Knowledge graph generation method and system for english teaching and storage medium

Abstract

The application discloses a knowledge graph generation method and system for English teaching and a storage medium, relates to the technical field of education informatization, and solves the problems of low accuracy of professional term extraction, incapability of constructing a hierarchical ability graph conforming to CEFR and low accuracy of entity relation recognition in the English education field during English teaching. The knowledge graph generation method comprises the following steps: constructing a subject knowledge dataset according to all English textbook teaching materials and teaching aid materials; cleaning, classifying and grading the subject knowledge dataset to generate a plurality of comprehensive subject content libraries; adopting an entity explicit-implicit correlation mechanism to structure all the subject content libraries to obtain a teaching field content table; adopting a dynamic knowledge fusion model in combination with a teaching field mode to construct the teaching field content table into an initial English teaching graph; and based on the training test feedback of students and in combination with an expert prior mechanism, optimizing and updating the initial English teaching graph to obtain an optimal English teaching graph.

Description

Technical Field

[0001] The present invention relates to the field of educational informatization technology, and in particular to a knowledge graph generation method, system, and storage medium for English teaching. Background Art

[0002] In recent years, the informatization of education has achieved remarkable results, and the deep integration of information technology with teaching and learning has become a key direction of educational reform. High school English instruction is also gradually becoming informatized, leveraging modern educational technology to improve teaching quality. Traditional English courses rely on a linear textbook structure, making it difficult to present the networked connections between vocabulary, grammar, and sentences. For example, students cannot intuitively understand the 15 possible semantic evolution paths of the word "get" in different contexts. Static knowledge point tracking models (such as IRT / BKT) assume that learners will not forget, resulting in an assessment bias rate of over 37%.

[0003] Knowledge graphs, derived from semantic web technology, are a method for structured representation of real-world entities, concepts, attributes, and relationships. In recent years, knowledge graphs have achieved remarkable results in fields such as natural language processing, recommendation systems, and search engines, laying the foundation for their application in education.

[0004] The knowledge graph categorizes and organizes English vocabulary according to semantic relationships, helping students systematically learn and memorize vocabulary. Through the knowledge graph, students can clearly see the connections between different words, improving vocabulary learning efficiency. The knowledge graph can structure complex grammatical knowledge, helping students understand and master grammatical rules. Through the knowledge graph, students can better understand the hierarchical relationships between grammatical knowledge, thereby improving grammar learning effectiveness.

[0005] Patent No. CN202410331524.2 discloses an AI school-based English textbook editing method and system, which includes: constructing a student cognitive knowledge graph based on the student's online and offline test data; constructing a student interest knowledge graph based on the student's online learning behavior data; constructing a teacher recommendation knowledge graph based on the teacher's historical teaching trajectory data; designing an AI text material recommendation engine based on the student cognitive knowledge graph, the student interest knowledge graph and the teacher recommendation knowledge graph, calculating the multi-dimensional matching degree of the text material, and automatically selecting teaching texts based on the multi-dimensional matching degree of the text material; based on the selected teaching text, matching and selecting relevant pre-class preview materials, knowledge point explanation courseware and after-class exercises from the digital resource library to form complete textbook content. The above application realizes the automation of textbook editing, significantly improves efficiency and quality, and ensures that the textbook content is consistent with the teaching syllabus and meets the needs of students and teachers.

[0006] Patent No. CN202411800226.X discloses an English course management system for English teaching, including collecting multi-source English teaching data and performing cleaning and feature extraction, and constructing a knowledge graph based on the extracted feature parameters; using feature filtering and weighting formulas based on the knowledge graph to calculate course configuration parameter factors; inputting the course configuration parameter factors into a deep reinforcement learning model for strategy training; outputting the optimal course content and difficulty configuration combination under the current teaching environment and dynamically updating the teaching strategy; the above scheme introduces a combination of knowledge graphs, multi-feature weighting and deep reinforcement learning, which not only accurately captures the complex feature relationships in the English teaching process, but also enables the scheme to continuously evolve through the adaptive optimization capabilities of reinforcement learning, thereby significantly improving the accuracy and flexibility of teaching resource allocation and course difficulty control.

[0007] Although large language models (such as GPT-4 and Llama3) have text parsing capabilities, they have two major flaws when directly applied to English teaching: the accuracy of professional terminology extraction is low, and it is impossible to build a hierarchical ability map that conforms to the CEFR (Common European Framework of Reference for Languages), and the accuracy of entity relationship recognition in the field of English education is low. Summary of the Invention

[0008] The purpose of the present invention is to provide a knowledge graph generation method, system and storage medium for English teaching, which can construct and optimize the learning path and teaching path of English teaching through the knowledge graph combined with a dynamic knowledge fusion model, so as to improve the accuracy of extracting English professional terms and improve the accuracy of entity relationship recognition in the field of English education.

[0009] The present invention utilizes the following technical solutions:

[0010] A knowledge graph generation method for English teaching, comprising the following steps:

[0011] S1: The multimodal acquisition module constructs a subject knowledge dataset based on all English textbooks and supplementary teaching materials;

[0012] S2: The information processing module cleans, categorizes, and grades the subject knowledge dataset to generate several comprehensive subject content libraries;

[0013] S3: The extraction and filling module uses the entity explicit and implicit association mechanism to structure all subject content libraries and obtain the teaching field content table;

[0014] S4: The graph construction module uses a dynamic knowledge fusion model combined with the teaching field model to construct the teaching field content table into an initial English teaching graph;

[0015] S5: The knowledge verification module optimizes and updates the initial English teaching map based on the students' training and test feedback combined with the expert prior mechanism to obtain the optimal English teaching map;

[0016] S6: The teaching analysis module combines training test feedback with learning level to visualize the domain knowledge blind spots of the optimal English teaching map, and at the same time selects teaching path branches in stages to generate a learning evaluation form.

[0017] Preferably, step S1 includes the following steps:

[0018] The multi-mode acquisition module parses all textbooks using optical character recognition rules to extract directory levels, knowledge points, and accompanying exercise sets;

[0019] At the same time, a hybrid character recognition engine is used to perform text recognition and image enhancement on supplementary teaching materials containing text and images to extract supplementary teaching content information;

[0020] At the same time, data crawling is used to crawl teaching materials containing multimedia resources, and record knowledge points and knowledge point timestamps;

[0021] The multi-mode acquisition module combines knowledge points according to the directory hierarchy, associates the supplementary teaching content information, knowledge point timestamps and supporting exercises, and obtains a knowledge point resource table for all versions;

[0022] According to the textbook version and semantic similarity, the knowledge point resource tables of different versions are matched to obtain the knowledge point equivalent mapping table;

[0023] The multi-mode acquisition module generates a unique identifier for a knowledge point based on the textbook version, the knowledge point page number, and an N-digit random number, and uses it as a joint index serial number.

