Method, program and apparatus for quantifying unstructured data and time-series forecasting based on large language model

Abstract

A method, program, and apparatus for quantifying unstructured data based on a large-scale language model and predicting a time series therethrough are disclosed according to one embodiment of the present disclosure. The method may include: acquiring unstructured data comprising at least one of news data, social media data, or data related to official reports; calculating a plurality of quantification indicators by analyzing text included in the unstructured data through a pre-trained large-scale language model; generating quantitative data based on the unstructured data and the calculated plurality of quantification indicators; generating an input data set by integrating the generated quantitative data with structured data including variables related to temporal order; and generating time series data including numerical values ​​regarding demand at a future point in time from the input data set through a pre-trained machine learning model.

Description

Technology Field

[0001] The present disclosure relates to a technology for time series forecasting, and specifically to a method and apparatus for generating time series data by utilizing both structured and unstructured data together, and performing time series forecasting of demand, etc., at a future point in time based thereon. Background Technology

[0003] Time series forecasting is utilized as a core decision-making tool in various industries, including economics, finance, and energy, and is one of the important research topics in the fields of machine learning and data science. Traditional time series forecasting models are trained based on structured data (e.g., historical prices, sales volume, etc.), which demonstrates high performance in stable environments. However, real-world data often contains uncertainty and volatility, and external events such as Black Swans—including wars, natural disasters, and political statements—act as major factors that drastically alter existing patterns. Existing models fail to reflect these unstructured events, revealing limitations in their realistic forecasting performance.

[0004] Unstructured data (e.g., news, social media) contains these external factors, and converting it into time-series data to incorporate into models has the potential to significantly improve predictive performance. However, existing research dealing with unstructured data has several limitations. First, there is a lack of systematic methodologies for quantifying unstructured data and converting it into time-series data. Second, existing approaches often fail to capture key information in unstructured data, such as event context, emotional intensity, and diffusion. Consequently, predictive models are unable to learn the complex external factors of the data.

[0005] Recent advancements in large-scale language models present new possibilities to overcome these limitations. Large-scale language models demonstrate superior performance in extracting key events from vast amounts of unstructured data, understanding text context and sentiment, and quantifying them. While existing natural language processing techniques have remained limited to keyword-based approaches, large-scale language models can reflect the contextual characteristics of data more precisely, making them a powerful tool for integrating unstructured data into time-series forecasting. Prior art literature

[0007] Korean Patent Publication No. 10-2786927 (March 25, 2025) The problem to be solved

[0008] The present disclosure is conceived in response to the aforementioned background technology and aims to provide a method for generating time series data in which unstructured and structured data are integrated using a large-scale language model, and for performing time series forecasting based thereon.

[0009] However, the problems to be solved in this disclosure are not limited to those mentioned above, and other unmentioned problems may be clearly understood based on the description below. means of solving the problem

[0011] A method for quantifying unstructured data based on a large-scale language model and predicting time-series data therethrough, performed by a computing device according to an embodiment of the present disclosure for realizing the aforementioned objectives, is disclosed. The method may include: acquiring unstructured data comprising at least one of news data, social media data, or data related to official reports; calculating a plurality of quantification indicators by analyzing text included in the unstructured data through a pre-trained large-scale language model; generating quantitative data based on the unstructured data and the calculated plurality of quantification indicators; generating an input data set by integrating the generated quantitative data with structured data including variables related to temporal order; and generating time-series data including a figure regarding demand at a future point in time from the input data set through a pre-trained machine learning model.

[0012] Alternatively, the step of calculating a plurality of quantitative indicators by analyzing text included in the unstructured data through the aforementioned pre-trained large-scale language model may include: a step of identifying events related to the demand forecast by analyzing text included in the unstructured data through the aforementioned large-scale language model; and a step of calculating a plurality of quantitative indicators for each identified event by analyzing the context of text included in the unstructured data through the aforementioned large-scale language model.

[0013] Alternatively, the plurality of quantification indicators may include a first indicator representing the number of times the identified event is mentioned in the unstructured data; a second indicator representing the importance and contextual impact of the identified event; a third indicator representing the degree of social diffusion of the identified event; and a fourth indicator regarding the temporal duration of the identified event.

[0014] Alternatively, the quantitative data may be time-series data generated by arranging the calculated plurality of quantification indicators in chronological order.

[0015] Alternatively, the input data set may be generated by mapping the generated quantitative data to variables related to the temporal order included in the structured data.

[0016] Alternatively, the machine learning model may include a first model that learns the correlation between the structured data and the quantitative data by integrating the quantitative data as an exogenous variable into the structured data.

[0017] Alternatively, the machine learning model may include a second model that receives the input data set as a sequence-type input and learns the effect of events included in the quantitative data on demand over time.

[0018] Alternatively, the machine learning model may include a third model that generates an embedding vector based on the input data set and performs a self-attention operation based on the embedding vector.

