Learning support system using search expansion generation

Abstract

A learning support system using search expansion generation is provided. [Solution] A learning support system that is equipped with a test question provision means, a page image acquisition means, a page text database recording means, a feature vector database recording means, a related feature vector extraction means, a duplicate reference address removal means, and a related text transcription means, and that uses search expansion generation to present image data that is difficult to convert to text to the user and enable additional information to be added to the question.As needed for learning support, multiple RAG databases are constructed for each related field to increase the storage capacity for additional information, and efficiency is improved by narrowing the target of the RAG search.Based on the group of chunks obtained by the RAG search, the system references the text or image of the entire original document page where related content is written, eliminating any missed references and reflecting background information and related information that may be written around the searched chunks in the context of the question.

Description

[Technical Field]

[0001] The present invention relates to aiding learning using search expansion generation, and in particular to understanding multiple choice exam questions. [Background technology]

[0002] In education, it is common practice to administer tests after lectures to test students' understanding of the content. Students who score low are required to undergo additional study to make up for their lack of understanding, while students who score high are required to receive additional study materials to further deepen their understanding. The application of machine learning is expected to address this issue. Machine learning has made remarkable progress in recent years, with large-scale language models (LLMs) becoming increasingly popular. Neural network parameters numbering in the tens of billions to trillions are trained on massive amounts of data amounting to terabytes, enabling tasks such as translation, speech and image recognition, and text summarization. Generative AI, which generates images, music, and documents based on prompts, is also becoming more practical. It has become known that performance improves by scaling models (scaling laws), and models are becoming increasingly large in scale.

[0003] As the range of applications of large-scale language models expands, knowledge data from various fields is required. Furthermore, with social conditions and technological advances in each field, there is a need to constantly incorporate the latest knowledge data. However, training large amounts of data requires large-scale computational resources, enormous amounts of electricity, and costs. For this reason, large-scale language models cannot be reconstructed frequently. When large-scale language models are queried, they can generate answers that are not based on facts, even if they do not have the necessary knowledge stored internally. This phenomenon, known as hallucination, has limited their widespread use in fields such as medicine, where errors can lead to serious accidents.

[0004] In order to utilize large-scale language models based on the latest information in the field, it is necessary to obtain additional up-to-date information, and two approaches are currently being used. The first is called additional training (fine tuning) or transfer learning, in which additional information is used to train part of the output layer of an existing large-scale language model (fine tuning) or only the final layer (transfer learning) to create a specialized large-scale language model. This is highly useful as it results in a domain-specific large-scale language model. However, while it is not as costly as training a large-scale language model from scratch, it does require a certain amount of cost and technical skill for training.

[0005] Another approach is the RAG Retrieval-Augmented Generation approach used in this invention. The large-scale language model itself is not modified. Additional information is stored in a separate database (the RAG database). The information necessary to solve a question is searched for in the RAG database, and the obtained information is added to the question to obtain an answer from the large-scale language model. While there is a large amount of potentially additional data, the additional information necessary to answer a given question is limited. Therefore, only this limited additional information is searched for in the RAG database, and the extracted additional information (context) is added to the question in the question input box of the large-scale language model. This approach does not involve any learning changes to the large-scale language model itself; it simply adds context information to the question input box, thereby reducing costs and technical hurdles. Furthermore, it offers excellent responsiveness, allowing for the immediate reflection of urgent information, such as information on emergency drug side effects.

[0006] The RAG database is created by cutting the document information to be added into small fragments (chunks), converting each chunk into a feature vector (chunk vector), and storing it in a database. When asking a question, the question itself is also vectorized, and chunk vectors with a high similarity to the question vector are searched for in the RAG database, and the contents of the obtained chunks are added to the question. This makes it possible to respond to specific fields and the latest information without the need for expensive additional training of large-scale language models, and by limiting the basis for responses to questions to the RAG database, it is possible to prevent hallucinations. Prior art documents relevant to this application include the following: [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] https: / / ja.wikipedia.org / wiki / %E5%A4%A7%E8%A6%8F%E6%A8%A1%E8%A8%80%E8%AA%9E%E3%83%A2%E3%83%87%E3%83%AB [Non-patent document 2] https: / / www.idnet.co.jp / column / page_308.html Summary of the Invention [Problem to be solved by the invention]