[0024] According to the material type, combined with the joint index serial number and the knowledge point equivalent mapping table, different types of knowledge points are associated and stored in different databases. At the same time, the traceability information is filled in according to the knowledge point timestamp to obtain the subject knowledge dataset;

[0025] The multi-mode acquisition module determines the number of resource types associated with each knowledge point in the subject knowledge dataset and the preset verification threshold:

[0026] If the number of resource types is less than the verification threshold, the current knowledge point is determined to be missing resource items, and the subject knowledge dataset is verified based on the joint index sequence number combined with the directory hierarchy, and the resource items are extracted and filled in based on the traceability information;

[0027] If the number of resource types is greater than or equal to the verification threshold, it is determined that there are no missing items in the current knowledge point, and information processing is performed on the subject knowledge dataset based on the knowledge point.

[0028] Preferably, step S2 includes the following steps:

[0029] The information processing module uses the conditional random field model to perform semantic segmentation on the subject knowledge dataset according to the directory level and the title level to obtain the chapter content knowledge table;

[0030] Using the named entity recognition model combined with the knowledge point equivalent mapping table, all knowledge points in the chapter content knowledge table are annotated, and grammatical points and vocabulary types are marked to obtain the chapter knowledge entity table;

[0031] The information processing module dynamically aligns multimedia resources with textbooks and supplementary teaching materials based on the knowledge point resource table, and links images and texts based on the knowledge point timestamps to obtain a multi-source entity knowledge association table;

[0032] The information processing module uses the isolation forest algorithm to filter noise from the multi-source entity knowledge association table, remove duplicate and irrelevant content, and obtain a noise-free multi-source knowledge association table;

[0033] The information processing module uses the contrast image pre-training model to perform anomaly detection and cross-modal similarity calculation on the noise-free multi-source knowledge association table, and compares it with the preset similarity threshold:

[0034] If the cross-modal similarity is greater than or equal to the similarity threshold, the current association pair is retained and abnormal matching is corrected; if the cross-modal similarity is less than the similarity threshold, the current association pair is synchronized according to the knowledge point timestamp, a timeline label is generated, and a noise-free multi-source knowledge association table is assigned;

[0035] The information processing module performs multimodal analysis on the subject knowledge dataset and the accompanying exercise set based on the noise-free multi-source knowledge association table to obtain linguistic and teaching features;

[0036] A bidirectional sentence embedding model is used to vectorize all knowledge points based on timeline labels, combining linguistic features and teaching features to obtain a multimodal knowledge feature vector.

[0037] A contrastive learning function is used to align all knowledge points according to the multimodal knowledge feature vector to obtain a cross-modal semantic mapping table;

[0038] The information processing module establishes a three-level label extraction branch based on the three-level encoding and the cross-module semantic mapping table;

[0039] A multimodal classification model is used in combination with a three-level label extraction branch and a confusion classification weight matrix to parse and classify knowledge points of all modalities and obtain a multimodal comprehensive weight matrix to complete the multimodal data classification.

[0040] The information processing module uses item response theory combined with teaching sequence and applicable scenarios to classify each modality of information into multiple dimensions. At the same time, it optimizes the cognitive difficulty of all knowledge points based on the answer feedback data and generates a knowledge learning path.

[0041] According to the knowledge learning path, the subject knowledge data set and the corresponding exercise set are clustered and allocated to obtain the micro-knowledge point library and the theme unit library. Combined with the popularity index and quality index of the knowledge points, several comprehensive subject content libraries are constructed.

[0042] Preferably, step S3 includes the following steps:

[0043] The extraction and filling module extracts knowledge point structured entities from the subject content library based on the three-level tags, and binds inherent attributes to the supporting exercise set through regular expressions to obtain auxiliary knowledge entities;

[0044] A graph neural network is used to mine the subject content library based on students' learning behavior, combined with supporting exercise sets and confusion classification weight matrices, to obtain students' behavior-derived entities and context-related entities of knowledge points;

[0045] According to the explicit and implicit entity association mechanism, the knowledge point structured entities and auxiliary knowledge entities are hierarchically associated, and topological sorting is performed in combination with the teaching logic chain to generate the teaching path and complete the explicit association modeling;

[0046] At the same time, cross-modal associations are performed between behavior-derived entities and context-related entities, and association corrections are performed based on the student's error rate to obtain multi-modal mapping association weights and complete implicit association reasoning.

[0047] The extraction and filling module constructs explicit nodes and explicit edges based on the teaching path and the multi-modal knowledge feature vector; at the same time, it constructs implicit nodes and implicit edges based on the multi-modal mapping association weight and the student identity;

[0048] The extraction and filling module automatically extracts the comprehensive subject content library based on the three-level labels, explicit nodes, explicit edges, implicit nodes and implicit edges, and constructs a three-dimensional tensor dynamic table;

[0049] Based on explicit association modeling and implicit association reasoning combined with cross-modal semantic mapping tables, conflict detection and teaching adaptation are performed on the three-dimensional tensor dynamic table to obtain the teaching field content table.

[0050] Preferably, step S4 includes the following steps:

[0051] The extended constraint layer of the dynamic knowledge fusion model uses expression rule language and combines it with network ontology description language to expand and constrain the teaching domain content table to obtain the difficulty-modality mapping library;

[0052] The multi-source incremental layer of the dynamic knowledge fusion model uses a stream processing algorithm and a conflict resolution strategy to perform heterogeneous fusion and knowledge alignment on the subject knowledge dataset and the supporting exercise set, and obtains a super-node aggregated explanation table.

