Learning support system

Abstract

A computer-implemented learning support system that assists user learning by providing appropriate questions based on incorrect answers. [Solution] A learning support system comprising: a first vectorization means for converting past exam questions of a specific exam into high-dimensional vectors to obtain high-dimensional vectors of past exam questions; a first dimensionality reduction means for performing dimensionality reduction processing on the high-dimensional vectors of past exam questions to obtain a visually recognizable past exam question mapping; a second vectorization means for converting incorrectly answered questions related to a specific exam for a target person into high-dimensional vectors to obtain high-dimensional vectors of incorrectly answered questions; a second dimensionality reduction means for performing dimensionality reduction processing on the high-dimensional vectors of incorrectly answered questions and overlaying information on incorrectly answered questions onto the past exam question mapping to obtain an incorrect answer mapping; and an incorrect answer related question output means for extracting past exam questions on the past exam question mapping that are close to the position of incorrectly answered questions in the incorrect answer mapping and outputting incorrect answer related questions.

Description

Technical Field

[0001] The present invention relates to a computer-implemented learning support system that supports learning in qualification examinations, national examinations, academic achievement tests, and the like, as well as related programs. For example, the learning support system of the present invention expresses past questions and incorrectly answered questions of a subject using high-dimensional vectors, visualizes and superimposes them through dimension reduction, and presents error-related questions, thereby realizing efficient review and reinforcement of weak points. Background Art

[0002] Japanese Patent No. 7303243 discloses an examination question prediction system. This system extracts words from the question texts of past examination questions, assigns numerical values to each vector element for the extracted words to generate vector data, performs vectorization, classifies examination questions into categories based on the vector data, and predicts the number of questions for each category of the next examination from the number of questions of the examination questions classified into said categories for each past year. This system is intended for generating predicted questions and cannot support a subject's learning. Prior Art Documents Patent Documents

[0003] Patent Document 1 Japanese Patent No. 7303243 Summary of the Invention Problem to be Solved by the Invention

[0004] An object of the present invention is to provide a computer-implemented learning support system and the like that support a user's learning by providing appropriate questions based on the user's incorrectly answered questions. Means for Solving the Problem

[0005] This invention is fundamentally based on the realization that by representing past exam questions and incorrect answers by the target student as high-dimensional vectors, visualizing and overlaying them through dimensionality reduction, and presenting related incorrect answers, efficient review and weakness reinforcement can be achieved.

[0006] The learning support system 10 of this invention is a computer-based learning support system for a specific test. The learning support system 10 includes a first vectorization means 20, a first dimensionality reduction means 30, a second vectorization means 40, a second dimensionality reduction means 50, and an incorrect answer related problem output means 60. The first vectorization means 20 is an element for converting past exam questions of a specific test into high-dimensional vectors and obtaining high-dimensional vectors of past exam questions. The first dimensionality reduction means 30 is an element for performing dimensionality reduction processing on the high-dimensional vectors of past exam questions to obtain a visually recognizable past exam question mapping. The second vectorization means 40 is an element for converting incorrect answers to questions related to a specific test for a subject into a high-dimensional vector, thereby obtaining a high-dimensional vector of the incorrect answers. The second dimensionality reduction means 50 is an element for obtaining an incorrect answer mapping by performing a dimensionality reduction process on the high-dimensional vector of the incorrect answer questions and superimposing the information of the incorrect answer questions onto the past question mapping. The incorrect answer related problem output means 60 is an element for outputting incorrect answer related problems by extracting past questions from the past question mapping that are close in position to the incorrect answer problem in the incorrect answer mapping. [Effects of the Invention]

[0007] This invention can provide a computer-implemented learning support system that assists user learning by providing appropriate problems based on the user's incorrect answers. [Brief explanation of the drawing]

[0008] [Figure 1] Figure 1 is a block diagram of the computers that make up the learning support system. [Figure 2]Figure 2 is a block diagram illustrating the learning support system. [Figure 3] Figure 3 is a conceptual diagram showing an example of past exam question mapping. [Figure 4] Figure 4 is a conceptual diagram showing an example of highlighting specific areas within a past exam question mapping. [Figure 5] Figure 5 is a conceptual diagram showing an example of a past exam question mapping that emphasizes a specific area different from that shown in Figure 4. [Figure 6] Figure 6 is a conceptual diagram showing an example of overlaying a mapping of incorrect answers onto a mapping of past exam questions. [Modes for carrying out the invention]

[0009] Figure 1 is a block diagram of a computer constituting a learning support system. The learning support system 10 of the present invention is implemented by a computer, which may be a server or a mobile terminal. Preferably, this learning support system 10 handles digital information using a computer. This learning support system 10 is a system for supporting user learning by outputting related problems based on the user's incorrect answers. This system can also be used to support user learning by analyzing past questions of a specific exam and providing the analyzed information. As shown in Figure 1, the learning support system 10 is implemented by a computer including a storage unit 11, an arithmetic processing unit 12, a display unit 13, an input / output unit 14, and a communication unit 15. Each element is connected by a bus or the like to enable the exchange of information. The display unit 13 may be a monitor of a computer or server, or a monitor of a user terminal. The arithmetic processing unit 12 may have an arithmetic unit and a control unit. The arithmetic processing unit 12 may be embodied by a program that executes one or more functional modules.

[0010] A "specific examination" refers to a particular examination. An examination is one in which a user's answer to one or more questions is judged as correct or incorrect, or the user provides required written answers. An examination may be a paper-based examination or a computer-based examination. Examples of specific examinations include the National Examination for Dentists, the National Examination for Physicians, various qualification examinations, various entrance examinations, language proficiency tests, and various periodic examinations. Among these, this invention can be preferably used in the National Examination for Dentists and the National Examination for Physicians. In particular, the National Examination for Dentists includes questions that span various fields, such as basic medicine, dental medicine, and clinical problems, in addition to mandatory questions in each field, so this invention can be preferably used. Furthermore, the National Examination for Dentists may include photographs such as X-ray images and intraoral photographs, as well as illustrations, and since examinees of the National Examination for Dentists are familiar with diagrams and charts, this invention is highly compatible.

[0011] The memory unit 11 may include, for example, a past exam question database 11a, a target student performance database 11b, and a model / parameter group memory unit 11c (embedded model, dimensionality reduction model, normalized parameters, trained model, etc.). The memory unit 11 may update information as appropriate when new information is input from the input unit. The past exam question database 11a stores, for example, each question ID, question text, question format (multiple choice / written answer, etc.), year of issue, subject tag, correct answer, and explanation. The past exam question database 11a may also store information such as the correct answer rate for each question in relation to each question ID. The subject tag may be various information related to the question, such as its classification or type. The participant performance database 11b stores, for example, participant ID, answer history (question ID, date and time, time taken, correct / incorrect), and type of incorrect answer (if labeling is possible, such as confusion of options or misunderstanding of concepts). The participant performance database 11b may also include, for example, information on the participant's affiliation (school name, affiliated institution), name, age, gender, and contact information (address, email address). The model parameter group storage unit 11c stores, for example, the model name, the random number seed used during training or fitting, the normalization coefficients, and the hyperparameters of UMAP / t-SNE / PCA, in order to ensure the reproducibility of vectorization and dimensionality reduction described later.

