• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.9-13
• /
• 2023

본 논문에서는 대규모 언어모델의 검색 기반 답변 생성능력을 평가하는 새로운 한국어 벤치마크, KFREB(Korean Fictional Retrieval Evaluation Benchmark)를 제안한다. KFREB는 모델이 사전학습 되지 않은 허구의 정보를 바탕으로 검색 기반 답변 생성 능력을 평가함으로써, 기존의 대규모 언어모델이 사전학습에서 보았던 사실을 반영하여 생성하는 답변이 실제 검색 기반 답변 시스템에서의 능력을 제대로 평가할 수 없다는 문제를 해결하고자 한다. 제안된 KFREB는 검색기반 대규모 언어모델의 실제 서비스 케이스를 고려하여 장문 문서, 두 개의 정답을 포함한 골드 문서, 한 개의 골드 문서와 유사 방해 문서 키워드 유무, 그리고 문서 간 상호 참조를 요구하는 상호참조 멀티홉 리즈닝 경우 등에 대한 평가 케이스를 제공하며, 이를 통해 대규모 언어모델의 적절한 선택과 실제 서비스 활용에 대한 인사이트를 제공할 수 있을 것이다.

• 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.

• The Journal of Society for e-Business Studies
• /
• v.8 no.3
• /
• pp.177-195
• /
• 2003

In a traditional customer support environment, mainly call centers or service centers are responsible for receiving inquiries from their customers via telephone calls. Due to the rapid growth of Internet with its widespread acceptance and accessibility, means of communication with customers in the traditional customer support center, such as telephones, letters, and direct-visiting, have been replaced by e-mails and bulletin board systems (BBSs) using the Internet constantly. BBSs are basically question and answer systems, they require some lead time to get answer from administrator. To reduce lead time, BBSs enable remote customers or users to log on and tap into a knowledge database that is generally formatted in the form of Frequently Asked Questions (FAQs) that provide answers and solutions to the common problems. And, many different types of the questions are mixed on the BBS. It is a burden to administrator. To build FAQs and to support BBS adminstrator, a supporting tool which is to categorize questions is helpful. In this research, we suggest an interactive question categorizing methodology which consists of steps to present question using keywords, identifying keywords' affinity, computing similarity among questions, and clustering questions. This methodology allows users to interact iteratively for clear manifestation of ambiguous questions. We also developed a prototype system, IQC (interactive question categorizer) and evaluated its performance using the comparison experiments with other systems. IQC is not a general purposed system, but it produces a good result in a given specific domain.

• Journal of KIISE:Databases
• /
• v.34 no.4
• /
• pp.283-303
• /
• 2007

Recently, the study of efficient way to store and retrieve enormous music data is becoming the one of important issues in the multimedia database. Most general method of MIR (Music Information Retrieval) includes a text-based approach using text information to search a desired music. However, if users did not remember the keyword about the music, it can not give them correct answers. Moreover, since these types of systems are implemented only for exact matching between the query and music data, it can not mine any information on similar music data. Thus, these systems are inappropriate to achieve similarity matching of music data. In order to solve the problem, we propose an Efficient Query-By-Humming System (EQBHS) with a content-based indexing method that efficiently retrieve and store music when a user inquires with his incorrect humming. For the purpose of accelerating query processing in EQBHS, we design indices for significant melodies, which are 1) frequent melodies occurring many times in a single music, on the assumption that users are to hum what they can easily remember and 2) melodies partitioned by rests. In addition, we propose an error tolerated mapping method from a note to a character to make searching efficient, and the frequent melody extraction algorithm. We verified the assumption for frequent melodies by making up questions and compared the performance of the proposed EQBHS with N-gram by executing various experiments with a number of music data.

• Annual Conference on Human and Language Technology
• /
• 1999.10e
• /
• pp.7-12
• /
• 1999