[0053] The explicit embedding layer of the dynamic knowledge fusion model uses a random walk algorithm to combine the three-dimensional tensor dynamic table with the knowledge point centrality based on explicit association modeling to generate a backbone learning path;

[0054] The implicit reasoning layer of the dynamic knowledge fusion model uses a relational graph convolutional network to combine the three-dimensional tensor dynamic table with the weight decay factor based on implicit association reasoning to predict potential teaching paths;

[0055] The graph building layer of the dynamic knowledge fusion model uses a distributed reasoning algorithm to construct a graph of the teaching domain content table based on the main learning path and potential teaching path, combined with the super-node aggregation explanation table, to obtain the initial English teaching graph;

[0056] The path verification layer of the dynamic knowledge fusion model uses betweenness centrality and the difficulty-modality mapping library to thin out the initial English teaching map, and then marks and archives all paths in the initial English teaching map.

[0057] Preferably, step S5 includes the following steps:

[0058] The knowledge verification module tracks students' learning trajectory based on the initial English teaching map, recording the length of time spent on knowledge points, the number of times wrong questions are redone, and resource click preferences;

[0059] A natural language network is used to capture the emotional signals of the teaching content table in combination with the authoritative standard mapping, and the knowledge emotional tendency and constraint node attributes are obtained;

[0060] The knowledge verification module uses a clustering algorithm based on the length of time spent on a knowledge point and the number of times a wrong question is redone to cluster the wrong question records in the accompanying exercise book and identify high-frequency error points.

[0061] Bayesian network is used to associate knowledge points with constraint node attributes based on high-frequency error points to obtain error-knowledge point association identifiers;

[0062] The node colors of the initial English teaching map are distributed and marked according to the error rate of knowledge points. At the same time, the high-frequency error points are traced back to determine the prerequisite knowledge points.

[0063] The nodes of the initial English teaching map are added and deleted based on the resource click preferences. At the same time, the teaching path is modified based on the students' knowledge mastery rate, and then resources are adapted to the students to obtain a modified English teaching map.

[0064] The knowledge verification module uses a graph traversal algorithm to traverse the nodes of the revised English teaching graph, and combines the Common European Framework of Reference for Languages ​​to perform graph splitting and edge adaptation to obtain the optimal English teaching graph.

[0065] A knowledge graph generation system for English teaching, and a knowledge graph generation method for English teaching, comprising a multi-modal acquisition module, an information processing module, an extraction and filling module, a graph construction module, a knowledge verification module, and a teaching analysis module; wherein,

[0066] The multimodal acquisition module is used to collect all English textbooks and supplementary teaching materials to obtain subject knowledge datasets;

[0067] Information processing module, used to pre-process subject knowledge data sets to obtain several comprehensive subject content libraries;

[0068] The extraction and filling module is used to structure all subject content libraries and obtain the teaching field content table;

[0069] A graph construction module is used to construct the teaching domain content table into an initial English teaching graph based on the dynamic knowledge fusion model;

[0070] The knowledge verification module is used to optimize the initial English teaching map based on the expert prior mechanism to obtain the optimal English teaching map;

[0071] The teaching analysis module is used to visualize the domain knowledge blind spots of the optimal English teaching map and generate a learning evaluation form.

[0072] A storage medium; the storage medium stores a computer program, and the computer program executes a knowledge graph generation method for English teaching.

[0073] Preferably, the storage medium includes a USB flash drive, a CF card, an SD card, an SDHC card, an MMC card, a SM card, a memory stick and an XD card.

[0074] The present invention constructs and optimizes the learning path and teaching path of English teaching by combining knowledge graph with dynamic knowledge fusion model, thereby improving the accuracy of extracting English professional terms and the accuracy of identifying entity relationships in the field of English education. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0076] Figure 1Schematic diagram of the knowledge graph generation system;

[0077] Figure 2 This is a block diagram of the knowledge graph generation method. DETAILED DESCRIPTION

[0078] The present invention will be described in detail below with reference to the accompanying drawings and embodiments:

[0079] like Figure 2 As shown, the method for generating a knowledge graph for English teaching according to the present invention includes the following steps:

[0080] S1: The multimodal acquisition module constructs a subject knowledge dataset based on all English textbooks and supplementary teaching materials;

[0081] S2: The information processing module cleans, categorizes, and grades the subject knowledge dataset to generate several comprehensive subject content libraries;

[0082] S3: The extraction and filling module uses the entity explicit and implicit association mechanism to structure all subject content libraries and obtain the teaching field content table;

[0083] S4: The graph construction module uses a dynamic knowledge fusion model combined with the teaching field model to construct the teaching field content table into an initial English teaching graph;

[0084] S5: The knowledge verification module optimizes and updates the initial English teaching map based on the students' training and test feedback combined with the expert prior mechanism to obtain the optimal English teaching map.

[0085] In the present invention, step S1 includes the following steps:

[0086] The multi-mode acquisition module parses all textbooks using optical character recognition rules to extract directory levels, knowledge points, and accompanying exercise sets;

[0087] At the same time, a hybrid character recognition engine is used to perform text recognition and image enhancement on supplementary teaching materials containing text and images to extract supplementary teaching content information;

[0088] At the same time, data crawling is used to crawl teaching materials containing multimedia resources, and record knowledge points and knowledge point timestamps;

[0089] The multi-mode acquisition module combines knowledge points according to the directory hierarchy, associates the supplementary teaching content information, knowledge point timestamps and supporting exercises, and obtains a knowledge point resource table for all versions;

[0090] According to the textbook version and semantic similarity, the knowledge point resource tables of different versions are matched to obtain the knowledge point equivalent mapping table;

[0091] The multi-mode acquisition module generates a unique identifier for a knowledge point based on the textbook version, the knowledge point page number, and an N-digit random number, and uses it as a joint index serial number.

[0092] According to the material type, combined with the joint index serial number and the knowledge point equivalent mapping table, different types of knowledge points are associated and stored in different databases. At the same time, the traceability information is filled in according to the knowledge point timestamp to obtain the subject knowledge dataset;

[0093] The multi-mode acquisition module determines the number of resource types associated with each knowledge point in the subject knowledge dataset and the preset verification threshold:

[0094] If the number of resource types is less than the verification threshold, the current knowledge point is determined to be missing resource items, and the subject knowledge dataset is verified based on the joint index sequence number combined with the directory hierarchy, and the resource items are extracted and filled in based on the traceability information;

[0095] If the number of resource types is greater than or equal to the verification threshold, it is determined that there are no missing items in the current knowledge point, and information processing is performed on the subject knowledge dataset based on the knowledge point.

