• Proceedings of the Korean Society of Computer Information Conference
• /
• 2023.07a
• /
• pp.541-544
• /
• 2023

LLM(Largescale Language Model)의 성능 향상을 위한 비용 효율적인 방법으로 ChatGPT, GPT-4와 같은 초거대 모델의 output에 대해 SLM(Small Language Model)을 finetune하는 방법이 주목받고 있다. 그러나, 이러한 접근법은 주로 범용적인 지시사항 모델을 위한 학습 방법으로 사용되며, 제한된 특정 도메인에서는 추가적인 성능 개선의 여지가 있다. 본 연구는 특정 도메인(Writing Assistant)에서의 성능 향상을 위한 새로운 방법인 Self-Guided Approach를 제안한다. Self-Guided Approach는 (1) LLM을 활용해 시드 데이터에 대해 도메인 특화된 metric(유용성, 관련성, 정확성, 세부사항의 수준별) 점수를 매기고, (2) 점수가 매겨진 데이터와 점수가 매겨지지 않은 데이터를 모두 활용하여 supervised 방식으로 SLM을 미세 조정한다. Vicuna에서 제안된 평가 방법인, GPT-4를 활용한 자동평가 프레임워크를 사용하여 Self-Guided Approach로 학습된 SLM의 성능을 평가하였다. 평가 결과 Self-Guided Approach가 Self-instruct, alpaca와 같이, 생성된 instruction 데이터에 튜닝하는 기존의 훈련 방법에 비해 성능이 향상됨을 확인했다. 다양한 스케일의 한국어 오픈 소스 LLM(Polyglot1.3B, PolyGlot3.8B, PolyGlot5.8B)에 대해서 Self-Guided Approach를 활용한 성능 개선을 확인했다. 평가는 GPT-4를 활용한 자동 평가를 진행했으며, Korean Novel Generation 도메인의 경우, 테스트 셋에서 4.547점에서 6.286점의 성능 향상이 발생했으며, Korean scenario Genration 도메인의 경우, 테스트 셋에서 4.038점에서 5.795 점의 성능 향상이 발생했으며, 다른 유사 도메인들에서도 비슷한 점수 향상을 확인했다. Self-Guided Approach의 활용을 통해 특정 도메인(Writing Assistant)에서의 SLM의 성능 개선 가능성을 확인했으며 이는 LLM에 비용부담을 크게 줄이면서도 제한된 도메인에서 성능을 유지하며, LLM을 활용한 응용 서비스에 있어 실질적인 도움을 제공할 수 있을 것으로 기대된다.

• Journal of Internet Computing and Services
• /
• v.25 no.3
• /
• pp.27-33
• /
• 2024

The environment of academic research is continuously changing due to the increase of information, which raises the need for an effective way to analyze and organize large amounts of documents. In this paper, we propose Performance Improvement of Topic Modeling using BART(Bidirectional and Auto-Regressive Transformers) based Document Summarization. The proposed method uses BART-based document summary model to extract the core content and improve topic modeling performance using LDA(Latent Dirichlet Allocation) algorithm. We suggest an approach to improve the performance and efficiency of LDA topic modeling through document summarization and validate it through experiments. The experimental results show that the BART-based model for summarizing article data captures the important information of the original articles with F1-Scores of 0.5819, 0.4384, and 0.5038 in Rouge-1, Rouge-2, and Rouge-L performance evaluations, respectively. In addition, topic modeling using summarized documents performs about 8.08% better than topic modeling using full text in the performance comparison using the Perplexity metric. This contributes to the reduction of data throughput and improvement of efficiency in the topic modeling process.

• Archives of design research
• /
• v.18 no.3 s.61
• /
• pp.45-54
• /
• 2005

Visual messages are fundamental elements for performing communication and indicate the signs which are delivered from the communicator to the communicatee via channels, Generally, we can classify visual messages into two groups; linguistic factors which are rational and deliver abstract concepts and unlinguistic factors which are mental and can be expressed concretely. Especially, web site with receivers' low attention and concentration need images which can attract their attention to visual messages. That is, web site is a medium which allows us to feel visual and emotional experiences. We can call it a standard of sign systems which are consisted of various styles of digital texts. The main purpose of this study lies in that we'll analyze how homepage introductory page as one of the forms of digital text conduces a meaning action to the receivers and that we'll apprehend the structures of images and different types of signs via a semiotic approach and analyze the underlying meaning of the messages. In order to survey the structures of images we'll look into the attitude toward perceiving messages by using semantic differential method which has been developed mostly by Osgood and analyze the visual images by adopting sign types of Fuss. As the signification is created by combining signs, it is significant that we'll analyze the meaning of sings between the transmitters and receivers from the semiotic viewpoint and study the signification systems.

