• 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.

• Korean Journal of Culture and Social Issue
• /
• v.26 no.3
• /
• pp.283-301
• /
• 2020

In Korean society, words related to misogyny are being created and spread out in the Internet communities and the Internet news posts comments. This study was conducted to investigate if exposure to misogynistic words affects misogynistic attitudes toward women. Study 1 examined the relationship between exposure of misogynistic words (the number of misogynistic words known and the level of Internet comments viewed) and explicit misogynistic attitudes. As a result, the greater the exposure of misogynistic words, the less explicit misogynistic attitudes were found among men. The result can be explained as a desensitization of stimuli caused by repetitive exposure to media. In Study 2, experiments were conducted to manipulate the exposure of misogynistic words and to identify the relationship between implicit misogynistic attitudes through implicit association tests. Results of analysis show that implicit misogyny attitude is stronger as male participants are exposed to misogynistic words. The finding of this study suggests that explicit and implicit attitudes toward misogyny can diverge. It also implies that the exposure to misogynistic words can affect men's and women's attitudes in a different manner.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.95-100
• /
• 2023

끊임없이 재생산되는 혐오 표현의 정확한 탐지를 위해서는 혐오란 무엇인가에 대한 본질적인 이해가 필요하다. 본 연구에서는 화용론에서 사용되는 적정성 조건이라는 분석 틀을 활용하여 모델이 '혐오하기' 화행을 어떻게 인식하고 있는지 평가하고자 했다. 혐오 화행의 적정성 조건을 명제 내용 조건, 예비 조건, 성실성 조건, 본질 조건으로 나누어 분석하였으며, 이를 진위형, 연결형, 단답형, 논술형 문항으로 구성했다. 그 결과 모든 문항 유형에서 50점이 넘는 점수를 받았으나 비교적 고차원인 사고 능력을 측정하는 단답형과 논술형 문항 유형의 점수가 가장 낮게 나타났다.

• Journal of Korean Philosophical Society
• /
• v.143
• /
• pp.189-214
• /
• 2017

Disgust is caused by human being incompleteness and also its denial. Therefore the understanding of disgust is concerned with the understanding of humans. Thinking critically about disgust, we can solve the problems of disgust and reflect on our limitations. I will refer to Nussbaum's view which finds out in "projective disgust" an antipathy to human fault and a wish for completeness. This article is interested in the fact that disgust is not only a rejective feeling of a person but also a collective emotion which is connected with an antagonistic relationship and power exercise. Specifically, this article focuses on the point that disgust is mobilized to maintain a dominative relationship between humans. This is associated with the inner characteristics of disgust because disgust in itself contains a one-sided perspective, exclusiveness, hierarchy, and domination. This article aims to reveal a collusion between disgust and domination. For the purpose we will pay attention to two basic inclinations which are immanent in disgust; purity and exclusion, while relying on Nussbaum's view of disgust. In accordance with this analysis, this article will specificity treat disgust which can be summarized as an ideological function of emotion and its violent tendency. Then, we can ensure that disgust threats the equal and dignitary worth of human beings and hinders the diversity and rationality of a liberal society. In addition, it will be emphasized that disgust should not be used as an ideology which discriminates and suppresses a specific group in a society.

• The Journal of the Korea Contents Association
• /
• v.20 no.2
• /
• pp.499-510
• /
• 2020

This study aims to analyze the causes of Internet hate speech, which has recently been emerging as a serious social problem and to seek for countermeasures. The experiences of hate speech are examined through the analysis of college students' essays and the causes and solutions of hate speech are suggested through the in-depth interviews with the experts. College students experience hate speech on the Internet on the basis of attributes such as age, gender, sexual orientation, and regionalism. Online comments on news, social media and online games are the main sources in spreading hate speech. On a personal level the lack of awareness of human dignity and the absence of media education are diagnosed as the reasons for online hate speech. The social reasons for online hate speech lie in the lack of human rights education and the problems of the media. In order to improve the problems of Internet hate speech, various suggestions are proposed on the legal, social and educational levels.

