• The Journal of Information Systems
• /
• v.28 no.4
• /
• pp.271-291
• /
• 2019

Purpose The purpose of this study is to propose a methodology to derive positive keywords and negative keywords through deep learning to classify reviews into positive reviews and negative ones, and then refine the results of topic modeling using these keywords. Design/methodology/approach In this study, we extracted topic keywords by performing LDA-based topic modeling. At the same time, we performed attention-based deep learning to identify positive and negative keywords. Finally, we refined the topic keywords using these keywords as filters. Findings We collected and analyzed about 6,000 English reviews of Gyeongbokgung, a representative tourist attraction in Korea, from Tripadvisor, a representative travel site. Experimental results show that the proposed methodology properly identifies positive and negative keywords describing major topics.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.15 no.10
• /
• pp.3498-3512
• /
• 2021

This article describes how events that make up text stories can be represented and extracted. We also address the results from our simple experiment on extracting and clustering events in terms of emotions, under the assumption that different emotional events can be associated with the classified clusters. Each emotion cluster is based on Plutchik's eight basic emotion model, and the attributes of the NLTK-VADER are used for the classification criterion. While comparisons of the results with human raters show less accuracy for certain emotion types, emotion types such as joy and sadness show relatively high accuracy. The evaluation results with NRC Word Emotion Association Lexicon (aka EmoLex) show high accuracy values (more than 90% accuracy in anger, disgust, fear, and surprise), though precision and recall values are relatively low.

• Annual Conference on Human and Language Technology
• /
• 2019.10a
• /
• pp.506-509
• /
• 2019

신문기사나 위키피디아와 같이 정보를 전달하는 텍스트와는 달리 사람의 감정 및 의도를 표현하는 텍스트는 다양한 형태의 노이즈를 포함한다. 본 논문에서는 data-driven 방법을 이용하여 노이즈와 단어들 사이의 관계를 LSTM을 이용하여 하나의 벡터로 요약하는 모델을 제안한다. 노이즈 문장 벡터를 표현하는 방식으로는 단방향 LSTM 인코더과 양방향 LSTM 인코더의 두 가지 모델을 이용하여 노이즈를 포함하는 영화 리뷰 데이터를 가지고 감정 분석 실험을 하였고, 실험 결과 단방향 LSTM 인코더보다 양방향 LSTM인 코더가 우수한 성능을 보여주었다.

• Annual Conference on Human and Language Technology
• /
• 2019.10a
• /
• pp.149-152
• /
• 2019

감성분석이란 입력된 텍스트의 감성을 분류하는 자연어처리의 한 분야로, 최근 CNN, RNN, Transformer등의 딥러닝 기법을 적용한 다양한 연구가 있다. 한국어 감성분석을 진행하기 위해서는 형태소, 음절 등의 추가 자질을 활용하는 것이 효과적이며 성능 향상을 기대할 수 있는 방법이다. 모델 생성에 있어서 아키텍쳐 구성도 중요하지만 문맥에 따른 언어를 컴퓨터가 표현할 수 있는 지식 표현 체계 구성도 상당히 중요하다. 이러한 맥락에서 BERT모델은 문맥을 완전한 양방향으로 이해할 수있는 Language Representation 기반 모델이다. 본 논문에서는 최근 CNN, RNN이 융합된 모델과 Transformer 기반의 한국어 KoBERT 모델에 대해 감성분석 task에서 다양한 성능비교를 진행했다. 성능분석 결과 어절단위 한국어 KoBERT모델에서 90.50%의 성능을 보여주었다.