• Journal of The Korean Association of Information Education
• /
• v.6 no.1
• /
• pp.64-74
• /
• 2002

The purpose of this paper is to review the importance of situated cognition and the features of MOO(Multi-user Object Oriented)environments for effective foreign language learning. Learning foreign languages is beyond simply recalling for the vocabularies or expression usages of targeted languages. As much the same as children naturally acquire their mother languages among active and social interactions with other surrounding people, foreign languages should be told in the circumstances and contexts for authentic applications of foreign languages. The MOO, one of the virtual realities with spatial metaphors on the text basis, has been gaining high attentions from educational fields, thanks to the strong functions of social contexts and learner interactions. This paper approaches the features of MOO as foreign language learning environments, in terms of activity, context and interaction.

• Journal of the Korean Society for information Management
• /
• v.24 no.2
• /
• pp.161-179
• /
• 2007

As increasingly diverse kinds of information materials are available on the Internet, it becomes a challenge to define an adequate level of metadata provision for each different type of material in the context of digital libraries. This study explores issues of metadata provision for a particular type of material, statistical tables. Statistical data always involves numbers and numeric values which should be interpreted with an understanding of underlying concepts and constructs. Because of the unique data characteristics, metadata in the statistical domain is essential not only for finding and discovering relevant data, but also for understanding and using the data found. However, in statistical metadata research, more emphasis has been put on the question of what metadata is necessary for processing the data and less on what metadata should be presented to users. In this study, a case study was conducted to gauge the status of metadata provision for statistical tables on the Internet. The websites of two federal statistical agencies in the United States were selected and a content analysis method was used for that purpose. The result showing insufficient and inconsistent provision of metadata demonstrate the need for more discussions on statistical metadata from the ordinary web users' perspective.

• Journal of Intelligence and Information Systems
• /
• v.21 no.2
• /
• pp.1-18
• /
• 2015

Collaborative filtering(CF) algorithm has been popularly used for recommender systems in both academic and practical applications. A general CF system compares users based on how similar they are, and creates recommendation results with the items favored by other people with similar tastes. Thus, it is very important for CF to measure the similarities between users because the recommendation quality depends on it. In most cases, users' explicit numeric ratings of items(i.e. quantitative information) have only been used to calculate the similarities between users in CF. However, several studies indicated that qualitative information such as user's reviews on the items may contribute to measure these similarities more accurately. Considering that a lot of people are likely to share their honest opinion on the items they purchased recently due to the advent of the Web 2.0, user's reviews can be regarded as the informative source for identifying user's preference with accuracy. Under this background, this study proposes a new hybrid recommender system that combines with users' review mining. Our proposed system is based on conventional memory-based CF, but it is designed to use both user's numeric ratings and his/her text reviews on the items when calculating similarities between users. In specific, our system creates not only user-item rating matrix, but also user-item review term matrix. Then, it calculates rating similarity and review similarity from each matrix, and calculates the final user-to-user similarity based on these two similarities(i.e. rating and review similarities). As the methods for calculating review similarity between users, we proposed two alternatives - one is to use the frequency of the commonly used terms, and the other one is to use the sum of the importance weights of the commonly used terms in users' review. In the case of the importance weights of terms, we proposed the use of average TF-IDF(Term Frequency - Inverse Document Frequency) weights. To validate the applicability of the proposed system, we applied it to the implementation of a recommender system for smartphone applications (hereafter, app). At present, over a million apps are offered in each app stores operated by Google and Apple. Due to this information overload, users have difficulty in selecting proper apps that they really want. Furthermore, app store operators like Google and Apple have cumulated huge amount of users' reviews on apps until now. Thus, we chose smartphone app stores as the application domain of our system. In order to collect the experimental data set, we built and operated a Web-based data collection system for about two weeks. As a result, we could obtain 1,246 valid responses(ratings and reviews) from 78 users. The experimental system was implemented using Microsoft Visual Basic for Applications(VBA) and SAS Text Miner. And, to avoid distortion due to human intervention, we did not adopt any refining works by human during the user's review mining process. To examine the effectiveness of the proposed system, we compared its performance to the performance of conventional CF system. The performances of recommender systems were evaluated by using average MAE(mean absolute error). The experimental results showed that our proposed system(MAE = 0.7867 ~ 0.7881) slightly outperformed a conventional CF system(MAE = 0.7939). Also, they showed that the calculation of review similarity between users based on the TF-IDF weights(MAE = 0.7867) leaded to better recommendation accuracy than the calculation based on the frequency of the commonly used terms in reviews(MAE = 0.7881). The results from paired samples t-test presented that our proposed system with review similarity calculation using the frequency of the commonly used terms outperformed conventional CF system with 10% statistical significance level. Our study sheds a light on the application of users' review information for facilitating electronic commerce by recommending proper items to users.