[0008] In order to understand learning topics, it is necessary to collect information scattered across numerous books and documents. Traditionally, teachers manually collected information for each learning topic, then edited it into a format that students could easily understand and provided it to them. However, the number of topics to be studied is enormous, and it is difficult to say that sufficient information is currently being provided for all topics. To solve this problem, it is expected that large-scale language models using search expansion and generation will be utilized. However, when actually constructing a RAG database for learning support and attempting to search for information related to a learning item, there are many unresolved issues, such as (1) the RAG database itself becomes huge, resulting in a relatively large amount of data unrelated to the question, (2) the content to be searched is not always properly contained within the chunk, resulting in missed searches of data related to the question, (3) background information and related information that may be written around the searched chunk is omitted, and (4) image recognition and search are still insufficient in current large-scale language models. Furthermore, repeated RAG searches for similar questions are a waste of resources such as money and time. Furthermore, exams often present multiple options and require students to select the appropriate answer. The explanations for all the options contain a mixture of different information, which not only increases the volume of explanations but also makes it difficult to focus on key points.

[0009] The present invention has been made to solve these conventional problems, and its purpose is to provide a learning support system that uses search expansion and generation, such as: building multiple RAG databases for each related field as needed to increase the storage capacity of additional information, and narrowing the scope of RAG searches to improve efficiency; referencing the text or images of the entire original document page where related content is described based on the chunks obtained from the RAG search, thereby eliminating missed references, and reflecting background and related information that may be written around the searched chunks in the context of the question; presenting image data that is difficult to convert to text to the user, allowing them to add it to the question; breaking down multiple-choice questions into individual options and generating answers to enable focused explanations; and storing records of questions and answers in a database that can be searched, eliminating the need for costly and time-consuming re-generation of answers. [Means for solving the problem]

[0010] In order to achieve the above object, the learning support system using search expansion generation as set forth in claim 1 comprises a test question providing means for recording test questions in a large-scale language model and providing the test questions as questions, (1) a page image acquisition means for acquiring an image of each page of an information source to be referenced separately from the large-scale language model during inference, together with a reference address to the image; (2) The above Individual page images a page text database recording means for extracting the text of the character string from the page text database and recording it together with the reference address; (3) A feature vector database recording means for dividing the extracted text into small sections (chunks), calculating feature vectors, and recording the calculated feature vectors together with the reference addresses, (4) Learning support system using search expansion generation a related feature vector extraction means for extracting a group of feature vectors that are highly related to a given test question together with the reference address, for the given test question feature vector, (5) A duplicate reference address removing means is provided for removing duplicate reference addresses from the extracted related feature vectors; (6) After removing duplicate reference addresses The page text specified by the reference address is Page text database recorder and then transcribes the extracted text into a question input field together with the question text, The answer to the test question can be obtained by performing the above operations (1) to (6).

[0011] The learning support system using search expansion generation as described in claim 2 is characterized in that, in the learning support system using search expansion generation as described in claim 1, it is provided with a page image database recording means for recording the acquired individual page image together with a reference address to the page image.

[0012] The learning support system using search expansion generation as described in claim 3 is characterized in that, in the learning support system using search expansion generation as described in claim 1 or 2, it is provided with a page image viewing means for displaying and viewing the page image specified by the reference address.

[0013] In the learning support system using search expansion generation according to claim 4, Claim 2 In a learning support system using search expansion generation, Referral Address The page image specified by Page image database recording means After extraction from The user Enter comments for figures and tables that are not fully translated into text. Based on that input The present invention is characterized by comprising a comment transcription means for transcribing a comment into the page text.

[0014] In the learning support system using search expansion generation according to claim 5, Claim 2 a page image acquisition means for acquiring a page image, a page text database recording means for recording the page image, and a page text database recording means for recording the page text. Feature vector database recording means and Feature vector database recording means The present invention is characterized by including a multiple RAG database search means for searching and extracting related feature vectors for the target object.