정보 검색 시스템은 사용자 질의의 키워드들과 문서들의 유사성(similarity)을 기준으로 관련 문서들을 순서화하여 사용자에게 제공한다. 그렇지만 인터넷 검색에 사용되는 질의는 일반적으로 짧기 때문에 보다 유용한 질의를 만들고자 하는 노력이 지금까지 계속되고 있다. 그러나 키워드에 포함된 정보가 제한적이기 때문에 이에 대한 보완책으로 사용자의 적합성 피드백을 이용하는 방법을 널리 사용하고 있다. 본 논문에서는 일반적인 적합성 피드백의 가장 큰 단점인 빈번한 사용자 참여는 지양하고, 시스템에 기반한 적합성 피드백에서 배제한 사용자 참여를 유도하는 검색 문서의 분류 정보에 기반한 용어 클러스터 질의 확장 모델(Term Cluster Query Expansion Model)을 제안한다. 이 방법은 검색 시스템에 의해 검색된 상위 n개의 문서에 대하여 분류기를 이용하여 각각의 문서에 분류 정보를 부여하고, 문서에 부여된 분류 정보를 이용하여 분류 정보의 수(m)만큼으로 문서들을 그룹을 짓는다. 적합성 피드백 알고리즘을 이용하여 m개의 그룹으로부터 각각의 용어 클러스터(Term Cluster)를 생성한다. 이 클러스터가 사용자에게 문서 대신에 피드백의 자료로 제공된다. 실험 결과, 적합성 알고리즘 중 Rocchio방법을 이용할 때 초기 질의보다 나은 성능을 보였지만, 다른 연구에서 보여준 성능 향상은 나타내지 못했다. 그 이유는 분류기의 오류와 문서의 특성상 한 영역으로 규정짓기 어려운 문서가 존재하기 때문이다. 그러나 검색하고자 하는 사용자의 관심 분야나 찾고자 하는 성향이 다르더라도 시스템에 종속되지 않고 유연하게 대처하며 검색 성능(retrieval effectiveness)을 향상시킬 수 있다.사용되고 있어 적응에 문제점을 가지기도 하였다. 본 연구에서는 그 동안 계속되어 온 한글과 한잔의 사용에 관한 논쟁을 언어심리학적인 연구 방법을 통해 조사하였다. 즉, 글을 읽는 속도, 글의 의미를 얼마나 정확하게 이해했는지, 어느 것이 더 기억에 오래 남는지를 측정하여 어느 쪽의 입장이 옮은 지를 판단하는 것이다. 실험 결과는 문장을 읽는 시간에서는 한글 전용문인 경우에 월등히 빨랐다. 그러나. 내용에 대한 기억 검사에서는 국한 혼용 조건에서 더 우수하였다. 반면에, 이해력 검사에서는 천장 효과(Ceiling effect)로 두 조건간에 차이가 없었다. 따라서, 본 실험 결과에 따르면, 글의 읽기 속도가 중요한 문서에서는 한글 전용이 좋은 반면에 글의 내용 기억이 강조되는 경우에는 한자를 혼용하는 것이 더 효율적이다.이 높은 활성을 보였다. 7. 이상을 종합하여 볼 때 고구마 끝순에는 페놀화합물이 다량 함유되어 있어 높은 항산화 활성을 가지며, 아질산염소거능 및 ACE저해활성과 같은 생리적 효과도 높아 기능성 채소로 이용하기에 충분한 가치가 있다고 판단된다.등의 관련 질환의 예방, 치료용 의약품 개발과 기능성 식품에 효과적으로 이용될 수 있음을 시사한다.tall fescue 23%, Kentucky bluegrass 6%, perennial ryegrass 8%) 및 white clover 23%를 유지하였다. 이상의 결과를 종합할 때, 초종과 파종비율에 따른 혼파초지의 건물수량과 사료가치의 차이를 확인할 수 있었으며, 레드 클로버 + 혼파 초지가 건물수량과 사료가치를 높이는데 효과적이었다.\ell}$ 이었으며 , yeast extract 첨가(添加)하여 배양시(培養時)는 yeast extract

• Journal of Digital Convergence
• /
• v.19 no.10
• /
• pp.469-489
• /
• 2021

This study explores consumers' responses to socially assistive robotics by using text-mining method focusing on Companion Pets from Hasbro as it gives emotional support. We conducted text frequency analysis, LDA analysis using R programming. The key findings are 1)the most frequently used words the mimicry of living pets and the appearance of companion pets, 2)the five topics were derived from the LDA analysis and classified keywords in each topic split between positive and negative, 3)user, product, environment affect the interaction between consumer and companion pets, 4)consumers who have difficulty in cognition and physical conditions use companion pets to replace living pets. This study provides an understanding of consumer responses in companion pets and gives practical implications that may improve the efficacy of usage for consumers and understand the companion robot, which provides emotional support in COVID-19.

