Data providing device, method, and computer program for generating answer to question using artificial intelligence technology

Abstract

According to embodiments of the present disclosure, disclosed is a method for generating an answer to a question using artificial intelligence technology, the method comprising: a step in which a data providing device obtains question data including a keyword and a question inputted by a user; a step in which the data providing device outputs first filtering data for the keyword and second filtering data for the question by inputting the question data into a filtering model in relation to the question data; a step in which the data providing device inputs the question data and the first and second filtering data into an answer generation model so as to output answer data for the question data; and a step in which the data providing device generates output data including the answer data and provides the output data to a user terminal.

Description

Data providing device, method and computer program for generating answers to questions using artificial intelligence technology

[0001] The present disclosure relates to a data providing device, method, and computer program that generate answers to questions using artificial intelligence technology, and more specifically, to a method for providing data that uses an artificial intelligence model to search for academic literature corresponding to the occurrence of a keyword and / or the occurrence of a question, and to search for and provide academic literature that includes data that can answer a question by identifying the context of the academic literature.

[0002] Current academic literature retrieval systems utilize keyword-based search algorithms. This technology returns documents containing specific keywords when a keyword is entered. Keywords are typically extracted from the title, abstract, or keyword section of a document. Conventional academic literature retrieval algorithms perform keyword-based searches. However, keyword-based retrieval methods generate search results based on the inclusion or absence of keywords, which limits the accuracy and relevance of the results.

[0003] More specifically, the basic operation of keyword-based searches is to identify documents containing words that match or are similar to the entered keyword. In such cases, even if a document does contain a keyword, the overall content of the document may be irrelevant to the keyword. For example, when searching for the keyword "cancer," documents frequently containing the word "cancer" will appear at the top, but this does not necessarily guarantee that these documents best reflect research findings on the topic.

[0004] Keyword-based searches cannot retrieve documents that truly capture the essence of the question or hypothesis being searched. Furthermore, keyword-based searches only consider the presence or absence of keywords, not analyzing the detailed content of the document to generate search results. Using keyword-based searches can result in researchers failing to find the documents necessary to verify or address their questions or hypotheses. Furthermore, because keyword-based searches do not consider the context of the document, the authors' claims, or the research results, researchers must manually review the content of each retrieved document.

[0005] There is a growing need for a service that accurately and efficiently searches for literature matching the questions or hypotheses researchers are seeking to address.

[0006] [Previous literature]

[0007] KR 10-2017-0142964 A (Published on December 28, 2017)

[0008] The embodiments disclosed in this specification aim to provide a device and method for searching academic literature corresponding to the occurrence of a keyword and / or the occurrence of a question using an artificial intelligence model, and for retrieving and providing academic literature containing data capable of answering a question by understanding the context of such academic literature.

[0009] A method according to embodiments of the present disclosure may include a step in which a data providing device obtains question data including keywords and questions input from a user; a step in which the data providing device inputs the question data into a filtering model in relation to the question data and outputs first filtering data for the keywords and second filtering data for the question; a step in which the data providing device inputs the question data and the first and second filtering data into an answer generating model and outputs answer data for the question data; and a step in which the data providing device generates output data including the answer data and provides the output data to a user terminal.

[0010] The above filtering model is a model trained using registered academic documents as input, and can be constructed to output the first filtering data including whether the keyword appears in the academic documents, or the second filtering data including whether the question can be answered.

[0011] The above answer generation model is a model trained using registered academic documents as input, and can detect academic documents related to the question data, analyze the detected academic documents, and output answer data corresponding to the question data.

[0012] The above filtering model may be learned by converting academic literature into embedding vectors, inputting the embedding vectors into an attention mechanism and a transformer, and obtaining output.

[0013] The above answer generation model may be one that converts academic literature into embedding vectors, inputs the embedding vectors into an attention mechanism and a transformer, and learns the obtained output.

[0014] The above filtering model or the above answer generation model may be a model trained with academic literature in the field to which the question data belongs.

[0015] A device according to embodiments of the present disclosure includes a communication unit, a memory, and a processor, wherein the processor obtains question data including a keyword and a question input from a user, inputs the question data into a filtering model in relation to the question data to output first filtering data for the keyword and second filtering data for the question, inputs the question data and the first and second filtering data into an answer generation model to output answer data for the question data, and generates output data including the answer data and transmits the output data to a user terminal.