[0019] A computer program stored on a computer-readable storage medium is disclosed in accordance with an embodiment of the present disclosure for realizing the aforementioned objectives. When the computer program is executed on one or more processors, it performs operations to perform quantification of unstructured data based on a large-scale language model and thereby perform time-series forecasting. At this time, the operations may include: an operation of acquiring unstructured data comprising at least one of news data, social media data, or data related to official reports; an operation of calculating a plurality of quantification indicators by analyzing text contained in the unstructured data through a pre-trained large-scale language model; an operation of generating quantitative data based on the unstructured data and the calculated plurality of quantification indicators; an operation of generating an input data set by integrating the generated quantitative data with structured data including variables related to temporal order; and an operation of generating time-series data including figures regarding demand at a future point in time from the input data set through a pre-trained machine learning model.

[0020] According to one embodiment of the present disclosure for realizing the aforementioned tasks, a computing device is disclosed for performing quantification of unstructured data based on a large-scale language model and thereby performing time series forecasting. The device may include a processor comprising at least one core; a memory comprising program codes executable on the processor; and a network unit for acquiring unstructured data comprising at least one of news data, social media data, or data related to official reports. In this case, the processor may calculate a plurality of quantification indicators by analyzing text included in the unstructured data through a pre-trained large-scale language model, generate quantitative data based on the unstructured data and the calculated plurality of quantification indicators, generate an input data set by integrating the generated quantitative data with structured data including variables related to temporal order, and generate time series data including figures regarding demand at a future point in time from the input data set through a pre-trained machine learning model. Effects of the invention

[0022] The present disclosure evaluates the contextual meaning, importance, and temporal persistence of unstructured data through a large-scale language model to quantify them multidimensionally, and integrates this with structured data to substantially improve the performance of existing time series forecasting models that relied solely on structured data. Brief explanation of the drawing

[0024] FIG. 1 is a block diagram of a computing device according to one embodiment of the present disclosure. FIG. 2 is a block diagram illustrating the process of quantifying unstructured data and predicting time series performed by a computing device according to one embodiment of the present disclosure. FIG. 3 is a flowchart illustrating a method for quantifying unstructured data and predicting time series according to one embodiment of the present disclosure. Specific details for implementing the invention

[0025] Embodiments of the present disclosure are described below with reference to the attached drawings so that those skilled in the art (hereinafter, those skilled in the art) can easily implement them. The embodiments presented in the present disclosure are provided to enable those skilled in the art to use or implement the contents of the present disclosure. Accordingly, various modifications to the embodiments of the present disclosure will be apparent to those skilled in the art. That is, the present disclosure may be embodied in various different forms and is not limited to the embodiments below.

[0026] Throughout the specification of the present disclosure, identical or similar reference numerals refer to identical or similar components. Additionally, to clearly explain the present disclosure, reference numerals in the drawings that are unrelated to the description of the present disclosure may be omitted.

[0027] The term “or” as used in this disclosure is intended to mean an implicit “or” rather than an exclusive “or.” That is, unless otherwise specified in this disclosure or its meaning is not clear from the context, “X uses A or B” should be understood to mean one of the natural implicit substitutions. For example, unless otherwise specified in this disclosure or its meaning is not clear from the context, “X uses A or B” may be interpreted as X using A, X using B, or X using both A and B.

[0028] The term “and / or” as used in this disclosure should be understood to refer to and include all possible combinations of one or more of the enumerated related concepts.

[0029] The terms “comprising” and / or “comprising” as used in this disclosure should be understood to mean the presence of certain features and / or components. However, the terms “comprising” and / or “comprising” should be understood not to exclude the presence or addition of one or more other features, other components and / or combinations thereof.

[0030] Where not otherwise specified in the present disclosure or where it is not clear from the context that the singular form indicates, the singular should generally be interpreted as including “one or more.”

[0031] The term “the N (N is a natural number)” used in this disclosure may be understood as an expression used to distinguish the components of this disclosure from one another according to certain criteria, such as functional perspectives, structural perspectives, or convenience of explanation. For example, components performing different functional roles in this disclosure may be distinguished as a first component or a second component. However, components that are substantially identical within the technical scope of this disclosure but must be distinguished for the convenience of explanation may also be distinguished as a first component or a second component.

[0032] The term “connection” as used in the present disclosure should be interpreted to include not only cases where the components are “directly connected,” but also cases where other components are “present” in between, and cases where they are “electrically connected” with other components in between.

[0033] The term “acquisition” as used in this disclosure may be understood to refer not only to receiving data through a wireless communication network with an external device or system, but also to generating or receiving data in an on-device form.

