Automatic question generation service system based on unstructured data

Abstract

Disclosed is an automatic question generation service system comprising: a mobile terminal; and a control server, which receives a document for generating questions through the mobile terminal, analyzes the document so as to form a document vector database of a language model for generating questions, and uses the language model so as to generate questions based on the document vector database.

Description

Automatic question generation service system based on unstructured data

[0001] The present invention relates to a service system that generates and provides questions based on a language model for unstructured data.

[0002] Recently, natural language generation technology is being applied to various applications that generate natural language using neural networks and support conversations between device terminals and users.

[0003] These neural networks are models that mathematically model the characteristics of human biological nerve cells, and use algorithms that mimic the human ability to learn.

[0004] Additionally, neural networks have the ability to generalize, allowing them to produce relatively correct outputs for input patterns that were not used in learning, based on learned results.

[0005] NLU module, Policy module, NLG module, and ChatLLM are well known in relation to artificial intelligence, natural language processing, and chatbots.

[0006] NLU modules often rely on rules based on word matching, such as regular expressions and morphological analysis, to understand customer speech. They also utilize AI models based on Bert (Bidirectional Encoder Representations from Transformers). While rule-based modules, like these, have clear criteria and can create controllable systems, they face limitations in processing the vast amount of natural language with all these rules.

[0007] For the Bert base model, the direction of improvement is clear due to the use of a smaller model compared to LLM.

[0008] Additionally, the Policy module is a module that provides directions on what kind of utterances should be made. Due to the rule base, it can provide appropriate directions based on conditions based on previous customer utterances and the current status, or it can utilize a reinforcement learning AI model.

[0009] In the case of rule-based systems, controllability is possible, but there is a disadvantage in that all conditional expressions must be directly modified in code whenever the scenario changes, and AI models based on reinforcement learning show high performance, but there is a problem in that accuracy decreases when the conversation length increases.

[0010] Regarding the NLG module, existing chatbots are structured as systems that output one of n candidate sets matching a policy once it's determined. Because this process is independent of the conversation, it's impossible to generate natural sentences.

[0011] ChatLLM can produce natural speech, but it is difficult to measure the reliability of the predicted behaviors of the step-by-step model, which can lead to hallucinations or incorrect information.

[0012] Meanwhile, the Retrieval-Augmented Generation (RAG) of a language model generally configures the system to search for documents and have the language model provide answers based on the search results, but there was a limitation in that it was difficult to obtain the exact documents required for answers through document search.

[0013] One aspect of the present invention discloses a service system that creates and provides questions in various forms through file upload or natural language input, and presents documents that serve as the basis for the questions.

[0014] The system of the present invention includes a mobile terminal; and a control server that receives a document for generating a question through the mobile terminal, analyzes the document to form a document vector database of a language model for generating a question, and generates a question based on the document vector database using the language model.

[0015] Meanwhile, the control server may include a document vector database configuration unit that generates a content vector, an attribute vector, and a document question vector for the document, and stores them in the document vector database so that each vector can have the ID of the document as a key; a question generation unit that generates a question based on the document vector database using a language model and outputs the question to the mobile terminal; and a question explanation generation unit that extracts a document that serves as a basis for the question output by the question generation unit and provides the document to the mobile terminal.

[0016] Meanwhile, the document vector database configuration unit includes a content vector generation unit that creates and embeds text combining attribute values ​​specified by a user; an attribute vector generation unit that creates and embeds text combining the title and attribute name of the document according to an attribute vector template that defines a method of combining the title and attribute name of the document; and a document question vector generation unit that embeds question text directly registered by the user for the document.

[0017] The above question generation unit includes a user question vector generation unit that generates a user question vector by embedding a user question text input through the mobile terminal; a vector search unit that extracts at least one of a content vector, an attribute vector, and a document question vector corresponding to the user question vector from the document vector database; and an answer generation unit that inputs at least one of the content vector, the attribute vector, and the document question vector extracted from the vector search unit into a language model and outputs an answer to the user question text input through the mobile terminal.

[0018] The above question explanation generation unit can extract a document that serves as the basis for an answer output by the answer generation unit by using the ID of a document stored in at least one vector among the content vector, attribute vector, and document question vector extracted by the vector search unit.