[0015] The learning support system using search expansion generation according to claim 6 is the learning support system using search expansion generation according to claim 1 or 2, further comprising a question and answer recording means for recording the question sentence and the obtained answer sentence, and for a new question sentence, the question and answer recording means first searches for a question sentence similar to the new question sentence, and if a similar question sentence is found, the answer record for the similar question sentence is stored. The question It features an F&Q database that answers questions.

[0016] The learning support system using search expansion generation according to claim 7 is the learning support system using search expansion generation according to claim 1 or 2, wherein the test question providing means In addition to the function to record test questions and provide them as questions, , the problem is divided into a question and a pair of individual options. Break it down and provide it as a question The system is characterized by being equipped with a means for decomposing question sentence choice pairs. [Effects of the Invention]

[0017] The learning support system using search expansion and generation as set forth in claim 1 is provided with a test question providing means, which records test questions and provides them as questions. The page image acquisition means is provided, and an image of each page of an information source to be referenced during inference, separately from the large-scale language model, is acquired together with a reference address to the image. Since it is equipped with a page text database recording means, Individual page images Extract the text of the string from and record it along with the reference address. The feature vector database recording means divides the extracted text into small sections (chunks), calculates feature vectors, and records them together with reference addresses. Since it is equipped with a related feature vector extraction means, Learning support system using search expansion generation For the feature vectors of the test questions given in the table, a group of feature vectors that are highly related to the test questions is extracted together with their reference addresses. The duplicated reference address removing means is provided, and duplicated reference addresses are removed from the extracted related feature vectors. It has a means of transcribing related text, After removing duplicate reference addresses The page text specified by the reference address is Page text database recorder After extracting it, it is transcribed into the question input field along with the question text.

[0018] The learning support system using search expansion and generation as set forth in claim 2 includes a page image database recording means, which records each acquired page image together with a reference address to the page image.

[0019] The learning support system using search expansion and generation as set forth in claim 3 includes a page image viewing means, which displays and allows viewing of the page image specified by the reference address.

[0020] The learning support system using search expansion generation described in claim 4 is equipped with a comment transcription means, so after extracting the page image specified by the reference address from the image data recording means, comments are entered for figures or tables that have not been sufficiently converted into text and transcribed into the page text.

[0021] In the learning support system using search expansion generation according to claim 5, a page image acquisition means, a page text database recording means for the page image database recording means, and the page text database recording means for the page text database recording means. Feature vector database recording means It has multiple. Also, any of the above Feature vector database recording means It is equipped with a multiple RAG database search means for searching and extracting related feature vectors for the target.

[0022] The learning support system using search expansion and generation according to claim 6 comprises a question and answer recording means for recording the question sentence and the obtained answer sentence. Regarding the new question, the question and answer recording means is first searched for a question similar to the new question, and if a similar question is found, the answer record for the similar question is recorded. Question It has a database of FAQs that answers your questions.

[0023] In the learning support system using search expansion generation according to claim 7, the test question providing means In addition to the function to record test questions and provide them as questions, , the problem is divided into a question and a pair of individual options. Break it down and provide it as a question It is equipped with a means for breaking down question choice pairs. [Brief explanation of the drawings]

[0024] [Figure 1] 1 illustrates an embodiment of a network configuration according to the present invention. [Figure 2]This is an example of a user interface for a large-scale language model (LLM). [Figure 3] The relationship between large-scale language models and the RAG database is shown. [Figure 4] 1 shows the overall structure of the expanded RAG database of the present invention. [Figure 5] FIG. 10 is a detailed diagram of the process of constructing an extended RAG database. [Figure 6] This shows how reference addresses are numbered. [Figure 7] This is the RAG database search procedure. [Figure 8] This is an example of an exam question. [Figure 9] This is an example of breaking down an exam question into a combination of question statements and answer options. [Figure 10] Construct multiple RAG databases for each area. [Figure 11] FIG. 1 is an explanatory diagram of an F&Q database. DETAILED DESCRIPTION OF THE INVENTION

[0025] A typical system configuration of the present invention is shown in Figure 1. Large-scale language models are constructed in huge servers such as cloud data centers and are provided via the web, as they consist of huge amounts of data, numerous CPUs (Central Processing Units), GPUs (Graphics Processing Units), and a high-speed network connecting them. In schools and other educational institutions, many PC terminals are connected via a local area network (LAN) connected to the web. Educational institutions also have servers that operate school databases that manage lecture and exam content, student grades, etc. In recent years, there has been an increasing number of cases where people use mobile devices such as smartphones and tablets to access cloud services such as large-scale language models and school database servers from inside and outside educational institutions.

