Earticle

Many methods of discovering social networking communities or clustering of features are based on the network structure or the content network. This paper proposes a community discovery method based on topic models using a time factor and an unsupervised clustering method. Online community discovery enables organizations and businesses to thoroughly understand the trend in users' interests in their products and services. In addition, an insight into customer experience on social networks is a tremendous competitive advantage in this era of ecommerce and Internet development. The objective of this work is to find clusters (communities) such that each cluster's nodes contain topics and individuals having similarities in the attribute space. In terms of social media analytics, the method seeks communities whose members have similar features. The method is experimented with and evaluated using a Vietnamese corpus of comments and messages collected on social networks and ecommerce sites in various sectors from 2016 to 2019. The experimental results demonstrate the effectiveness of the proposed method over other methods.

The exponential increase in publications and the interconnected nature of sub-domains make traditional methods of information extraction and organization inadequate. This inefficiency can impede scientific progress and innovation. To address these challenges, this research leverages the ability of Bidirectional Encoder Representations from Transformers for keyword extraction (KeyBERT) and integrates with K-Means clustering to organize topics from large datasets effectively. Analyzing a dataset of 47,627 articles from SCOPUS in the domains of Reinforcement Learning and Computer Vision. An ablation study demonstrates the generalizability of the approach across these fields, with the optimal number of clusters determined to be three using the Elbow Method. The results demonstrate that KeyBERT is effective in extracting and organizing topics within these domains, with a particular focus on applications such as medical imaging, autonomous driving, and real-time detection systems. This methodology offers a scalable solution for organizing vast academic datasets, enabling researchers to extract meaningful insights efficiently and apply this approach to other domains.

To deal with the issues like existing common data sparseness in weibo social network and the phenomena of cold start, this paper puts forward a two-stage clustering based on the recommendation algorithm GCCR. The algorithm firstly selects users’ focused nodes which have higher number, so as to extract a dense subset of sparse data, and by using the method of graph paper, similar concerned interested core clustering is formed to this dense subset. Then, it is extracted that weibo content features of seed clustering and the whole data set other users. Then the entire user group is clustered based on content similarity. Finally the clustering results are used in subject recommendation. Through clustering the two phases of dense data subset and the whole data set, the clustering effect of extreme sparse data sets are improved. At the same time, because of fuzziness of graph clustering, this thesis retains a certain diversity in the process of user interest clustering, so as to avoid convergence too fast when cold start. This method is verified through the real social network data, and the experimental results show that this algorithm can effectively solve the problems such as data sparseness and cold start phenomenon.

Time is an important dimension of information space. It plays important roles in Web search, because most Web pages contain time information and many Web queries are time-related. Therefore, exploiting temporal information in Web pages has been a hotspot in the research on Web search. In this paper, we focus on the time-enhanced topic clustering issue for news search results. Traditional clustering algorithms are usually based on the common phrases of Web pages, and they have little consideration about using the temporal information of Web pages. From this perspective, we propose a time-enhanced topic clustering algorithm for news Web pages. It improves traditional algorithms which only consider textual clustering, and applies a temporal clustering procedure on the topics returned by a textual clustering algorithm, which is to arrange every Web page in a cluster along a timeline based on the update time in Web pages. We conduct experiments on a real dataset crawled from Google News, and compare our algorithm with other competitors including K-Means, STC, TFIC, and Minhash Clustering in terms of different metrics such as precision and recall. The experimental results show that the proposed algorithm has better performance under both offline and online clustering test.

In recent years, topic detection has become a hot research point of the social network, which can be very good to find the key factors from the massive information and thus discover the topics. The traditional label propagation-based topic discovery algorithm (LPA) is widely concerned because of its approximate linear time complexity and there is no need to define the target function. However, LPA algorithm has the uncertainty and the randomness, which affects the accuracy and the stability of the topic discovery. In this paper, a method for clustering label words based on mutual information analysis is presented to find the current topic. Firstly, through filtering the stop words and extracting keywords with TF-IDF, topic words are been extracted out, and then a common word matrix is built, a topic discovery algorithm based on mutual information and label clustering is put forward. Finally, extensive experiments on two real datasets validate the effectiveness of the proposed MI-LC (Mutual information-Label clustering) algorithm against other well-established methods LPA and LDA in terms of running time, NMI value and perplexity value.