• Knowledge Management Research
• /
• v.22 no.4
• /
• pp.217-235
• /
• 2021

As 49% of the world's population uses social media platforms, communication and content sharing within social media are becoming more active than ever. In this environmental base, the one-person media market grew rapidly and formed public opinion, creating a new trend called sell-sumer. This study defined new types of influencers by product category by analyzing the subject concentration of the commercial/non-commercial keywords of influencers and the impact of the ratio of commercial postings on sales. It is hoped that influencers working within social media will be helpful to new sales strategies that are transformed into sell-sumers. The method of this study classifies influencers' commercial/non-commercial posts using Python, performs text mining using KoNLPy, and calculates similarity between FastText-based words. As a result, it has been confirmed that the higher the keyword theme concentration of the influencer's commercial posting, the higher the sales. In addition, it was confirmed through the cluster analysis that the influencer types for each product category were classified into four types and that there was a significant difference between groups according to sales. In other words, the implications of this study may suggest empirical solutions of social media sales strategies for influencers working on social media and marketers who want to use them as marketing tools.

• Journal of Internet of Things and Convergence
• /
• v.8 no.2
• /
• pp.13-24
• /
• 2022

In order to explore the context of fraudulent claims and the measures for preventing them targeting the long-term care institutions for the elderly, which is increasing every year in Korea, this study conducted the text mining analysis using the media report articles. The media report articles were collected from the news big data analysis system called 'BIG KINDS' for about 15 years from July 2008 when the Long-Term Care Insurance for the Elderly took effect, to February 28^th 2022. During this period of time, total 2,627 articles were collected under keywords like 'elderly care+fraudulent claims' and 'long-term care+fraudulent claims', and among them, total 946 articles were selected after excluding overlapped articles. In the results of the text mining analysis in this study, first, the top 10 keywords mentioned in the highest frequency in every section(July 1^st 2008-February 28^th 2022) were shown in the order of long-term care institution for the elderly, fraudulent claims, National Health Insurance Service, Long-Term Care Insurance for the Elderly, long-term care benefits(expenses), elderly care facilities, The Ministry of Health & Welfare, the elderly, report, and reward(payment). Second, in the results of the N-gram analysis, they were shown in the order of long-term care benefits(expenses) and fraudulent claims, fraudulent claims and long-care institution for the elderly, falsehood and fraudulent claims, report and reward(payment), and long-term care institution for the elderly and report. Third, the analysis of TF-IDF was similar to the results of the frequency analysis while the rankings of report, reward(payment), and increase moved up. Based on such results of the analysis above, this study presented the future direction for the prevention of fraudulent claims of long-term care institutions for the elderly.

• Journal of the Korea Society for Simulation
• /
• v.18 no.2
• /
• pp.39-48
• /
• 2009

This study reviews the limit of image search by considering on the image search methods related to content-based image retrieval and suggests a user interface for more efficient content-based image retrieval and the ways to utilize image properties. For now, most studies on image search are being performed focusing on content-based image retrieval; they try to search based on the image's colors, texture, shapes, and the overall form of the image. However, the results are not satisfactory because there are various technological limits. Accordingly, this study suggests a new retrieval system which adapts content-based image retrieval and the conventional keyword search method. This is about a way to attribute properties to images using texts and a fast way to search images by expressing the attribute of images as keywords and utilizing them to search images. Also, the study focuses on a simulation for a user interface to make query language on the Internet and a search for clothes in an online shopping mall as an application of the retrieval system based on image attribute. This study will contribute to adding a new purchase pattern in online shopping malls and to the development of the area of similar image search.

• The Journal of the Convergence on Culture Technology
• /
• v.10 no.2
• /
• pp.499-503
• /
• 2024

Broadcasting programs are provided to various media such as Internet replay, OTT, and IPTV services as well as self-broadcasting. In this case, it is very important to provide keywords for search that represent the characteristics of the content well. Broadcasters mainly use the method of manually entering key keywords in the production process and the archive process. This method is insufficient in terms of quantity to secure core metadata, and also reveals limitations in recommending and using content in other media services. This study supports securing a large number of metadata by utilizing closed caption data pre-archived through the DTV closed captioning server developed in EBS. First, core metadata was automatically extracted by applying Google's natural language AI technology. The next step is to propose a method of finding core metadata by reflecting priorities and content characteristics as core research contents. As a technology to obtain differentiated metadata weights, the importance was classified by applying the TF-IDF calculation method. Successful weight data were obtained as a result of the experiment. The string metadata obtained by this study, when combined with future string similarity measurement studies, becomes the basis for securing sophisticated content recommendation metadata from content services provided to other media.