[0026] Servers, terminals, and mobile devices all consist of memory for recording programs and data, recording media such as hard disks that permanently record said programs and data as needed, a CPU that reads and processes said programs and data, a GPU that performs high-speed parallel processing as needed, and a communication module. As cloud services become more stable and cheaper, there are increasing cases of schools migrating some or all of their school database server functions to the cloud. Conversely, there is also a movement underway to move part or all of large-scale language models to devices with increased processing capabilities and memory capacity (edge ​​computing). Furthermore, the development of small language models (SLMs) that limit the number of parameters in a language model is also progressing. However, even though they are small, they are still sufficiently large compared to the models before the advent of large-scale language models, and any embodiment, including this form, is included in the present invention.

[0027] Figure 2 shows an example of a user interface for a large-scale language model (LLM). LLM is currently being developed rapidly, with numerous models being developed, including ChatG PT (a registered trademark of OpenAI), Bard, LaMDA (a registered trademark of Google), and LLaMA (a registered trademark of Meta). Naturally, the user interface will differ, but typically, as shown in Figure 2, it consists of a box (prompt input box) for inputting prompts to instruct and inquire of the LLM, a box (answer display box) for displaying the answers to those prompts, and a box (usage history box) for displaying the history of prompts and answers as a usage log.

[0028] Recently, in addition to using LLMs alone as described above, there has been an increase in cases where LLMs themselves have APIs (Application Programming Interfaces) and LLM functions are used from external software. In this case, prompts, answers, and history are input and output to and from external software via API, so the display format is not limited to that shown in Figure 2 and is controlled by the external software.

[0029] In large-scale language models, to represent a certain vocabulary, a one-hot vector is used, which is a long vector consisting of zeros with the same number of dimensions as the number of types of vocabulary used, with a single 1 placed in the position corresponding to the vocabulary in question. All the vocabulary in a large amount of literature is replaced with vectors of this format, and deep learning is used to determine the associations (attention) between each vocabulary vector. In response to a query (prompt), vocabulary that is likely to appear next to the query and the part of the answer that has already been generated is generated and added one by one to create an answer. If the information that serves as the basis for generating an answer is contained in the large amount of literature, it is expected that a rational, useful, and correct sentence will be generated. However, if the information on which the answer is generated is not contained in the large amount of literature, the answer generation process will proceed by mechanically adopting vocabulary with a high probability, which is known to generate false answers that are not based on any evidence (hallucination). If this hallucination occurs in settings such as medical care or education, it could put patients' lives at risk or cause students to believe incorrect information, which is one of the reasons why the application of large-scale language models to core business processes has not progressed.

[0030] Training large-scale language models requires large-scale servers equipped with a large number of parallel computing units (GPUs), as well as a large amount of electricity and money, in order to read a large amount of documents and use deep learning to find relationships between vocabulary words. New documents are generated every day, and it is not realistic to reflect all of them in a large-scale language model without any time delay. Furthermore, the large amount of documents that are scanned are mainly public documents that are available on the web, etc., but they do not include sensitive information such as internal company documents or hospital electronic medical records, so it is said that the documents that can be collected are only a small portion of the total documents that exist on Earth.