[0096] In the present invention, step S2 includes the following steps:

[0097] The information processing module uses the conditional random field model to perform semantic segmentation on the subject knowledge dataset according to the directory level and the title level to obtain the chapter content knowledge table;

[0098] Using the named entity recognition model combined with the knowledge point equivalent mapping table, all knowledge points in the chapter content knowledge table are annotated, and grammatical points and vocabulary types are marked to obtain the chapter knowledge entity table;

[0099] The information processing module dynamically aligns multimedia resources with textbooks and supplementary teaching materials based on the knowledge point resource table, and links images and texts based on the knowledge point timestamps to obtain a multi-source entity knowledge association table;

[0100] The information processing module uses the isolation forest algorithm to filter noise from the multi-source entity knowledge association table, remove duplicate and irrelevant content, and obtain a noise-free multi-source knowledge association table;

[0101] The information processing module uses the contrast image pre-training model to perform anomaly detection and cross-modal similarity calculation on the noise-free multi-source knowledge association table, and compares it with the preset similarity threshold:

[0102] In this embodiment, cross-modal similarity is calculated to detect image-text / audio-text mismatches (e.g., the illustration shows "apple" but the audio reads "orange"). The CLIP model is used to calculate image-text similarity, and association pairs with a score > 0.85 are retained.

[0103] If the cross-modal similarity is greater than or equal to the similarity threshold, the current association pair is retained and abnormal matching is corrected; if the cross-modal similarity is less than the similarity threshold, the current association pair is synchronized according to the knowledge point timestamp, a timeline label is generated, and a noise-free multi-source knowledge association table is assigned;

[0104] The information processing module performs multimodal analysis on the subject knowledge dataset and the accompanying exercise set based on the noise-free multi-source knowledge association table to obtain linguistic and teaching features;

[0105] In this embodiment, linguistic features: extract text complexity (lexical diversity, sentence length standard deviation), grammatical structure (clause nesting depth) and semantic density (proportion of named entities);

[0106] Teaching features: marking the cognitive level of knowledge points (memory / understanding / application), common error types (confusion of tenses / missing articles) and related teaching aids (word cards / grammar diagrams).

[0107] A bidirectional sentence embedding model is used to vectorize all knowledge points based on timeline labels, combining linguistic features and teaching features to obtain a multimodal knowledge feature vector.

[0108] A contrastive learning function is used to align all knowledge points according to the multimodal knowledge feature vector to obtain a cross-modal semantic mapping table;

[0109] The information processing module establishes a three-level label extraction branch based on the three-level encoding and the cross-module semantic mapping table;

[0110] In this embodiment, three-level coding refers to the classification and coding of information in a certain field for easy management and query. For English, the three-level coding can be divided according to different classification standards. The following is a common example of three-level coding: First-level coding: division according to the main classification areas of English, such as grammar, vocabulary, reading, writing, etc.; Second-level coding: further subdividing each first-level coding, for example, in terms of grammar, it can include tense, voice, sentence pattern, etc.; in terms of vocabulary, it can include nouns, verbs, adjectives, etc.; Third-level coding: further refinement under the second-level coding, for example, in terms of nouns, it can include person names, place names, object names, etc.; in terms of verbs, it can include tense and voice, etc.

[0111] Level 3 label branch: major category: grammar / vocabulary; subcategory: tense / voice; subcategory: past perfect / subjunctive mood;

[0112] A multimodal classification model is used in combination with a three-level label extraction branch and a confusion classification weight matrix to parse and classify knowledge points of all modalities and obtain a multimodal comprehensive weight matrix to complete the multimodal data classification.

[0113] The information processing module uses item response theory combined with teaching sequence and applicable scenarios to classify each modality of information into multiple dimensions. At the same time, it optimizes the cognitive difficulty of all knowledge points based on the answer feedback data and generates a knowledge learning path.

[0114] In this embodiment, item response theory (IRT), also known as latent trait theory or latent trait model, is a modern psychometric theory, and its significance lies in guiding item screening and test compilation.

[0115] According to the knowledge learning path, the subject knowledge data set and the corresponding exercise set are clustered and allocated to obtain the micro-knowledge point library and the theme unit library. Combined with the popularity index and quality index of the knowledge points, several comprehensive subject content libraries are constructed.

[0116] In this embodiment, the micro-knowledge point library: stores atomic knowledge points (such as "the usage of the indefinite article a / an"), and comes with more than three types of teaching resources (text analysis / video explanation / interactive exercises);

[0117] Thematic unit library: Aggregates related knowledge points to form teaching modules (e.g., the "travel scene unit" includes vocabulary for asking for directions, present continuous tense, and directional prepositions).

[0118] Popularity indicators: number of clicks on knowledge points and frequency of citations by teachers;

[0119] Quality indicators: student mastery rate, completeness score of related resources;

[0120] In the present invention, step S3 includes the following steps:

[0121] The extraction and filling module extracts knowledge point structured entities from the subject content library based on the three-level tags, and binds inherent attributes to the supporting exercise set through regular expressions to obtain auxiliary knowledge entities;

[0122] A graph neural network is used to mine the subject content library based on students' learning behavior, combined with supporting exercise sets and confusion classification weight matrices, to obtain students' behavior-derived entities and context-related entities of knowledge points;

[0123] According to the explicit and implicit entity association mechanism, the knowledge point structured entities and auxiliary knowledge entities are hierarchically associated, and topological sorting is performed in combination with the teaching logic chain to generate the teaching path and complete the explicit association modeling;

[0124] At the same time, cross-modal associations are performed between behavior-derived entities and context-related entities, and association corrections are performed based on the student's error rate to obtain multi-modal mapping association weights and complete implicit association reasoning.

[0125] The extraction and filling module constructs explicit nodes and explicit edges based on the teaching path and the multi-modal knowledge feature vector; at the same time, it constructs implicit nodes and implicit edges based on the multi-modal mapping association weight and the student identity;

[0126] The extraction and filling module automatically extracts the comprehensive subject content library based on the three-level labels, explicit nodes, explicit edges, implicit nodes and implicit edges, and constructs a three-dimensional tensor dynamic table;

[0127] Based on explicit association modeling and implicit association reasoning combined with cross-modal semantic mapping tables, conflict detection and teaching adaptation are performed on the three-dimensional tensor dynamic table to obtain the teaching field content table.

