• Annual Conference on Human and Language Technology
• /
• 2022.10a
• /
• pp.465-469
• /
• 2022

자동 음성 인식 (Automatic Speech Recognition) 기술이 발달함에 따라 자동 음성 인식 시스템의 성능을 높이기 위한 방법 중 하나로 자동 후처리기 연구(automatic post-processor)가 진행되어 왔다. 후처리기를 훈련시키기 위해서는 오류 유형이 포함되어 있는 병렬 말뭉치가 필요하다. 이를 만드는 간단한 방법 중 하나는 정답 문장에 오류를 삽입하여 오류 문장을 생성하여 pseudo 병렬 말뭉치를 만드는 것이다. 하지만 이는 실제적인 오류가 아닐 가능성이 존재한다. 이를 완화시키기 위하여 Back TranScription (BTS)을 이용하여 후처리기 모델 훈련을 위한 병렬 말뭉치를 생성하는 방법론이 존재한다. 그러나 해당 방법론으로 생성 할 경우 노이즈가 적을 수 있다는 관점이 존재하다. 이에 본 연구에서는 BTS 방법론과 인위적으로 노이즈 강도를 추가한 방법론 간의 성능을 비교한다. 이를 통해 BTS의 정량적 성능이 가장 높은 것을 확인했을 뿐만 아니라 정성적 분석을 통해 BTS 방법론을 활용하였을 때 실제 음성 인식 상황에서 발생할 수 있는 실제적인 오류를 더 많이 포함하여 병렬 말뭉치를 생성할 수 있음을 보여준다.

• Journal of the Korea Society of Computer and Information
• /
• v.28 no.8
• /
• pp.31-38
• /
• 2023

Style transfer is one of deep learning-based image processing techniques that has been actively researched recently. These research efforts have led to significant improvements in the quality of result images. Style transfer is a technology that takes a content image and a style image as inputs and generates a transformed result image by applying the characteristics of the style image to the content image. It is becoming increasingly important in exploiting the diversity of digital content. To improve the usability of style transfer technology, ensuring stable performance is crucial. Recently, in the field of natural language processing, the concept of Transformers has been actively utilized. Attention maps, which forms the basis of Transformers, is also being actively applied and researched in the development of style transfer techniques. In this paper, we analyze the representative techniques SANet and AdaAttN and propose a novel attention map-based structure which can generate improved style transfer results. The results demonstrate that the proposed technique effectively preserves the structure of the content image while applying the characteristics of the style image.

• Journal of the Korea Society of Computer and Information
• /
• v.28 no.9
• /
• pp.35-46
• /
• 2023

In this study, we propose a Dual Approach methodology to enhance the accuracy of document classifiers by utilizing both contextual and keyword information. Firstly, contextual information is extracted using Google's BERT, a pre-trained language model known for its outstanding performance in various natural language understanding tasks. Specifically, we employ KoBERT, a pre-trained model on the Korean corpus, to extract contextual information in the form of the CLS token. Secondly, keyword information is generated for each document by encoding the set of keywords into a single vector using an Autoencoder. We applied the proposed approach to 40,130 documents related to healthcare and medicine from the National R&D Projects database of the National Science and Technology Information Service (NTIS). The experimental results demonstrate that the proposed methodology outperforms existing methods that rely solely on document or word information in terms of accuracy for document classification.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2022.05a
• /
• pp.2-5
• /
• 2022

With the advent of deep learning and the rapid development of ICT (Information and Communication Technology), research using artificial intelligence is being actively conducted in various fields of society such as politics, economy, and culture and so on. Deep learning-based artificial intelligence technology is subdivided into various domains such as natural language processing, image processing, speech processing, and recommendation system. In particular, as the industry is advanced, the need for a recommendation system that analyzes market trends and individual characteristics and recommends them to consumers is increasingly required. In line with these technological developments, this paper extracts and classifies attribute information from structured or unstructured text and image big data through deep learning-based technology development of 'language processing intelligence' and 'image processing intelligence', and We propose an artificial intelligence-based 'customized fashion advisor' service integration system that analyzes trends and new materials, discovers 'market-consumer' insights through consumer taste analysis, and can recommend style, virtual fitting, and design support.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2022.05a
• /
• pp.206-208
• /
• 2022

As the big data industry has recently developed significantly, interest in privacy violations caused by personal information leakage has increased. There have been attempts to automate this through named entity recognition in natural language processing. In this paper, named entity recognition data is constructed semi-automatically by identifying sentences with de-identification information from de-identification information in Korean Wikipedia. This can reduce the cost of learning about information that is not subject to de-identification compared to using general named entity recognition data. In addition, it has the advantage of minimizing additional systems based on rules and statistics to classify de-identification information in the output. The named entity recognition data proposed in this paper is classified into twelve categories. There are included de-identification information, such as medical records and family relationships. In the experiment using the generated dataset, KoELECTRA showed performance of 0.87796 and RoBERTa of 0.88.