[0031] In order to utilize large-scale language models based on the latest information in the field, it is necessary to obtain additional up-to-date information, and two approaches are currently being taken. One is called additional learning (fine tuning) or transfer learning, which uses additional information to train part of the output layer of an existing large-scale language model (fine tuning) or only the final layer (transfer learning), resulting in a specialized large-scale language model. It is a domain-specific large-scale language model and is therefore highly useful, but it does require a certain amount of cost and technical skill for learning, although not as much as training a large-scale language model from scratch. Furthermore, because the additional information contains sensitive information such as personal information and descriptions of medical conditions, there is a risk that it may be accessed outside the organization. To prevent this, it is necessary to build and operate a large-scale, additionally trained language model exclusively within your own company or hospital.

[0032] The other approach is the Retrieval Expansion and Generation (RAG) approach used in this invention. The large-scale language model itself is not changed. Additional information is stored in a separate database (RAG database), and the information required to solve the question is searched for in the RAG database. The text information of the obtained chunks is added to the question, and the answer is obtained from the large-scale language model. Although there is a large amount of potentially additional data, the additional information required to answer a question is limited, so only this limited additional information is searched for in the RAG database, and the extracted additional information is added to the question (context) in the question input box of the large-scale language model. This approach does not involve any changes to the large-scale language model itself through learning, but simply adds contextual information to the question input box, which reduces costs and technical hurdles.Furthermore, it is highly responsive, as it can immediately reflect information of high urgency, such as information on emergency side effects of drugs.

[0033] Figure 3 shows the flow of large-scale language model processing using a generalized search expansion generation (RAG). A question posed by a user is received by a broker program such as a chat app (1), and information related to the question is searched against the RAG database (2). The broker program sends the search results (3) along with the query to the LLM (4), receives a response from the LLM (5), and presents the response to the user (6). However, although the length of the text string of a chunk can be changed by setting, there is a certain limit, and it may not be possible to include all the necessary information. Conversely, if a chunk is too long, the focus of the feature vector will become blurred, and there is a risk that it will not be able to respond appropriately to searches. Also, if a keyword overlaps the junction of chunks, it will be impossible to create an appropriate feature vector, making searches difficult. In addition, scanned images such as pictures and illustrations have not yet been fully recognized and converted into text, so there is a possibility that they may not be found in searches. Furthermore, building a practical RAG database would require a large scale database based on a large number of document chunks. The set of documents required varies greatly depending on the area of ​​interest. For example, the set of documents required in the medical field is very different from those in the historical or literary fields. Constructing a RAG database from a collection of documents covering all fields would result in a redundant structure, and there is a risk that it would place an unnecessary burden on database construction and search.

[0034] FIG. 4 shows the overall structure of the expanded RAG database of the present invention. FIG. 5 is a detailed diagram of the processing flow of FIG. From additional potential information sources (information resources), such as books, e-books, PDFs of web documents, etc., scanned images of books, images of PDF-converted web documents, etc., the images of individual pages (page images) are assigned a reference address such as "information resource name + page" and recorded in a database (page image database). Next, text is extracted from each page image (page text) and recorded together with the reference address (page text database). The text of each page is cut into small fragments (chunks), a feature vector is calculated for each chunk, and the reference address is assigned and recorded in the feature vector database. If necessary, for example, when there are strict restrictions on recording capacity, the page text may be compressed and recorded using a summarization function of a large-scale language model.

[0035] Here, the reference address is numbered in units of the familiar book name + page, as shown in Figure 6, but unlike books which have physical constraints, with documents on the web, there is no need to be particular about the display, so depending on the granularity of the information, it may be numbered in units of paragraphs, or conversely, in larger units such as sections or chapters. The page text can be unstructured plain text, but if the content is complex, it is preferable to use a tagged notation such as XML, JSON, or Markdown to make the document structure clear. Recommendations vary depending on the software used to calculate feature vectors. While the present invention uses JSON notation, any notation may be used.

[0036] At present, it is not possible to perfectly translate images, videos, illustrations, etc. contained in page images into text. For this reason, ultimately, judgment and understanding by human eyes is required. It is also useful for the user to add explanatory text to the corresponding page text based on the image, video, illustration, etc., as needed. Although the page image database after extracting the page text is not essential, it may be useful to refer to the page image as needed.