[0034] Meanwhile, the terms "module" or "unit" used in this disclosure may be understood as referring to an independent functional unit that processes computing resources, such as a computer-related entity, firmware, software or a part thereof, hardware or a part thereof, or a combination of software and hardware. In this case, "module" or "unit" may be a unit composed of a single element, or a unit expressed as a combination or set of multiple elements. For example, in a narrow sense, "module" or "unit" may refer to a hardware element of a computing device or a set thereof, an application program that performs a specific function of software, a procedure implemented through software execution, or a set of instructions for program execution. Furthermore, in a broad sense, "module" or "unit" may refer to the computing device itself that constitutes the system, or an application executed on the computing device. However, since the above-described concept is merely an example, the concept of "module" or "part" may be defined in various ways within the scope understandable to those skilled in the art based on the contents of this disclosure.

[0035] As used in this disclosure, the term “model” may be understood as a system implemented using mathematical concepts and language to solve a specific problem, a set of software units to solve a specific problem, or an abstract model regarding a process to solve a specific problem. For example, a machine learning “model” may refer to an overall system that possesses problem-solving capabilities through learning based on data.

[0036] The explanation of the foregoing terms is intended to aid in understanding the present disclosure. Accordingly, it should be noted that unless a foregoing term is explicitly stated as a matter limiting the content of the present disclosure, it is not to be used in the sense of limiting the technical concept of the content of the present disclosure.

[0037] FIG. 1 is a block diagram of a computing device according to one embodiment of the present disclosure.

[0038] A computing device (100) according to one embodiment of the present disclosure may be a hardware device or a part of a hardware device that performs comprehensive processing and computation of data, or it may be a software-based computing environment connected to a communication network. For example, the computing device (100) may be a server that performs intensive data processing functions and shares resources through communication with clients. Additionally, the computing device (100) may be a cloud system that enables multiple servers and clients to process data comprehensively. Since the above description is merely one example regarding the type of computing device (100), the type of computing device (100) may be configured in various ways within a range understandable to those skilled in the art based on the contents of the present disclosure.

[0039] Referring to FIG. 1, a computing device (100) according to one embodiment of the present disclosure may include a processor (110), a memory (120), and a network unit (130). However, since FIG. 1 is merely an example, the computing device (100) may include other configurations for implementing a computing environment. Additionally, only some of the disclosed configurations may be included in the computing device (100).

[0040] A processor (110) according to one embodiment of the present disclosure may be understood as a constituent unit comprising hardware and / or software for performing computing operations. For example, the processor (110) may process instructions generated as a result of user interaction through a user interface. Additionally, the processor (110) may read a computer program and perform data processing for machine learning. The processor (110) may process computational processes such as processing input data for machine learning, feature extraction for machine learning, and error calculation based on backpropagation. A processor (110) for performing such data processing and operations may include a central processing unit (CPU), a general purpose graphics processing unit (GPGPU), a tensor processing unit (TPU), an application specific integrated circuit (ASIC), or a field programmable gate array (FPGA). Since the above-described type of processor (110) is merely an example, the type of processor (110) can be configured in various ways within a range understandable to those skilled in the art based on the contents of the present disclosure.

[0041] The processor (110) can quantify unstructured data for demand forecasting to generate structured quantitative data. The processor (110) can analyze text contained in unstructured data using a pre-trained large-scale language model. The processor (110) can calculate multiple quantification indicators based on the results of analyzing text contained in unstructured data using a pre-trained large-scale language model. Then, the processor (110) can generate quantitative data based on the multiple quantification indicators. At this time, the quantitative data may be data structured so that unstructured data can be integrated into time series forecasting based on structured data. By generating structured quantitative data from unstructured data using a large-scale language model, the processor (110) can provide an environment in which unstructured data can be reflected in time series forecasting.

[0042] For example, the processor (110) can identify at least one event related to demand forecasting by inputting unstructured data into a pre-trained large-scale language model and analyzing the text contained in the unstructured data. At this time, the event related to demand forecasting may include an event corresponding to an external factor affecting demand forecasting. The processor (110) can operate the large-scale language model to identify the event by analyzing the text contained in the unstructured data based on a pre-set system prompt. The processor (110) can calculate multiple quantification indicators for each event identified through the large-scale language model. At this time, the multiple quantification indicators may be numerical values ​​obtained by analyzing the context of the text containing the event to evaluate the sentiment, importance, relevance, etc. of the event. That is, the processor (110) can produce multiple quantification indicators that quantify the influence of the event by analyzing the context of the text containing the event through the large-scale language model. Furthermore, the processor (110) can generate structured quantitative data by organizing the multiple quantification indicators calculated for each event along a time axis. The processor (110) can generate quantitative data having time-series characteristics by arranging events identified in unstructured data and multiple quantification indicators calculated for each event in time units. Here, the time unit may be daily, weekly, monthly, etc. That is, the processor (110) can extract key events related to demand forecasting from unstructured data through a pre-trained large-scale language model, and convert the unstructured data into numerical data having time-series characteristics through quantification indicators related to the key events, thereby creating a form that is analyzable (i.e., integrable with structured data).