[0016] A computer program according to an embodiment of the present invention can be stored in a medium to execute any one of the methods according to an embodiment of the present invention using a computer.

[0017] In addition, other methods for implementing the present invention, other systems, and computer-readable recording media recording a computer program for executing the above methods are further provided.

[0018] Other aspects, features and advantages other than those described above will become apparent from the following drawings, claims and detailed description of the invention.

[0019] According to any one of the aforementioned problem solving means, an artificial intelligence model can be used to search for academic literature corresponding to the occurrence of a keyword and / or the occurrence of a question, and the context of such academic literature can be identified to search for and provide academic literature containing data that can answer the question.

[0020] FIG. 1 is a drawing for explaining a detailed configuration of a data providing device that searches for and provides academic literature using artificial intelligence technology according to embodiments of the present disclosure.

[0021] Figure 2 is a block diagram of the data provision unit.

[0022] Figure 3 is a block diagram of a learning device (300) that learns a filtering model and / or an answer generation model.

[0023] Figure 4 is a flowchart of a method for generating an answer according to embodiments of the present disclosure.

[0024] Figure 5 is a diagram of input data and output data (Predict answer) of the answer generation model.

[0025] FIG. 6a is a diagram showing the structure of an answer generation model generated according to embodiments of the present disclosure.

[0026] Figure 6b is an example drawing of an attention mask modified by the answer generation model.

[0027] Figure 6c is a diagram comparing a general attention vector and an attention vector generated by an answer generation model.

[0028] FIG. 7 is an exemplary diagram of academic literature data according to embodiments of the present disclosure.

[0029] FIG. 8 is an example drawing of extracted text according to embodiments of the present disclosure.

[0030] FIG. 9 is an example diagram of context data for text according to embodiments of the present disclosure.

[0031] Figure 10 is an example diagram of question data obtained and processed from academic literature.

[0032] FIG. 11 is a flowchart of a method for generating an answer according to embodiments of the present disclosure.

[0033] The configuration and operation of the present invention will be described in detail with reference to embodiments of the present invention illustrated in the attached drawings below.

[0034] The present invention is capable of various modifications and embodiments. Specific embodiments are illustrated in the drawings and described in detail in the detailed description. The effects and features of the present invention, as well as the methods for achieving them, will become clearer with reference to the embodiments described in detail below, along with the drawings. However, the present invention is not limited to the embodiments disclosed below and can be implemented in various forms.

[0035] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings. When describing with reference to the drawings, identical or corresponding components are given the same drawing reference numerals, and redundant descriptions thereof will be omitted.

[0036] Hereinafter, the term “upper” or “upper” may include not only things that are directly above in contact, but also things that are above in a non-contact manner.

[0037] In the following examples, the terms first, second, etc. are not used in a limiting sense, but are used for the purpose of distinguishing one component from another.

[0038] In the examples below, singular expressions include plural expressions unless the context clearly indicates otherwise.

[0039] In the following examples, terms such as “include” or “have” mean that a feature or component described in the specification is present, and do not preclude the possibility that one or more other features or components may be added.

[0040] For convenience of explanation, the sizes of components in the drawings may be exaggerated or reduced. For example, the sizes and thicknesses of each component shown in the drawings are arbitrarily indicated for convenience of explanation, and thus the present invention is not necessarily limited to what is shown.

[0041] Additionally, terms such as “…part”, “…area”, etc. described in this specification may mean a unit that processes at least one function or operation.

[0042] In embodiments of the present disclosure, relevant documents may be obtained from a keyword database and / or a domain database. Here, the keyword database may read document data, extract keywords corresponding to the document data, and store these keywords and the document data as structured information. Here, the domain database may extract information about the document data, such as the title, journal name, publication date, abstract, and body of the document, and store this information and the document data as structured information.

[0043] In embodiments of the present disclosure, the response generation model is a type of artificial intelligence trained on a large-scale text data set to generate human-like responses to natural language input. It is trained to understand and generate natural language using a very large model and a very large amount of text. It may be trained based on an attention processor or a transformer.

[0044] FIG. 1 is a drawing for explaining a detailed configuration of a data providing device that searches for and provides academic literature using artificial intelligence technology according to embodiments of the present disclosure.