[0128] In the present invention, step S4 includes the following steps:

[0129] The extended constraint layer of the dynamic knowledge fusion model uses the expression rule language ERL and combines it with the network ontology description language OWL to expand and constrain the teaching domain content table D, and obtain the difficulty-modality mapping library E;

[0130] In this embodiment, E=OWL(ERL(D));

[0131] in, φ(d): Difficulty rule mapping function (such as ); ψ(d): modal rule mapping function (such as ); Rule combination operator (Cartesian product constraint).

[0132] The multi-source incremental layer of the dynamic knowledge fusion model uses a stream processing algorithm and combines it with the conflict resolution strategy CR to perform heterogeneous fusion and knowledge alignment on the subject knowledge dataset K and the matching exercise set Q, and obtains the super-node aggregated explanation table F;

[0133] In this embodiment, F = Stream(K||Q)⊙CR(K,Q);||: parallel input of heterogeneous data streams; ⊙: conflict resolution operator Stream(·): stream processing algorithm;

[0134] The explicit embedding layer Embed of the dynamic knowledge fusion model adopts the random walk algorithm RW, which combines the three-dimensional tensor dynamic table with the knowledge point centrality C according to the explicit association modeling to generate the trunk learning path P main In this embodiment, P main =RW(Tdynamic,C); where Tdynamic: 3D tensor dynamic table (T ijk represents the association strength between knowledge point i, modality j, and difficulty k); λ: normalization factor (usually the total number of nodes); Auv: explicit association weight from node u to v in the adjacency matrix;

[0135] The implicit reasoning layer Infer of the dynamic knowledge fusion model uses the relational graph convolutional network (RGCN). Based on implicit association reasoning, it combines the three-dimensional tensor dynamic table with the weight attenuation factor β to predict the potential teaching path Platent.

[0136] In this embodiment, Platform = RGCN (Tdynamic, β), where β(t) = e -γt , t: path node;

[0137] The graph building layer of the dynamic knowledge fusion model uses the distributed inference algorithm DI to construct a graph of the teaching domain content table based on the main learning path and potential teaching path, combined with the super-node aggregation explanation table, to obtain the initial English teaching graph Ginit;

[0138] In this embodiment, Path fusion operator;

[0139] The distributed inference algorithm is Spark-based graph partitioning and merging;

[0140] The path validation layer Validate of the dynamic knowledge fusion model uses the betweenness centrality B(v) and the difficulty-modality mapping library to sparsely represent the initial English teaching graph, and then marks and archives all the paths in the initial English teaching graph.

[0141] In this embodiment, Gfinal = Sparsify(Ginit, E) Sparsify(·): sparsification function (keep the key nodes with B(v)>θ); σst(v): node v in the path node s→

[0142] The number of occurrences in t: θ: threshold parameter (dynamically adjusted by the difficulty-modality mapping library E);

[0143] Gfinal=Validate(Build(Embed(T)+Infer(T),F),E), Gfinal: the initial English teaching graph after sparsification.

[0144] In the present invention, step S5 includes the following steps:

[0145] The knowledge verification module tracks students' learning trajectory based on the initial English teaching map, recording the length of time spent on knowledge points, the number of times wrong questions are redone, and resource click preferences;

[0146] A natural language network is used to capture the emotional signals of the teaching content table in combination with the authoritative standard mapping, and the knowledge emotional tendency and constraint node attributes are obtained;

[0147] The knowledge verification module uses a clustering algorithm based on the length of time spent on a knowledge point and the number of times a wrong question is redone to cluster the wrong question records in the accompanying exercise book and identify high-frequency error points.

[0148] Bayesian network is used to associate knowledge points with constraint node attributes based on high-frequency error points to obtain error-knowledge point association identifiers;

[0149] The node colors of the initial English teaching map are distributed and marked according to the error rate of knowledge points. At the same time, the high-frequency error points are traced back to determine the prerequisite knowledge points.

[0150] The nodes of the initial English teaching map are added and deleted based on the resource click preferences. At the same time, the teaching path is modified based on the students' knowledge mastery rate, and then resources are adapted to the students to obtain a modified English teaching map.

[0151] The knowledge verification module uses a graph traversal algorithm to traverse the nodes of the revised English teaching graph, and combines the Common European Framework of Reference for Languages ​​to perform graph splitting and edge adaptation to obtain the optimal English teaching graph.

[0152] In this embodiment, the Common European Framework of Reference for Languages ​​divides language proficiency into six levels:

[0153] A1: Beginner, able to understand and use simple daily expressions.

[0154] A2: Basic user, able to understand common sentences and expressions.

[0155] B1: Independent user, able to express simple opinions on familiar topics.

[0156] B2: Independent user, able to understand the main idea and details, and able to communicate in familiar areas.

[0157] C1: Skilled user, able to understand complex texts and express themselves fluently and freely.

[0158] C2: Proficient user, able to easily understand almost everything they hear or read.

[0159] In the present invention, the knowledge graph generation method further includes the following steps:

[0160] S6: The teaching analysis module combines training test feedback with learning level to visualize the domain knowledge blind spots of the optimal English teaching map, and at the same time selects teaching path branches in stages to generate a learning evaluation form.

[0161] like Figure 1As shown, a knowledge graph generation system for English teaching, used for the knowledge graph generation method, includes a multi-modal acquisition module, an information processing module, an extraction and filling module, a graph construction module, a knowledge verification module and a teaching analysis module; wherein,

[0162] The multimodal acquisition module is used to collect all English textbooks and supplementary teaching materials to obtain subject knowledge datasets;

[0163] Information processing module, used to pre-process subject knowledge data sets to obtain several comprehensive subject content libraries;

[0164] The extraction and filling module is used to structure all subject content libraries and obtain the teaching field content table;

[0165] A graph construction module is used to construct the teaching domain content table into an initial English teaching graph based on the dynamic knowledge fusion model;

[0166] The knowledge verification module is used to optimize the initial English teaching map based on the expert prior mechanism to obtain the optimal English teaching map;

[0167] The teaching analysis module is used to visualize the domain knowledge blind spots of the optimal English teaching map and generate a learning evaluation form.

[0168] A storage medium stores a computer program, which executes the knowledge graph generation method for English teaching. The storage medium includes a USB flash drive, a CF card, an SD card, an SDHC card, an MMC card, an SM card, a memory stick, and an XD card.

