• Annual Conference of KIPS
• /
• 2022.05a
• /
• pp.381-383
• /
• 2022

최근 자연어처리 분야에서는 Bert, GPT 등 Transformer기반의 언어모델 연구가 활발히 이뤄지고 있다. 이러한 언어모델은 대용량의 말뭉치 데이터와 많은 파라미터를 이용하여 사전학습을 진행하여 다양한 자연어처리 테스트에서 높은 성능을 보여주고 있다. 이에 본 논문에서는 Transformer기반의 언어모델인 Bert와 GPT-2의 성능평가를 진행한다. 성능평가는 '네이버 영화 리뷰' 데이터 셋을 통해 긍정 부정의 정확도와 학습시간을 측정한다. 측정결과 정확도에서는 GPT-2가 Bert보다 최소 4.16%에서 최대 5.32% 높은 정확도를 나타내었지만 학습시간에서는 Bert가 GPT-2보다 최소 104초에서 116초 빠르게 나타났다. 향후 성능 비교는 더 많은 데이터와 다양한 조건을 통해 구체적인 성능 비교가 필요하다.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.17 no.4
• /
• pp.275-287
• /
• 2024

Research on the integration of Knowledge Graphs (KGs) and Language Models (LMs) has been consistently explored over the years. However, studies focusing on the automatic generation of text using the structured knowledge from KGs have not been as widely developed. In this study, we propose a methodology for automatically generating specific domain-related research items (Related Work) at a level comparable to existing papers. This methodology involves: 1) selecting optimal prompts, 2) extracting triples through a four-step refinement process, 3) constructing a knowledge graph, and 4) automatically generating related research. The proposed approach utilizes GPT-4, one of the large language models (LLMs), and is desigend to automatically generate related research by applying the four-step refinement process. The model demonstrated performance metrics of 17.3, 14.1, and 4.2 in Triple extraction across #Supp, #Cont, and Fluency, respectively. According to the GPT-4 automatic evaluation criteria, the model's performamce improved from 88.5 points vefore refinement to 96.5 points agter refinement out of 100, indicating a significant capability to automatically generate related research at a level similar to that of existing papers.

• Journal of the Korea Society of Computer and Information
• /
• v.29 no.2
• /
• pp.43-50
• /
• 2024

This paper evaluates the Korean sentiment analysis performance of large language models like GPT-3.5 and GPT-4 using a zero-shot approach facilitated by the ChatGPT API, comparing them to pre-trained Korean models such as KoBERT. Through experiments utilizing various Korean sentiment analysis datasets in fields like movies, gaming, and shopping, the efficiency of these models is validated. The results reveal that the LMKor-ELECTRA model displayed the highest performance based on F1-score, while GPT-4 particularly achieved high accuracy and F1-scores in movie and shopping datasets. This indicates that large language models can perform effectively in Korean sentiment analysis without prior training on specific datasets, suggesting their potential in zero-shot learning. However, relatively lower performance in some datasets highlights the limitations of the zero-shot based methodology. This study explores the feasibility of using large language models for Korean sentiment analysis, providing significant implications for future research in this area.

• International Journal of Internet, Broadcasting and Communication
• /
• v.16 no.1
• /
• pp.182-190
• /
• 2024

Various artificial intelligence neural network models that have emerged since 2014 enable the creation of new content beyond the existing level of information discrimination and withdrawal, and the recent generative artificial intelligences such as ChatGPT and Gall-E2 create and present new information similar to actual data, enabling natural interaction because they create and provide verbal expressions similar to humans, unlike existing chatbots that simply present input content or search results. This study aims to present a model that can improve the ChatGPT communication skills of university students through curriculum research on ChatGPT, which can be participated by students from all departments, including engineering, humanities, society, health, welfare, art, tourism, management, and liberal arts. It is intended to design a way to strengthen competitiveness to embody the practical ability to solve problems through ethical attitudes, AI-related technologies, data management, and composition processes as knowledge necessary to perform tasks in the artificial intelligence era, away from simple use capabilities. It is believed that through creative education methods, it is possible to improve university awareness in companies and to seek industry-academia self-reliant courses.

• Journal of Korea Society of Digital Industry and Information Management
• /
• v.19 no.2
• /
• pp.135-145
• /
• 2023

ChatGPT is an artificial intelligence-based conversation agent developed by OpenAI using natural language processing technology. In this study, an empirical study was conducted on incumbent teachers on the intention to use the newly emerged Chat GPT. First, we studied how accuracy, entertainment, system accessibility, perceived usefulness, and perceived ease of use affect ChatGPT's acceptance intention. In addition, we analyzed whether perceived usefulness and perceived ease of use differ in the intention to accept depending on the digital technology of teachers. As a result of the study, the suitability of the structural equation model was generally good. Accuracy and entertainment were found to have a significant effect on perceived usefulness, and system accessibility was found to have a significant effect on perceived ease of use. In the analysis of teachers' digital technology control effects, it was found that perceived usefulness and perceived ease of use had a control effect between acceptance intentions. It was found that the group with high digital skills of teachers was strongly intended to accept the service regardless of perceived usefulness and ease of use. In the group with low digital skills of teachers, it is thought that ChatGPT's service shows the acceptance intention only when the perceived usefulness and ease of use are high. Therefore, in the group with low digital technology, it is necessary to seek teaching activities such as the development of instructional models using ChatGPT.