• Journal of Creative Information Culture
• /
• v.5 no.3
• /
• pp.215-223
• /
• 2019

The wave of the information age gradually accelerates, and fusion analysis solutions that can utilize these knowledge data according to accumulation of various forms of big data such as large capacity texts, sounds, movies and the like are increasing, Reduction in the cost of storing data accordingly, development of social network service (SNS), etc. resulted in quantitative qualitative expansion of data. Such a situation makes possible utilization of data which was not trying to be existing, and the potential value and influence of the data are increasing. Research is being actively made to present future-oriented education systems by applying these fusion analysis systems to the improvement of the educational system. In this research, we conducted a big data analysis on Twitter, analyzed the natural language of the data and frequency analysis of the word, quantitative measure of how domestic windows education problems and outcomes were done in it as a solution.

• Journal of Creative Information Culture
• /
• v.5 no.2
• /
• pp.183-191
• /
• 2019

The purpose of this paper is to develop an unplugged education program that develops the 4C (Creativity, Critical thinking, Communication ability, Collaboration) and CT (Computational Thinking) competencies required in modern society. This study discovered "Fine Dust Robot" as a theme suitable for the unplugged education program, and designed the Unplugged 4-hour education program which can develop 4C and CT competencies. The first stage motivates learning, and the second and third stages develop unplugged activity to develop CT. In the fourth stage, the algorithms created through unplugged activities were programmed through the natural language instruction card and produced the output. We developed educational materials that can be utilized in the unplugged education program. Finally, education programs were conducted for elementary school students, and pre- and post-tests of computational thinking were conducted for general students and gifted students. Educational effective was found in both groups.

• Journal of Broadcast Engineering
• /
• v.28 no.3
• /
• pp.333-336
• /
• 2023

Recently, the transformer model used in natural language processing is also applied to the image super resolution field, showing good performance. However, these transformer based models have a disadvantage that they are difficult to use in small mobile devices because they are complex and have many learning parameters and require high hardware resources. Therefore, in this paper, we propose a knowledge distillation technique that can effectively reduce the size of a transformer based super resolution model. As a result of the experiment, it was confirmed that by applying the proposed technique to the student model with reduced number of transformer blocks, performance similar to or higher than that of the teacher model could be obtained.

• Journal of the Korea Society of Computer and Information
• /
• v.14 no.8
• /
• pp.65-71
• /
• 2009

Natural language processing is necessary in order to efficiently perform filtering tremendous information produced in information retrieval of world wide web. This paper suggested an algorithm for meaning of numerals in the text. The algorithm for meaning of numerals utilized context-free grammars with the chart parsing technique, interpreted affixes connected with the numerals and was designed to disambiguate their meanings systematically supported by the n-gram based words. And the algorithm was designed to use POS (part-of-speech) taggers, to automatically recognize restriction conditions of trigram words, and to gradually disambiguate the meaning of the numerals. This research performed experiment for the suggested system of the numeral interpretation. The result showed that the frequency-proportional method recognized the numerals with 86.3% accuracy and the condition-proportional method with 82.8% accuracy.

• Journal of the Korea Society of Computer and Information
• /
• v.14 no.1
• /
• pp.1-8
• /
• 2009

Learning user's preference is a key issue in intelligent system such as personalized service. The study on user preference model has adapted simple user preference model, which determines a set of preferred keywords or topic, and weights to each target. In this paper, we recommend multi-perspective user preference model that factors sentiment information in the model. Based on the topicality and sentimental information processed using natural language processing techniques, it learns a user's preference. To handle timc-variant nature of user preference, user preference is calculated by session, short-term and long term. User evaluation is used to validate the effect of user preference teaming and it shows 86.52%, 86.28%, 87.22% of accuracy for topic interest, keyword interest, and keyword favorableness.