[0043] Meanwhile, a large-scale language model used for the analysis of structured data may be a neural network model specialized for natural language processing, composed of a large number of parameters. A large-scale language model used for the analysis of unstructured data may be constructed based on a transformer-based neural network structure. A large-scale language model used for the analysis of unstructured data may perform self-supervised learning and / or reinforcement learning to extract meaning from a series of texts and understand the relationships between words and phrases within the text. For example, a large-scale language model used for the analysis of unstructured data may be a model such as a generative pre-trained transformer (GPT). However, the large-scale language model of the present disclosure is not limited to these examples and may be composed of various forms of models applicable to those skilled in the art based on the description above.

[0044] The processor (110) can generate an input data set for time series forecasting by integrating quantitative data generated from unstructured data with structured data for demand forecasting. At this time, the structured data for demand forecasting may include variables related to temporal order, such as day, week, month, and date. Since the quantitative data generated from unstructured data is time series data in which quantitative indicators for each event affecting demand forecasting are arranged in temporal order, the processor (110) can generate an input data set to be used for time series forecasting by matching the quantitative data and the structured data based on the time axis.

[0045] For example, the processor (110) can match structured data, in which demand-related variables such as price and trading volume are organized in chronological order, with quantitative data, in which quantitative indicators for each event are arranged in time units, based on a time axis. If the unit of chronological order in the structured data is a date, the processor (110) can combine the structured data and the quantitative data by matching the event-specific quantitative indicators included in the quantitative data based on the sameness of the date. That is, the processor (110) can create an expanded data set containing new features by adding quantitative indicators for unstructured data to the existing quantitative data in accordance with time units.

[0046] The processor (110) can perform time series forecasting based on an input data set in which unstructured data and structured data are integrated, using a pre-trained machine learning model. The processor (110) can generate time series data containing figures regarding demand at a future point in time by inputting the input data set in which unstructured data and structured data are integrated into the pre-trained machine learning model. The machine learning model can generate data representing changes in demand over time by learning the correlation between event-specific quantitative indicators generated from unstructured data and demand variables included in structured data. At this time, the learning method of the machine learning model may vary depending on the type of model. For example, if the machine learning model is a recurrent neural network-based model, the machine learning model can learn sequential changes and characteristics of variables over time by performing supervised learning based on training data including labels. If the machine learning model is a transformer-based model that performs self-attention operations, the machine learning model can model complex relationships between variables by performing self-supervised learning based on training data that does not include labels.

[0047] The processor (110) can generate a time series prediction report based on time series data generated through a machine learning model. The processor (110) can also include in the time series prediction report the results of interpreting the impact of unstructured data on time series prediction through data comparison evaluation using a machine learning model. In this case, the data comparison evaluation using a machine learning model may be the result of comparing and analyzing the generation of time series data performed based only on structured data with the generation of time series data performed based on an input data set in which unstructured data and structured data are integrated. For example, the processor (110) can calculate how much the accuracy of the result of time series prediction performed based on integrated data of unstructured and structured data has improved compared to the accuracy of the result of time series prediction performed based only on structured data. In this case, the processor (110) can analyze, event by event, how much the event identified in the unstructured data influenced the improvement of accuracy. At this time, the analysis may be performed through a large-scale language model used for quantification, or through a separate machine learning model prepared for data comparison evaluation. The processor (110) may include these comparison and analysis results in a time series forecast report along with the time series data. Since time series forecasting predicts changes in demand over time, the user can easily check the predicted changes in demand and external variables affecting demand forecasting in the future after the present through the time series forecast report generated by the processor (110).

[0048] A memory (120) according to one embodiment of the present disclosure may be understood as a configuration unit comprising hardware and / or software for storing and managing data processed by a computing device (100). That is, the memory (120) may store data of any form generated or determined by a processor (110) and data of any form received by a network unit (130). For example, the memory (120) 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, RAM (random access memory), SRAM (static random access memory), ROM (read-only memory), EEPROM (electrically erasable programmable read-only memory), PROM (programmable read-only memory), magnetic memory, a magnetic disk, or an optical disk. Additionally, the memory (120) may include a database system that controls and manages data in a predetermined system. Since the above-described type of memory (120) is merely an example, the type of memory (120) can be configured in various ways within a range understandable to those skilled in the art based on the contents of the present disclosure.

[0049] Memory (120) can structure and organize data, combinations of data, and program code executable by the processor (110) that are necessary for the processor (110) to perform calculations. For example, memory (120) can store program code that causes the processor (110) to process data, program code that causes the processor (110) to create a machine learning model and perform demand forecasting calculations using the created machine learning model, and various data generated as the program code is executed.