• Journal of Intelligence and Information Systems
• /
• v.25 no.3
• /
• pp.89-115
• /
• 2019

This study was conducted to present the stylistic differences between Arthur Conan Doyle and Agatha Christie, famous as writers of classical mystery novels, through data analysis, and further to present the analytical methodology of the study of style based on text mining. The reason why we chose mystery novels for our research is because the unique devices that exist in classical mystery novels have strong stylistic characteristics, and furthermore, by choosing Arthur Conan Doyle and Agatha Christie, who are also famous to the general reader, as subjects of analysis, so that people who are unfamiliar with the research can be familiar with them. The primary objective of this study is to identify how the differences exist within the text and to interpret the effects of these differences on the reader. Accordingly, in addition to events and characters, which are key elements of mystery novels, the writer's grammatical style of writing was defined in style and attempted to analyze it. Two series and four books were selected by each writer, and the text was divided into sentences to secure data. After measuring and granting the emotional score according to each sentence, the emotions of the page progress were visualized as a graph, and the trend of the event progress in the novel was identified under eight themes by applying Topic modeling according to the page. By organizing co-occurrence matrices and performing network analysis, we were able to visually see changes in relationships between people as events progressed. In addition, the entire sentence was divided into a grammatical system based on a total of six types of writing style to identify differences between writers and between works. This enabled us to identify not only the general grammatical writing style of the author, but also the inherent stylistic characteristics in their unconsciousness, and to interpret the effects of these characteristics on the reader. This series of research processes can help to understand the context of the entire text based on a defined understanding of the style, and furthermore, by integrating previously individually conducted stylistic studies. This prior understanding can also contribute to discovering and clarifying the existence of text in unstructured data, including online text. This could help enable more accurate recognition of emotions and delivery of commands on an interactive artificial intelligence platform that currently converts voice into natural language. In the face of increasing attempts to analyze online texts, including New Media, in many ways and discover social phenomena and managerial values, it is expected to contribute to more meaningful online text analysis and semantic interpretation through the links to these studies. However, the fact that the analysis data used in this study are two or four books by author can be considered as a limitation in that the data analysis was not attempted in sufficient quantities. The application of the writing characteristics applied to the Korean text even though it was an English text also could be limitation. The more diverse stylistic characteristics were limited to six, and the less likely interpretation was also considered as a limitation. In addition, it is also regrettable that the research was conducted by analyzing classical mystery novels rather than text that is commonly used today, and that various classical mystery novel writers were not compared. Subsequent research will attempt to increase the diversity of interpretations by taking into account a wider variety of grammatical systems and stylistic structures and will also be applied to the current frequently used online text analysis to assess the potential for interpretation. It is expected that this will enable the interpretation and definition of the specific structure of the style and that various usability can be considered.

• Journal of Korea Society of Industrial Information Systems
• /
• v.27 no.6
• /
• pp.115-126
• /
• 2022

The purpose of this study is to try to IPA(Importance-Performance Analysis) by applying text mining approaches to user review data for korea mobile banking applications, and to derive priorities for improvement. User review data on mobile banking applications of korea commercial banks (Kookmin Bank, Shinhan Bank, Woori Bank, Hana Bank), local banks (Gyeongnam Bank, Busan Bank), and Internet banks (Kakao Bank, K-Bank, Toss) that gained from Google playstore were used. And LDA topic modeling, frequency analysis, and sentiment analysis were used to derive key attributes and measure the importance and satisfaction of each attribute. Result, although 'Authorizing service', 'Improvement of Function', 'Login', 'Speed/Connectivity', 'System/Update' and 'Banking Service' are relatively important attributes when users use mobile banking applications, their satisfaction is not at the average level, indicating that improvement is urgent.

• Journal of the Korean Institute of Landscape Architecture
• /
• v.49 no.3
• /
• pp.1-10
• /
• 2021