[0169] Example:

[0170] The multi-mode acquisition module uses optical character recognition rules to parse all textbooks and materials, extracting directory levels, knowledge points, and accompanying exercise books. It also uses a hybrid character recognition engine to perform text recognition and image enhancement on supplementary teaching materials containing text and images, extracting supplementary teaching content information. Furthermore, it uses data crawling methods to capture supplementary teaching materials containing multimedia resources, recording knowledge points and their timestamps.

[0171] The multi-mode acquisition module combines knowledge points according to the directory hierarchy, associates the supplementary teaching content information, knowledge point timestamps and supporting exercises, and obtains a knowledge point resource table of all versions. It also matches the knowledge point resource tables of different versions according to the textbook version and semantic similarity to obtain a knowledge point equivalent mapping table.

[0172] The multi-mode acquisition module generates a unique identifier for a knowledge point based on the textbook version, knowledge point page number, and an N-digit random number, and uses it as a joint index serial number. Based on the material type, the joint index serial number is combined with the knowledge point equivalent mapping table to associate different types of knowledge points using different databases for storage. At the same time, traceability information is filled in based on the knowledge point timestamp to obtain a subject knowledge dataset.

[0173] The multi-mode acquisition module judges the number of resource types associated with each knowledge point in the subject knowledge dataset against the preset verification threshold: if the number of resource types is less than the verification threshold, it is determined that the current knowledge point is missing resource items, and the subject knowledge dataset is verified based on the joint index serial number combined with the directory hierarchy, and the resource items are extracted and filled in combination with the traceability information; if the number of resource types is greater than or equal to the verification threshold, it is determined that the current knowledge point has no missing items, and the subject knowledge dataset is processed according to the knowledge point.

[0174] The information processing module uses a conditional random field model to perform semantic segmentation on the subject knowledge dataset according to the directory hierarchy and the title level to obtain a chapter content knowledge table. It also uses a named entity recognition model combined with a knowledge point equivalent mapping table to annotate all knowledge points in the chapter content knowledge table, while also marking grammatical points and vocabulary types to obtain a chapter knowledge entity table.

[0175] The information processing module dynamically aligns multimedia resources with textbooks and supplementary teaching materials based on the knowledge point resource table. It also links images and text based on knowledge point timestamps to generate a multi-source entity knowledge association table. The information processing module uses the isolation forest algorithm to filter noise from the multi-source entity knowledge association table, eliminating duplicate and irrelevant content to generate a noise-free multi-source knowledge association table.

[0176] The information processing module uses the contrast image pre-training model to perform anomaly detection and cross-modal similarity calculation on the noise-free multi-source knowledge association table, and compares it with the preset similarity threshold: if the cross-modal similarity is greater than or equal to the similarity threshold, the current association pair is retained and the abnormal match is corrected; if the cross-modal similarity is less than the similarity threshold, the current association pair is synchronized according to the knowledge point timestamp, a timeline label is generated, and the noise-free multi-source knowledge association table is assigned.

[0177] The information processing module performs multimodal analysis on the subject knowledge dataset and accompanying exercise set based on a noise-free multi-source knowledge association table to obtain linguistic and pedagogical features. A bidirectional sentence embedding model is used to vectorize all knowledge points based on timeline labels, combining linguistic and pedagogical features to obtain a multimodal knowledge feature vector. A contrastive learning function is used to align all knowledge points based on the multimodal knowledge feature vector to obtain a cross-modal semantic mapping table.

[0178] The information processing module establishes a three-level label extraction branch based on the three-level coding and cross-modal semantic mapping table. It uses a multimodal classification model combined with the three-level label extraction branch and the confusion classification weight matrix to parse and classify knowledge points of all modalities, and obtains a multimodal comprehensive weight matrix to complete the multimodal data classification.

[0179] The information processing module uses item response theory combined with teaching sequence and applicable scenarios to classify each modal material into multiple dimensions. At the same time, it optimizes the cognitive difficulty of all knowledge points based on the answer feedback data and generates knowledge learning paths; and clusters and allocates subject knowledge data sets and supporting exercise sets to obtain micro-knowledge point libraries and thematic unit libraries, and combines the popularity indicators and quality indicators of knowledge points to construct several comprehensive subject content libraries.

[0180] The extraction and filling module extracts structured entities of knowledge points from the subject content library based on the three-level tags, and binds inherent attributes to the accompanying exercise set through regular expressions to obtain auxiliary knowledge entities. A graph neural network is used to mine the subject content library based on students' learning behavior, combined with the accompanying exercise set and the confusion classification weight matrix, to obtain student behavior-derived entities and context-related entities of knowledge points.

[0181] Based on the explicit and implicit entity association mechanism, the knowledge point structured entities and auxiliary knowledge entities are hierarchically associated, and topological sorting is performed in combination with the teaching logic chain to generate a teaching path and complete explicit association modeling. At the same time, the behavior-derived entities and context-related entities are cross-modally associated, and the association is corrected based on the student error rate to obtain the multi-modal mapping association weight and complete the implicit association reasoning.

[0182] The extraction and filling module constructs explicit nodes and explicit edges based on the teaching path and the multi-modal knowledge feature vector; at the same time, it constructs implicit nodes and implicit edges based on the multi-modal mapping association weight and the student identity;

[0183] The extraction and filling module automatically extracts the comprehensive subject content library according to the three-level labels combined with explicit nodes, explicit edges, implicit nodes and implicit edges to construct a three-dimensional tensor dynamic table; based on explicit association modeling and implicit association reasoning combined with cross-modal semantic mapping tables, the three-dimensional tensor dynamic table is subjected to conflict detection and teaching adaptation to obtain the teaching field content table.

[0184] The graph construction module uses the expression rule language through the extended constraint layer of the dynamic knowledge fusion model and combines it with the network ontology description language to expand and constrain the teaching domain content table to obtain the difficulty-modality mapping library;

[0185] The multi-source incremental layer of the dynamic knowledge fusion model adopts a stream processing algorithm, combined with a conflict resolution strategy, to perform heterogeneous fusion and knowledge alignment of subject knowledge datasets and supporting exercise sets, and obtain a super-node aggregation explanation table; the explicit embedding layer adopts a random walk algorithm, and combines the three-dimensional tensor dynamic table with the knowledge point centrality based on explicit association modeling to generate the main learning path; the implicit reasoning layer adopts a relational graph convolutional network, and combines the three-dimensional tensor dynamic table with the weight attenuation factor based on implicit association reasoning to predict the potential teaching path; the graph establishment layer adopts a distributed reasoning algorithm to construct a graph of the teaching field content table based on the main learning path and potential teaching path, combined with the super-node aggregation explanation table, to obtain the initial English teaching graph; the path verification layer uses the betweenness centrality and the difficulty-modality mapping library to sparse the initial English teaching graph, and then marks and archives all paths in the initial English teaching graph.