[0050] A network unit (130) according to one embodiment of the present disclosure may be understood as a configuration unit that transmits and receives data through any known form of wired or wireless communication system. For example, the network unit (130) may perform data transmission and reception using wired or wireless communication systems such as a local area network (LAN), wideband code division multiple access (WCDMA), long term evolution (LTE), wireless broadband internet (WiBro), 5th generation mobile communication (5G), ultra-wide-band wireless communication, ZigBee, radio frequency (RF) communication, wireless LAN, wireless fidelity (Wi-Fi), near field communication (NFC), or Bluetooth. Since the communication systems described above are merely examples, wired or wireless communication systems for data transmission and reception of the network unit (130) may be applied in various ways other than those described above.

[0051] The network unit (130) can receive data necessary for the processor (110) to perform calculations through wired or wireless communication with any system or any client. Additionally, the network unit (130) can transmit data generated through the processor (110)'s calculations through wired or wireless communication with any system or any client. The network unit (130) can transmit and receive data generated through interaction with any entity through communication with the entity. For example, the network unit (130) can obtain structured data containing variables indicating temporal order or unstructured data containing information regarding events affecting demand forecasting through communication with a database. Furthermore, the network unit (130) can transmit data processed by the processor (110) to a database or a client for verifying demand forecasting results.

[0052] FIG. 2 is a block diagram illustrating the process of quantifying unstructured data and time series forecasting performed by a computing device according to one embodiment of the present disclosure. The demand forecasting framework, composed of the preprocessing module (200), large-scale language model (300), and machine learning model (400) disclosed in FIG. 2, can be performed by a processor (110) of a computing device (100) according to one embodiment of the present disclosure.

[0053] A preprocessing module (200) according to one embodiment of the present disclosure may perform preprocessing to convert unstructured data related to demand forecasting into a form suitable for processing. The preprocessing module (200) may receive first unstructured data (10) produced from various sources and generate second unstructured data (15) processed into a form suitable for subsequent work (e.g., quantification). At this time, the first unstructured data (10) may include at least one of news data, social media data, or data related to official reports. For example, news data may be data produced by recording economic, political, or environmental events in the field for which demand is to be forecasted on a news platform. News data may be collected through a web crawling method. Social media data may be data reflecting the real-time public reactions and sentiments in the field for which demand is to be forecasted on a social media platform. Social media data may be collected in a real-time streaming manner based on an API (application programming interface) provided by a social media platform. Data related to official reports may include reports produced by credible institutions related to the field for which demand is to be forecasted. Data related to official reports may be collected by downloading reports provided on the websites of credible institutions in response to user commands.

[0054] The preprocessing module (200) can generate refined second unstructured data (15) by performing the task of filtering unnecessary or duplicate information from the first unstructured data (10) and the task of tokenizing the text included in the first unstructured data (10). For example, the preprocessing module (200) can remove HTML tags, special characters, duplicate characters, etc. included in the first unstructured data (10), and remove unnecessary metadata such as advertising phrases or time information. In addition, the preprocessing module (200) can divide the text included in the first unstructured data (10) into sentence units and tokenize each sentence to process it into a form suitable for processing by artificial intelligence. Through this process, the preprocessing module (200) can generate refined second unstructured data (15) from the first unstructured data (10). If the first unstructured data (10) is written in a language other than English, the preprocessing module (200) may perform the task of translating the first unstructured data (10) according to the user's needs.

[0055] Meanwhile, if the first unstructured data (10) is collected in a state where it has already been processed, such as the second unstructured data (15), the first unstructured data (10) may be input directly into a large-scale language model (300) for quantification without passing through the preprocessing module (200). That is, the preprocessing module (200) can operate selectively depending on the collection state of the first unstructured data (10).

[0056] A large-scale language model (300) according to one embodiment of the present disclosure can perform a quantification operation on unstructured data by analyzing text included in unstructured data. The large-scale language model (300) can identify at least one event related to demand forecasting by analyzing text included in the second unstructured data (15). The large-scale language model (300) can extract text related to the event identified in the second unstructured data (15). Furthermore, the large-scale language model (300) can calculate a plurality of quantification indicators for each event by analyzing the context of the text extracted in relation to the event. At this time, the plurality of quantification indicators may include a first indicator (21) indicating the number of times the event is mentioned in the unstructured data, a second indicator (23) indicating the importance and contextual impact of the event, a third indicator (25) indicating the degree of social diffusion of the event, and a fourth indicator (27) regarding the temporal duration of the event. Each indicator may be a quantification of the emotions, importance, etc. represented by the event, by analyzing the context of the text extracted in relation to the event.