[0019] According to one aspect of the present invention described above, by generating and providing various types of questions for files or natural language, it can be easily utilized in various fields such as employment, education, entertainment, and research, for example, judging the authenticity of self-introductions, creating test questions based on textbooks and self-study books, creating problems in games, and creating review questions for research projects.

[0020] Figure 1 is a conceptual diagram of a system according to one embodiment of the present invention.

[0021] Figure 2 is a conceptual diagram of a control server according to one embodiment of the present invention.

[0022] The advantages and features of the present invention, and the methods for achieving them, will become clearer with reference to the embodiments described in detail below together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various different forms. These embodiments are provided only to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims. Like reference numerals designate like elements throughout the specification.

[0023] The terminology used herein is for the purpose of describing embodiments and is not intended to limit the present invention. In this specification, the singular also includes the plural unless specifically stated otherwise. As used herein, the terms "comprises" and / or "comprising" do not exclude the presence or addition of one or more other components, steps, or operations.

[0024]

[0025] Figure 1 is a conceptual diagram of a system according to one embodiment of the present invention.

[0026] Referring to FIG. 1, a system (1) according to one embodiment of the present invention may include a control server (100) and a mobile terminal (200).

[0027] The control server (100) can execute the automatic question generation service according to the present invention.

[0028] The control server (100) may be its own server for the automatic question generation service according to the present invention, a cloud server, or a p2p (peer-to-peer) collection of distributed nodes.

[0029] The control server (100) can perform one or more of the operations, storage, reference, input / output, and control functions of a general computer, and can include an artificial neural network to be described later based on input data.

[0030] The control server (100) may include a processor and memory. The processor may perform the smart interior service according to the present invention and may include devices capable of performing the same. The processor may also execute a program or control the control server (100). The program code executed by the processor may be stored in the memory. The memory may store relevant information for performing the service according to the present invention or a program for implementing the method. The memory may be volatile or non-volatile memory.

[0031] The control server (100) can transmit data to an external device or receive data from an external device using a network.

[0032] The control server (100) can train an artificial neural network and utilize a trained artificial neural network. The processor can train or execute an artificial neural network stored in memory, and the memory can store a trained artificial neural network. The electronic device that trains the artificial neural network and the electronic device that utilizes the artificial neural network may be the same or separate.

[0033] Artificial intelligence (AI) is a computer system that embodies some of the functions of the human brain, capable of learning, making inferences, and making judgments on its own. As learning progresses, the probability of extracting answers increases. AI can be comprised of learning and the underlying technologies that utilize it. AI learning is an algorithmic technology that classifies and learns features based on input data, while the underlying technologies can utilize learning algorithms to partially emulate the functions of the human brain.

[0034] AI is a technology that easily approaches problems with multiple probabilistic answers, enabling it to logically and probabilistically infer optimal cycles, methods, and plans based on input data. AI inference techniques can include evaluating input data, making optimal predictions, knowledge- and probability-based inferences, and preference-based planning.

[0035] An artificial neural network (ANN) is a learning algorithm in the field of machine learning. It implements the connections between neurons and synapses in the brain through a program. An ANN can be programmed to create a neural network structure and then train it to achieve a desired function. While errors may exist, it can learn from massive amounts of data, producing appropriate output data based on input data. Its advantages include the ability to obtain output data that yields statistically positive results and its resemblance to human reasoning.

[0036] A network is a high-speed backbone network of a large communications network capable of providing large-capacity, long-distance voice and data services, and may be a next-generation wired or wireless network to provide the Internet or high-speed multimedia services.

[0037] If the network is a mobile communication network, it may be a synchronous mobile communication network or an asynchronous mobile communication network. An example of an asynchronous mobile communication network is a Wideband Code Division Multiple Access (WCDMA) network. In this case, although not shown in the diagram, the network may include a Radio Network Controller (RNC). While a WCDMA network is mentioned as an example, it could also be a 3G LTE network, a 4G network, a next-generation communication network such as 5G, or any other IP-based network.

[0038] A mobile terminal (200) may include any terminal capable of transmitting and receiving data over a network, such as a desktop computer, laptop, tablet, or smartphone.

[0039] The mobile terminal (200) may include one or more of the computational functions, storage functions, reference functions, input / output functions, and control functions of a computer to perform the service according to the present invention.

[0040] A mobile terminal (200) can access a website or install an application to receive services according to the present invention. A mobile terminal (300) can exchange data via a website or application.

