading behavior event. The viewing duration, highlighted text length, note length, question length, number of analysis clicks, and exercise error rate are normalized to obtain the viewing duration vector, highlighted text vector, note vector, question vector, analysis click vector, and exercise error vector. Feature aggregation is performed on the current page, current paragraph, current knowledge unit, current chapter, and current learning session in the current reading content to obtain a reading vector; The learning state features are obtained by combining the viewing dwell vector, the highlighted text vector, the note vector, the question vector, the parsing click vector, the exercise error vector, and the reading vector.

3. The intelligent learning companion method as described in claim 2, characterized in that, Determining the posterior probability of the checkpoint based on the learning state features includes: The expected dwell time is determined based on the text length of the current knowledge unit, the user's historical reading speed, and the difficulty of the unit knowledge. The dwell anomaly is determined based on the viewing dwell time and the expected dwell time. The dwell evidence strength is determined based on the dwell anomaly. Obtain the number of times a user rewatches a video, and determine the rewatch intensity based on the number of rewatches; The system obtains the user's knowledge question in the current knowledge unit, calculates the semantic similarity between the knowledge question and the current knowledge unit, and determines the intensity of repeated questions based on the semantic similarity. The strength of the explanatory dependency is determined based on the number of clicks, the user's historical mastery of the current knowledge unit is obtained, and the mastery gap is determined based on the historical mastery. The knowledge difficulty and conceptual confusion of the current knowledge unit are obtained, and the observation likelihood is determined based on the strength of evidence of dwell time, the intensity of revisiting, the intensity of repeated questioning, the intensity of explanation dependence, the mastery gap, the knowledge difficulty, and the conceptual confusion. The prior probability is determined based on the probability of being stuck in the previous time window, the difficulty of knowledge, the gap in mastery, and the degree of conceptual confusion. The posterior probability of being stuck is then determined based on the prior probability and the observation likelihood.

4. The intelligent learning companion method as described in claim 3, characterized in that, The formulas used to determine the prior probability based on the probability of getting stuck in the previous time window, the difficulty of the knowledge, the gap in mastery, and the degree of conceptual confusion include: π_t =P_stuck(t1)(1-μ)+(1-p_stuck(t1))•[1-(1-β0)(1-βg g_t)(1-βm m_t)(1-βcc_t)(1-βi i_t)] Wherein, π_t represents the prior probability, β0, βg, βm, βc, and βi represent preset prior triggering parameters, p_stuck(t1) represents the probability of being stuck in the previous time window t1, μ represents the state retention parameter, g_t represents the knowledge difficulty, m_t represents the mastery gap, c_t represents the concept confusion degree, and i_t represents the knowledge importance.

5. The intelligent learning companion method as described in claim 4, characterized in that, Determining the posterior probability of the checkpoint based on the prior probability and the observed likelihood includes: p_stuck(t)=π_t•L1_t / [π_t•L1_t+(1-π_t)•L0_t] L1_t=ε+(1-ε)[1-∏_{s∈S_t}(1-ρ_s•s)] L0_t=ε+(1-ε)∏_{s∈S_t}(1-ρ_s·s) Where p_stuck(t) represents the posterior probability of the checkpoint, L1_t represents the strong likelihood, L0_t represents the weak likelihood, ε is the smoothing term, ρ_s is the confidence weight of different observation evidence, S_t represents the set of evidence in the set of evidence, and s represents the evidence in the set of evidence in the set of evidence.

6. The intelligent learning companion method as described in claim 3, characterized in that, Identify the type of checkpoint at the current checkpoint, including: The intent of the knowledge question is classified to obtain the question intent, and the unit type of the current knowledge unit is obtained; When the degree of abnormality of the stay is greater than the degree of abnormality threshold, and the intention of the question is the first preset intention, then the type of the stuck point is determined to be conceptual incomprehension. When the review intensity is greater than the intensity threshold and / or the mastery gap is greater than the gap threshold, the type of the bottleneck is determined to be the prior knowledge deficiency type. When the unit type is a preset type and the knowledge question contains a first preset word, the obstacle type is determined to be either the derivation process difficulty type or the formula symbol difficulty type. If the knowledge question contains a second preset word, then the type of obstacle is determined to be conceptual confusion.

7. The intelligent learning companion method as described in claim 1, characterized in that, Active learning prompts are provided based on the posterior probability of the checkpoint, the current learning support mode, the current checkpoint, and the checkpoint type, including: The first probability threshold, the second probability threshold, and the third probability threshold are determined based on the current learning companion mode, and learning prompt information is determined based on the current checkpoint and the checkpoint type. If the posterior probability of the checkpoint is less than the first probability threshold, then the prompting stops; If the posterior probability of the checkpoint is greater than or equal to the first probability threshold and less than the second probability threshold, then an information prompt port is displayed at the current checkpoint. If a confirmation instruction is received for the information prompt port, the learning prompt information is displayed; If the posterior probability of the checkpoint is greater than or equal to the second probability threshold and less than the third probability threshold, then the checkpoint type and explanation button are displayed on the current checkpoint. If a confirmation instruction is received for the explanation button, the learning prompt information is displayed; If the posterior probability of the checkpoint is greater than or equal to the third probability threshold, the learning prompt information is displayed directly.

8. An intelligent learning companion system, characterized in that, The system includes: The feature analysis module is used to acquire reading behavior events of users when learning to read, and to perform feature analysis on the reading behavior events to obtain learning state features; The checkpoint verification module is used to determine the posterior probability of the checkpoint based on the learning state features, and to identify the checkpoint type of the current checkpoint when the posterior probability of the checkpoint meets the preset triggering conditions. The learning prompt module is used to obtain the current learning companion mode and provide proactive learning prompts based on the posterior probability of the stuck point, the current learning companion mode, the current stuck point, and the stuck point type.

9. The intelligent learning companion system as described in claim 8, characterized in that, The feature analysis module is also used to: obtain the viewing duration, highlighted text length, note length, question length, number of clicks on the analysis, exercise error rate, and current reading content in the reading behavior event; The viewing duration, highlighted text length, note length, question length, number of analysis clicks, and exercise error rate are normalized to obtain the viewing duration vector, highlighted text vector, note vector, question vector, analysis click vector, and exercise error vector. Feature aggregation is performed on the current page, current paragraph, current knowledge unit, current chapter, and current learning session in the current reading content to obtain a reading vector; The learning state features are obtained by combining the viewing dwell vector, the highlighted text vector, the note vector, the question vector, the parsing click vector, the exercise error vector, and the reading vector.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images