Toxic Comment Detection in Online Discussions

Risch, Julian; Krestel, Ralf

doi:10.1007/978-981-15-1216-2_4

Cited by 41 publications

(18 citation statements)

References 42 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Two independent study team members coded the comments to the original posts, in terms of 7 coding categories: (1) tone of the comment (positive, negative, or neutral), (2) nature of contribution (toxic, healthy, or unclear/not applicable), (3) agreement with prevention message (agree, disagree, or seek clarification or advice), (4) mentions of government agency (yes or no), (5) policy/regulation (proregulation, antiregulation, neutral-regulation, or not applicable), (6) promotion/spam (yes or no), and (7) format (text only, meme/sticker/emoji/emoticon only, or both). Toxic contributions were defined as “a rude, disrespectful, or unreasonable comment that is likely to make other users leave a discussion” [ 27 ], whereas healthy contributions were defined as those using non-toxic language, those that were unclear or for which the classification was not applicable, or those using vague terms or emojis/stickers/emoticons, for which toxicity could not be determined.…”

Section: Methodsmentioning

confidence: 99%

Public Response to a Social Media Tobacco Prevention Campaign: Content Analysis

Majmundar¹,

Le²,

Moran³

et al. 2020

JMIR Public Health Surveill

View full text Add to dashboard Cite

Background Prior research suggests that social media–based public health campaigns are often targeted by countercampaigns. Objective Using reactance theory as the theoretical framework, this research characterizes the nature of public response to tobacco prevention messages disseminated via a social media–based campaign. We also examine whether agreement with the prevention messages is associated with comment tone and nature of the contribution to the overall discussion. Methods User comments to tobacco prevention messages, posted between April 19, 2017 and July 12, 2017, were extracted from Twitter, Facebook, and Instagram. Two coders categorized comments in terms of tone, agreement with message, nature of contribution, mentions of government agency and regulation, promotional or spam comments, and format of comment. Chi-square analyses tested associations between agreement with the message and tone of the public response and the nature of contributions to the discussions. Results Of the 1242 comments received (Twitter: n=1004; Facebook: n=176; Instagram: n=62), many comments used a negative tone (42.75%) and disagreed with the health messages (39.77%), while the majority made healthy contributions to the discussions (84.38%). Only 0.56% of messages mentioned government agencies, and only 0.48% of the comments were antiregulation. Comments employing a positive tone (84.13%) or making healthy contributions (69.11%) were more likely to agree with the campaign messages (P=0.01). Comments employing a negative tone (71.25%) or making toxic contributions (36.26%) generally disagreed with the messages (P=0.01). Conclusions The majority of user comments in response to a tobacco prevention campaign made healthy contributions. Our findings encourage the use of social media to promote dialogue about controversial health topics such as smoking. However, toxicity was characteristic of comments that disagreed with the health messages. Managing negative and toxic comments on social media is a crucial issue for social media–based tobacco prevention campaigns to consider.

show abstract

Section: Methodsmentioning

confidence: 99%

Public Response to a Social Media Tobacco Prevention Campaign: Content Analysis

Majmundar¹,

Le²,

Moran³

et al. 2020

JMIR Public Health Surveill

View full text Add to dashboard Cite

show abstract

“…Word2Vec (Mikolov et al, 2013) and GloVe fail to find a good representation of these words because words never occurred in training time. These words are out-of-vocabulary (OOV) (Risch and Krestel, 2020). However, we can take advantage of failing representations to represent toxic tokens that are not in vocabulary by setting their representations to zero.…”

Section: Multi-embedding Layermentioning

confidence: 99%

UoT-UWF-PartAI at SemEval-2021 Task 5: Self Attention Based Bi-GRU with Multi-Embedding Representation for Toxicity Highlighter

Giglou

Rahgooy

Rahgouy³

et al. 2021

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021)

View full text Add to dashboard Cite

Toxic Spans Detection(TSD) task is defined as highlighting spans that make a text toxic. Many works have been done to classify a given comment or document as toxic or non-toxic. However, none of those proposed models work at the token level. In this paper, we propose a self-attention-based bidirectional gated recurrent unit(BiGRU) with a multi-embedding representation of the tokens. Our proposed model enriches the representation by a combination of GPT-2, GloVe, and RoBERTa embeddings, which led to promising results. Experimental results show that our proposed approach is very effective in detecting span tokens.

show abstract

“…Dalam komentar pada sosial media, seringkali ditemukan sekelompok orang jahat dimana ia menghalangi diskusi yang saling menghormati dengan komentarnya yang beracun (toxic comment) yang didominasi oleh remaja. Komentar beracun (toxic comment) didefinisikan sebagai komentar yang kasar, tidak sopan, tidak masuk akal, atau bahkan sampai mempermalukan seseorang di media sosial yang cenderung membuat pengguna lain merasa tidak nyaman (Risch & Krestel, 2020). Dalam pendeteksian komentar dapat menggunakan pendekatan machine learning yaitu analisis sentimen yang sangat diperlukan dalam menyaring komentar di media sosial.…”

Section: Pendahuluanunclassified

Sentimen Analisis Komentar Toxic pada Grup Facebook Game Online Menggunakan Klasifikasi Naïve Bayes

Sidiq

Dermawan

Umaidah

2020

JIUP

View full text Add to dashboard Cite

Toxic comments are comments made by social media users that contain expressions of hatred, condescension, threatening, and insulting. Social media users who are on average still teenagers with a nature that still cannot be controlled completely becomes a matter of great concern when they comment, their comments can be studied as text processing. Sentiment analysis can be used as a solution to identifying toxic comments by dividing them into two classifications. Where the data used amounted to 1,500 taken from social media Facebook in the private group Arena of Valor community. The dataset is divided into 2 classes: toxic and non-toxic. This research uses Naive Bayes with TF-IDF transformation and Information Gain feature selection and use distribution ratio 80:20. It will be compared the results of the evaluation where Naive Bayes without transformation, using TF-IDF transformation, and TF-IDF using Information Gain feature selection. The results of the comparison of evaluations from confusion matrix that have been carried out obtained the best classification model is to use the ratio of training and testing data 80:20 with TF-IDF transformation resulting in an accuracy of 75%, precision of 63%, recall of 67%, and F-measure of 64%.

show abstract

Toxic Comment Detection in Online Discussions

Cited by 41 publications

References 42 publications

Public Response to a Social Media Tobacco Prevention Campaign: Content Analysis

Public Response to a Social Media Tobacco Prevention Campaign: Content Analysis

UoT-UWF-PartAI at SemEval-2021 Task 5: Self Attention Based Bi-GRU with Multi-Embedding Representation for Toxicity Highlighter

Sentimen Analisis Komentar Toxic pada Grup Facebook Game Online Menggunakan Klasifikasi Naïve Bayes

Contact Info

Product

Resources

About