[0045] The data providing device (100) automatically generates answer data corresponding to question data entered by a researcher using an artificial intelligence technology using a model that has been trained and fine-tuned with medical academic literature, and the data providing device (100) can provide answer data for question data based on text data from academic literature related to the question data entered using artificial intelligence technology. The data providing device (100) can generate and provide answer data for question data using an answer generation model that has learned the context of each academic literature. The data providing device (100) can contextually analyze previously registered academic literature to generate first filtering data regarding the appearance of keywords in the question data and second filtering data regarding the appearance of questions. The data providing device (100) can determine whether an answer can be generated for the question data based on the first filtering data and the second filtering data.

[0046] As illustrated in FIG. 1, a data providing device (100) according to some embodiments may include a processor (110), an input / output unit (130), a memory (140), a communication unit (150), and a data providing unit (200). However, not all of the components illustrated in FIG. 1 are essential components of the data providing device (100). The data providing device (100) may be implemented with more components than the components illustrated in FIG. 1, or may be implemented with fewer components than the components illustrated in FIG. 1. The data providing device (100) may be a user terminal, a server, a game network system, or a separate device.

[0047] According to one embodiment of the present disclosure, the processor (110) typically controls the overall operation of the data providing device (100). For example, the processor (110) may control the components included in the data providing device (100) by executing a program stored in the data providing device (100).

[0048] According to one embodiment of the present disclosure, the processor (110) can generate and provide answer data for input question data by executing commands stored in the data provision unit (200). The processor (110) can search for documents related to the question data using an artificial intelligence model and generate answer data based on the text of the searched documents. The artificial intelligence model may be a large-scale language model that analyzes documents related to the question data using a multilayer neural network. Depending on the artificial intelligence model, the texts of the documents may be analyzed through an embedding generation step, an attention processor, and a transformer.

[0049] The processor (110) may extract keywords and questions from input question data, generate first filtering data for the keywords, generate second filtering data for the questions, and input the first filtering data, the second filtering data, and the question data into an answer generation model to output answer data for the question data. As the answer generation model is executed, the processor (110) may use first academic literature searched based on keywords and question data to divide each of one or more first academic literatures into a plurality of context data, generate individual query data, and provide answer data according to the learning results. The processor (110) may output answer data for the question data using output values ​​of the answer generation model. The processor (110) may generate first filtering data on whether keywords appear and second filtering data on whether questions appear in relation to the first academic literature. The processor (110) may output the first filtering data and / or the second filtering data using the filtering model. Here, the filtering model may be a model trained through machine learning, but is not limited thereto. It may also be a model that provides search results using a fixed algorithm. The first filtering data may further include data on academic literature in which the keyword appears. The second filtering data may further include data on academic literature in which the question appears.

[0050] The processor (110) is a component for controlling the overall operation of the data provision device (100). Specifically, the processor (110) controls the overall operation of the data provision device (100) using various programs stored in the storage medium (150) of the data provision device (100). For example, the processor (110) may include a CPU, RAM, ROM, and a system bus. Here, the ROM is a component in which a command set for system booting is stored, and the CPU copies the stored operating system of the data provision device (100) to the RAM according to the command stored in the ROM and executes the O / S to boot the system. When the system booting is complete, the CPU can copy various stored applications to the RAM and execute them to perform various operations. Although the service provision server (100) has been described as including only one CPU in the above, it may be implemented with multiple CPUs (or DSPs, SoCs, etc.) during implementation.

[0051] According to one embodiment of the present invention, the processor (110) may be implemented as a digital signal processor (DSP), a microprocessor, or a time controller (TCON) that processes digital signals. However, the present invention is not limited thereto, and may include one or more of a central processing unit (CPU), a micro controller unit (MCU), a micro processing unit (MPU), a controller, an application processor (AP), a communication processor (CP), or an ARM processor, or may be defined by the corresponding terms. In addition, the processor (110) may be implemented as a system on chip (SoC) or large scale integration (LSI) having a built-in processing algorithm, or may be implemented in the form of a field programmable gate array (FPGA).

[0052] According to one embodiment of the present disclosure, the input / output unit (130) can display an interface generated by the memory (140) of the data providing device (100). According to one embodiment of the present invention, the input / output unit (130) can display a user interface for input user input. The input / output unit (130) can output stored graphic data, visual data, auditory data, and vibration data under the control of the memory (140).

