• Journal of KIISE:Software and Applications
• /
• v.33 no.12
• /
• pp.1090-1102
• /
• 2006

This paper suggests a new algorithm for detecting the plagiarism among a set of source codes, constrained to be functionally equivalent, such are submitted for a programming assignment or for a programming contest problem. The typical algorithms largely exploited up to now are based on Greedy-String Tiling, which seeks for a perfect match of substrings, and analysis of similarity between strings based on the local alignment of the two strings. This paper introduces a new method for detecting the similar interval of the given programs based on an adaptive similarity matrix, each entry of which is the logarithm of the probabilities of the keywords based on the frequencies of them in the given set of programs. We experimented this method using a set of programs submitted for more than 10 real programming contests. According to the experimental results, we can find several advantages of this method compared to the previous one which uses fixed similarity matrix(+1 for match, -1 for mismatch, -2 for gap) and also can find that the adaptive similarity matrix can be used for detecting various plagiarism cases.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.14 no.2
• /
• pp.142-149
• /
• 2004

It is important to retrieve exact information coinciding with user's need from lots of Usenet news and filter desired information quickly. Differently from email system, we must previously register our interesting news group if we want to get the news information. However, it is not easy for a novice to decide which news group is relevant to his or her interests. In this work, we present a service classifying user preferred news groups among various news groups by the use of Kohonen network. We first extract candidate terms from example documents and then choose a number of representative keywords to be used in Kohonen network from them through fuzzy inference. From the observation of training patterns, we could find the sparsity problem that lots of keywords in training patterns are empty. Thus, a new method to train neural network through reduction of unnecessary dimensions by the statistical coefficient of determination is proposed in this paper. Experimental results show that the proposed method is superior to the method using every dimension in terms of cluster overlap defined by using within cluster distance and between cluster distance.

• 한국HCI학회:학술대회논문집
• /
• 2008.02b
• /
• pp.298-303
• /
• 2008

Watching a program on IPTV among the numerous choices from the internet requires a burden of searching and browsing for a favorite one. This paper introduces a new concept called Mosaic Map and presents how it provides preview information of image map links to other programs. In Mosaic Map the pixels in the still image are used both as shading the background and as thumbnails which can link up with other programs. This kind of contextualized preview of choices can help IPTV users to associate the image with related programs without making visual saccades between watching IPTV and browsing many choices. The experiments showed that the Mosaic Map reduces the time to complete search and browsing, comparing to the legacy menu and web search.

• Journal of the Korean Society for information Management
• /
• v.19 no.1
• /
• pp.5-22
• /
• 2002

Users usually search for the required information via search engines which contain locations of the information. However. as the amount of data gets large, the result of the search is often not the information that users actually want. In this paper a web guide is proposed in order to resolve this problem. The web guide uses case-based learning method which stores and utilizes cases based on the keywords of user's action and agent's visit. The proposed agent system learns the user's visiting actions following the input of the data to be searched, and then helps rapid searches of the data wanted.

• Journal of the Korea Society of Computer and Information
• /
• v.15 no.4
• /
• pp.1-9
• /
• 2010

In the new generation of Web, commonly called "Web 2.0", blogging has facilitated the publishing information or his/her opinion on the web. Various blog retrieval algorithms have been proposed to search for blogs more effectively. However, actually keyword-based searching or link-analysis blog ranking system cannot satisfy the user's requirement. In this paper, we suggest a topic-based weighting blog retrieval system in which the links between blog writings and searching words are considered to improve the search results. Our system extracts topics from each blog and weights them much higher than other guide words. In the comparison with other systems, we see that the proposed topic-base system has better recall rate of search results.

• Annual Conference on Human and Language Technology
• /
• 2020.10a
• /
• pp.184-187
• /
• 2020

다중 지문 기계독해는 질문과 여러 개의 지문을 입력받고 입력된 지문들에서 추출된 정답 중에 하나의 정답을 출력하는 문제이다. 다중 지문 기계독해에서는 정답이 있을 단락을 선택하는 순위화 방법에 따라서 성능이 크게 달라질 수 있다. 본 논문에서는 단락 안에 정답이 있을 확률을 예측하는 단락 재순위화 모델과 선택된 단락에서 서술형 정답을 위한 세부적인 정답의 경계를 예측하는 세부 단락 선별 기법을 제안한다. 단락 순위화 모델 학습의 경우 모델 학습을 위해 각 단락의 출력에 softmax와 cross-entroy를 이용한 손실 값과 sigmoid와 평균 제곱 오차의 손실 값을 함께 학습하고 키워드 매칭을 함께 적용했을 때 KorQuAD 2.0의 개발셋에서 상위 1개 단락, 3개 단락, 5개 단락에서 각각 82.3%, 94.5%, 97.0%의 재현율을 보였다. 세부 단락 선별 모델의 경우 입력된 두 단락을 비교하는 duoBERT를 이용했을 때 KorQuAD 2.0의 개발셋에서 F1 83.0%의 성능을 보였다.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2023.07a
• /
• pp.3-4
• /
• 2023