[0037] Figure 7 shows the search procedure for the RAG database. When a question is entered in the prompt input box, the broker program calculates the feature vector of the question and extracts a list of approximate feature vectors from the vector database (RAG database). The degree of similarity between feature vectors is often expressed by the magnitude of the inner product between the vectors (cosine similarity), but the Manhattan distance, which is the sum of the absolute values ​​of the differences between the elements of both vectors, may also be used. The threshold value of the degree of similarity that serves as the extraction standard is variable depending on the situation. If the extracted list is too large, you can narrow it down by raising the criteria, or if the list is too small, you can lower the criteria to increase the size of the list. In some cases, the criteria may be automatically adjusted so that the number of extracted items is equal to the set value, such as the top 10 items.

[0038] When a list of feature vectors corresponding to multiple chunks is extracted, conventional RAG adds the text of the chunk that each feature vector originated from to the question. However, as mentioned above, there is a limit to the length of the text string of a chunk, and it may not be possible to include all the necessary information. Furthermore, if a keyword overlaps the boundary between chunks, proper feature vectorization is not possible. Furthermore, scanned images such as pictures and illustrations may not be converted into text satisfactorily. To overcome this drawback, the present invention takes the following steps.

[0039] The reference addresses assigned to each feature vector are sorted, duplicate reference addresses are removed, and the entire page text indicated by the remaining reference addresses is transcribed into the prompt input box along with the question text, prompting the large-scale language model to answer based on this. This procedure solves the problem of the joins, which was a problem with the fragmented chunks mentioned above, where a chunk cannot contain all the necessary information, and also allows the text between chunks to be added to the question. This function is useful when educating students, as it not only shows them how to solve a specific problem, but also provides background information on the problem, encouraging deeper understanding.

[0040] If the text conversion of an image, video, illustration, etc. is insufficient, the image of the page is searched from the page image database and displayed to the user, and the user can add a description of the image, video, illustration, etc. to the question. If necessary, it can also be added to the text of the page. This allows information that has not been fully utilized in the past, such as images, videos, and illustrations, to be utilized. As described above, by using the procedure shown in Figure 7, when a question about a certain learning topic is asked, information can be collected across a wide range of information resources, including many books and documents, and provided to the LLM for organization and summary, making it easy to create explanatory texts, which previously required lengthy manual work.

[0041] Figure 8 shows an example of a typical exam question. Although it would be possible to post a question in this format and ask for an answer and an explanation, since it contains a large number of independent items, it is easy for a lengthy explanation to be generated. Furthermore, because there are countless combinations of options, there is a lot of overlap in the explanations for each question. To avoid this, as shown in Figure 9, if we break down the question into pairs of questions and individual options, and then ask questions about each option, we can expect to get focused and concise explanations (problem-option pair decomposition method). With this configuration, options similar to the option in question can be easily searched for, and confirmation questions can be easily created to check whether a student has fully understood an option that was incorrectly chosen the first time.

[0042] Building a practical RAG database would require a large scale database based on a large number of document chunks. For example, in the field of medicine, there are easily over 50 major fields, such as internal medicine, surgery, obstetrics and gynecology, and as medicine advances, there is a tendency for further subdivision. If we were to consolidate all of these into a single RAG database, it would result in a huge burden on data storage and retrieval. The required document set varies greatly depending on the field of interest. For example, even in the medical field, there is a large gap between the document sets of abdominal surgery and psychiatry. Furthermore, the set of documents required for medicine is very different from those in the fields of history and literature. Therefore, as shown in Figure 10, by constructing multiple RAG databases for each field of interest and having the broker program search one or more RAG databases related to the question as needed, it is possible to avoid inefficiencies in searches due to the RAG databases becoming too large. Of course, if the RAG database is too fragmented, it will be necessary to perform multiple searches, which will reduce search efficiency, so it goes without saying that it is necessary to build a RAG database by aggregating documents from highly related fields.

[0043] In educational settings, when generating explanatory text for answers to test questions, explanatory text based on a general RAG database is appropriate for students who answer a question incorrectly and whose grades are not considered good, but this poses the risk of not satisfying the intellectual curiosity of students with good grades. In this case, it would be useful to build a RAG database containing theoretically more advanced content, and then generate and provide explanatory text with more advanced and in-depth content from it.