• Journal of the Korean Regional Science Association
• /
• v.40 no.1
• /
• pp.37-51
• /
• 2024

To respond to local extinction, South Korea, under the leadership of the Ministry of the Interior and Safety, identified depopulated areas in 2021 and launched the Local Extinction Response Fund in 2022. However, due to its early stage of implementation, analyzing the characteristics and changes of policy response to local extinction at the central government level remains a challenge. In contrast, Japan, facing similar issues of local extinction as South Korea, has established a robust central government-led response system based on the Regional Revitalization Act and the Comprehensive Strategy and Basic Policy on Community-Population-Job Creation. Hence, this study examines Japan's policy responses to local extinction by analyzing the first and second periods of the Comprehensive Strategy and Basic Policy on Community-Population-Job Creation. For the analysis, topic modeling was employed to enhance text analysis efficiency and accuracy, complemented by expert interviews for validation. The results revealed that the first-period strategy's topics encompassed economy and society, start-up, local government, living condition, service, and industry. Meanwhile, the second-period strategy's topics included resource, the New Normal, woman, digital transformation, industry, region, public-private partnership, and population. The analysis highlights that the policy target, policy direction, and environmental change significantly influenced these policy shifts.

• Journal of Intelligence and Information Systems
• /
• v.23 no.3
• /
• pp.69-94
• /
• 2017

Recently, increase of demand for big data analysis has been driving the vigorous development of related technologies and tools. In addition, development of IT and increased penetration rate of smart devices are producing a large amount of data. According to this phenomenon, data analysis technology is rapidly becoming popular. Also, attempts to acquire insights through data analysis have been continuously increasing. It means that the big data analysis will be more important in various industries for the foreseeable future. Big data analysis is generally performed by a small number of experts and delivered to each demander of analysis. However, increase of interest about big data analysis arouses activation of computer programming education and development of many programs for data analysis. Accordingly, the entry barriers of big data analysis are gradually lowering and data analysis technology being spread out. As the result, big data analysis is expected to be performed by demanders of analysis themselves. Along with this, interest about various unstructured data is continually increasing. Especially, a lot of attention is focused on using text data. Emergence of new platforms and techniques using the web bring about mass production of text data and active attempt to analyze text data. Furthermore, result of text analysis has been utilized in various fields. Text mining is a concept that embraces various theories and techniques for text analysis. Many text mining techniques are utilized in this field for various research purposes, topic modeling is one of the most widely used and studied. Topic modeling is a technique that extracts the major issues from a lot of documents, identifies the documents that correspond to each issue and provides identified documents as a cluster. It is evaluated as a very useful technique in that reflect the semantic elements of the document. Traditional topic modeling is based on the distribution of key terms across the entire document. Thus, it is essential to analyze the entire document at once to identify topic of each document. This condition causes a long time in analysis process when topic modeling is applied to a lot of documents. In addition, it has a scalability problem that is an exponential increase in the processing time with the increase of analysis objects. This problem is particularly noticeable when the documents are distributed across multiple systems or regions. To overcome these problems, divide and conquer approach can be applied to topic modeling. It means dividing a large number of documents into sub-units and deriving topics through repetition of topic modeling to each unit. This method can be used for topic modeling on a large number of documents with limited system resources, and can improve processing speed of topic modeling. It also can significantly reduce analysis time and cost through ability to analyze documents in each location or place without combining analysis object documents. However, despite many advantages, this method has two major problems. First, the relationship between local topics derived from each unit and global topics derived from entire document is unclear. It means that in each document, local topics can be identified, but global topics cannot be identified. Second, a method for measuring the accuracy of the proposed methodology should be established. That is to say, assuming that global topic is ideal answer, the difference in a local topic on a global topic needs to be measured. By those difficulties, the study in this method is not performed sufficiently, compare with other studies dealing with topic modeling. In this paper, we propose a topic modeling approach to solve the above two problems. First of all, we divide the entire document cluster(Global set) into sub-clusters(Local set), and generate the reduced entire document cluster(RGS, Reduced global set) that consist of delegated documents extracted from each local set. We try to solve the first problem by mapping RGS topics and local topics. Along with this, we verify the accuracy of the proposed methodology by detecting documents, whether to be discerned as the same topic at result of global and local set. Using 24,000 news articles, we conduct experiments to evaluate practical applicability of the proposed methodology. In addition, through additional experiment, we confirmed that the proposed methodology can provide similar results to the entire topic modeling. We also proposed a reasonable method for comparing the result of both methods.