기존의 웹서비스가 정적이고 수동적인데 반해 최근의 웹 서비스는 점차 동적이고 능동적으로 변화하고 있다. 이러한 웹서비스 변화의 흐름을 잘 반영하는 것이 웹 2.0이다. 웹 2.0에서 대부분의 정보는 사용자에 의해 생산되고, 사용자가 붙인 태그(tag)에 의해 분류되어진다. 그러나 현재 태그에 관한 서비스 및 연구들은 태깅(tagging) 방법에 대한 연구를 비롯해 이를 표현하기 위한 tag cloud에 초점이 맞춰져 진행됨에 따라, 다양한 태그 정보자원 간의 체계와 연결 관계인 지식체계를 제공하지 못하고 있다. 이에 본 논문에서는 체계화된 지식표현을 위해 웹상에 편재되어 있는 학습 관련 리소스(resources) 및 태그들를 수집한다. 이를 사용자가 요청한 검색 키워드와 연관성이 있는 태그 정보들을 맵핑 및 클러스터링하여 최적화된 표현 형식인 토픽 맵(topic map)화하기 위한 시스템을 제안하며, 이 중 토픽 맵 생성을 위한 초기 연구 단계로서, 연관 태그들 간의 맵핑 및 클러스터링을 위한 알고리즘 제시를 중심으로 소개한다.

Fast and high-quality document clustering algorithms play an important role in providing data exploration by organizing large amounts of information into a small number of meaningful clusters. Hierarchical clustering improves the performance of retrieval and makes that users can understand easily. For outperforming of clustering, we implemented hierarchical structure with variety and readability, by careful selection of cluster topic words and deciding the number of clusters dynamically. It is important to select topic words because hierarchical clustering structure is summarizes result of searching. We made choice of noun word as a cluster topic word. The quality of topic words is increased $33\%$ as follows. As the topic word of each cluster, the only noun word is extracted for the top-level cluster and the used topic words for the children clusters were not reused.

본 연구는 행정부가 제출하는 법안은 어떠한 특징을 보이고 있으며, 어떠한 내용을 담고 있는지 알아보는 것을 목적으로 하였다. 행정부가 국회에 제출하는 법안은 행정부의 특징에 따라 다를 것이라 예상해 볼 수 있다. 이 를 위해 노무현정부부터 문재인정부까지 각 정부가 국회에 제출한 법안들을 수집하여 분석하였다. 법안에 대 한 분석은 행정부가 제출한 법안명을 대상으로 자연어 처리, 군집 분석, 토픽 모델링을 활용하여 법안의 구체 적 내용의 주요 주제와 세부 주제가 무엇인지 알아보았다. 네 개 정부가 제출한 전체 법안에 대한 분석을 통해 조세･재정 및 공공행정 운영, 사회복지 및 규제정책, 산업･경제 진흥 및 환경정책으로 나뉘고 있음을 알 수 있 었다. 노무현정부와 이명박정부는 모두 산업･경제 진흥 및 재정 개혁을 중점적으로 추진하였다. 그러나 세부 정책을 보면 차이가 있음을 알 수 있다. 반면, 박근혜정부와 문재인정부는 사회복지 및 공공안전 정책을 가장 중요한 우선순위로 설정하였다는 공통점이 있다. 그러나 구체적인 세부 정책에서는 차이가 있음을 발견할 수 있다. 본 연구는 정부 제출 법안을 통해 각 정부의 정책적 우선순위가 단순한 이념적 구분을 넘어 경제･사회적 으로 요구되는 정책 공급을 위한 노력이었다.

This study examines the characteristics and content of bills submitted by the executive branch to the National Assembly. The nature of these bills is expected to vary depending on the characteristics of the administration. To investigate this, the study collected and analyzed bills submitted by the administrations of Roh Moo-hyun, Lee Myung-bak, Park Geun-hye, and Moon Jae-in. Using natural language processing (NLP), cluster analysis, and topic modeling, the study identifies the key themes and subtopics within these bills. The findings reveal that the bills submitted by these four administrations primarily fall into three broad categories: (1) Taxation, Fiscal Policy, and Public Administration, (2) Social Welfare and Regulatory Policies, and (3) Industrial and Economic Promotion and Environmental Policies. While both the Roh and Lee administrations prioritized economic development and fiscal reform, their approaches differed in specific policy areas. Conversely, the Park and Moon administrations shared a common emphasis on social welfare and public safety policies, yet significant differences were observed in their specific policy priorities. By analyzing government-submitted bills, this study demonstrates that each administration’s policy priorities were not solely determined by ideological divisions but also reflected broader economic and social demands for policy implementation.

