• Journal of Information Technology Services
• /
• v.19 no.5
• /
• pp.83-91
• /
• 2020

Language Models such as BERT has been an important factor of deep learning-based natural language processing. Pre-training the transformer-based language models would be computationally expensive since they are consist of deep and broad architecture and layers using an attention mechanism and also require huge amount of data to train. Hence, it became mandatory to do fine-tuning large pre-trained language models which are trained by Google or some companies can afford the resources and cost. There are various techniques for fine tuning the language models and this paper examines three techniques, which are data augmentation, tuning the hyper paramters and partly re-constructing the neural networks. For data augmentation, we use no-answer augmentation and back-translation method. Also, some useful combinations of hyper parameters are observed by conducting a number of experiments. Finally, we have GRU, LSTM networks to boost our model performance with adding those networks to BERT pre-trained model. We do fine-tuning the pre-trained korean-based language model through the methods mentioned above and push the F1 score from baseline up to 89.66. Moreover, some failure attempts give us important lessons and tell us the further direction in a good way.

• Korean Journal of Cognitive Science
• /
• v.15 no.3
• /
• pp.35-44
• /
• 2004

As the use of internet becomes proliferated, a great amount of information is provided through diverse channels. Users require effective information providing service and we have studied the conversational agent that exchanges information between users and agents using natural language dialogue. In this paper, we develop a question answering agent providing the corresponding answer by analyzing the user's intention using artificial intelligence techniques such as pattern matching and Bayesian network We work out various problems in knowledge representation of users by constructing keyword synonym database. The proposed method is applied to designing an agent for the introduction of a fashion web site, which confirms that it responds more flexibly to the user's queries.

• Journal of muscle and joint health
• /
• v.11 no.1
• /
• pp.42-49
• /
• 2004

Purpose: To establish more effective health counselling system by internet and to provide basic data for developing on-line education program for lupus patients. Method: This survey analyzed the contents of health counselling cases from September, 2002 to March, 2004. Among 108 questioners visited questioning/answering(Q/A) site in one lupus center home page. Contents of Q/A were classified and analyzed using SAS program. Results: Most frequent questioners were female and twenties, they needed to know diagnosing process, treatment principles and information regarding the clinic visits. During the study period, the average number of user's questioning were 166.2 cases and answering was 171.1 cases. Conclusions: These findings suggest that computer mediated health counselling may be useful supplement of health education for lupus patients. More research is needed to develop better health counselling service and education program on the web.

• Journal of muscle and joint health
• /
• v.7 no.1
• /
• pp.40-52
• /
• 2000

Recently with the development in computer technology and its communication system, many changes in medical informatics enable us to use various medical information regardless of time or place. There are many home pages on the web, which provide medical counseling and hospital information. On May 11th 1999, Korean Rheumatology Health Professions Society began its new service as a home page on the web with various rheumatologic health information, questioning/ answering, and so on. This study was undertaken to examine the content and the purpose of health counseling on the web. The data was collected from 173 questioners who visited questioning/ answering site in the KRHPS home page for May 11th, 1999 through November 10th, 1999. Most of the questioners consulted the health problems of their or their families. Over two thirds of them were already diagnosed medically. Rheumatoid arthritis was the most frequent one. Other diseases, such as, osteoarthritis, ankylosing spondylitis, Still's disease were also on the list. Most of the questioners wanted to know treatment strategies, to consult about their symptoms, and to make diagnosis. And many questioners wanted detailed explanations about their diseases or the informations regarding the hospital. These findings suggest that the health counseling on the web may be used to supplement the lack of direct medical interviews with doctors. It also is expected to guide the patients to the right direction.

• Proceedings of the IEEK Conference
• /
• 2003.11b
• /
• pp.133-136
• /
• 2003

Due to recent explosion of information available online, question- answering (Q&A) systems are becoming a compelling framework for finding relevant information in a variety of domains. Question-answering system is one of the best ways to introduce a novice customer to a new domain without making him/her to obtain prior knowledge of its overall structure improving search request with specific answer. However, the current web poses serious problem for finding specific answer for many overlapped meanings for the same questions or duplicate questions also retrieved answer for many overlapped meanings fer the same questions or duplicate questions also retrieved answer is slow due to enhanced network traffic, which leads to wastage of resource. In order to avoid wrong answer which occur due to above-mentioned problem we propose the system using ontology by RDF, RDFS and mobile agent based on JAVA. We also choose wireless internet based embedded device as our test bed for the system and apply the system in E-commerce information domain. The mobile agent provides agent routing with reduced network traffic, consequently helps us to minimize the elapsed time for answers and structured ontology based on our proposed algorithms sorts out the similarity between current and past question by comparing properties of classes.