[0044] Questions and answers could be created each time, but this would require a certain amount of computational resources and cost. As shown in FIG. 11, pairs of questions and answers to the questions are stored as an FAQ database in the RAG database. When a question is entered, the FAQ database is first searched for answers that are closely related to the question, and only if the user is dissatisfied with the content of the searched answer is a new answer created using the procedure of the present invention, and the result is similarly registered in the FAQ database. In this case, it is more likely that a more versatile explanation without duplication will be obtained if the question is broken down into multiple choice answers as in Figure 9 rather than the format in Figure 8. When registering this FAQ database, it is possible to either register only the feature vector of the question in the RAG database and provide a link to the answer, or to chunk the question and answer pair and register the feature vector.

[0045] Note that there is a risk of information leakage if the content entered in the prompt is used to train a large-scale language model, so it is useful to explicitly declare in the prompt that training is prohibited, or to use a paid version of a large-scale language model that guarantees that training will not be used.

[0046] Although the embodiments have been described above, the specific configuration of the present invention is not limited to the above-described embodiments, and the present invention also includes design changes and the like that do not deviate from the gist of the invention. For example, although questions are provided from the test question providing means, test questions or study items may also be entered directly into the prompt input box. While the present invention has been described primarily with reference to a chunk feature vector database, the present invention also includes the addition of information useful for generation, such as information extracted from a normal relational database using SQL statements or reference information from the Web, to the prompt as appropriate, in addition to the feature vector database.

Claims

1. In the large-scale language model, a test question providing means is provided for recording test questions and providing them as questions; (1) a page image acquisition means for acquiring an image of each page of an information source to be referenced separately from the large-scale language model during inference, together with a reference address to the image; (2) a page text database recording means for extracting text strings from the image of each page and recording the extracted text together with the reference address; (3) A feature vector database recording means for dividing the extracted text into small sections (chunks), calculating feature vectors, and recording the calculated feature vectors together with the reference addresses, (4) A related feature vector extraction unit extracts a group of feature vectors that are highly related to a test question, together with the reference addresses, from the feature vectors of a test question provided to the learning support system using search expansion and generation, (5) A duplicate reference address removing means is provided for removing duplicate reference addresses from the extracted related feature vectors; (6) A related text transcription means is provided for extracting page text designated by the reference address after removing duplicate reference addresses from the page text database recording means, and then transcribing the page text together with the question text into a question input field; A learning support system using search expansion and generation, characterized in that answers to the test questions are obtained by the operations (1) to (6) above.

2. 2. A learning support system using search expansion and generation according to claim 1, further comprising a page image database recording means for recording each of the acquired page images together with a reference address to the page image.

3. 3. A learning support system using search expansion and generation according to claim 1, further comprising page image viewing means for displaying and viewing the page image designated by said reference address.

4. 3. The learning support system using search expansion generation according to claim 2, further comprising a comment transcription means for, after extracting a page image specified by a reference address from a page image database recording means, a user inputting a comment for a figure or table that has not been sufficiently converted into text, and transcribing the comment into the page text based on the input.

5. 3. A learning support system using search expansion and generation as claimed in claim 2, further comprising: a plurality of page image acquisition means, a plurality of page text database recording means for said page image database recording means, and a plurality of feature vector database recording means for said page text database recording means; and a plurality of RAG database search means for searching for and extracting related feature vectors from any of said feature vector database recording means.

6. 3. A learning support system using search expansion and generation as claimed in claim 1 or 2, further comprising a question and answer recording means for recording the question sentence and the obtained answer sentence, and for a new question sentence, first searching the question and answer recording means for a question sentence similar to the new question sentence, and if a similar question sentence is found, a F&Q database for taking the answer record to the similar question sentence as the answer to the question sentence.

7. 3. A learning support system using search expansion and generation as described in claim 1 or 2, characterized in that the test question providing means, in addition to the function of recording test questions and providing them as questions, is provided with a question statement option pair decomposition means for decomposing questions into pairs of question statements and individual options and providing them as questions.

Images