[0012] When predetermined information is input from the input unit, the control unit reads the control program stored in the memory unit. The control unit then reads the information stored in the memory unit and transmits it to the arithmetic unit. The control unit also transmits the input information to the arithmetic unit. The arithmetic unit performs calculations using the received information and stores it in the memory unit. The control unit reads the calculation results stored in the memory unit and outputs them from the output unit. In this way, various processes and steps are executed. These various processes are executed by the various units and means. A computer may have a processor, and the processor may implement various functions and processes. A computer may be standalone. A computer may have some of its functions distributed between a server and terminals. In that case, it is preferable that the server and terminals can exchange information via a network such as the internet or an intranet. A computer may have a processor and memory connected to the processor. The memory stores instructions, and when executed by the processor, these instructions may cause the computer to perform various processes or to function as various elements. A computer may build a learning model by providing various training data and perform various calculations using machine learning. In this case, the computer may perform various analyses and interpretations using the learning model created by AI (artificial intelligence) machine learning and deep learning.

[0013] Figure 2 is a block diagram illustrating a learning support system. As shown in Figure 2, the learning support system 10 includes a first vectorization means 20, a first dimensionality reduction means 30, a second vectorization means 40, a second dimensionality reduction means 50, and an incorrect answer-related problem output means 60. Also, as shown in Figure 2, the learning support system 10 may further include a target user analysis means 70 and a display control means 80. In addition to the above, the learning support system 10 may further include various means that are present in known systems. Note that "means" may be read as "parts" or "elements". The display control means 80 may be means that perform various controls related to the display when outputting various information as a GUI to the display unit 13.

[0014] The first vectorization means 20 is an element for converting past exam questions of a specific test into high-dimensional vectors and obtaining high-dimensional vectors of past exam questions. The first vectorization means 20 can be implemented, for example, by an engine that converts text into vectors using a large-scale language processing model (LLM). More specifically, text can be converted into vectors using embedding processing in a computer or server that provides artificial intelligence. The first vectorization means 20 reads, for example, past exam questions input from the input unit, or information about past exam questions for a specific exam stored in the memory unit. The information about past exam questions may include two or more of the following: the year in which the past exam question was given, the type of exam the past exam question was given, the question number of the past exam question, the correct answer rate of the past exam question, the question text of the past exam question, the correct answer of the past exam question, the explanation of the past exam question, the field to which the past exam question belongs, the type of past exam question, and the correct answer rate of the past exam question. This information may be categorized by the user, or it may be analyzable by the system into which the past exam questions were input. Examples of past exam question types include written questions, multiple-choice questions, fill-in-the-blank questions, matching questions, reading, listening, and speaking. In this way, the first vectorization means 20 obtains the text of the past exam questions (text information obtained by OCR or image embedding of figures and tables as necessary). For subsequent processing, the first vectorization means 20 may aggregate the text of the past exam questions into CSV format, or obtain past exam question information in CSV format. The first vectorization means 20 may perform morphological analysis and normalization processing suitable for natural language processing on the question text, answer choices, and, if applicable, the explanatory text of figures and tables of incorrect answers. For example, if the specific test is a national dental licensing examination, the first vectorization means 20 may use a morphological analyzer such as MeCab or Sudachi to divide the text into phrases and perform removal of unnecessary words and word normalization. A distributed representation model (e.g., Word2Vec, FastText, BERT, etc.) may be applied to the preprocessed text to obtain vector representations at the word or sentence level. In this case, the following methods may be used to convert the entire question text into a single fixed-length vector: averaging of word vectors (simple average or weighted average), and vector synthesis based on the sentence representation of a BERT-type model using the -[CLS] token and weighting such as TF-IDF.

[0015] The first vectorization means 20 may perform preprocessing on the information related to past questions before performing embedding in order to convert the past questions into high-dimensional vectors. Examples of preprocessing include notation normalization, morphological analysis, stop word processing, or unit normalization. The first vectorization means 20 converts past question information into high-dimensional vectors. The high-dimensional vectors obtained in this way may be used as high-dimensional vectors of past questions. The first vectorization means 20 may map the obtained high-dimensional vectors to vectors in a Hilbert space, develop them into a higher-dimensional feature space, and obtain a final high-dimensional vector (the high-dimensional vector of the past question). The high-dimensional vector is, for example, a vector in a Hilbert space where an inner product is defined, such as a Euclidean space. These spaces are spaces where an inner product is defined, and may be feature spaces mainly used in the context of machine learning. The high-dimensional vector may be a vector with 100 or more and 100,000 or less dimensions, a vector with 1,000 or more and 10,000 or less dimensions, or an infinite-dimensional vector. The distance index for the vector is, for example, cosine distance or Euclidean distance, and L 2 normalization may be performed so that the inner product functions as a similarity. In this way, the first vectorization means 20 can convert past questions of a specific test into high-dimensional vectors, and obtain high-dimensional vectors of the past questions. The obtained high-dimensional vectors of past questions may be readably stored in the past question database 11a in the storage unit 11 as appropriate.

[0016] The first dimension reduction means 30 is an element for performing dimension reduction processing on the high-dimensional vectors of past questions to obtain a visible mapping of past questions. The first dimension reduction means 30 has a function of, for example, receiving as input the high-dimensional vector group of past questions obtained by the first vectorization means 20, and mapping them to a low-dimensional space such as two-dimensional or three-dimensional space. The result obtained by mapping to the low-dimensional space is also referred to as a map. Furthermore, the processing of mapping to a low-dimensional space is also referred to as mapping. The high-dimensional vector obtained by the first vectorization means 20 is, for example, a result of converting the entire test question text into an embedded representation, and usually has feature quantities of hundreds to thousands of dimensions. In this state, it is difficult for humans to intuitively grasp the content. For this reason, the learning support system 10 performs dimension reduction processing to visualize past question information and the relationship between past questions and incorrect answers.

[0017] The first dimension reduction means 30 is configured, for example, as a software module in the arithmetic processing unit 12, and any one of the following processing algorithms or a combination of a plurality of processing algorithms can be used. (1) Principal Component Analysis (PCA) Principal component analysis is a process of extracting main component axes using a linear transformation that maximizes variance. Since principal component analysis has low computational cost, it can efficiently process large-scale past question databases. (2) t-SNE (t-Distributed Stochastic Neighbor Embedding) t-SNE is a process that models the similarity between data points in a high-dimensional space as a probability distribution, and arranges the data so as to maintain this similarity in a low-dimensional space. By means of t-SNE, semantically similar past questions are plotted close to each other, making it easy to visually grasp the cluster structure. (3) UMAP (Uniform Manifold Approximation and Projection) UMAP is an algorithm that efficiently maps data from a high-dimensional space to a low-dimensional space while preserving the topological structure of the high-dimensional space. Compared with t-SNE, UMAP has a higher computation speed and is suitable for large-scale data. These methods are selectively applied according to the type of examination and the number of past question data. For example, when there are more than thousands of past question data such as in the National Dental Examination, UMAP may be used for high-speed mapping, or t-SNE may be used. On the other hand, when it is desired to emphasize local relationships in error answer analysis, it is effective to use t-SNE.