최근 소셜 네트워크 서비스의 사용량이 급증함에 따라, 이를 효과적으로 분석하고 이용하려는 연구가 활발히 진행 중이다. 소셜 미디어 데이터는 짧은 시간에 많은 양이 생산되기 때문에 효과적인 분석을 위해서는 이를 적절히 클러스터링하고, 요약(레이블링)해 줄 수 있는 기술이 필요하다. 그러나 기존의 문서 클러스터링 모델들은 여러 문장으로 구성된 일반적인 문서를 대상으로 하기 때문에 한 두 문장으로 구성된 짧은 소셜 미디어 데이터의 특징을 제대로 반영하지 못한다. 본 연구에서는 주제에 따라 긍정, 부정으로 분류된 소셜 미디어 데이터를 세부 주제별로 클러스터링하고 이를 요약하는 새로운 클러스터링 모델을 제안한다. 제안된 모델은 FTC(Frequent Term-based Clustering)와 GAAC(Group Average Agglomerative Clustering) 를 2단계로 적용한 것으로 정보가 부족한 데이터에 적합하며, 여러 주제에 대하여 중복으로 클러스터링할 수 있기 때문에 소셜 미디어 데이터 분석에 보다 효과적이다.

As the use of Social Network Service (SNS) is sharply increasing, many researches have been done to analyze SNS data. Because much social media data is produced in a very short time, we need effective techniques for clustering and summarizing (or labeling) of the data. However, existing document clustering models are developed for general documents which usually consist of many sentences, they do not adequately reflect characteristics of social media data of which document is mostly composed of one or two sentences. In this paper, we propose a new clustering model which deals with pre-processed twitter data that is boardly classified into positive and negative opinion documents for board topics. The model clusters the pre-processed twitter data into sub topics and labels with short key words, using the proposed two-level clustering model based on FTC (Frequent Term-based Clustering) and GAAC (Group Average Agglomerative Clustering). The experiment showed that it is appropriate for sparse data and more effective for social media data analysis than other models.

4차산업혁명으로 인해 이미지 콘텐츠의 양이 급속히 증가함에 따라 다양한 이미지를 저장 및 관리하기 위한 필요성이 커지고 있다. 이를 위해서 이미지 내용 및 주제의 파악이 선행되어야 하며, 이미지의 세부 객체를 판별하기 위한 방법론들이 제안되었다. 이와 같은 방법론들은 이미지 내의 세부 객체를 주석으로 제공하게 되는데, 이를 통해 전체 이미지의 주제를 설명하기에는 한계점이 있다. 본 논문에서는 이미지의 주석을 이용하여 주제를 파악하기 위해 잠재토픽모형을 이용하도록 한다. 잠재토픽모형으로는 Hofmann(2001)이 제안한 PLSA 모형을 이용하며, 이를 통해 각 이미지 및 주석을 잠재집단에 할당하도록 한다. 각 잠재집단에서 추정된 주석의 확률을 이용하여 잠재집단을 규명하고, 각 이미지 또는 주석이 잠재집단에 속할 사후확률을 추정하도록 한다. 본 논문에서는 Visual Genome에서 제공하는 300개의 이미지 및 주석을 분석하였으며, 다양한 잠재집단의 수를 설정하고 PLSA 모형을 적합한 후, 적합도 및 동질성 척도 등을 이용하여 모형을 평가하였다. 마지막으로, 수작업으로 분류한 결과와 모형을 통해 분류된 결과를 비교하여 오분류율을 구하고 적합도 및 동질성 척도의 결과와 비교하였다.