[0041] The system (1) according to one embodiment of the present invention generates and provides various types of questions for files or natural language input through a mobile terminal (200), so that it can be easily utilized in various fields such as employment, education, entertainment, and research, for example, in determining the authenticity of self-introductions, creating test questions based on textbooks and self-study books, creating problems in games, and creating review questions for research projects.

[0042]

[0043] Figures 2 to 4 are conceptual diagrams of the control server illustrated in Figure 1.

[0044] Referring to FIG. 2, a control server (100) according to one embodiment of the present invention may include a document vector database configuration unit (110), a question generation unit (120), and a question explanation generation unit (130).

[0045] The document vector database configuration unit (110) can receive documents for item creation through a mobile terminal (200), analyze the documents, and configure a document vector database of a language model for item creation.

[0046] In this embodiment, a document means data in a key-value format such as JSON, and may also include data such as table columns commonly used in RDB.

[0047] The document vector database configuration unit (110) can generate content vectors, attribute vectors, and document question vectors for documents, and can store them in the document vector database by including the document ID in the key.

[0048] A language model can generate items based on a document vector database like this, and when generating items based on any of the content vectors, attribute vectors, and document question vectors, the document that became the basis for item generation can be accurately extracted through the document ID stored in the corresponding vector.

[0049] Referring to FIG. 3, the document vector database configuration unit (110) may include a content vector generation unit (111), an attribute vector generation unit (112), and a document question vector generation unit (113).

[0050] The content vector generation unit (111) can create and embed text by concatenating attribute values ​​specified by the user.

[0051] For example, the content vector generation unit (111) can generate a text that combines the values ​​of attributes that can represent the document among the attributes of the document as a content vector.

[0052] The attribute vector generation unit (112) can generate and embed text combining the document title and attribute name according to an attribute vector template that defines a method of combining the document title and attribute name.

[0053] For example, the attribute vector generation unit (112) can generate a text combining the document title and attribute name as an attribute vector according to the attribute vector template of “What is the [attribute name] of [document title]?”

[0054] The document question vector generation unit (113) can embed question text directly registered by the user for the document.

[0055] For example, the document question vector generation unit (113) can limit the documents to be shown for questions directly registered by the user.

[0056] Table 1 below is an example of a document created in the document vector database configuration unit (110).

[0057] [Table 1]

[0058] [Correction pursuant to Rule 91, August 2025]

[0059] Table 2 below is an example of a content vector, attribute vector, and document question vector generated from a document according to Table 1 in the document vector database configuration unit (110).

[0060] [Table 2]

[0061] [Correction pursuant to Rule 91, August 2025]

[0062] [Correction pursuant to Rule 91, August 2025]

[0063] Meanwhile, the document vector database configuration unit (110) can receive input from a mobile terminal (100) in the form of a file (Hwp, PDF, etc.) or a document in natural language.

[0064] When a document in file form is input, the document vector database configuration unit (110) can parse the file to preprocess unstructured data, analyze unstructured morphemes, and then execute the embedding step of the content vector, attribute vector, and document question vector as described above.

[0065] When a document in natural language format is input, the document vector database configuration unit (110) can preprocess unstructured data, analyze unstructured morphemes, and then execute the embedding step of the content vector, attribute vector, and document question vector as described above.

[0066] The question generation unit (120) can generate questions based on a document vector database using a language model and output them to a mobile terminal (200).

[0067] Referring to FIG. 4, the question generation unit (120) may include a user question vector generation unit (121), a vector search unit (122), and an answer generation unit (123).

[0068] The user question vector generation unit (121) can generate a user question vector by embedding the user question text input through the mobile terminal (200).

[0069] The vector search unit (122) can extract at least one of a content vector, an attribute vector, and a document question vector corresponding to a user question vector from a document vector database.

[0070] The answer generation unit (123) can generate an answer by inputting at least one of the content vector, attribute vector, and document question vector extracted from the vector search unit (122) into the language model.

[0071] For example, the language model can be applied to various common language models such as the NLU module, the Policy module, the NLG module, and the ChatLLM.

[0072] The question explanation generation unit (130) can extract a document that serves as the basis for the question output from the question generation unit (120) and provide it to the mobile terminal (200).

[0073] For example, the question explanation generation unit (130) can extract a document using the ID of the document stored in at least one vector among the content vector, attribute vector, and document question vector extracted from the question generation unit (120).