[0018] As a result of the dimensionality reduction process by the first dimensionality reduction means 30, each past question is represented as a point in a low-dimensional space. For example, a "past question mapping" can be obtained by plotting the obtained points in the low-dimensional space on a two-dimensional coordinate system. A "past question mapping" refers to arrangement information obtained by mapping a group of high-dimensional vectors of past questions to two or three dimensions and visualizing it. The past question mapping is visualized in scatter plot format on the display unit 13, allowing for an intuitive understanding of the structure of the entire exam scope. It is preferable that the past question mapping can be displayed by the display unit. For example, questions from similar fields naturally form clusters, making it possible to grasp the question trends and the distribution of questions by field at a glance. The past question mapping may be stored in the memory unit 11 as appropriate, or it may be stored in the past question database 11a.

[0019] Furthermore, it is preferable that the first dimensionality reduction means 30 includes a function to control random number seeds and hyperparameters in order to maintain processing consistency. For example, the fitting results of the dimensionality reduction model (basis of the embedding space, random seed, etc.) are stored in the model parameter group storage unit 11c and aligned to the same space in the subsequent error answer mapping. This ensures that similar mappings are obtained even between different learning sessions, and facilitates comparative analysis with the error answer mapping (results from the second dimensionality reduction means). As described above, the first dimensionality reduction means 30 plays a fundamental role in representing past exam data in a form that is intuitively understandable to humans, thereby providing learning support effects to the target audience.

[0020] Figure 3 is a conceptual diagram showing an example of past exam question mapping. Figure 3 shows a 2D mapping of past questions from the National Dental Examination by performing embedding processing using generative AI and dimensionality reduction processing. In this example, each past question and each set of past questions are represented as a dot. The past questions are clustered by subject area and can be displayed in different colors for each subject area. Depending on the display medium, each past question may be displayed with varying shades of gray for each subject area (classification). In the example of Figure 3, past questions are classified into the following fields: implantology, restorative dentistry, complete denture, internal medicine, crown and bridge denture, oral surgery, oral hygiene, pediatric dentistry, microbiology, endodontics, periodontology, dental radiology, dental materials science, orthodontics, dental anesthesiology, biochemistry, physiology, epidemiology, pathology, histology, pharmacology, anatomy, partial denture, and geriatrics. For the National Dental Examination, questions may be classified into all of these fields, or into classifications that include two or more of these fields. The vertical and horizontal axes are the result of dimensionality reduction and do not have any particular significance. Furthermore, by inputting the year of past exam questions, you can analyze trends in past questions by year or within a specific range of years.

[0021] Figure 4 is a conceptual diagram showing an example of highlighting a specific field in the past exam question mapping. As described above, since the past exam questions are clustered, when input specifying a particular group is entered into the learning support system 10, the points on the past exam question mapping corresponding to the past exam questions in the entered specific field are highlighted. In Figure 4, the × in the center of the diagram indicates the centroid of the entire past exam question system, and the × on the right indicates the centroid of the past exam questions in a specific field (for example, dental materials science). On the past exam question mapping, the centroid of dental materials science is located to the right of the centroid of the entire past exam question system. In this example, when the user specifies a field and information regarding the specified field is entered into the system, the past exam question information belonging to the entered field is highlighted. When the user selects or indicates a point related to a past exam question displayed on the screen, information regarding the past exam question corresponding to the selected or indicated point may be output (for example, displayed on the display unit 13).

[0022] Figure 5 is a conceptual diagram showing an example of a past exam question mapping that emphasizes a specific field different from that in Figure 4. The example of a different specific field is orthodontics. On the past exam question mapping, the centroid of orthodontics lies below the centroid of the entire past exam question system.

[0023] The second vectorization means 40 is an element for converting incorrect answers to questions related to a specific test for the subject into high-dimensional vectors, thereby obtaining high-dimensional vectors of the incorrect answers. The second vectorization means 40 has the function of, for example, receiving incorrect answers to test questions answered by the subject as input and processing them to represent them as high-dimensional vectors. The second vectorization means 40 can be any means capable of performing similar processing to the first vectorization means 20. For this reason, the learning support system 10 can implement the second vectorization means 40 in the same way as the first vectorization means 20.

[0024] When a user answers practice problems or mock exam questions, their answer history is recorded in an answer management unit (not shown). The answer management unit maintains an answer log that includes the determination of whether the answer was correct or incorrect, and the second vectorization means 40 extracts incorrectly answered questions by referring to this log. An incorrectly answered question may be a question number, or it may include the text of the question that was answered incorrectly. If the system stores information about the question text corresponding to the question number, the system can obtain information about the incorrectly answered question using the question number of the incorrectly answered question. Preferably, the second vectorization means 40 is designed so that the incorrectly answered question vector obtained by the second vectorization means 40 is mapped to the same vector space as the past question vector obtained by the first vectorization means 20. This makes it possible to compare past questions and incorrectly answered questions using the same criteria and to superimpose them on the same mapping after dimensionality reduction.

[0025] The second vectorization means 40 can not only create a simple question vector, but also incorporate the subject's answer patterns and error tendencies in the answer choices as additional features and add them to the vector. For example, it is possible to correctively embed tendencies such as "incorrect answers are concentrated in questions related to anatomy" or "the error rate is high in questions related to X-ray images" to improve the accuracy of learning support. The second vectorization means 40 may also be able to represent the subject's incorrect answers in a format that can be directly compared with past questions in a high-dimensional space. As a result, in the error mapping generation process by the second dimensionality reduction means 50, it becomes possible to intuitively understand where the incorrect answers belong within the entire scope of the exam.

[0026] The second vectorization means 40 can track the subject's error tendencies over time by continuously accumulating vectors of incorrect answers. For example, it can analyze learning progress, such as noticing that there are many errors in the "basic dental medicine field" in the initial learning stage, and that errors in the "clinical dentistry field" increase in the later stages. As described above, the second vectorization means 40 plays a central role in semantically converting the subject's error data into mathematical representations and forms the basis for error analysis in the entire system.

[0027] The second dimensionality reduction means 50 is an element for obtaining an incorrect answer mapping by performing dimensionality reduction processing on the high-dimensional vector of the incorrect answer question and superimposing the information of the incorrect answer question onto the past question mapping. The second dimensionality reduction means 50 has the function of, for example, receiving the high-dimensional vector of the incorrect answer question obtained by the second vectorization means 40 as input and mapping it to the same low-dimensional space as the past question mapping generated by the first dimensionality reduction means 30.

[0028] The second dimensionality reduction means 50 is preferable to arrange in a low-dimensional space in a manner consistent with the past question mapping, rather than performing dimensionality reduction alone, in order to enable the analysis of the characteristics of incorrect answers in relation to past questions. For this reason, the second dimensionality reduction means 50 has the following characteristics.

[0029] Use of the same algorithm By applying the same algorithm (e.g., PCA, t-SNE, UMAP) used in the first dimensionality reduction method 30, past questions and incorrect answers are arranged based on the same spatial criteria. This gives meaning to the relative positional relationship between incorrect answers and past questions, and allows for a quantitative interpretation of proximity.

[0030] Application of a pre-trained mapping function Some dimensionality reduction methods allow new data to be projected onto an existing mapping space (e.g., the transformation functions of PCA and UMAP). The learning support system 10 retains the mapping function and weight parameters obtained when generating past question mappings, and the second dimensionality reduction means 50 can project the incorrect answer vectors into a lower-dimensional space using the same criteria as the past question mappings by inputting them into that function. This allows the incorrect answer mappings to be superimposed on the past question mappings in a consistent manner. "Incorrect answer mapping" refers to information that allows a set of high-dimensional vectors of incorrect answer questions to be mapped onto the same low-dimensional space as the past question mappings and superimposed on the past question mappings.

