• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.24 no.12
• /
• pp.1595-1603
• /
• 2020

Online defamation incidents such as Internet news comments on portal sites, SNS, and community sites are increasing in recent years. Bias and hate expressions threaten online service users in various forms, such as invasion of privacy and personal attacks, and defamation issues. In the past few years, academia and industry have been approaching in various ways to solve this problem The purpose of this study is to build a dataset and experiment with deep learning classification modeling for detecting various bias expressions as well as hate expressions. The dataset was annotated 7 labels that 10 personnel cross-checked. In this study, each of the 7 classes in a dataset of about 137,111 Korean internet news comments is binary classified and analyzed through deep learning techniques. The Proposed technique used in this study is multi-channel CNN model with attention. As a result of the experiment, the weighted average f1 score was 70.32% of performance.

• International Journal of Fuzzy Logic and Intelligent Systems
• /
• v.6 no.1
• /
• pp.27-32
• /
• 2006

Verbal communication is the most commonly used mean of communication. A spoken word carries a lot of informations about speakers and their emotional states. In this paper we designed a model to recognize emotional states in a speech, a first phase of two phases in developing a toy machine that recognizes emotional states in a speech. We conducted an experiment to extract and analyse the emotional state of a speaker in relation with speech. To analyse the signal output we referred to three characteristics of sound as vector inputs and they are the followings: frequency, intensity, and period of tones. Also we made use of eight basic emotional parameters: surprise, anger, sadness, expectancy, acceptance, joy, hate, and fear which were portrayed by five selected students. In order to facilitate the differentiation of each spectrum features, we used the wavelet transform analysis. We applied ANFIS (Adaptive Neuro Fuzzy Inference System) in designing an emotion recognition model from a speech. In our findings, inference error was about 10%. The result of our experiment reveals that about 85% of the model applied is effective and reliable.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2021.11a
• /
• pp.606-609
• /
• 2021

최근 한국에서도 이루다의 윤리 이슈를 기점으로 딥러닝 모델의 윤리적 언어학습 필요성이 대두되었다. 그럼에도 불구하고 영어 데이터에 비해 한국어 데이터는 Korean Hate Speech Detection Dataset 이 유일하다. 이번 연구에서는 기존 데이터 세트의 유연성이 떨어지고 세부 라벨이 제한적이라는 문제를 개선한 새로운 데이터 세트를 제안하고, 해당 데이터 세트에 대하여 다양한 신경망 분류 모델을 적용한 벤치마크 결과를 공개한다.

• The Journal of the Acoustical Society of Korea
• /
• v.22 no.1
• /
• pp.23-27
• /
• 2003

Recently attempts to we the syllable as the recognition unit to enhance performance in continuous speech recognition hate been reported. However, syllables are worse in their trainability than phones and the former have a disadvantage in that contort-dependent modeling is difficult across the syllable boundary since the number of models is much larger for syllables than for phones. In this paper, we propose a method to enhance the trainability for the syllables in Korean and phoneme-context dependent syllable modeling across the syllable boundary. An experiment in which the proposed method is applied to word recognition shows average 46.23% error reduction in comparison with the common syllable modeling. The right phone dependent syllable model showed 16.7% error reduction compared with a triphone model.

• Journal of the Korea Society of Computer and Information
• /
• v.23 no.9
• /
• pp.27-34
• /
• 2018

Personal media centered on blogs, Twitter, and Facebook has opened up a personal broadcasting area while meeting platforms such as YouTube and Africa TV. Due to the many advantages and disadvantages of personal broadcasting, a study on it was necessary and statistical survey was conducted. The study conducted opinion survey of 118 university students on personal broadcasting. As a result, we are getting news using smartphones and mainly watching videos through YouTube, and watching videos type in the order of games, music videos and sports. Satisfaction rate of video was 72.4%, 80.2% of survey did not use paid services, experiences about personal broadcasting was 96.6% and 90.5% of survey the prospect of person broadcasting of the opinion that "it will be expanded". The first thing we want to be improved in personal broadcasting is the prevention of abusive language and hate speech. Second, we were reluctant to sensational content. Third, the survey results are the improvement of excessive advertising.

• Asian Journal for Public Opinion Research
• /
• v.7 no.3
• /
• pp.169-198
• /
• 2019

This study focused on the increase in refugee-related discourse in Korean society with the recent inflow of asylum seekers to Jeju Island. The purpose of our study was to understand the trends in public opinion concerning the acceptance of refugees by analyzing the content of refugee-related video commentary on YouTube. Topic modeling was conducted to analyze the main points, context, and ideas in the comments. The results indicated that the media mainly focus on the pros and cons of refugees, restricting the refugee issue to the problem of acceptance with a narrow focus on the case of Jeju Island. Refugee acceptance was treated as overwhelmingly unacceptable in the comments. We found that commenters often used negative discourse in the comments as a device for reproducing and amplifying hate speech.