• Communications of Mathematical Education
• /
• v.37 no.2
• /
• pp.233-256
• /
• 2023

This study conducted foundational research to derive ways to use ChatGPT in mathematics education by analyzing ChatGPT's responses to questions from the National Assessment of Educational Achievement (NAEA) and the College Scholastic Ability Test (CSAT). ChatGPT, a generative artificial intelligence model, has gained attention in various fields, and there is a growing demand for its use in education as the number of users rapidly increases. To the best of our knowledge, there are very few reported cases of educational studies utilizing ChatGPT. In this study, we analyzed ChatGPT 3.5 responses to questions from the three-year National Assessment of Educational Achievement and the College Scholastic Ability Test, categorizing them based on the percentage of correct answers, the accuracy of the solution process, and types of errors. The correct answer rates for ChatGPT in the National Assessment of Educational Achievement and the College Scholastic Ability Test questions were 37.1% and 15.97%, respectively. The accuracy of ChatGPT's solution process was calculated as 3.44 for the National Assessment of Educational Achievement and 2.49 for the College Scholastic Ability Test. Errors in solving math problems with ChatGPT were classified into procedural and functional errors. Procedural errors referred to mistakes in connecting expressions to the next step or in calculations, while functional errors were related to how ChatGPT recognized, judged, and outputted text. This analysis suggests that relying solely on the percentage of correct answers should not be the criterion for assessing ChatGPT's mathematical performance, but rather a combination of the accuracy of the solution process and types of errors should be considered.

• 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.

• International journal of advanced smart convergence
• /
• v.13 no.2
• /
• pp.205-213
• /
• 2024

This study uses a large language model (LLM) to identify Aristotle's rhetorical principles (ethos, pathos, and logos) in COVID-19 information on Naver Knowledge-iN, South Korea's leading question-and-answer community. The research analyzed the differences of these rhetorical elements in the most upvoted answers with random answers. A total of 193 answer pairs were randomly selected, with 135 pairs for training and 58 for testing. These answers were then coded in line with the rhetorical principles to refine GPT 3.5-based models. The models achieved F1 scores of .88 (ethos), .81 (pathos), and .69 (logos). Subsequent analysis of 128 new answer pairs revealed that logos, particularly factual information and logical reasoning, was more frequently used in the most upvoted answers than the random answers, whereas there were no differences in ethos and pathos between the answer groups. The results suggest that health information consumers value information including logos while ethos and pathos were not associated with consumers' preference for health information. By utilizing an LLM for the analysis of persuasive content, which has been typically conducted manually with much labor and time, this study not only demonstrates the feasibility of using an LLM for latent content but also contributes to expanding the horizon in the field of AI text extraction.

• The Mathematical Education
• /
• v.63 no.2
• /
• pp.123-138
• /
• 2024

This study described the utilization of ChatGPT in teaching and students' learning processes for the course "Introductory Mathematics for Artificial Intelligence (Math4AI)" at 'S' University. We developed a customized ChatGPT and presented a learning model in which students supplement their knowledge of the topic at hand by utilizing this model. More specifically, first, students learn the concepts and questions of the course textbook by themselves. Then, for any question they are unsure of, students may submit any questions (keywords or open problem numbers from the textbook) to our own ChatGPT at https://math4ai.solgitmath.com/ to get help. Notably, we optimized ChatGPT and minimized inaccurate information by fully utilizing various types of data related to the subject, such as textbooks, labs, discussion records, and codes at http://matrix.skku.ac.kr/Math4AI-ChatGPT/. In this model, when students have questions while studying the textbook by themselves, they can ask mathematical concepts, keywords, theorems, examples, and problems in natural language through the ChatGPT interface. Our customized ChatGPT then provides the relevant terms, concepts, and sample answers based on previous students' discussions and/or samples of Python or R code that have been used in the discussion. Furthermore, by providing students with real-time, optimized advice based on their level, we can provide personalized education not only for the Math4AI course, but also for any other courses in college math education. The present study, which incorporates our ChatGPT model into the teaching and learning process in the course, shows promising applicability of AI technology to other college math courses (for instance, calculus, linear algebra, discrete mathematics, engineering mathematics, and basic statistics) and in K-12 math education as well as the Lifespan Learning and Continuing Education.

• Transactions on Semiconductor Engineering
• /
• v.1 no.1
• /
• pp.23-30
• /
• 2023

With the recent development of Generative AI technology by IT giants, the size of the transformer model is increasing exponentially over trillion won. In order to continuously enable these AI services, it is essential to reduce the weight of the model. In this paper, we find a hardware-friendly structured pruning pattern and propose a lightweight method of the transformer model. Since compression proceeds by utilizing the characteristics of the model algorithm, the size of the model can be reduced and performance can be maintained as much as possible. Experiments show that the structured pruning proposed when pruning GPT-2 and BERT language models shows almost similar performance to fine-grained pruning even in highly sparse regions. This approach reduces model parameters by 80% and allows hardware acceleration in structured form with 0.003% accuracy loss compared to fine-tuned pruning.