[0074]

[0075] Meanwhile, the question explanation generation unit (130) can calculate a user evaluation score for a question using the following mathematical formula 1. For example, the question explanation generation unit (130) can calculate a user evaluation score based on the number of documents referenced in question generation, the number of times the user viewed the document, and the time it took the user to input a response to the question.

[0076] [Mathematical Formula 1]

[0077] [Correction pursuant to Rule 91, August 2025]

[0078] R in mathematical equation 1 s is the user evaluation score, n is the number of documents referenced in generating the question, D n 0 if the nth document was viewed, 1 if not viewed, S r refers to the time it takes for a user to enter a response to a question.

[0079] For example, if the number of documents referenced in question generation is 2, And, if the user's response input time for the question is 1 minute, it can be calculated as Rs=50.

[0080] If the user evaluation score is calculated to be below a predetermined score, the item explanation generation unit (130) may consider that there is difficulty in explaining the item and provide additional related documents through web crawling.

[0081] Meanwhile, the item explanation generation unit (130) can calculate an automatic item generation score using the mathematical expression 2 below, and can update the document vector database according to the automatic item generation score.

[0082] [Equation 2]

[0083] [Correction pursuant to Rule 91, August 2025]

[0084] In mathematical expression 2, St is the automatically generated score, s s1is the number of items output by the system during the time between t1 and t2, s s2 RI is the number of questions input by the user through the mobile terminal (200) during the time between t1 and t2. t is the average time from the system output item to the user's question input through the mobile terminal (200) during the time between t1 and t2, P is the number of pre-stored positive words included in the user input question text, a is the time weight, and S0 is the system-set basic score.

[0085] For example, if ss1=3, ss2=4, RIt=70, P=50, a=0.5, S0=1, can be produced as

[0086] The item explanation generation unit (130) can update the document vector database when the item automatic generation score is below a predetermined score.

[0087]

[0088] In one embodiment, the document vector database configuration unit (110) calculates the importance index of each word included in the sentences constituting the learning question data (or document) using the following mathematical expression 3 to estimate the meaning of the learning question data and documents, and extracts the word with the highest calculated importance index by setting it as a keyword.

[0089]

[0090] [Equation 3]

[0091] [Correction pursuant to Rule 91, August 2025]

[0092] Here, I_n is an importance index for word n, r_n is the number of repeated occurrences of word n ​​in the document data, v_n is the size of the embedding vector for word n, v_i is the size of the embedding vector for the word with the i-th highest number of occurrences in the document data, and t is a variation constant, which is set to a first constant value when word n ​​is a word included in a preset important word dictionary, and is set to a second constant value smaller than the first constant value when word n ​​is a word not included in a preset important word dictionary.

[0093] To this end, the system according to the present invention can build a neural network that extracts contextual information for input data by learning training data using the Word2Vec algorithm to understand or estimate the meaning of data in text form.

[0094] The Word2Vec algorithm can incorporate a Neural Network Language Model (NNLM). A NNLM is essentially a neural network consisting of an input layer, a projection layer, a hidden layer, and an output layer. NNLM is used to vectorize words. Because NNLM is a well-known technology, a detailed description will be omitted.

[0095] The Word2vec algorithm, designed for text mining, determines proximity by examining the preceding and following relationships between words. The Word2vec algorithm is an unsupervised learning algorithm. As its name suggests, the Word2vec algorithm is a metric technique that represents the meaning of words in vector form. The Word2vec algorithm can represent each word as a vector in a space of approximately 200 dimensions. Using the Word2vec algorithm, a vector corresponding to each word can be derived.

[0096] The Word2vec algorithm can dramatically improve accuracy in natural language processing compared to other conventional algorithms. Word2vec can learn the meaning of words by leveraging the relationships between words and adjacent words in sentences within an input corpus. The Word2vec algorithm is based on artificial neural networks and assumes that words with similar contexts have similar meanings. The Word2vec algorithm trains on text documents, and it trains the artificial neural network on words that appear nearby (5 to 10 words before or after a word) as related words. Because words with related meanings are more likely to appear close together in a document, the two words can gradually have similar vectors through repeated training.

[0097] The Word2vec algorithm's learning methods include the Continuous Bag of Words (CBOW) method and the skip-gram method. The CBOW method predicts a target word using the context created by surrounding words. The skip-gram method predicts potential surrounding words based on a single word. The skip-gram method is known to be more accurate for large-scale datasets.