[0031] Position correction mechanism The second dimensionality reduction means 50 is preferably capable of scaling and correcting the mapping results so that even when there are only a small number of incorrect answers, they can be appropriately visualized. For example, if the distribution of incorrect answers is extremely concentrated in a specific area of ​​the past question mapping, a function to highlight that area can be provided to visually clarify the weaknesses of the target individual.

[0032] Incorrect Answer Mapping Generation As a result of the above processing, for example, incorrectly answered questions are displayed overlaid on the past question mapping. On the display unit 13, past questions are plotted as scatter points, and incorrectly answered questions are displayed with different markers or colors. It is preferable that the past question mapping and incorrect answer mapping can be displayed by the display unit. Furthermore, it is preferable that the past question mapping and incorrect answer mapping are adjusted so that the positions of each coordinate in the mapping are the same. In other words, it is preferable that the past question mapping and incorrect answer mapping have a corresponding relationship. Also, by color-coding, for example, displaying past questions as gray dots and incorrectly answered questions as red dots, the subject's tendency to make mistakes can be grasped at a glance. By comparing the spatial distance between incorrectly answered questions and past questions in the incorrect answer mapping, it is possible to extract related incorrect answer questions. Furthermore, by analyzing the distribution and centroid of the incorrect answer question group, it is possible to quantitatively identify the areas in which the subject is weak. In addition, by arranging the temporal progression of incorrect answers as multiple incorrect answer mappings in chronological order, the progress of learning can be dynamically visualized.

[0033] High-dimensional vectors in past exam questions and incorrect answers belong to a real inner product space, and their similarity may be calculated using the inner product or cosine similarity. In this case, the high-dimensional vectors in past exam questions and incorrect answers are constructed as belonging to an inner product space over the field of real numbers. This allows the similarity between two vectors to be calculated based on the inner product value. In particular, cosine similarity can be calculated by normalizing the vectors by their norms and then calculating the inner product. In relation to evaluation, using cosine similarity allows for a similarity evaluation that reflects the semantic closeness of the problem statement, independent of the magnitude of the vectors. Therefore, by adopting this method, it becomes possible to more appropriately extract past exam questions that are closely related to the incorrect answers.

[0034] In this case, it is preferable that the actual inner product space is a Hilbert space that is complete with respect to the norm induced by the inner product. That is, it is preferable that the actual inner product space is configured as a space that is complete with respect to the norm induced by the inner product, i.e., a Hilbert space. This ensures the convergence and continuity of vectors representing past questions and incorrect answers, guaranteeing mathematical consistency even when extending vector representations to infinite dimensions or performing dimensionality reduction operations. Furthermore, introducing analysis based on Hilbert spaces increases compatibility with existing machine learning algorithms and kernel methods, enabling more accurate extraction of incorrect answer-related questions.

[0035] The vectorization process in this invention may be a method that directly maps the input text to a finite-dimensional vector space using an embedding model. For example, it may be a method that directly maps the input text to a finite-dimensional Euclidean space or a finite-dimensional Hilbert space, such as the “text-embedding-3-small” model provided by OpenAI (registered trademark).

[0036] Another embodiment involves representing the higher-order vectors of past exam questions and incorrect answers as vectors expanded in an infinite-dimensional Hilbert space. These expanded vectors are then projected onto predetermined finite-dimensional axes and further visualized using two-dimensional mapping. By mapping the target data to an infinite-dimensional Hilbert space and projecting the resulting infinite-dimensional expansion vectors onto predetermined finite-dimensional axes, it becomes possible to effectively handle nonlinear features and extract the relationship between incorrect answers and past exam questions with greater precision. In this embodiment, the vector representations of past exam questions and incorrect answers can be defined, for example, as basis expansions in an infinite-dimensional Hilbert space. For example, a mapping from the input space to the infinite-dimensional feature space can be realized using a kernel function. Since the infinite-dimensional vectors thus expanded are difficult to visualize directly, they are projected onto predetermined finite-dimensional axes. After the vectors are projected onto the finite-dimensional axes, past exam questions and incorrect answers can be visualized on a two-dimensional plane by using a two-dimensional mapping method (for example, the one described earlier). This allows learners to intuitively and visually understand the relationship between incorrect answers and existing past questions, enabling them to appropriately identify areas that they should focus on reviewing.

[0037] Examples of representative methods for mapping to infinite-dimensional Hilbert spaces and for finite-dimensional projection include kernel principal component analysis (Kernel PCA), support vector machines (SVM), kernel ridge regression, Gaussian processes, random Fourier features, Nystrom approximation, deep kernel learning, and finite approximation using Mercer expansions. By using these methods, it is possible to utilize higher-order feature representations that cannot be obtained with finite-dimensional direct embedding methods. Furthermore, applying these methods to the system of the present invention yields the following remarkable advantages that are not normally conceivable: namely, it becomes possible to extract potential conceptual misunderstandings inherent in learners' incorrect answers in a high-dimensional feature space, going beyond mere superficial word similarity in the problem statement. As a result, even if the problems appear to be of different forms, groups of incorrect answers based on the same misconception can be accurately identified and presented as closely related problems. Furthermore, by tracking the chronological shift of the center of error, it is possible to dynamically visualize which concepts learners' understanding is converging on, enabling a precise diagnosis of learning achievement.

[0038] In particular, methods such as Kernel PCA and Random Fourier Features are relatively easy to implement and allow for approximate use of information from an infinite-dimensional feature space in a finite-dimensional space, making them easily applicable to the learning support system of this invention. Thus, the present invention is not limited to embodiments based on finite-dimensional embedding representations, but also includes embodiments in which the data is first expanded into an infinite-dimensional Hilbert space and then projected onto a finite-dimensional space for use. This further improves the accuracy of extracting past questions related to learners' incorrect answers and enhances the intuitiveness of visualization.

[0039] Figure 6 is a conceptual diagram showing an example of overlaying an incorrect answer mapping onto a past question mapping. In the example in Figure 6, points on the incorrect answer mapping are displayed in a more emphasized manner compared to points on the past question mapping. The learning support system 10 can display a past question mapping overlaid with an incorrect answer mapping, as shown in Figure 6, on the display unit 13. For example, if a user moves the pointer using a cursor and inputs an instruction, the system may read and display past questions or similar past questions related to the inputted position. This would allow the user to easily obtain past questions and other information related to incorrect answers. Also, for example, if a user creates a new question (a mock exam question or a draft exam for a new year), the new question can be input into the system, the system can perform mapping similar to the past question mapping, and the resulting new question mapping can be compared with the past question mapping to check the trends of the newly created question and the trends of past questions. In this case, instead of incorrect answer questions, information about the new question should be input into the system. For example, if the distance between the centroid of past questions for a specific year or a specific range of years (described later) and the centroid of a new question is within a threshold, the system may determine that the new question is consistent with the trends of past questions. Furthermore, the vector relating to the difference between the centroid of the new question and the centroid of past questions can be used as a reference for determining which areas should be increased (or decreased) in the new question to bring it closer to the trends of past questions. For this reason, the system may use the vector relating to the difference between the centroid of the new question and the centroid of past questions to determine and output a policy for modifying the new question. In this case, the system functions as a question creation support system.