[0053] The input / output unit (130) may be implemented as a display panel of various forms. For example, the display panel may be implemented as a display technology of various forms, such as a liquid crystal display (LCD), an organic light emitting diode (OLED), an active-matrix organic light-emitting diode (AM-OLED), a liquid crystal on silicon (LCoS), or a digital light processing (DLP). In addition, the input / output unit (130) may be coupled to at least one of the front area, the side area, and the rear area of ​​the display panel in the form of a flexible display.

[0054] The input / output unit (130) may be implemented as a touch screen with a layer structure. The touch screen may have a function of detecting not only a display function but also a touch input location, a touched area, and even a touch input pressure, and may also have a function of detecting not only a real touch but also a proximity touch.

[0055] The input / output unit (130) may include a user interface for inputting various information into the data provision device (100).

[0056] According to one embodiment of the present disclosure, the memory (140) may store a program for processing and controlling the processor (110) and / or the data providing unit (200), and may also store data input to or output from the data providing device (100). According to one embodiment of the present disclosure, the memory (140) may store information on documents, and may also store information on keywords and questions. The memory (140) may include a database storing the above information.

[0057] According to one embodiment of the present disclosure, the memory (140) may include at least one type of storage medium among a flash memory type, a hard disk type, a multimedia card micro type, a card type memory (e.g., SD or XD memory, etc.), a RAM (Random Access Memory), a SRAM (Static Random Access Memory), a ROM (Read-Only Memory), an EEPROM (Electrically Erasable Programmable Read-Only Memory), a PROM (Programmable Read-Only Memory), a magnetic memory, a magnetic disk, and an optical disk. In addition, according to one embodiment of the present disclosure, programs stored in the memory (140) may be classified into a plurality of modules according to their functions.

[0058] According to one embodiment of the present disclosure, the communication unit (150) can communicate with an external device of the processor (110). For example, the communication unit (150) can communicate with an external device, such as a payment server or an authentication server, under the control of the processor (110). In addition, the communication unit (150) can also obtain user information or user input through communication with an external interface.

[0059] The data provision unit (200) may include a data input unit (210), a filtering unit (220), a search provision unit (230), and a data output unit (240), as illustrated in FIG. 2.

[0060] The data input unit (110) can input question data input by a user. The data input unit (110) can extract keywords and questions from the question data. The data input unit (110) can extract keywords and / or questions from the question data using natural language processing technology. The data input unit (110) can extract keywords and / or questions as a result of learning using a large-scale data set. Here, the question data is text data written in natural language and may further include pictures, drawings, videos, sound data, etc. The data input unit (110) can extract keywords and / or questions for the question data through a process of analyzing the pictures, drawings, videos, and sound data included in the question data.

[0061] The filtering unit (120) can determine filtering data related to the question data. The filtering unit (120) can output first filtering data related to keywords in the question data. The first filtering data can include the presence or absence of academic literature in which the keyword appears, data on academic literature in which the keyword appears, etc. The academic literature may be, but is not limited to, literature used to train the answer generation model. The academic literature may be registered and published in real time in an academic literature database, such as PubMed.

[0062] The filtering unit (120) can output second filtering data related to the question of the question data. The filtering unit (120) can output second filtering data regarding the question of the question data. The filtering unit (120) can output second filtering data including data regarding whether the question of the question data appears in literature in a field related to the question data and / or related literature. The second filtering data can include whether the question is answerable. Whether the question is answerable can be output as a result of learning the frequency of the question in the academic literature in which the question appears or the characteristics between the sentence containing the question and other sentences.

[0063] Optionally, the filtering unit (120) may obtain first and second filtering data for the question data using a filtering model. The filtering model is trained using documents from related fields as input and may output first filtering data for keywords in the question data and second filtering data for the question. The filtering model is trained using a data set of documents as input and may be designed to output whether a keyword appears or whether an answer to the question is possible.

[0064] The filtering model can be trained by inputting documents registered in multiple databases using a Large Language Model (LLM), but is not limited thereto. The filtering model trains the text of academic documents with the LLM to output information such as whether keywords appear and whether questions can be answered. Through this process, the filtering model can further output data on academic documents in which keywords appear and data on academic documents that can answer questions. If the number of academic documents in which keywords appear is 1 or greater, the filtering model can output whether keywords appear as 1. If the number of academic documents that can answer questions is 1 or greater, the filtering model can output whether questions can be answered as 1.