[0057] For example, the first indicator (21) may be measured by summing the number of times an event affecting demand forecasting is mentioned within the second unstructured data (15). The second indicator (23) may be determined by combining the first indicator (21) and the sentiment score regarding the event. The sentiment score regarding the event may be a numerical value calculated by a large-scale language model (300) analyzing the emotional tone of text related to the event. In this case, the combination of the first indicator (21) and the sentiment score may correspond to an operation of multiplying the first indicator (21) and the sentiment score, but various other operations that can be used to normalize the numerical values ​​by combining them may be applied. The third indicator (25) may be derived by calculating the average value of the results of quantifying people's reactions related to the event. People's reactions related to the event may be generated through actions such as retweets and likes on social media platforms. Accordingly, the result of quantifying people's reactions related to the event may correspond to the number of actions extracted from social media data included in the second unstructured data (15). The fourth indicator (27) may be derived by calculating the duration based on the start date and end date of the event mentioned in the second unstructured data (15). In other words, the fourth indicator (27) may represent a value quantified from the start time of the event to the end time. Such indicators may be calculated for each event identified in the second unstructured data (15).

[0058] Meanwhile, multiple quantitative indicators may additionally include a fifth indicator, which is a composite index that integrates the first indicator (21) to the fourth indicator (27). Although it is necessary to individually reflect the quantitative characteristics represented by the first indicator (21) to the fourth indicator (27), if the number of events is large, the amount of information stored increases, which may not be efficient in terms of data management. Therefore, the large-scale language model (300) can generate summarized information by calculating a composite index that integrates each indicator. For example, the fifth indicator can be calculated by calculating the weighted sum of the first indicator (21), the second indicator (22), the third indicator (25), and the fourth indicator (27). At this time, the weights assigned to each indicator can be adjusted through expert feedback in the field where demand forecasting is to be performed.

[0059] Meanwhile, multiple quantitative indicators may additionally include a sixth indicator indicating the importance of an event and a seventh indicator indicating the time at which the event occurred. For example, the sixth indicator may be measured based on an evaluation by a large-scale language model (300) regarding an event that affects demand forecasting. The large-scale language model (300) may determine how important a specific event is to a specific demand based on text included in the second unstructured data (15) as well as a knowledge database regarding demand forecasting. The sixth indicator may be a numerical value based on this evaluation. In addition to the large-scale language model (300), the sixth indicator may also be calculated through a feedback model of a domain expert related to specific demand forecasting. The seventh indicator may be calculated by determining the time at which the event occurred based on a timestamp recorded when the first unstructured data (10) was acquired.

[0060] The processor (110) can generate structured quantitative data (30) by arranging multiple indicators generated through a large-scale language model (300) in time units. For example, the processor (110) can group multiple quantitative indicators calculated per event based on the date or period of occurrence of the event. The processor (110) can map the grouped event-specific indicators to a time axis so that the events and their corresponding indicators are sorted in chronological order. Then, the processor (110) can generate quantitative data (30) containing the events and their corresponding quantitative indicators sorted in chronological order. Since the quantitative data (30) is generated based on information listed in chronological order, it can be considered to correspond to time-series data itself. That is, the processor (110) can convert the second unstructured data (15) into quantitative data (30) corresponding to time-series data based on multiple indicators generated through a large-scale language model (300).

[0061] The processor (110) can generate an input data set (50) for time series forecasting by integrating structured quantitative data (30) and structured data (40) containing demand-related variables. At this time, the structured data (40) may be time series data containing variables related to temporal order. Accordingly, the processor (110) can generate an input data set (50) by matching the structured quantitative data (30) and the structured data (40) based on the time axis. For example, the processor (110) can map information about prices on a specific date included in the structured data (40) to information about events and quantification indicators on the same date included in the quantitative data (30). In this way, the processor (110) can generate an input data set (50) by mapping the quantitative data (30) to the unit of the variables related to temporal order included in the structured data (40).

[0062] A machine learning model (400) according to one embodiment of the present disclosure can generate time series data (60) including numerical values ​​regarding demand at a future point in time by analyzing an input data set (50) generated by integrating quantitative data (30) and structured data (40). At this time, the machine learning model (400) may include sub-models that can operate selectively or integrally depending on the field and purpose of time series forecasting. That is, the machine learning model (400) may be a single model for performing a specific time series forecast, or it may be a hybrid model in which a plurality of models are combined.

[0063] For example, the machine learning model (400) may include a first model that learns the correlation between the structured data (40) and the quantitative data (30) by integrating the quantitative data (30) as an exogenous variable into the structured data (40). The first model may be a model suitable for performing short-term predictions based on past data patterns. For example, the first model may include an ARIMA (autoregressive integrated moving average) model that aims to perform short-term predictions based on the correlation between past structured data and unstructured data. The machine learning model (400) may include a second model that receives the input data set (50) as a sequence-type input and learns the impact of events included in the quantitative data (30) on demand over time. For example, the second model may be a model capable of learning long-term dependencies and non-linear relationships, and may be an LSTM (long short-term memory) model. The machine learning model (400) may include a third model that generates embedding vectors based on an input data set (50) and performs self-attention operations based on the embedding vectors. For example, the third model may include a transformer model capable of learning complex interactions between time series in large-scale data through a multi-head attention mechanism.