[0040] As described above, the second dimensionality reduction means 50 plays a crucial role in integrating incorrect answer questions into the past question mapping, enabling an intuitive and objective understanding of the learner's learning tendencies. In other words, the dimensionality reduction process of the second dimensionality reduction means 50 is not "mere dimensionality reduction," but rather a process for generating incorrect answer mappings that maintains consistency with the past question mapping.

[0041] The incorrect answer-related problem output means 60 is an element for outputting incorrect answer-related problems by extracting past questions from the past question mapping that are close to the location of the incorrect answer problem in the incorrect answer mapping. The incorrect answer-related problem output means 60 has the function of receiving, for example, the incorrect answer mapping generated by the second dimensionality reduction means 50 as input, extracting past questions from the past question mapping that are close to each incorrect answer problem, and presenting them to the user as "incorrect answer-related problems". The incorrect answer-related problem output means 60 is provided with the aim of efficiently reinforcing the user's weak areas through learning. The incorrect answer-related problem output means 60 performs, for example, the following processing.

[0042] Nearest neighbor search process The incorrect answer-related question output means 60 calculates the distance between the incorrect answer question and each past question on the past question mapping. Examples of distance indicators include Euclidean distance, cosine similarity, and Manhattan distance. In particular, cosine similarity is preferable to take into account the semantic closeness in vector representation. Through this calculation, the incorrect answer-related question output means 60 can extract a certain number of past questions (e.g., the top 5 questions) that are in the vicinity of the incorrect answer question. The number of questions to be extracted can be input from the input means and can be changed as appropriate by being stored in the memory unit. Close relationship means a relationship where the distance on the mapping is short. Past questions that are within an assumed distance on the mapping may be extracted, or, as described above, a predetermined number of past questions that are close on the mapping may be extracted.

[0043] The extracted questions may be actual past exam questions from a specific test, or similar questions automatically generated or edited based on past questions. The generation of similar questions is achieved by using natural language processing to paraphrase or rearrange answer choices, for example, based on past questions. This allows test-takers to deepen their understanding by solving multiple questions related to the same topic.

[0044] Weighted filtering The incorrect answer-related question output means 60 may not simply extract past questions based on proximity on the map, but may also weight them based on the user's past answer history. For example, the incorrect answer-related question output means 60 may lower the priority of questions that have been answered correctly once in the past, and raise the priority of questions in areas that have been repeatedly answered incorrectly. This makes it possible to provide an optimized set of incorrect answer-related questions for each individual user.

[0045] Incorrect answer-related questions are presented on the display unit 13 in the following format. On the incorrect answer mapping, related questions are highlighted around the incorrect answer. The recommended review questions will be presented in a text list format. The interface includes hyperlinks, allowing users to immediately access the practice problem screen with a single click. Furthermore, the output goes beyond simply presenting the questions; it can also display supplementary information such as "the field to which the incorrect answer belongs" and "the relative distance from past questions."

[0046] In the example in Figure 6, the centroids of past questions and the centroids of incorrect answers are indicated by "x", and a straight line connects them, showing the vector. The system may use the information about the centroids of past questions and the centroids of incorrect answers on the mapping to analyze the user's incorrect answer trends and output the user's incorrect answer trends (for example, by printing or displaying them on a screen). Furthermore, the system may use the information about the centroids of past questions and the centroids of incorrect answers on the mapping to output past questions that it recommends to the user.

[0047] The incorrect answer-related problem output means 60 allows users to efficiently review problems conceptually similar to those they answered incorrectly, enabling rapid and effective knowledge reinforcement. Compared to conventional question-and-answer type review, it becomes possible to systematically learn related knowledge. Furthermore, by presenting incorrect answer-related problems, it is possible to promote comprehensive learning of the user's weak areas. As described above, the incorrect answer-related problem output means 60 does not merely record incorrect answers, but uses those incorrect answers as a starting point to present a group of related problems, thereby realizing efficient learning support.

[0048] Target audience analysis method 70 The subject analysis means 70 is an element for analyzing the subject by comparing the centroid point of each set of positions for past questions in the past question mapping with the centroid point of each set of positions for each incorrect answer question in the incorrect answer mapping. The subject analysis means 70 may also perform analyses other than those described above. For example, the subject analysis means 70 receives the past question mapping obtained by the first dimensionality reduction means 30 and the incorrect answer mapping obtained by the second dimensionality reduction means 50 as input, compares the two to analyze the subject's learning tendencies, and reflects the analysis results in learning support. The subject analysis means 70 may include a function to perform one or more of the following processes.

[0049] Center of gravity comparison process The subject analysis means 70 calculates the centroid of the set of positions of all past questions in the past question mapping and the centroid of the set of positions of incorrect answers in the incorrect answer mapping, and measures the distance (or direction) between them. This distance (and direction) serves as an indicator of how biased the subject's tendency to make incorrect answers is within the entire scope of the exam. For example, if the difference in centroids is heavily biased in a particular direction, the subject analysis means 70 can determine that the subject has weaknesses in a particular field (e.g., anatomy, pharmacology, diagnostic imaging, etc.).

[0050] Clustering analysis The target analysis method 70 clusters the question sets on the past question mapping according to subject, subject area, etc., and analyzes which clusters the incorrect answers are concentrated in. This allows the target analysis method 70 to specifically identify "which areas of weakness exist," rather than simply identifying who is "weak."

[0051] Time course analysis The subject analysis method 70 accumulates error mappings in chronological order when subjects repeatedly take practice tests and tracks their changes. For example, it can visualize learning progress, such as a high number of errors in "basic subjects" in the early stages and a shift to "clinical subjects" in the later stages. This function is useful for adjusting learning plans and checking the progress of overcoming weaknesses.

[0052] Difficulty analysis method The learning support system 10 may also include an analysis function that considers whether the questions answered incorrectly are of high or low difficulty within the overall exam. For example, if a subject answers a difficult question incorrectly that many test-takers also answer incorrectly, it may not be considered a serious weakness, while if the incorrect answers are concentrated on basic questions, it may be judged as a significant weakness. The subject analysis means 70 may also have such a difficulty level analysis means.

[0053] Method of presenting analysis results The analysis results from the subject analysis means 70 may be output in the following form. For example, the subject analysis means 70 may connect the "center of error" and the "center of past questions" on the error answer mapping with lines and output their direction, distance, and vector. The target analysis method 70 may also output a "List of Areas with Concentrated Incorrect Answers" in a table format, showing the area name, number of incorrect answers, and percentage relative to the total number of questions. The participant analysis tool 70 may also output specific feedback messages as learning advice, such as, "Since the rate of incorrect answers in the field of clinical radiology is high, we recommend reviewing related questions."

[0054] With the above configuration, the target user can objectively and quantitatively understand their weak areas and create an efficient review plan. The learning support system 10 functions as a computer-implemented system to support the user's learning in this way. Unlike conventional analysis that relies solely on "correct answer rate display," by utilizing distribution information on the mapping space, it becomes possible to intuitively understand learning biases. Furthermore, by using the learning progress trend analysis function, the degree of learning improvement can be tracked over time, which also contributes to maintaining motivation. Furthermore, the learning support system 10 may obtain past question mappings in advance, and when information regarding incorrect answers to questions related to a specific exam for the target individual is entered, it may obtain incorrect answer mappings in the same or similar manner as when the past question mappings were obtained. Incorrect answers to questions related to a specific exam for the target individual refer to the question number or question text of the questions related to the specific exam that the target individual answered incorrectly. The questions related to the specific exam may be from mock exams related to the specific exam, or they may be past questions.