This study aimed to evaluate preferences of Bukhansan dulegil using sentiment analysis, a natural language processing technique, to derive preferred and non-preferred factors. Therefore, we collected blog articles written in 2019 and produced sentimental scores by the derivation of positive and negative words in the texts for 21 dulegil courses. Then, content analysis was conducted to determine which factors led visitors to prefer or dislike each course. In blogs written about Bukhansan dulegil, positive words appeared in approximately 73% of the content, and the percentage of positive documents was significantly higher than that of negative documents for each course. Through this, it can be seen that visitors generally had positive sentiments toward Bukhansan dulegil. Nevertheless, according to the sentiment score analysis, all 21 dulegil courses belonged to both the preferred and non-preferred courses. Among courses, visitors preferred less difficult courses, in which they could walk without a burden, and in which various landscape elements (visual, auditory, olfactory, etc.) were harmonious yet distinct. Furthermore, they preferred courses with various landscapes and landscape sequences. Additionally, visitors appreciated the presence of viewpoints, such as observation decks, as a significant factor and preferred courses with excellent accessibility and information provisions, such as information boards. Conversely, the dissatisfaction with the dulegil courses was due to noise caused by adjacent roads, excessive urban areas, and the inequality or difficulty of the course which was primarily attributed to insufficient information on the landscape or section of the course. The results of this study can serve not only serve as a guide in national parks but also in the management of nearby forest green areas to formulate a plan to repair and improve dulegil. Further, the sentiment analysis used in this study is meaningful in that it can continuously monitor actual users' responses towards natural areas. However, since it was evaluated based on a predefined sentiment dictionary, continuous updates are needed. Additionally, since there is a tendency to share positive content rather than negative views due to the nature of social media, it is necessary to compare and review the results of analysis, such as with on-site surveys.

• Journal of Internet Computing and Services
• /
• v.17 no.5
• /
• pp.111-119
• /
• 2016

In recent years, the opinion analysis is one of the key research fronts of any domain. Opinion target extraction is an essential process of opinion analysis. Target is usually referred to noun or noun phrase in an entity which is deliberated by the opinion holder. Extraction of opinion target facilitates the opinion analysis more precisely and in addition helps to identify the opinion polarity i.e. users can perceive opinion in detail of a target including all its features. One of the most commonly employed algorithms is a sequence labeling algorithm also called Conditional Random Fields. In present article, recent opinion target extraction approaches are reviewed based on sequence labeling algorithm and it features combinations by analyzing and comparing these approaches. The good selection of features combinations will in some way give a good or better accuracy result. Features combinations are an essential process that can be used to identify and remove unneeded, irrelevant and redundant attributes from data that do not contribute to the accuracy of a predictive model or may in fact decrease the accuracy of the model. Hence, in general this review eventually leads to the contribution for the opinion analysis approach and assist researcher for the opinion target extraction in particular.

• The Journal of Society for e-Business Studies
• /
• v.22 no.4
• /
• pp.141-158
• /
• 2017

Along with IT development, customers are getting more easily to express their opinions using various IT channels. In this situation, complaint management is a pressing issue for companies to acquire and maintain loyal customers with low cost. Most of previous studies have investigated customer complaint information by quantitative variables such as demographic information, transaction information, or complaint frequency, but studies focusing on qualitative aspects of complaint information are limited. Therefore, this paper considers the possibility for customers to leave even when they complain occasionally or briefly. This paper analyzes the quantitive aspects as well as the qualitative aspects using sentiment analysis with Exit-voice theory. The dataset contains 268,364 inquiries of 46,235 customers obtained from a contact center of a private security company in Korea. This paper carries out logistic regression and the results imply that the customers's explicit response and their implicit sentiment have different effect on customers leave. This study is expected to provide useful suggestions for the effective complaint management.

• Journal of Digital Convergence
• /
• v.18 no.2
• /
• pp.277-284
• /
• 2020

Text mining is a representative method of big data analysis that extracts meaningful information from unstructured and large amounts of text data. Social media such as Twitter generates hundreds of thousands of data per second and acts as a one-person media that instantly and directly expresses public opinions and ideas. The traditional media are delivering informations, criticizing society, and forming public opinions. For this, we compare the responses of SNS with the responses of media on the issue of the termination of the Korea-Japan GSOMIA (General Security of Military Information Agreement), one of the domestic issues in the second half of 2019. Data collected from 201,728 tweets and 20,698 newspaper articles were analyzed by sentiment analysis, association keyword analysis, and cluster analysis. As a result, SNS tends to respond positively to this issue, and the media tends to react negatively. In association keyword analysis, SNS shows positive views on domestic issues such as "destruction, decision, we," while the media shows negative views on external issues such as "disappointment, regret, concern". SNS is faster and more powerful than media when studying or creating social trends and opinions, rather than the function of information delivery. This can complement the role of the media that reflects public perception.