[0065] Optionally, the filtering model can be trained using reinforcement learning by re-inputting user-entered question data and filtering data for the question data. The filtering model is characterized by being trained using a data set of unlabeled documents. The training process of the filtering model is described in detail in Figure 3.

[0066] The search provider (230) can input question data into the answer generation model based on the first filtering data and the second filtering data. The search provider (230) can input the first and second filtering data and data of academic literature related to the question into the answer generation model to output answer data. If the presence or absence of a keyword is true (TRUE, 1) and the presence or absence of a question is true (TRUE, 1), the question data can be input into the answer generation model to output answer data.

[0067] The search provider (230) may, in response to the input format of the answer generation model, encode keywords and / or questions included in the question data and convert them into embedding vectors. The embedding vectors may encode each of the determined keywords into a predetermined value. The determined keywords may be determined as a result of the answer generation model learning academic literature. The keywords may be determined based on the frequency and / or relevance to the main text of the literature among keywords appearing in the academic literature. For example, the search provider (230) may generate an embedding vector having 10 columns or rows including whether the first to tenth keywords determined based on the frequency and / or relevance to the main text of the literature appear in the question data. The answer generation model may receive the embedding vector as input, detect an area related to the input question, and perform an operation (R-step, FIG. 6a) of modifying an attention mask corresponding to the area (position, length) related to the question. The answer generation model can be implemented to generate an attention vector (AM2, see Fig. 6c) through an attention mask including a modified area (R-area, see Fig. 6b) in response to a question, and input the attention vector into a transformer to output answer data.

[0068] The search provider (230) can input an embedding vector into the answer generation model to output answer data for question data.

[0069] Here, the answer generation model is a model trained by inputting previously published literature data and can be trained using supervised learning. The answer generation model can be trained by inputting literature data into a transformer and using the processed data as training data. At this time, the answer generation model can process the data using a modified attention processor. The answer generation model is trained using unlabeled literature data, and can be trained without a data labeling process. The training process of the answer generation model is described in detail in Figure 3.

[0070] Figure 3 is a block diagram of a learning device (300) that learns a filtering model and / or an answer generation model.

[0071] The learning device (300) can train a filtering model and / or an answer generation model using currently published academic literature as input. The learning device (300) can build a filtering model that outputs filtering data such as the presence of keywords related to the existence of academic literature that can answer question data, whether the question can be answered, and the like. The filtering model can additionally output data on academic literature related to the question. The learning device (300) can build an answer generation model that outputs data on academic literature useful for generating answers to the input question data and / or answer data for the question data. The learning device (300) can build an answer generation model that searches for academic literature corresponding to the question data and generates an answer to the question by training a large amount of academic literature using a large-scale language model.

[0072] The dataset generation unit (310) can generate a training data set using a predetermined method. More specifically, the dataset generation unit (310) can obtain documents related to keywords from a database and extract text from the documents using text extraction technology (OCR). The dataset generation unit (310) can generate contextual data (keywords, questions, text, abstract, author, publication date, etc.) from the extracted text (see FIG. 8). The dataset generation unit (310) can divide academic literature into multiple contextual data and convert each contextual data (see FIG. 9) into individual question-and-answer data (semantic search QA query, see FIG. 10) to be converted into the form of learning data.

[0073] The model learning unit (320) may include a first learning unit (321) that learns a filtering model and a second learning unit (322) that learns an answer generation model.

[0074] The model learning unit (320) can input the training data set generated by the dataset generation unit (310) into the first learning unit (321) to train the filtering model. The model learning unit (320) can input the training data set generated by the dataset generation unit (310) into the second learning unit (322) to train the answer generation model.

[0075] The first learning unit (321) can learn a filtering model by inferring relationships between keywords, questions, etc., and the textual content of documents as a training data set is input. The first learning unit (321) can infer parameters that determine the correlation between keywords appearing in academic documents and the textual content while learning a large training data set without labeling. In addition, the first learning unit (321) can infer parameters that determine the correlation between questions in academic documents and the likelihood of the textual content answering the question.