[0055] The learning support system 10 may be configured to output similar past questions extracted from the incorrect answer mapping in a format that can be stored in an existing national examination question database system. In this case, the learning support system 10 of the present invention can output the incorrect answer-related questions extracted by the incorrect answer-related question output means 60 in a format that can be stored in an existing national examination question database system, rather than merely displaying them on the screen. Specifically, the incorrect answer-related question output means 60 extracts past questions that are located near the incorrect answer questions on the incorrect answer mapping, and generates an "incorrect answer-related question data record" that includes the question text, answer choices, correct answer, and explanation information, as well as additional metadata such as a similarity score, extraction date and time, and target person ID. This incorrect answer-related question data record is output in a format such as the following.

[0056] (1) Structured data format The data records related to incorrect answers are converted to JSON (JavaScript Object Notation), XML (Extensible Markup Language), or CSV (Comma Separated Values) format. This makes it easy to import the data into external national examination question databases.

[0057] (2) Examples of data items The incorrect answer-related question data record includes, for example, the following fields: Question ID (a unique identifier that can be associated with a key in an existing database), Question text (text data), Choice options (multiple choices stored in an array format), Correct answer information, Explanation information, Classification (subject name, field name, difficulty level, etc.), Similarity score (a numerical value indicating the semantic distance from the incorrect answer question), Source incorrect answer ID, Extraction date and time, and Target ID (anonymized user identifier).

[0058] (3) Database storage means The generated incorrect answer-related problem data records are sent to an external national examination question database system via an API (Application Programming Interface) and stored according to the database's existing schema. This allows the incorrect answer-related problems extracted by the learning support system 10 to be integrated and used in the conventional national examination question search and management system.

[0059] (4) Usage form Users can retrieve incorrect answer-related questions stored in the national examination question database from the search screen and use them as a set of questions to overcome weaknesses in specific areas. Furthermore, it is possible to automatically generate "incorrect answer-related question sets" based on the similarity score output by the learning support system 10 and reconstruct them as mock exams on the national examination question database. This configuration allows for easy integration of incorrect answer-related questions uniquely extracted by the learning support system 10 into existing national examination question database systems. This ensures compatibility between conventional systems and the present invention system, making individually optimized question sets tailored to the user's error tendencies available even in conventional question search environments.

[0060] The learning support system 10 of the present invention can automatically generate new similar problems using a generation AI (for example, a natural language generation algorithm including a large-scale language model) based on existing similar past problems extracted by the incorrect answer related problem output means 60, and add such problems to the national examination problem database. Specifically, the generation AI processing unit takes the past question data records extracted by the incorrect answer related question output means 60 as input and performs the following processing. (1) Input processing The extracted question text, answer choices, correct answer information, explanation information, and classification information (subject, field, etc.) are provided to the generating AI as prompts. (2) Problem generation The generative AI, based on the semantic features and classification information of the input problem, Paraphrasing problems that involve changing the vocabulary and expressions of the original text. Applied problems that set up new case studies belonging to the same field. Alternative answer formats with changes in question format (e.g., multiple-choice → fill-in-the-blank), These will be newly generated.

[0061] (3) Quality control The generated problems are checked by an automated verification module. Specifically, Grammatical consistency (verification through syntactic analysis), consistency between the question's intent and the correct answer choices, You may evaluate the similarity to past questions (threshold management) and only include questions that meet the criteria. (4) Database addition format Newly generated similar questions are converted into a data format (JSON, XML, CSV, etc.) that can be stored in the existing national examination question database and saved as a data record containing the following items: Question ID (automatically assigned), Question text (content generated by the generation AI), Choice options, Correct answer information, Explanation information (generated or reused from existing information), Classification (subject, field, difficulty level), Original question ID (linked to the past question from which it was extracted), Generation flag (information identifying that it is an AI-generated question), and Generation date and time. (5) Usage form By selecting the "Learning Mode Including AI-Generated Questions" on the national examination question database, participants can tackle new, similar questions tailored to their own error tendencies. This provides additional practice opportunities that were not available through traditional past question study alone. With this configuration, the learning support system 10 not only relies on the existing past question database but can also generate new, similar questions according to the participant's error tendencies and store them in the database. As a result, the question database is dynamically expanded, creating a learning environment optimized for the participant. Furthermore, because the generating AI produces diverse expressions and applied questions, participants can improve their ability to handle unfamiliar questions in the actual exam.

[0062] This specification also provides a program for causing a computer to function as one of the learning support systems described above. This program causes a computer to function as a learning support system for a specific test by implementing each of the means described above. This specification also provides a non-temporary information recording medium that can be read by a computer storing such a program. Examples of information recording media include CDs, CD-ROMs, DVDs, USB memory sticks, hard disks, and disks on servers.

[0063] The learning support system 10 described above can be used as a computer-based learning support method for a specific test. This method includes a first vectorization step, a first dimensionality reduction step, a second vectorization step, a second dimensionality reduction step, and an error-related problem output step. A learning support method according to one embodiment of the present invention is performed using the learning support system 10 and supports the subject in effectively studying past questions for a specific test. This method includes the following steps.

[0064] (1) First vectorization process The first vectorization step is a step implemented, for example, by the first vectorization means 20, and is a step for converting past exam questions of a specific exam into high-dimensional vectors to obtain high-dimensional vectors of past exam questions. Past exam question data of a specific exam is collected, and each past exam question is converted into a high-dimensional vector by the first vectorization means 20. This step is also called the past exam question vectorization step. During the conversion, for example, natural language processing techniques can be used to extract semantic features of the question text and answer choices, and embedding models such as Word2Vec, BERT, and Doc2Vec can be applied. As a result, each past exam question is represented as a high-dimensional vector on Hilbert space.

[0065] (2) First dimensionality reduction process The first dimensionality reduction process is implemented, for example, by the first dimensionality reduction means 30, and is a process to perform dimensionality reduction on the high-dimensional vectors of past questions to obtain a visually recognizable past question mapping. This process is also called the past question mapping generation process because it generates a past question mapping. In this process, the obtained set of past question vectors is input to the first dimensionality reduction means 30 and projected onto a low-dimensional space. This provides a "past question mapping" that is visually recognizable to the target user. PCA, t-SNE, UMAP, etc. can be used for dimensionality reduction, and sets of questions with similar content are placed close together on the mapping.

[0066] (3) Second vectorization process The second vectorization step is a step implemented, for example, by the second vectorization means 40, and is a step to convert incorrect answers to questions related to a specific test for the subject into high-dimensional vectors, thereby obtaining high-dimensional vectors of the incorrect answers. This step is also called the incorrect answer vectorization step. When the subject answers practice problems, the incorrect answers are input to the second vectorization means 40, and those problems are converted into high-dimensional vectors. This yields the incorrect answer problem vectors.

[0067] (4) Second dimensionality reduction process The second dimensionality reduction process is implemented, for example, by the second dimensionality reduction means 50. This process involves performing dimensionality reduction on the high-dimensional vectors of incorrect answers and overlaying the information of the incorrect answers onto the past question mapping to obtain an incorrect answer mapping. This process is also called the incorrect answer mapping generation process. The second dimensionality reduction means 50 projects the incorrect answer question vectors into the same low-dimensional space as the past question mapping and overlays them onto the past question mapping to generate the "incorrect answer mapping". Incorrect answers are displayed with markers of different colors and shapes than the past questions, allowing the user to intuitively grasp their own tendency to make mistakes.