• Journal of Intelligence and Information Systems
• /
• v.29 no.1
• /
• pp.249-263
• /
• 2023

Collaborative Filtering, a representative recommendation system methodology, consists of two approaches: neighbor methods and latent factor models. Among these, the latent factor model using matrix factorization decomposes the user-item interaction matrix into two lower-dimensional rectangular matrices, predicting the item's rating through the product of these matrices. Due to the factor vectors inferred from rating patterns capturing user and item characteristics, this method is superior in scalability, accuracy, and flexibility compared to neighbor-based methods. However, it has a fundamental drawback: the need to reflect the diversity of preferences of different individuals for items with no ratings. This limitation leads to repetitive and inaccurate recommendations. The Adaptive Deep Latent Factor Model (ADLFM) was developed to address this issue. This model adaptively learns the preferences for each item by using the item description, which provides a detailed summary and explanation of the item. ADLFM takes in item description as input, calculates latent vectors of the user and item, and presents a method that can reflect personal diversity using an attention score. However, due to the requirement of a dataset that includes item descriptions, the domain that can apply ADLFM is limited, resulting in generalization limitations. This study proposes a Generalized Adaptive Deep Latent Factor Recommendation Model, G-ADLFRM, to improve the limitations of ADLFM. Firstly, we use item ID, commonly used in recommendation systems, as input instead of the item description. Additionally, we apply improved deep learning model structures such as Self-Attention, Multi-head Attention, and Multi-Conv1D. We conducted experiments on various datasets with input and model structure changes. The results showed that when only the input was changed, MAE increased slightly compared to ADLFM due to accompanying information loss, resulting in decreased recommendation performance. However, the average learning speed per epoch significantly improved as the amount of information to be processed decreased. When both the input and the model structure were changed, the best-performing Multi-Conv1d structure showed similar performance to ADLFM, sufficiently counteracting the information loss caused by the input change. We conclude that G-ADLFRM is a new, lightweight, and generalizable model that maintains the performance of the existing ADLFM while enabling fast learning and inference.

• Journal of Food Hygiene and Safety
• /
• v.31 no.3
• /
• pp.153-166
• /
• 2016

The purpose of this study was to investigate consumers' perceptions on monosodium L-glutamate (MSG) in social media. Data were collected from Naver blogs and Naver web communities (Korean representative portal web-site), and media reports including comment sections on a Yonhap news website (Korean largest news agency). The results from Naver blogs and Naver web communities showed that it was primarily mentioned MSG-use restaurant reviews, 'MSG-no added' products, its safety, and methods of reducing MSG in food. When TV shows on current affairs, newspaper, or TV news reported uses and side effects of MSG, search volume for MSG has increased in both PC and mobile search engines. Search volume has increased especially when TV shows on current affairs reported it. There are more periods with increased search volume for Mobile than PC. Also, it was mainly commented about safety of MSG, criticism of low-quality foods, abuse of MSG, and distrust of government below the news on the Yonhap news site. The label of MSG-no added products in market emphasized "MSG-free" even though it is allocated as an acceptable daily intake (ADI) not-specified by the Joint FAO/WHO Expert Committee on Food Additives (JECFA). When consumers search for MSG (monosodium L-glutamate) or purchase food on market, they might perceive that 'MSG-no added' products are better. Competent authorities, offices of education and local government provide guidelines based on no added MSG principle and these policies might affect consumers' perceptions. TV program or news program could be a powerful and effective consumer communication channel about MSG through Mobile rather than PC. Therefore media including TV should report item on monosodium L-glutamate with responsibility and information based on scientific background for consumers to get reliable information.