[0064] The machine learning model (400) may include various models suitable for time series prediction in addition to the examples described above, and may generate time series data (60) based on a combination of these models. For example, if a short-term prediction needs to be performed by a user command, the machine learning model (400) may generate time series data (60) by running the first model, which is suitable for short-term prediction among the first model, the second model, and the third model. If a long-term prediction needs to be performed by a user command, the machine learning model (400) may generate time series data by running both the second model and the third model. Then, the machine learning model (400) may generate time series data (60) to be used for prediction by integrating the time series data generated by the second model and the third model, respectively.

[0065] FIG. 3 is a flowchart illustrating a method for quantifying unstructured data and predicting time series according to one embodiment of the present disclosure.

[0066] Referring to FIG. 3, a computing device (100) according to one embodiment of the present disclosure may acquire unstructured data including at least one of news data, social media data, or data related to official reports (S100). For example, the computing device (100) may collect unstructured data regarding the raw materials market from various data sources, such as news platforms, social media platforms, or the homepage of an organization issuing official reports. At this time, the collection may be carried out in various ways, such as crawling the data source or API integration. News data that can be collected as unstructured data may be article data dealing with events and price volatility related to raw materials such as oil, gold, and copper. Social media data that can be collected as unstructured data may be data extracted from opinions related to specific raw materials using hashtags on social media platforms. Data related to official reports may be data extracted from reports related to the raw materials market published by organizations such as the UN and the IMF.

[0067] A computing device (100) can perform quantification of unstructured data using a large-scale language model (S200). For example, the computing device (100) can calculate multiple quantification indicators by analyzing text included in unstructured data regarding raw material markets through a pre-trained large-scale language model. Specifically, the computing device (100) can identify events related to raw material markets by analyzing text included in unstructured data regarding raw material markets through a large-scale language model. At this time, events related to raw material markets can be classified into any one of the following: production site issues, which include events occurring in countries that produce raw materials; consumer issues, which include changes in demand occurring in countries that consume raw materials; and macroeconomic issues, which include economic factors affecting the entire raw material market. Production site issues may include various factors affecting production activities, such as policy changes in producing countries, political instability, and natural disasters. Consumer issues may include the economic conditions, environmental regulations, and policy changes of major raw material consuming countries, and may include various consumption factors that directly affect the demand for raw materials. Macroeconomic issues may include changes in international interest rates, global economic recession, and trade agreements between major economic blocs. Based on these classifications, the computing device (100) can identify and extract events related to the raw materials market from the text included in the unstructured data. Additionally, the computing device (100) can calculate multiple quantitative indicators for each event by analyzing the context of the text included in the unstructured data through a large-scale language model. A detailed explanation of the quantitative indicators is substituted with the description of FIG. 2 described above. The computing device (100) can generate quantitative data based on the unstructured data regarding the raw materials market and the multiple quantitative indicators calculated above.The computing device (100) can generate structured quantitative data having time series characteristics by arranging the quantification indicators calculated for each event according to the temporal order of the events.

[0068] The computing device (100) can generate an input data set for time series forecasting by integrating unstructured data quantified through S200 with structured data regarding demand forecasting (S300). For example, the computing device (100) can generate an input data set for demand forecasting by integrating quantitative data generated from unstructured data with structured data containing information on changes in raw material prices over time. Specifically, the computing device (100) can generate an expanded data set containing various factors affecting raw material prices by mapping quantified indicators included in the quantitative data to information on raw material prices included in the structured data based on a specific date. The input data set generated through the steps described above can contribute to increasing the accuracy of the forecast by reflecting various events or external factors that could not be reflected by the structured data alone into the forecast.

[0069] The computing device (100) can generate time series data for demand forecasting using a pre-trained machine learning model (S400). For example, the computing device (100) can input the input data set generated through S300 into the machine learning model to generate time series data containing forecast information on raw material prices at a future point in time. Here, the machine learning model may include all statistical models, recurrent neural network-based models, tree-based models, etc., that can be used for time series forecasting.

[0070] The computing device (100) can analyze the contribution of events included in unstructured data to the demand forecast result by comparing the result of generating time series data based only on structured data through a pre-trained machine learning model with the result of generating time series data based on an input data set. For example, the computing device (100) can generate time series data by inputting only structured data into a pre-trained machine learning model. The computing device (100) can compare the accuracy of the result of inputting only structured data with the result of inputting an input data set. Through this comparison, the computing device (100) can analyze the contribution of events related to the raw material market to the price forecast result. If the unstructured data includes events related to production site issues, such as "Indonesia labor policy tightening," the computing device (100) can calculate through comparative analysis how much the event related to these production site issues improved the forecast accuracy. If the accuracy of the result obtained by inputting the input data set is calculated to be improved by 25% compared to the result obtained by inputting only structured data, the computing device (100) can normalize the calculated value and convert it into a contribution of an event related to production site issues such as "Indonesia labor policy tightening." Such a contribution can be used as an indicator to interpret the impact of a specific event included in the quantified unstructured data on the prediction result. Meanwhile, the calculation of the contribution is not mandatory and can be performed optionally according to user commands.