[0068] (5) Process for outputting incorrect answer-related questions The incorrect answer-related problem output process is a process implemented, for example, by the incorrect answer-related problem output means 60, and is a process for extracting past questions on the past question mapping that are close to the position of the incorrect answer question in the incorrect answer mapping, and outputting the incorrect answer-related problems. The incorrect answer-related problem output means 60 searches for past questions that are close to the position of the incorrect answer question in the incorrect answer mapping and extracts them as "incorrect answer-related problems". The extracted problems are preferably either actual past questions from a specific exam, or similar questions generated based on past questions. By reviewing these, the target user can efficiently learn a group of problems that are conceptually similar to the questions they answered incorrectly.

[0069] (6) Target analysis process The target analysis process involves comparing the centroid of each set of locations for past questions in the past question mapping with the centroid of each set of locations for incorrect answers in the incorrect answer mapping to analyze the target individuals. Other analyses may also be performed in the target analysis process. The target analysis means 70 quantitatively evaluates the target individuals' weak areas and level of understanding by comparing the distribution and centroid of the incorrect answer mapping with the past question mapping. Furthermore, if incorrect answers are concentrated in a particular area, feedback is provided to encourage reviewing that area intensively. In addition, by accumulating the incorrect answer mapping chronologically, the progress of learning and the effectiveness of overcoming weaknesses can be visualized.

[0070] (7) Output process The results of the above processing are output to the display unit 13. Examples of the information output include past question mapping, incorrect answer mapping, a diagram (scatter plot) overlaying the past question mapping and incorrect answer mapping, a list of related incorrect answers, analysis results, and learning advice based on the analysis results (e.g., "Incorrect answers are concentrated in the clinical pharmacology field. Please prioritize reviewing related questions."). The target user can refer to this while reviewing.

[0071] (8) Iterative learning process The participants re-solve the incorrect answer-related problems presented to them and input the results back into the learning support system 10, thereby updating the incorrect answer mapping and analysis. Through this iterative process, participants can overcome their weaknesses and progress in their learning. According to the learning support method described above, learning is not based solely on true / false judgments, but rather on systematic review that considers the semantic relationships between problems. As a result, students can efficiently reinforce their weaknesses and increase their chances of passing specific examinations (e.g., the National Dental Examination, the National Medical Examination). [Examples]

[0072] As one embodiment of the present invention, a method for applying the learning support system 10 to studying for the national dental licensing examination will be described. (1) Preparation of past exam question database First, all questions (approximately 3,000 questions) from the past 10 years of the National Dental Examination are input into the learning support system 10. The first vectorization means 20 converts each question and answer choice into a high-dimensional vector based on a natural language processing model. For example, using an embedding model based on a Transformer-type language model, each question is represented as a feature vector of approximately 1,500 dimensions.

[0073] (2) Generation of past exam question mapping The first dimensionality reduction means 30 projects approximately 1500-dimensional vectors into a two-dimensional space. Here, t-SNE is adopted with an emphasis on preserving the relationships between subjects in the national dental licensing examination. As a result, questions from each field, such as oral anatomy, dental materials science, prosthodontics, and oral surgery, naturally form clusters on the mapping. The display unit 13 presents each field as a scatter plot with different colors.

[0074] (3) Exercises by the target participants and error vectorization The participant logs into the learning support system 10 and answers randomly generated practice exam questions (including past questions). Questions that the participant answers incorrectly are similarly converted into vectors of approximately 1500 dimensions by the second vectorization means 40.

[0075] (4) Generation of incorrect answer mappings The second dimensionality reduction means 50 projects the incorrect answer vectors into two dimensions and displays them superimposed as red markers on the past question mapping. In this case, the past question mapping can be displayed as, for example, a gray marker, as the background for the incorrect answer mapping. The user can intuitively confirm, for example, that their incorrect answer questions are concentrated in the "dental materials science" field.

[0076] (5) Extraction of questions related to incorrect answers The incorrect answer related question output means 60 searches for past questions located near the incorrect answer question marker and lists highly relevant questions. For example, if the subject gives an incorrect answer to "Operating procedures for amalgam fillings," questions related to "Operating procedures for composite resin restorations" and "Principles of cavity preparation" located nearby are extracted as incorrect answer related questions and presented to the subject.

[0077] (6) Participant analysis The subject analysis method 70 compares the overall centroid of the past exam question mapping with the centroid of the incorrectly answered questions. As a result, it was found that the centroid of incorrect answers among the subjects was heavily skewed towards the "Dental Materials Science" cluster. Furthermore, the analysis results showed that the proportion of incorrectly answered questions in the "Dental Materials Science" field was 20%, which is twice the overall average.

[0078] (7) Provide feedback The display unit 13 shows the following information: - Past exam question mapping (display of clusters by subject area + red marker for incorrectly answered questions) - List of incorrect answer-related questions (e.g., "Cavity preparation," "Adhesive resin," "Properties of dental alloys") - Study advice (Example: "There are many incorrect answers in the dental materials science field. In particular, your understanding of restorative materials is insufficient, so please review related problems in detail.")

[0079] (8) Repetitive learning Participants re-solve the presented incorrect answer-related problems and input the results back into the system. The process of decreasing incorrect answers is reflected in the incorrect answer mapping, and the bias in the center of gravity gradually shrinks. This change is recorded chronologically, making the progress of learning visible. According to this implementation, participants can intuitively grasp their weak areas in the national dental licensing examination and review them efficiently. In particular, by presenting incorrect answer-related problems based on the semantic relationships between problems, it is possible to promote a systematic understanding that cannot be obtained by simply repeating incorrect answers. [Examples]

[0080] As another embodiment of the present invention, a method for applying the learning support system 10 to language proficiency test preparation will be described. Here, the English language proficiency test (TOEFL iBT) will be used as an example. (1) Preparation of past exam question database The learning support system 10 is input with reading, listening, speaking, and writing questions from official TOEFL iBT practice tests from the past several years. The first vectorization means 20 converts each question, question, and answer choice into a high-dimensional vector. For reading and listening questions, text embedding based on BERT or Sentence-BERT is used, while for speaking and writing, semantic vectors are constructed by referring to model answers and scoring criteria. As a result, each question is represented as a high-dimensional or infinite-dimensional vector on a Hilbert space that reflects its lexical, syntactic, and semantic characteristics.

[0081] (2) Generation of past exam question mapping The first dimensionality reduction means 30 reduces problem vectors that extend to high dimensions or infinite dimensions to two or three dimensions and presents them as a mapping on the display unit 13. On the mapping, the sections "Reading," "Listening," "Speaking," and "Writing" form clusters, and further sub-clusters such as "Academic Papers" and "General Topics" within Reading, and "Lectures" and "Conversations" within Listening are naturally distributed.

[0082] (3) Exercises for the target participants and error vectorization The target user takes a mock exam via the system, and questions for which they answered incorrectly or received a low score are vectorized by the second vectorization means 40. For example, if a user fails to understand a conversation in a listening comprehension question, that question is extracted as a vector with characteristics such as "everyday conversation" or "short sentence comprehension."