• Journal of Public Diplomacy
• /
• v.1 no.2
• /
• pp.49-74
• /
• 2021

In this paper, I applied grounded theory in exploring how Twitter became the battlefield for China's public diplomacy campaign. China's new move to global social media platforms, such as Twitter and Facebook, has been a controversial strategy in public diplomacy. This study analyzes Chinese Foreign Spokesperson Zhao Lijian's Twitter posts and comments. It models China's recent diplomatic move to Twitter as a "war of words" model, with features including "leadership," "polarization," and "aggression," while exerting possible effects as "resistance," "hatred," and "sarcasm" to the global community. Our findings show that by failing to gage public opinion and promote the country's positive image, China's current digital diplomacy strategy reflected by Zhao Lijian's tweets has instead constructed a polarized political public sphere, contradictory to the country's promoted "shared human destiny." The "war of words" model extends our understanding of China's new digital diplomacy move as a hybrid of state propaganda and self-performance. Such a strategy could spread hate speech and accelerate political polarization in cyberspace, despite improvements to China's homogenous network building on Twitter.

• Annual Conference on Human and Language Technology
• /
• 2020.10a
• /
• pp.298-303
• /
• 2020

온라인 공간에서 특정인, 혹은 특정 집단의 사람들을 대상으로 한 혐오 표현은 당사자에게 정신적 고통을 미칠 뿐 아니라 이를 보는 이에게도 간접적인 불쾌함을 유발한다. 이에 관한 문제의식은 사회적으로 공감대가 형성된 바 있지만, 아직 한국어에서는 많은 연구들이 혐오 표현 자체의 논의에 집중하고 있으며, 이는 실제로 관찰되는 혐오 표현들의 자동 탐지 및 예방에는 효과적인 정보를 제공하지 못하는 것이 사실이다. 이에 우리는 실제 온라인 댓글들을 탐구하여 혐오, 모욕 및 사회적 편견을 탐지할 수 있는 모델 학습에 필요한 코퍼스 구축 가이드라인을 제작하였다. 구체적인 사례를 동반한 가이드라인과 크라우드소싱을 바탕으로 약 9천 3백 문장 가량의 코퍼스를 구축하였으며, 해당 데이터에 관한 개요와 함께 우리의 접근 방식이 어떤 점에서 기존의 담론과 연관되어 있는지에 대한 분석을 제시한다.

• Annual Conference on Human and Language Technology
• /
• 2021.10a
• /
• pp.365-370
• /
• 2021

대형 코퍼스로 학습한 언어 모델은 코퍼스 안의 사회적 편견이나 혐오 표현까지 학습한다. 본 연구에서는 한국어 오픈 도메인 대화 모델에서 혐오 표현 생성을 완화하는 방법을 제시한다. Seq2seq 구조인 BART [1]를 기반으로 하여 컨트롤 코드을 추가해 혐오 표현 생성 조절을 수행하였다. 컨트롤 코드를 사용하지 않은 기준 모델(Baseline)과 비교한 결과, 컨트롤 코드를 추가해 학습한 모델에서 혐오 표현 생성이 완화되었고 대화 품질에도 변화가 없음을 확인하였다.

• Annual Conference on Human and Language Technology
• /
• 2021.10a
• /
• pp.518-523
• /
• 2021

챗봇과 같은 대화형 에이전트 사용이 증가하면서 채팅에서의 혐오 표현 사용도 더불어 증가하고 있다. 혐오 표현을 자동으로 탐지하려는 노력은 다양하게 시도되어 왔으나, 챗봇 데이터를 대상으로 한 혐오 표현 탐지 연구는 여전히 부족한 실정이다. 이 연구는 혐오 표현을 포함한 챗봇-사용자 대화 데이터 35만 개에 한국어 말뭉치로 학습된 KoELETRA 기반 혐오 탐지 모델을 적용하여, 챗봇-사람 데이터셋에서의 혐오 표현 탐지의 성능과 한계점을 검토하였다. KoELECTRA 혐오 표현 분류 모델은 챗봇 데이터셋에 대해 가중 평균 F1-score 0.66의 성능을 보였으며, 오탈자에 대한 취약성, 맥락 미반영으로 인한 편향 강화, 가용한 데이터의 정확도 문제가 주요한 한계로 포착되었다. 이 연구에서는 실험 결과에 기반해 성능 향상을 위한 방향성을 제시한다.