As the amount of image content rapidly increases due to the 4th industrial revolution, the need for storing and managing images is increasing. To figure out the subject of images for efficient storing and managing, various methods to provide the annotations describing the detailed object in the images has been proposed. However, it is hard to figure out the subject of images by using the annotations. In this paper, to understand the topic of the image PLSA model which is the latent topic model suggested by Hofmann(2001) is used. PLSA model assigns the annotations of each image to a latent group which can be identified by the estimated probabilities of annotation. In addition, by using PLSA model, the posterior probabilities of a latent group given by words or document are also provided. In this paper, 300 images and their annotations provided by Visual Genome are analyzed. To find out the appropriate the number of latent group, various models with different number of latent group are compared by using the goodness of fit such as AIC and BIC, coherence measures such as UCI and UMass coherence, and mis-specification rate.

정보의 양이 많아지면서 정보 검색 시스템에 검색 결과를 자동으로 구조화하는 계층적 클러스터링을 적용하는 시도가 늘고 있다. 계층적 클러스터링은 문서 간의 유사도를 통해 클러스터를 계층 구조로 만들어 검색 성능을 높이고 결과를 사용자에게 이해하기 쉽게 보여준다. 계층 구조는 검색 결과를 요약하는 것이기 때문에 클러스터의 내용을 효과적으로 함축할 수 있는 대표어의 선정이 중요하다. 각 클러스터의 대표어를 선정하기 위해 대표어에 명사인 단어만 추출하고 상위 클러스터 대표어에 사용된 단어는 하위 클러스터에 사용하지 않는 방법을 적용하여 대표어의 질을 높였다.

정보의 양이 많아지면서 정보 검색 시스템에 검색 결과를 자동으로 구조화하는 계층적 클러스터링을 적용하는 시도가 늘고 있다. 계층적 클러스터링은 문서 간의 유사도를 통해 클러스터를 계층 구조로 만들어 검색 성능을 높이고 결과를 사용자에게 이해하기 쉽게 보여준다. 계층 구조는 검색 결과를 요약하는 것이기 때문에 클러스터의 내용을 효과적으로 함축할 수 있는 대표어의 선정이 중요하다. 각 클러스터의 대표어를 선정하기 위해 대표어에 명사인 단어만 추출하고 상위 클러스터 대표어에 사용된 단어는 하위 클러스터에 사용하지 않는 방법을 적용하여 대표어의 질을 높였다.

The expansion of the range of ruling disclosure has significantly increased the number of accessible cases. As a result, the potential of using quantitative research methods in legal research has expanded, but quantitative methods are yet to be employed frequently in legal research. As quantitative research can reveal what previous studies could not detect, attempts at quantitative research in law are needed. This study collected Supreme Court cases related to rape crimes from the 1960s to 2022 to analyze the Supreme Court's position on the “degree of assault or threat,” an element of rape crimes. Using LDA topic modeling and analysis of adjective frequency for the collected cases, the increase or decrease in the topic proportion of precedents by decade and changes in the Supreme Court’s description of the behavioral aspect of rape were studied. Through LDA, this study focused on the diachronic change of frequently appearing topics. It has been revealed that the change in the proportion of precedents’ topics is deeply related to the laws enacted or revised at that time (such as the adoption of electronic tracking devices, abolishment of Antragsdelikt, or revision of the Act of Juvenile Sex Protection). In addition, the fact that the proportion of cases disputing for ‘the establishment of rape crimes’ rebounded from the 2010s and soared in the 2020s suggests that there is an implicit change in the attitude of precedents worthy to contend with. The analysis of the adjective frequency has shown how the attitude toward ‘the degree of assault or threat’ of rulings has changed. In the past, ‘difficult(‘곤란하게’ in Korean)’ was the most important adjective in rape-related precedents, but today, the frequency has decreased. On the contrary, ‘against with(‘반하는’ or ‘반하여’ in Korean)’ (which is used to examine the victim's intention) was not often used in rulings in the past, but has become the most frequently used adjective in recent rulings. This study concludes that the Supreme Court's key point of judgment gradually changed from how difficult the victim resisted to what was the victim's intention when considering the ‘degree of assault or threat.’ The adoption of quantitative methods in legal research enables fast and efficient access to a large amount of text data and has the advantage of detecting minute changes that are difficult to detect only by qualitative research methods. Thus, I expect various attempts on quantitative approaches to jurisprudence to be conducted after this study.