[0071] The computing device (100) can generate a time series forecast report containing time series data generated using a machine learning model (S500). For example, the computing device (100) can graphically visualize time series data containing forecast information on raw material prices at a future point in time and reflect it in the raw material price forecast report. The computing device (100) can also generate interpretation results for the time series data through a large-scale language model and reflect them in the raw material price forecast report. Furthermore, if the contribution of events related to the raw material market to the price forecast result is calculated, the computing device (100) can selectively include information regarding the contribution in the raw material price forecast report according to user commands. That is, the computing device (100) can synthesize the information necessary for time series forecasting of raw material prices to generate a report that the user can intuitively and easily understand, and can provide this to the user terminal.

[0072] The various embodiments of the present disclosure described above may be combined with additional embodiments and modified to the extent understandable to those skilled in the art in light of the detailed description above. The embodiments of the present disclosure are illustrative in all respects and should be understood as not restrictive. For example, each component described as a single unit may be implemented in a distributed manner, and components described as distributed may likewise be implemented in a combined form. Accordingly, all modifications or variations derived from the meaning, scope, and equivalents of the claims of the present disclosure should be interpreted as being included within the scope of the present disclosure.

Claims

Claim 1 A method for quantifying unstructured data based on a large-scale language model and predicting a time series therefrom, performed by a computing device including at least one processor, comprising: a step of acquiring unstructured data including at least one of news data, social media data, or data related to official reports; a step of calculating a plurality of quantification indicators by analyzing text included in the unstructured data through a pre-trained large-scale language model; a step of generating quantitative data based on the unstructured data and the calculated plurality of quantification indicators; a step of generating an input data set by integrating the generated quantitative data with structured data including variables related to temporal order; and a step of generating time series data including figures regarding demand at a future point in time from the input data set through a pre-trained machine learning model. Claim 2 A method according to claim 1, wherein the step of calculating a plurality of quantitative indicators by analyzing text included in the unstructured data through the pre-trained large-scale language model comprises: a step of identifying events related to the demand forecast by analyzing text included in the unstructured data through the large-scale language model; and a step of calculating a plurality of quantitative indicators for each identified event by analyzing the context of text included in the unstructured data through the large-scale language model. Claim 3 A method according to claim 1, wherein the plurality of quantification indicators comprises: a first indicator representing the number of times the identified event is mentioned in the unstructured data; a second indicator representing the importance and contextual impact of the identified event; a third indicator representing the degree of social diffusion of the identified event; and a fourth indicator regarding the temporal duration of the identified event. Claim 4 A method according to claim 1, wherein the quantitative data is time-series data generated by arranging the calculated plurality of quantification indicators in chronological order. Claim 5 A method according to claim 1, wherein the input data set is generated by mapping the generated quantitative data to a variable related to the temporal order included in the structured data. Claim 6 A method according to claim 1, wherein the machine learning model comprises a first model that learns the correlation between the structured data and the quantitative data by integrating the quantitative data as an exogenous variable into the structured data. Claim 7 A method according to claim 1, wherein the machine learning model receives the input data set as a sequence-type input and includes a second model that learns the effect of events included in the quantitative data on demand over a temporal flow. Claim 8 A method according to claim 1, wherein the machine learning model comprises a third model that generates an embedding vector based on the input data set and performs a self-attention operation based on the embedding vector. Claim 9 A computer program stored on a computer-readable storage medium, wherein, when the computer program is executed on one or more processors, it performs operations to perform quantification of unstructured data based on a large-scale language model and thereby perform time-series forecasting, the operations comprising: an operation of acquiring unstructured data including at least one of news data, social media data, or data related to official reports; an operation of calculating a plurality of quantification indicators by analyzing text included in the unstructured data through a pre-trained large-scale language model; an operation of generating quantitative data based on the unstructured data and the calculated plurality of quantification indicators; an operation of generating an input data set by integrating the generated quantitative data with structured data including variables related to temporal order; and an operation of generating time-series data including figures regarding demand at a future point in time from the input data set through a pre-trained machine learning model. Claim 10 A computing device for performing quantification of unstructured data based on a large-scale language model and thereby performing time series forecasting, comprising: a processor including at least one core; a memory including program codes executable on said processor; and a network unit for acquiring unstructured data including at least one of news data, social media data, or data related to official reports; wherein the processor analyzes text included in said unstructured data through a pre-trained large-scale language model to calculate a plurality of quantification indicators, generates quantitative data based on said unstructured data and said calculated plurality of quantification indicators, integrates said generated quantitative data with structured data including variables related to temporal order to generate an input data set, and generates time series data including figures regarding demand at a future point in time from said input data set through a pre-trained machine learning model.

Images