• Journal of the Korean Society for information Management
• /
• v.21 no.3
• /
• pp.289-303
• /
• 2004

This experimental study proposes a multi-document summarization method that produces optimal summaries in which users can find answers to their queries. In order to identify the most effective method for this purpose, the performance of the three summarization methods were compared. The investigated methods are sentence clustering, passage extraction through spreading activation, and clustering-passage extraction hybrid methods. The effectiveness of each summarizing method was evaluated by two criteria used to measure the accuracy and the redundancy of a summary. The passage extraction method using the sequential bnb search algorithm proved to be most effective in summarizing multiple documents with regard to summarization precision. This study proposes the passage extraction method as the optimal multi-document summarization method.

• Journal of the Korean Operations Research and Management Science Society
• /
• v.24 no.1
• /
• pp.99-117
• /
• 1999

Since query language is used as a handy tool to obtain information from a database, a more intelligent query answering system is needed to provide user-friendly and fault-tolerant human-machine Interface. Frequently, database users prefer less rigid querying structure, one which allows for vagueness in composing queries, and want the system to understand the intent behind a query. When there is no matching data available, users would rather receive approximate answers than a null information response. This paper presents a knowledge abstraction database that facilitates the development of such a fault-tolerant and intelligent database system. The proposed knowledge abstraction database adepts a multilevel knowledge representation scheme called the knowledge abstraction hierarchy(KAH), extracts semantic data relationships from the underlying database, and provides query transformation mechanisms using query generalization and specialization steps. In cooperation with the underlying database, the knowledge abstraction database accepts vague queries and allows users to pose approximate queries as well as conceptually abstract queries. Specifically. four types of vague queries are discussed, including approximate selection, approximate join, conceptual selection, and conceptual Join. A prototype system has been implemented at KAIST and is being tested with a personnel database system to demonstrate the usefulness and practicality of the knowledge abstraction database in ordinary database application systems.

• ETRI Journal
• /
• v.41 no.6
• /
• pp.771-781
• /
• 2019

We propose the VS³-NET model to solve the task of question answering questions with machine-reading comprehension that searches for an appropriate answer in a given context. VS³-NET is a model that trains latent variables for each question using variational inferences based on a model of a simple recurrent unit-based sentences and self-matching networks. The types of questions vary, and the answers depend on the type of question. To perform efficient inference and learning, we introduce neural question-type models to approximate the prior and posterior distributions of the latent variables, and we use these approximated distributions to optimize a reparameterized variational lower bound. The context given in machine-reading comprehension usually comprises several sentences, leading to performance degradation caused by context length. Therefore, we model a hierarchical structure using sentence encoding, in which as the context becomes longer, the performance degrades. Experimental results show that the proposed VS³-NET model has an exact-match score of 76.8% and an F1 score of 84.5% on the SQuAD test set.

• The Journal of Korean Association of Computer Education
• /
• v.6 no.4
• /
• pp.23-30
• /
• 2003

This paper proposes a passage retrieval system based on Okapi's BM25 for developing an efficient QA system and evaluates performances of the passage retrieval system. The test collection of TREC Q&A track which is composed of about one million documents was indexed and a hundred queries of TREC Q&A track are used as testing queries. The experimental results shows that the proposed passage retrieval system can reach to 100% recall rate by searching in only 1700 sentences while the conventional document retrieval system have to search about 120 thousands sentences which are about 70 times more than the proposed passage retrieval system.

• Journal of Fisheries and Marine Sciences Education
• /
• v.28 no.5
• /
• pp.1508-1516
• /
• 2016

This research intended to analyse the perception level of seafarers in order to propose enhancement of health management for seafarers and improvement of marine telemedicine assistance system. To this end, questionnaire survey was conducted for 422 seafarers in relation to current status of utilizing Emergency Medical Advice by radio(EMAR). By trading area, experience of benefiting EMAR by coastal seafarers was higher than that of ocean-going seafarers. By kinds of ships the rate of utilizing EMAR by fisherman was higher than that of seafarers of merchant ships. The level of satisfaction on the EMAR was 14.9% and regarded to be very low while the answering rate of necessity and urgency toward introducing marine telemedicine assistance system for the future through utilizing telemedicine equipment what is called U-helath including motion pictures information was very high. The answering rate of willingness to provide their personal health information positively when introducing marine telemedicine system was very high(97.4%). It is thus necessary to permit marine telemedicine service system, establish service model of marine telemedicine through conducting demonstration projects, promote marine telemedicine actively at the level of government, etc.