[0076] The first learning unit (321) converts academic documents into embedding vectors, focusing on keywords, questions, and texts of academic documents, in order to input them into a large-scale language model, and processes the embedding vectors with an attention processor to generate a second embedding vector that weights important parts (questions and answers), and inputs the second embedding vector into a converter implemented in the form of a multilayer neural network to train a filtering model. The converter can model the relationship between the second embedding vector and the output sequence, such as whether an answerable keyword appears or whether a question is answerable. The first learning unit (321) can further perform a fine-tuning process to train the filtering model (OD1).

[0077] The second learning unit (322) can learn a filtering model by inferring relationships between keywords, questions, etc. and the main text of documents as a training data set is input. The second learning unit (322) can infer parameters indicating the relationship between keywords appearing in academic documents and the main text while learning a large training data set without labeling. In addition, the second learning unit (322) can divide each academic document into multiple context data and learn each context data as learning data. The second learning unit (322) can infer parameters that determine the relationship between questions, keywords, and one or more sentences corresponding to answers to the questions in the main text.

[0078] The second learning unit (322) divides academic documents into a plurality of context data, focusing on keywords, questions, and texts of academic documents, in order to input them into a large language model, converts the context data into embedding vectors, processes the embedding vectors with an attention processor to generate second embedding vectors that have weights assigned to important parts, and inputs the second embedding vectors into a converter in the form of a multilayer neural network to train an answer generation model.

[0079] The transformer can be modeled by inferring the relationship between the second embedding vector and the output sequences such as sentences corresponding to answers to questions, combinations of sentences, etc. The second learning unit (322) can further perform a fine-tuning process to learn the answer generation model (OD2). The second learning unit (322) can learn keywords, questions, and texts of documents through unsupervised learning, and can learn the answer generation model based on the feature values ​​of sentences that can be included in answers among the texts of documents and are related to questions and keywords.

[0080] The second learning unit (322) can construct an answer generation model including an encoder and a decoder, as illustrated in FIG. 6A. The decoder can process input data using an attention processor that has been modified by learning from academic literature in the medical field. The modified attention processor can set the attention value corresponding to the location of the part corresponding to the question among the input data to 0. By applying the modified attention processor to the decoder, questions and question-related data can be utilized to generate answers, regardless of their importance.

[0081] The learning device (300) can build a filtering model and / or an answer generation model learned from large amounts of data through a model learning unit (320).

[0082] The learning device (300) can learn from academic literature in each field and build a filtering model and / or an answer generation model for each field. The learning device (300) can learn from academic literature related to cancer and build a filtering model that outputs whether cancer-related keywords appear, whether answers to cancer-related questions are possible, etc., and an answer generation model that generates answers to cancer-related questions. The learning device (300) can learn from academic literature limited to specific fields such as cancer, brain disease, heart disease, dementia, etc. and build a filtering model and / or an answer generation model for each field.

[0083] The learning device (300) may be a computing device including a processor, a communication unit, and a memory, and may further include a dataset generation unit (310) and a model learning unit (320).

[0084] Figure 4 is a flowchart of a method for generating an answer according to embodiments of the present disclosure.

[0085] The data providing device (100) can obtain input data including keywords and questions from publicly available web pages of documents (DI).

[0086] The data provision device (100) can input keywords and questions into the filtering unit (Filtering) to obtain filtering data (FT). The filtering unit (FT) can be designed to output data on academic literature corresponding to keywords and questions. The filtering unit (FT) can infer and output filtering data on keywords (whether keywords appear, information on academic literature in which keywords appear, etc.) and filtering data on questions (whether questions can be answered, information on academic literature containing information that can answer the question, etc.) by using a filtering model that has learned academic literature using a large language model and has even performed a fine-tuning process. The filtering data on keywords and the filtering data (OG) on questions inferred and output in this way can be input into the answer generation model (M2) to output answer data on input data including keywords and questions.

[0087] The answer generation model (M2) receives filtering data for keywords in the question data and filtering data related to the question, and can infer information about academic literature related to the answer to the question data. It can also generate answer data for the question data based on the text of the academic literature related to the answer. Upon user request, the answer generation model (M2) can also output information about academic literature related to the answer to the question data. Information about academic literature related to the answer may include the title, abstract, author, publication date, and keywords of the academic literature.

[0088] The data provision device (100) can request answer data for input data from the answer generation model if it is possible to generate an answer through the answer generation model by taking into account the filtering data.