[0098] Therefore, in embodiments of the present invention, the Word2vec algorithm utilizing the skip-gram method is used. For example, if training is successfully completed using the Word2vec algorithm, similar words can be located nearby in a high-dimensional space. According to the Word2vec algorithm described above, the closer the distribution of surrounding words in a learning document is to a word, the more similar the resulting vector values ​​can be. In addition, words with similar resulting vector values ​​can be considered similar. Since the Word2vec algorithm is a well-known technology, a detailed description of the vector value calculation will be omitted.

[0099] The control server (100) can input data collected into a neural network and extract an evaluation result vector value representing contextual information.

[0100] The control server (100) can calculate the similarity between the evaluation result vector value and each of a plurality of reference vector values, and extract the reference vector value having the highest similarity with the evaluation result vector value among the plurality of reference vector values. At this time, the similarity calculation method may employ Euclidean distance, cosine similarity, Tanimoto coefficient, etc.

[0101] The control server (100) can extract the word corresponding to the reference vector value with the highest similarity to the evaluation result vector value as the word corresponding to the recognized text.

[0102] In addition, the control server (100) can train an artificial neural network and utilize an artificial neural network that has completed training. The processor can train or execute an artificial neural network stored in memory, and the memory can store an artificial neural network that has completed training. The electronic device that trains the artificial neural network and the electronic device that uses the artificial neural network may be the same, but may be separate. Artificial intelligence is a computer system that partially implements the functions of the human brain and can learn, guess, and judge on its own. As learning progresses, the probability of extracting an answer may increase. Artificial intelligence may be composed of learning and component technologies that utilize it. Artificial intelligence learning is an algorithm technology that classifies and learns features based on input data, and the component technologies may be technologies that partially implement the functions of the human brain using learning algorithms.

[0103] AI is a technology that easily approaches problems with multiple probabilistic answers, enabling it to logically and probabilistically infer optimal cycles, methods, and plans based on input data. AI inference techniques can include evaluating input data, making optimal predictions, knowledge- and probability-based inferences, and preference-based planning.

[0104] An artificial neural network (ANN) is a learning algorithm in the field of machine learning. It implements the connections between neurons and synapses in the brain through a program. An ANN can be programmed to create a neural network structure and then train it to achieve a desired function. While errors may exist, it can learn from massive amounts of data, producing appropriate output data based on input data. Its advantages include the ability to obtain output data that yields statistically positive results and its resemblance to human reasoning.

[0105] The control server (100) can build a query / metric dataset required for learning using an artificial intelligence algorithm built based on big data, and for this purpose, can include a number of pre-learned artificial neural networks.

[0106] The system according to the present invention may include multiple pre-trained artificial neural networks for performing machine learning algorithms. Machine learning allows the system to output data based on input data and utilize the results to learn independently, thereby enhancing its data processing capabilities. The artificial neural network extracts features from input data, infers patterns, and outputs result data. As these processes accumulate, the reliability of the result data increases.

[0107] In this embodiment, the artificial neural network may be an algorithm that outputs text data from at least one feature data item, including the shape, length, number, and height difference of objects recognized as text. The artificial neural network can infer optimal output data by using big data as input data directly or after processing it to remove unnecessary data.

[0108] Artificial intelligence machine learning models can be categorized into Supervised Learning, Unsupervised Learning, Semisupervised Learning, and Reinforcement Learning, depending on the learning type. Machine learning algorithms that can be used include Decision Trees, K-Nearest Neighbor, Artificial Neural Networks, Support Vector Machines, Ensemble Learning, Gradient Descent, Naive Bayes Classifiers, Hidden Markov Models, and K-Means Clustering.

[0109] An artificial neural network may be pretrained on various input values ​​that may be included in the input data. An artificial neural network may be trained using reinforcement learning, a learning method. Reinforcement learning is a method that gradually increases the probability of obtaining a correct result by setting rewards and constraints. An artificial neural network may also be modeled based on a convelutional neural network (CNN) or a recurrent neural network (RNN).

[0110] In this way, the system according to the present invention can estimate the meaning of text data using big data and an artificial neural network.

[0111] In some other embodiments, the system according to the present invention calculates a validity index for each collected document data using the following mathematical expression 4 to select and store only reliable document data among a plurality of document data.