[0186] The knowledge verification module tracks students' learning trajectories based on the initial English teaching map, recording the length of time spent on knowledge points, the number of times they redo wrong questions, and resource click preferences. It also uses a natural language network to capture sentiment signals from the teaching domain content table in conjunction with authoritative standard mapping, obtaining knowledge sentiment tendencies and constraint node attributes.

[0187] The knowledge verification module uses a clustering algorithm based on the length of time spent on a knowledge point and the number of times a wrong question is redone to cluster the wrong question records in the accompanying exercise book and identify high-frequency error points. A Bayesian network is used to associate knowledge points with constraint node attributes based on high-frequency error points to obtain error-knowledge point association identifiers.

[0188] The knowledge verification module distributes and annotates the node colors of the initial English teaching map based on the error rate of knowledge points, traces the source of high-frequency error points, and determines the prerequisite knowledge points; it adds and deletes nodes in the initial English teaching map based on resource click preferences, and modifies the teaching path based on the students' knowledge point mastery rate, and then adapts resources to the students to obtain a revised English teaching map;

[0189] The knowledge verification module uses a graph traversal algorithm to traverse the nodes of the revised English teaching graph, and combines the Common European Framework of Reference for Languages ​​to perform graph splitting and edge adaptation to obtain the optimal English teaching graph.

[0190] The teaching analysis module combines training and test feedback with learning levels to visualize the domain knowledge blind spots of the optimal English teaching map, while selecting teaching path branches in stages and generating a learning evaluation form.

Claims

1. A knowledge graph generation method for English teaching, characterized in that: The following steps are involved: S1: The multimodal acquisition module constructs a subject knowledge dataset based on all English textbooks and supplementary teaching materials; S2: The information processing module cleans, categorizes, and grades the subject knowledge dataset to generate several comprehensive subject content libraries; S3: The extraction and filling module uses the entity explicit and implicit association mechanism to structure all subject content libraries and obtain the teaching field content table; Among them, S3 includes the following steps: The extraction and filling module extracts knowledge point structured entities from the subject content library based on the three-level tags, and binds inherent attributes to the supporting exercise set through regular expressions to obtain auxiliary knowledge entities; A graph neural network is used to mine the subject content library based on students' learning behavior, combined with supporting exercise sets and confusion classification weight matrices, to obtain students' behavior-derived entities and context-related entities of knowledge points; According to the explicit and implicit entity association mechanism, the knowledge point structured entities and auxiliary knowledge entities are hierarchically associated, and topological sorting is performed in combination with the teaching logic chain to generate the teaching path and complete the explicit association modeling; At the same time, cross-modal associations are performed between behavior-derived entities and context-related entities, and association corrections are performed based on the student's error rate to obtain multi-modal mapping association weights and complete implicit association reasoning. S4: The graph construction module uses a dynamic knowledge fusion model combined with the teaching field model to construct the teaching field content table into an initial English teaching graph; Among them, S4 includes the following steps: The extended constraint layer of the dynamic knowledge fusion model uses expression rule language and combines it with network ontology description language to expand and constrain the teaching domain content table to obtain the difficulty-modality mapping library; The explicit embedding layer of the dynamic knowledge fusion model uses a random walk algorithm to combine the three-dimensional tensor dynamic table with the knowledge point centrality based on explicit association modeling to generate a backbone learning path; The implicit reasoning layer of the dynamic knowledge fusion model uses a relational graph convolutional network to combine the three-dimensional tensor dynamic table with the weight decay factor based on implicit association reasoning to predict potential teaching paths; S5: The knowledge verification module optimizes and updates the initial English teaching map based on students' training and test feedback combined with expert prior mechanisms, and integrates the Common European Framework of Reference for Languages ​​to perform map splitting and edge adaptation to obtain the optimal English teaching map.

2. The method for generating a knowledge graph for English teaching according to claim 1, wherein: The step S1 includes the following steps: the multi-mode acquisition module parses all textbooks and extracts the directory hierarchy, knowledge points and supporting exercise sets; at the same time, the text recognition and image enhancement are performed on the teaching materials containing text and images to extract the teaching content information; at the same time, the teaching materials containing multimedia resources are captured and the knowledge points and knowledge point timestamps are recorded; the multi-mode acquisition module combines the knowledge points according to the directory hierarchy, associates the teaching content information, the knowledge point timestamps and the supporting exercises, and obtains the knowledge point resource table of all versions; according to the textbook version and the semantic similarity, the knowledge point resource tables of different versions are matched to obtain the knowledge point equivalent mapping table; the multi-mode acquisition module generates a unique knowledge point identifier according to the textbook version, the knowledge point page number and the N-bit random number, and makes a is a joint index serial number; according to the material type combined with the joint index serial number and the knowledge point equivalent mapping table, different types of knowledge points are associated and stored using different databases, and the traceability information is filled in according to the knowledge point timestamp to obtain the subject knowledge data set; the multi-mode acquisition module judges the number of resource types associated with each knowledge point in the subject knowledge data set and the preset verification threshold: if the number of resource types is less than the verification threshold, it is determined that the current knowledge point is missing resource items, and the subject knowledge data set is verified according to the joint index serial number combined with the directory hierarchy, and the resource items are extracted and filled in in combination with the traceability information; if the number of resource types is greater than or equal to the verification threshold, it is determined that the current knowledge point has no missing items, and the subject knowledge data set is processed according to the knowledge point.