[0089]

[0090] *Figure 5 is a diagram of the input data and output data (Predict answer) of the answer generation model.

[0091] The answer generation model according to embodiments of the present disclosure can be trained using input data including a question, context, and target (answer). The question is extracted from the input text format (see FIG. 8) through natural language processing and can be extracted in a format similar to that shown in FIG. 10.

[0092] The context is generated from the text format for each document (see Figure 8) and may have a form as shown in Figure 9.

[0093]

[0094] *The answer generation model learned in this way can output answer data (Answer data) to questions in response to input data.

[0095] The answer generation model trained in this way can output answer data corresponding to a question based on input data. The answer generation model can generate answer data for a question using academic literature related to keywords and the text of academic literature relevant to the question. Using the results of analyzing academic literature, the answer generation model can detect sentences, paragraphs, sentence fragments, and paragraph fragments that correspond to answers to the question and generate an answer based on this data.

[0096] FIG. 6a is a diagram showing the structure of an answer generation model generated according to embodiments of the present disclosure.

[0097] The answer generation model may include an encoder that processes input data and a decoder that outputs output data. When outputting the output data, which is the answer data to the question, the answer generation model may further include a step (R-setp) of applying a modified attention mask to the area corresponding to the question.

[0098] Figure 6b is an example drawing of an attention mask modified by the answer generation model.

[0099] The answer generation model according to embodiments of the present disclosure can utilize academic literature data to generate output data. The answer generation model can detect an area (R-area) corresponding to a user-entered question in academic literature data containing keywords or questions, and modify an attention mask so that an attention value of "0" is applied to this area. This allows data in the area (R-area) corresponding to the question to be sufficiently considered in answer generation.

[0100] Figure 6c is a diagram comparing a general attention vector and an attention vector generated by an answer generation model.

[0101] As shown in Fig. 6c, compared to the attention vector (AM1) generated by a general attention mask, the attention vector (AM2) generated by the answer generation model has different result values ​​in the area corresponding to the question.

[0102] FIG. 11 is a flowchart of a method for generating an answer according to embodiments of the present disclosure.

[0103] In S110, the data providing device (100) can input / obtain question data including keywords and / or questions input by a user. The data providing device (100) can extract keywords and questions from the question data. The data providing device (100) can extract keywords and / or questions from the question data using natural language processing technology. The data providing device (100) can extract keywords and / or questions using a large-scale data set. Here, the question data is text data written in natural language and may further include pictures, drawings, videos, sound data, etc. The data providing device (100) can extract keywords and / or questions for the question data through a process of analyzing pictures, drawings, videos, and sound data included in the question data.

[0104] In S120, the data provision device (100) can determine filtering data related to the question data. The data provision device (100) can output first filtering data related to keywords in the question data. The first filtering data can include the presence or absence of academic literature in which the keywords appear. The academic literature can be, but is not limited to, literature used to train the filtering model and / or the answer generation model. The academic literature can be registered and published in real time in an academic literature database, such as PubMed.

[0105] The data providing device (100) can output second filtering data related to a question in the question data. The data providing device (100) can output second filtering data regarding a question in the question data. The data providing device (100) can output second filtering data based on whether the question in the question data appears in literature in a field related to the question data. The second filtering data can include whether the question is answerable. Whether the question is answerable can be output as a result of learning the frequency of the question in the academic literature in which the question appears or the characteristics between the sentence containing the question and other sentences.

[0106] In S130, the data providing device (100) can input question data into the answer generation model based on the first filtering data and the second filtering data. If the presence or absence of a keyword is true (TRUE, 1) and the presence or absence of a question is true (TRUE, 1), the question data can be input into the answer generation model to output answer data.

[0107] The data providing device (100) may, in response to the input format of the answer generation model, encode keywords and / or questions included in the question data and convert them into an embedding vector. The embedding vector may encode each of the determined keywords into a predetermined value. The determined keywords may be determined as a result of the answer generation model learning academic literature. The keywords may be determined based on the frequency and / or relevance to the main text of the literature among keywords appearing in the academic literature. For example, the data providing device (100) may generate an embedding vector having 10 columns or rows including whether the first to tenth keywords determined based on the frequency and / or relevance to the main text of the literature appear in the question data.