• Journal of the Korean Society for information Management
• /
• v.39 no.1
• /
• pp.331-350
• /
• 2022

This study aims to understand topics of incivility related to COVID-19 from analyzing Twitter posts including COVID-19-related hate speech. To achieve the goal, a total of 63,802 tweets that were created between December 1st, 2019, and August 31st, 2021, covering three targets of hate speech including region and public facilities, groups of people, and religion were analyzed. Frequency analysis, dynamic topic modeling, and keyword co-occurrence network analysis were used to explore topics and keywords. 1) Results of frequency analysis revealed that hate against regions and public facilities showed a relatively increasing trend while hate against specific groups of people and religion showed a relatively decreasing trend. 2) Results of dynamic topic modeling analysis showed keywords of each of the three targets of hate speech. Keywords of the region and public facilities included "Daegu, Gyeongbuk local hate", "interregional hate", and "public facility hate"; groups of people included "China hate", "virus spreaders", and "outdoor activity sanctions"; and religion included "Shincheonji", "Christianity", "religious infection", "refusal of quarantine", and "places visited by confirmed cases". 3) Similarly, results of keyword co-occurrence network analysis revealed keywords of three targets: region and public facilities (Corona, Daegu, confirmed cases, Shincheonji, Gyeongbuk, region); specific groups of people (Coronavirus, Wuhan pneumonia, Wuhan, China, Chinese, People, Entry, Banned); and religion (Corona, Church, Daegu, confirmed cases, infection). This study attempted to grasp the public's anti-citizenship public opinion related to COVID-19 by identifying domestic COVID-19 hate targets and keywords using social media. In particular, it is meaningful to grasp public opinion on incivility topics and hate emotions expressed on social media using data mining techniques for hate-related to COVID-19, which has not been attempted in previous studies. In addition, the results of this study suggest practical implications in that they can be based on basic data for contributing to the establishment of systems and policies for cultural communication measures in preparation for the post-COVID-19 era.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.05a
• /
• pp.621-623
• /
• 2022

최근 코로나19로 인해 비대면으로 소통할 수 있는 플랫폼에 대한 관심이 증가하고 있으며, 가상 세계의 개념을 도입한 메타버스 플랫폼이 MZ세대의 새로운 SNS로 떠오르고 있다. 아바타를 통해 상호 교류가 가능한 메타버스는 텍스트 기반의 소통뿐만 아니라 음성과 동작 시선 등을 활용하여 변화된 의사소통 방식을 사용한다. 음성을 활용한 소통이 증가함에 따라 다른 이용자에게 불쾌감을 주는 혐오 발언에 대한 신고가 증가하고 있다. 그러나 기존 혐오 발언 탐지 시스템은 텍스트를 기반으로 하여 사전에 정의된 혐오 키워드만 특수문자로 대체하는 방식을 사용하기 때문에 음성 혐오 발언에 대해서는 탐지하지 못한다. 이에 본 논문에서는 인공지능을 활용한 음성 혐오 표현 탐지 시스템을 제안한다. 제안하는 시스템은 음성 데이터의 파형을 통해 은유적 혐오 표현과 혐오 발언에 대한 감정적 특징을 추출하고 음성 데이터를 텍스트 데이터로 변환하여 혐오 문장을 탐지한 결과와 결합한다. 향후, 제안하는 시스템의 현실적인 검증을 위해 시스템 구축을 통한 성능평가가 필요하다.

• Journal of the Korean Society for Library and Information Science
• /
• v.56 no.4
• /
• pp.431-453
• /
• 2022

Hate is a collective expression of exclusivity toward others and it is fostered and reproduced through false public perception. This study aims to explore the objects and issues of hate discussed in our society using text mining techniques. To this end, we collected 17,867 news data published from 1990 to 2020 and constructed a co-word network and cluster analysis. In order to derive an explicit co-word network highly related to hate, we carried out sentence split and extracted a total of 52,520 sentences containing the words 'hate', 'prejudice' and 'discrimination' in the preprocessing phase. As a result of analyzing the frequency of words in the collected news data, the subjects that appeared most frequently in relation to hate in our society were women, race, and sexual minorities, and the related issues were related laws and crimes. As a result of cluster analysis based on the co-word network, we found a total of six hate-related clusters. The largest cluster was 'genderphobic', accounting for 41.4% of the total, followed by 'sexual minority hatred' at 28.7%, 'racial hatred' at 15.1%, 'selective hatred' at 8.5%, 'political hatred' accounted for 5.7% and 'environmental hatred' accounted for 0.3%. In the discussion, we comprehensively extracted all specific hate target names from the collected news data, which were not specifically revealed as a result of the cluster analysis.

• 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.