자연어처리는 수많은 정보가 수집된 전자의무기록의 비정형 데이터에서 유의미한 정보나 패턴 등을 추출해 의료진의 의사결정을 지원하고, 환자에게 더 나은 진단이나 치료 등을 지원할 수 있어 큰 잠재력을 가지고 있다. 그러나 전자의무기록은 개인정보와 같은 민감한 정보가 다수 포함되어 있어 접근하기 어렵고, 이로 인해 충분한 양의 데이터를 확보하기 어렵다. 따라서 본 논문에서는 신뢰할 수 있는 의료 합성 텍스트를 생성하기 위해 ChatGPT 기반 의료 텍스트 증강 도구를 개발하였다. 이는 사용자가 입력한 실제 의료 텍스트로 의료 합성 데이터를 생성한다. 이를 위해, 적합한 프롬프트와 의료 텍스트에 대한 전처리 방법을 탐색하였다. ChatGPT 기반 의료 텍스트 증강 도구는 입력 텍스트의 핵심 키워드를 잘 유지하였고, 사실에 기반한 의료 합성 텍스트를 생성할 수 있다는 것을 확인할 수 있었다.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.11a
• /
• pp.514-517
• /
• 2023

본 연구는 인 컨택스트 러닝 (In-Context Learning)을 오디오-언어 작업에 적용하기 위한 멀티모달 (Multi-Modal) 딥러닝 모델을 다룬다. 해당 모델을 통해 학습 단계에서 오디오와 텍스트의 소통 가능한 형태의 표현 (Representation)을 학습하고 여러가지 오디오-텍스트 작업을 수행할 수 있는 멀티모달 딥러닝 모델을 개발하는 것이 본 연구의 목적이다. 모델은 오디오 인코더와 언어 인코더가 연결된 구조를 가지고 있으며, 언어 모델은 6.7B, 30B 의 파라미터 수를 가진 자동회귀 (Autoregressive) 대형 언어 모델 (Large Language Model)을 사용한다 오디오 인코더는 자기지도학습 (Self-Supervised Learning)을 기반으로 사전학습 된 오디오 특징 추출 모델이다. 언어모델이 상대적으로 대용량이기 언어모델의 파라미터를 고정하고 오디오 인코더의 파라미터만 업데이트하는 프로즌 (Frozen) 방법으로 학습한다. 학습을 위한 과제는 음성인식 (Automatic Speech Recognition)과 요약 (Abstractive Summarization) 이다. 학습을 마친 후 질의응답 (Question Answering) 작업으로 테스트를 진행했다. 그 결과, 정답 문장을 생성하기 위해서는 추가적인 학습이 필요한 것으로 보였으나, 음성인식으로 사전학습 한 모델의 경우 정답과 유사한 키워드를 사용하는 문법적으로 올바른 문장을 생성함을 확인했다.

• Asia-Pacific Journal of Business Venturing and Entrepreneurship
• /
• v.17 no.1
• /
• pp.191-211
• /
• 2022

As the value of highly sustainable companies increases, ESG(Environmental, Social, and Governance) has emerged as the biggest topic of discussion for companies around the world. In addition, as domestically, more research is being done on ESG in line with global trends, it is necessary to examine ESG research trends. Accordingly, ESG academic papers that have been published for the past 10 years were collected for each year, and frequency analysis was conducted using text mining techniques regarding key themes and thesis titles. This paper analyzed the number of selected publications by year and the cumulated number of studies through bibliometric analysis. The findings suggested that the number of ESG papers is increasing each year and that academic interest in ESG-related issues continues to abound. Next, according to the results of frequency analysis of the keywords and titles of the research papers, the words- "ESG", "company", "society", "responsibility", "management", "investment", and "sustainability"- were extracted. This analysis identified the research fields and keywords that have been relevant to ESG in the past 10 years. As a result of comparing the major ESG issues presented in recent overseas studies and the common factors of the ESG key keywords presented in this study, it was confirmed that the environment is the focus of recent studies compared to previous studies. Third, it was found that the data used by domestic ESG studies mainly include the KEJI index, the KRX index, and the KCGS ESG evaluation index. After identifying the main research subjects of ESG papers, research found that 8 out of 152 domestic ESG studies were focused on SMEs. Through this study, it was possible to confirm the ESG research trend and increase in research, and future researchers divided the research topics and research keywords and presented basic data for selecting more diverse research topics. Based on both, the arguments of previous ESG studies conducted on SMEs and the results of this study, there is a lack of studies on guidelines for ESG practice and their application to SMEs, and more ESG research regarding SMEs will need to be conducted in the future.

• Journal of Advanced Navigation Technology
• /
• v.15 no.2
• /
• pp.287-297
• /
• 2011

As electronic mails are being widely used for facility and speedness of information communication, as the amount of spam mails which have malice and advertisement increase and cause lots of social and economic problem. A number of approaches have been proposed to alleviate the impact of spam. These approaches can be categorized into pre-acceptance and post-acceptance methods. Post-acceptance methods include bayesian filters, collaborative filtering and e-mail prioritization which are based on words or sentances. But, spammers are changing those characteristics and sending to avoid filtering system. In the case of Korean, the abnormal usages can be much more than other languages because syllable is composed of chosung, jungsung, and jongsung. Existing formal expressions and learning algorithms have the limits to meet with those changes promptly and efficiently. So, we present an methods for recognizing Korean abnormal language(Koral) to improve accuracy and efficiency of filtering system. The method is based on syllabic than word and Smith-waterman algorithm. Through the experiment on filter keyword and e-mail extracted from mail server, we confirmed that Koral is recognized exactly according to similarity level. The required time and space costs are within the permitted limit.