[0108] An answer generation model can output answer data corresponding to the question data using data on academic literature related to the question data. The answer generation model can convert data on academic literature into an embedding vector, generate an attention vector corresponding to the input question data, and process this attention vector through a converter to output answer data for the question data. The answer generation model can generate a modified attention mask corresponding to the question data, which applies a modified attention value to the region where the question appears, and generate an attention vector by applying this modified attention mask. An attention value of '0' can be applied to prevent missing academic literature data for the region where the question appears.

[0109] In S140, the data providing device (100) can generate output data including answer data and provide the output data to a user terminal.

[0110] In addition, although the preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the specific embodiments described above, and various modifications can be made by a person having ordinary skill in the art to which the present invention pertains without departing from the gist of the present invention claimed in the claims, and such modifications should not be understood individually from the technical idea or prospect of the present invention.

[0111] Therefore, the spirit of the present invention should not be limited to the embodiments described above, and not only the scope of the patent claims described below, but also all scopes equivalent to or equivalently modified from the scope of the patent claims are considered to fall within the scope of the spirit of the present invention.

Claims

1. A step in which a data providing device obtains question data including keywords and questions input by a user; A step in which the data providing device, in relation to the question data, inputs the question data into a filtering model to output first filtering data for the keyword and second filtering data for the question; A step in which the data providing device inputs the question data and the first and second filtering data into an answer generation model, and outputs answer data for the question data; A method for generating an answer to a question using artificial intelligence technology, comprising a step of the data providing device generating output data including the answer data and providing the output data to a user terminal.

2. In paragraph 1, The above filtering model is, A method for generating an answer to a question using artificial intelligence technology, wherein the model is constructed to output the first filtering data including whether the keyword has appeared in academic literature, data on academic literature in which the keyword has appeared, or the second filtering data including whether the question can be answered, data on academic literature in which the question has appeared.

3. In paragraph 1, The above answer generation model is, A method for generating an answer to a question using artificial intelligence technology, which detects academic documents related to the question data using a model trained with registered academic documents as input, analyzes the detected academic documents, and outputs answer data corresponding to the question data.

4. In paragraph 1, The above answer generation model is, A method for generating answers to questions using artificial intelligence technology, which comprises converting academic literature into embedding vectors, inputting the embedding vectors into an attention mechanism and a transformer, and learning through the obtained output.

5. In paragraph 4, The above answer generation model is, A method for generating an answer to a question using artificial intelligence technology, wherein the method detects a question-related area of ​​input question data, processes it with a modified attention mask corresponding to the question, generates an attention vector, and outputs answer data.

6. In paragraph 1, The above filtering model or the above answer generation model, A method for generating answers to questions using artificial intelligence technology, wherein the model is trained with academic literature in the field to which the above question data belongs and is constructed to generate attention vectors centered on questions and keywords related to the field.

7. Includes communication unit, memory and processor, The above processor, Obtaining question data including keywords and questions entered by users, In relation to the above question data, the question data is input into a filtering model to output first filtering data for the keyword and second filtering data for the question, By inputting the above question data and the first and second filtering data into the answer generation model, answer data for the question data is output, A data providing device that generates output data including the above response data and transmits the output data to a user terminal.

8. In paragraph 7, The above filtering model is, A data providing device, which is a model constructed to output the first filtering data including whether the keyword has appeared in academic literature, data on academic literature in which the keyword has appeared, or the second filtering data including whether the question can be answered, data on academic literature in which the question has appeared.

9. In paragraph 7, The above answer generation model is, A data providing device that detects academic documents related to the question data using a model trained with registered academic documents as input, analyzes the detected academic documents, and outputs answer data corresponding to the question data.

10. In paragraph 7, The above answer generation model is, A data providing device that converts academic literature into embedding vectors, inputs the embedding vectors into an attention mechanism and a transformer, and learns through the output obtained.

11. In paragraph 7, The above answer generation model is, A data providing device that detects a question-related area of ​​input question data and processes it with a modified attention mask corresponding to the question to generate an attention vector, thereby outputting answer data.

12. In paragraph 7, The above filtering model or the above answer generation model, A data providing device, which is a model built to generate attention vectors centered on questions and keywords related to the field by learning from academic literature in the field to which the above question data belongs.

13. A computer program stored in a computer-readable storage medium to execute the method of paragraph 1 using a computer.

Images