[0083] (4) Generation of incorrect answer mappings The second dimensionality reduction method 50 projects the vectors of incorrect answers onto the same space as the past question mapping and displays them superimposed as red markers. This allows the user to visually understand that their incorrect answers are concentrated in "listening conversation questions" and "reading vocabulary inference questions."

[0084] (5) Extraction of questions related to incorrect answers The incorrect answer related question output means 60 searches for past questions near the incorrect answer question marker and extracts and presents similar questions. For example, if a participant answers an incorrect "reading vocabulary question," nearby "synonym selection" and "meaning estimation from context" questions will be listed as incorrect answer related questions. Also, if the incorrect answer was a listening question, nearby questions with similar speed and pronunciation tendencies will be included, allowing the participant to practice questions that directly address their weaknesses.

[0085] (6) Participant analysis The participant analysis method 70 calculates the centroid of the incorrect answer mapping and compares it with the overall distribution of past question mappings. For example, if the centroid of incorrect answers is significantly skewed towards the "listening-conversation cluster," it is inferred that there is a "lack of understanding of fast responses and abbreviated expressions specific to conversational formats." Furthermore, if incorrect answers are concentrated in "sentence structure selection for writing tasks," it is judged that there is a problem with grammatical application ability. These analysis results are quantified and presented to the participants.

[0086] (7) Provide feedback The following information is displayed on the display unit 13. Error distribution map by section in error mapping A list of extracted questions related to incorrect answers (e.g., "Academic reading reasoning questions," "Listening short conversation questions") Study advice (Example: "You are making a lot of mistakes on the listening comprehension conversation questions. We recommend repeatedly answering similar questions to get used to the audio speed.")

[0087] (8) Repetitive learning and progress management Participants re-solve the problems related to their incorrect answers and re-enter the results into the system. As the number of incorrect answers decreases, the red markers on the incorrect answer mapping disappear, and the center of error approaches the overall center of error. This change is recorded chronologically, and the process of the participant overcoming their weaknesses is presented as a graph. According to this embodiment, participants can intuitively grasp their weak areas in language proficiency tests (e.g., vocabulary inference, conversational listening, grammatical application, etc.) and systematically review problems related to their incorrect answers. In particular, by extracting problems related to incorrect answers using semantic proximity between problems, learning that focuses on weak areas rather than mere rote practice is realized. This increases the participant's scoring ability and their chances of passing the test. [Industrial applicability]

[0088] Since this invention relates to a learning support system, it can be used in various learning support institutions and the information provision industry. [Explanation of symbols]

[0089] 10…Learning support system 11...Storage section 11a... Past Exam Database 11b…Participant performance database 11c...Model parameter group storage unit 12…Processing Unit 13…Display section 14…Input / output section 15… Communications Department 20...First vectorization means 30...First dimensionality reduction method 40...Second vectorization means 50...Second dimensionality reduction method 60…Incorrect answer related problem output method 70…Methods for analyzing target audience 80...Display control means

Claims

1. A computer-based learning support system for specific exams, A first vectorization means for converting past exam questions of the aforementioned specific examination into high-dimensional vectors and obtaining high-dimensional vectors of past exam questions, A first dimensionality reduction means for performing dimensionality reduction processing on the high-dimensional vector of the aforementioned past exam questions to obtain a visually recognizable past exam question mapping, A second vectorization means for converting incorrect answers to the aforementioned specific test questions of the subject into high-dimensional vectors, and obtaining high-dimensional vectors of the incorrect answers, A second dimensionality reduction means for obtaining an incorrect answer mapping, which involves performing a dimensionality reduction process on the high-dimensional vector of the incorrect answer question and superimposing the information of the incorrect answer question onto the past question mapping, An error-related question output means extracts past questions from the past question mapping that are close to the position of the error-answered question in the error-answer mapping and outputs error-answer related questions, A learning support system having, A learning support system further comprising a subject analysis unit that analyzes the subject by comparing the centroid point of each set of positions for past questions in the past question mapping with the centroid point of each set of positions for each incorrect answer question in the incorrect answer mapping.

2. A learning support system according to claim 1, A learning support system in which the past exam question mapping and the incorrect answer mapping can be displayed by a display unit.

3. A learning support system according to claim 1, A learning support system in which the vectors of the past exam questions and the vectors of the incorrect answers are infinite-dimensional vectors in Hilbert space.

4. A learning support system according to claim 1, wherein the high-dimensional vectors of past questions and the high-dimensional vectors of incorrect answers belong to a real inner product space, and the similarity is calculated by the inner product or cosine similarity.

5. A learning support system according to claim 4, A learning support system in which the actual inner product space is a Hilbert space complete with respect to the norm induced by the inner product.

6. A learning support system according to claim 1, A learning support system in which the higher-order vectors from past exam questions and the higher-order vectors from incorrectly answered questions are represented as vectors expanded in an infinite-dimensional Hilbert space, the results of this expansion are projected onto predetermined finite-dimensional axes, and further visualized by two-dimensional mapping.

7. A learning support system according to claim 1, The aforementioned incorrect answer-related questions are (1) past questions from the aforementioned specific examination, (2) A learning support system consisting of similar questions created using past questions from the aforementioned specific examination.

8. A learning support system according to claim 1, A learning support system in which the aforementioned specific examination is the national examination for dentists or the national examination for physicians.

9. Computers, A learning support system for a specific exam, A first vectorization means for converting past exam questions of the aforementioned specific examination into high-dimensional vectors and obtaining high-dimensional vectors of past exam questions, A first dimensionality reduction means for performing dimensionality reduction processing on the high-dimensional vector of the aforementioned past exam questions to obtain a visually recognizable past exam question mapping, A second vectorization means for converting incorrect answers to the aforementioned specific test questions of the subject into high-dimensional vectors, and obtaining high-dimensional vectors of the incorrect answers, A second dimensionality reduction means for obtaining an incorrect answer mapping, which involves performing a dimensionality reduction process on the high-dimensional vector of the incorrect answer question and superimposing the information of the incorrect answer question onto the past question mapping, An error-related question output means extracts past questions from the past question mapping that are close to the position of the error-answered question in the error-answer mapping and outputs error-answer related questions, A learning support system having, A program for functioning as a learning support system, further comprising a subject analysis means for analyzing the subject by comparing the centroid point of each set of positions for past questions in the past question mapping with the centroid point of each set of positions for incorrect answers in the incorrect answer mapping.

10. A non-temporary information recording medium that can be read by a computer storing the program described in Claim 9.

11. A computer-based learning support method for a specific exam, A first vectorization step to convert past exam questions of the aforementioned specific exam into high-dimensional vectors and obtain high-dimensional vectors of past exam questions, A first dimensionality reduction step is performed on the high-dimensional vector of the aforementioned past exam questions to obtain a visually recognizable past exam question mapping. A second vectorization step to obtain high-dimensional vectors of incorrect answers to the aforementioned specific test for the subject, A second dimensionality reduction step to obtain an incorrect answer mapping is performed by applying a dimensionality reduction process to the high-dimensional vector of the incorrect answer question and superimposing the information of the incorrect answer question onto the past question mapping. An error-related problem output step which extracts past questions from the past question mapping that are close to the position of the incorrect answer question in the incorrect answer mapping and outputs the incorrect answer-related problem, A subject analysis step involves comparing the centroid point of each set of positions for past questions in the past question mapping with the centroid point of each set of positions for each incorrect answer question in the incorrect answer mapping to analyze the subject, Learning support methods, including those mentioned above.

Images