[0112]

[0113] [Equation 4]

[0114] [Correction pursuant to Rule 91, August 2025]

[0115] Here, S is the effective index, n is the size (bite) of the document data, pd_i is the difference between the preset reference value and the embedding vector constituting the document data i, g is the first weight that is set differently for each category of the document, and the difference between the average embedding vector value of other document data classified into the same category as document i and the embedding vector constituting the document data i.

[0116] For example, if document data is 16 bytes, pd_i is 3.5, sd_k is 1.2, g is 14, and c is 50, the validity index S can be calculated as 76. At this time, if the preset standard validity value is 70, document data with a validity index of 76 can be stored as data valid (necessary) for monitoring because it exceeds the standard validity value.

[0117] In this way, by calculating the validity index for each document data and storing only the document data whose calculated validity index is greater than a preset standard value, unnecessary data is prevented from being stored and taking up storage space, thereby enabling efficient use of storage space.

[0118] In some other embodiments, the system according to the present invention may further include a security processing unit (not shown) that encrypts and manages stored document data.

[0119] Specifically, the security processing unit calculates a security need index for each collected document data using the following mathematical formula 5.

[0120]

[0121] [Equation 3]

[0122] [Correction pursuant to Rule 91, August 2025]

[0123] Here, R is a security need index, wv_n is a size value of a word embedding vector for a specific keyword n ​​among keywords extracted from document data, wv_m is a size value of a word embedding vector for a specific keyword m among keywords extracted from other document data that has been secured, sv_o is an average size value of sentence embedding vectors for sentences extracted from document data, sv_p is an average size value of sentence embedding vectors for sentences extracted from other document data that has been secured, a is a first weight value set in proportion to the number of sentences constituting the document data, and b is a second weight value set differently for each category of document data.

[0124] In this way, the security processing unit calculates a security need index for each document data using the method described above. Data with a security need index exceeding a preset threshold is encrypted and stored, while data with a security need index lower than the threshold is stored without a separate encryption process. In this way, by selectively performing data security operations according to established security policies, the resources and time required for security operations can be reduced.

[0125] Meanwhile, the document encryption algorithm may use any one of various known encryption techniques, including symmetric key encryption algorithms such as AES, 3DES, and SEED, asymmetric key encryption algorithms such as RSA, ECC, and DSA, and encryption techniques using hash algorithms.

[0126]

[0127] Although the embodiments of the present invention have been described with reference to the attached drawings, those skilled in the art will appreciate that the present invention can be implemented in other specific forms without altering the technical concept or essential characteristics thereof. Therefore, the embodiments described above should be understood to be illustrative in all respects and not restrictive.

Claims

1. Mobile terminal; and An automatic question generation service system, comprising a control server that receives a document for question generation through the mobile terminal, analyzes the document to form a document vector database of a language model for question generation, and generates questions based on the document vector database using the language model.

2. In paragraph 1, The above control server, A document vector database configuration unit that generates a content vector, an attribute vector, and a document question vector for the above document, and stores them in the document vector database so that each vector can have the ID of the above document as a key; A question generation unit that generates questions based on the document vector database using a language model and outputs them to the mobile terminal; and An automatic question generation service system, comprising a question explanation generation unit that extracts documents that serve as the basis for questions output from the above question generation unit and provides them to the mobile terminal.

3. In paragraph 2, The above document vector database configuration part is, A content vector generation unit that embeds text by generating text that combines user-specified attribute values; An attribute vector generation unit that creates and embeds text combining the title and attribute name of the document according to an attribute vector template that defines a method of combining the title and attribute name of the document; and Includes a document question vector generation unit that embeds question text directly registered by the user for the above document; The above question generation section, A user question vector generation unit that generates a user question vector by embedding a user question text input through the mobile terminal; A vector search unit that extracts at least one of a content vector, an attribute vector, and a document question vector corresponding to the user question vector from the document vector database; and An answer generation unit that inputs at least one of the content vector, attribute vector, and document question vector extracted from the above vector search unit into a language model and outputs an answer to a user question text input through the mobile terminal; The above question explanation generation section, An automatic question generation service system that extracts documents that serve as the basis for answers output from the answer generation unit by using the ID of documents stored in at least one vector among the content vector, attribute vector, and document question vector extracted from the above vector search unit.

Images