3. The method for generating a knowledge graph for English teaching according to claim 1, wherein: The step S2 includes the following steps: the information processing module performs semantic segmentation on the subject knowledge data set according to the title level according to the directory level to obtain a chapter content knowledge table; according to the knowledge point equivalent mapping table, all knowledge points in the chapter content knowledge table are annotated, and grammatical points and vocabulary types are marked to obtain a chapter knowledge entity table; the information processing module dynamically aligns multimedia resources with textbooks and teaching materials according to the knowledge point resource table, and links images and texts according to the knowledge point timestamps to obtain a multi-source entity knowledge association table; the information processing module performs noise filtering on the multi-source entity knowledge association table to eliminate duplicate content and irrelevant content to obtain a noise-free multi-source knowledge association table; the information processing module performs anomaly detection and cross-modal similarity calculation on the noise-free multi-source knowledge association table, and compares it with a preset similarity threshold: if the cross-modal similarity is greater than or equal to the similarity threshold, the current association pair is retained and the abnormal match is corrected; if the cross-modal similarity is less than the similarity threshold, the current association pair is synchronized according to the knowledge point timestamp, a time axis label is generated, and the noise-free multi-source knowledge association table is assigned.

4. The method for generating a knowledge graph for English teaching according to claim 1, wherein: The step S2 also includes the following steps: the information processing module performs multimodal analysis on the subject knowledge data set and the supporting exercise set according to the noise-free multi-source knowledge association table to obtain linguistic features and teaching features; according to the timeline label, all knowledge points are vectorized in combination with the linguistic features and teaching features to obtain a multimodal knowledge feature vector; according to the multimodal knowledge feature vector, all knowledge points are aligned to obtain a cross-modal semantic mapping table; the information processing module establishes a three-level label extraction branch based on the three-level coding and the cross-modal semantic mapping table; according to the three-level label extraction branch and the confusion classification weight matrix, the knowledge points of all modalities are analyzed and classified to obtain a multimodal comprehensive weight matrix to complete the multimodal data classification; the information processing module uses item response theory combined with teaching sequence and applicable scenarios to multi-dimensionally grade each modal data, and at the same time optimizes the cognitive difficulty of all knowledge points based on the answer feedback data to generate a knowledge learning path; according to the knowledge learning path, the subject knowledge data set and the supporting exercise set are clustered and allocated to obtain a micro knowledge point library and a theme unit library, and combined with the popularity index and quality index of the knowledge points, a number of comprehensive subject content libraries are constructed.

5. The method for generating a knowledge graph for English teaching according to claim 1, wherein: The step S3 includes the following steps: the extraction and filling module constructs explicit nodes and explicit edges according to the teaching path in combination with the multimodal knowledge feature vector; at the same time, the implicit nodes and implicit edges are constructed according to the multimodal mapping association weight in combination with the student identification; the extraction and filling module automatically extracts the comprehensive subject content library according to the three-level labels in combination with the explicit nodes, explicit edges, implicit nodes and implicit edges to construct a three-dimensional tensor dynamic table; according to the explicit association modeling and implicit association reasoning combined with the cross-modal semantic mapping table, the three-dimensional tensor dynamic table is subjected to conflict detection and teaching adaptation to obtain the teaching field content table.

6. The method for generating a knowledge graph for English teaching according to claim 1, wherein: The step S4 includes the following steps: the multi-source incremental layer of the dynamic knowledge fusion model adopts a stream processing algorithm, combined with a conflict resolution strategy, to perform heterogeneous fusion and knowledge alignment on the subject knowledge data set and the supporting exercise set, and obtains a super-node aggregation explanation table; the graph establishment layer of the dynamic knowledge fusion model adopts a distributed reasoning algorithm to construct a graph of the teaching field content table based on the trunk learning path and the potential teaching path, combined with the super-node aggregation explanation table, to obtain an initial English teaching graph; the path verification layer of the dynamic knowledge fusion model performs sparseness on the initial English teaching graph based on the betweenness centrality and the difficulty-modality mapping library, and then marks and archives all paths in the initial English teaching graph.

7. The method for generating a knowledge graph for English teaching according to claim 1, wherein: The step S5 includes the following steps: the knowledge verification module tracks the student's learning trajectory according to the initial English teaching map, records the length of time the student stays at the knowledge point, the number of times the wrong questions are redone, and the resource click preference; uses a natural language network to capture the emotional signal of the teaching field content table in combination with the authoritative standard mapping to obtain the knowledge emotional tendency and constraint node attributes; the knowledge verification module uses a clustering algorithm based on the length of time the student stays at the knowledge point and the number of times the wrong questions are redone to cluster the wrong question records of the matching exercise book and identify high-frequency error points; uses a Bayesian network to associate the knowledge points with the constraint node attributes based on the high-frequency error points The error-knowledge point association identifier is obtained; the node colors of the initial English teaching map are distributed and marked according to the knowledge point error rate, and the high-frequency error points are traced to determine the prerequisite knowledge points; the nodes of the initial English teaching map are added and deleted according to the resource click preference, and the teaching path is modified according to the students' knowledge point mastery rate, and then the resources are adapted to the students to obtain the revised English teaching map; the knowledge verification module uses a graph traversal algorithm to traverse the nodes of the revised English teaching map, and combines the European Common Framework for Languages ​​to perform graph splitting and edge adaptation to obtain the optimal English teaching map.

8. The method for generating a knowledge graph for English teaching according to claim 1, wherein: The knowledge graph generation method also The following steps are involved: S6: The teaching analysis module combines training test feedback with learning level to visualize the domain knowledge blind spots of the optimal English teaching map, and at the same time selects teaching path branches in stages to generate a learning evaluation form.

9. A knowledge graph generation system for English teaching, used in the knowledge graph generation method according to any one of claims 1 to 8, characterized in that: It includes a multi-modal acquisition module, an information processing module, an extraction and filling module, a graph construction module, a knowledge verification module and a teaching analysis module; among them, the multi-modal acquisition module is used to collect all English textbooks and teaching materials to obtain subject knowledge data sets; the information processing module is used to pre-process the subject knowledge data sets to obtain several comprehensive subject content libraries; the extraction and filling module is used to structure all subject content libraries to obtain a teaching field content table; the graph construction module is used to construct the teaching field content table into an initial English teaching graph based on a dynamic knowledge fusion model; the knowledge verification module is used to optimize the initial English teaching graph based on the expert prior mechanism to obtain the optimal English teaching graph; the teaching analysis module is used to visualize the domain knowledge blind spots of the optimal English teaching graph and generate a learning evaluation table.

10. A storage medium, characterized in that: The storage medium stores a computer program, which executes the knowledge graph generation method according to any one of claims 1 to 8. The storage medium includes a USB flash drive, a CF card, an SD card, an SDHC card, an MMC card, an SM card, a memory stick and an XD card.

Images