NLP measures for cyber environment safety: Detection, mitigation and intervention

Zhu, Wanzheng

Permalink

https://hdl.handle.net/2142/116198 Copy

Description

Title

NLP measures for cyber environment safety: Detection, mitigation and intervention

Author(s)

Zhu, Wanzheng

Issue Date

2022-07-10

Director of Research (if dissertation) or Advisor (if thesis)

Bhat, Suma

Doctoral Committee Chair(s)

Bhat, Suma

Committee Member(s)

• Han, Jiawei
• Hasegawa-Johnson, Mark
• Fanti, Giulia

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Content Moderation
• Hate Speech Mitigation
• Cyberbullying Intervention

Abstract

In the profound digital world, users have the liberty to express their views about almost any topic on social media platforms. Such content may contain sensitive or illegal textual messages (e.g., drug dealing, weapon trafficking, prostitution, hate speech, etc.), which are not favorable for many people viewing. Very importantly, these contents are typically visible to all, including children, teenagers, and vulnerable people. To ensure a safe and clean cyber environment, people have actively participated in manual content moderation and counterspeech generation. However, such manual effort is expensive, inhumane, and ineffective. This dissertation contributes to cyber environment safety in an automatic manner by three main stages: detection, mitigation, and intervention. The first stage aims at improving content moderation through better detection of non-standard language use (e.g., euphemism---an innocuous word or expression used in place of one that may be found offensive or suggest something unpleasant) on social media platforms. We focus on sensitive or illegal topics wildly spread (e.g., drug, weapon, sexuality), through (1) euphemism detection, and (2) euphemism identification---understanding what a euphemism refers to. The second stage involves developing a hate speech mitigation model by transferring the hate speech to a non-hateful one while preserving the content of the original text. Those users who plan to post an offensive message could be encouraged to have a change of mind and avoid the profanity, if one could not only alert that the content is offensive, but also offer a non-hateful version of the message that is acceptable. Lastly, the third stage focuses on cyberbullying intervention by developing a conversational Artificial Intelligence (AI) system that acts like a human bystander to generate counterspeech. Such a system creates messages that look appropriate in the context of a given comment thread and that contain psychologically valid bystander interventions, with the aim of mitigating the hate speech online, educating the hate speakers, and fostering a more harmonious conversation on social media platforms. By having accomplished the aforementioned three stages, this dissertation could have real impacts in reducing online aggression on social media platforms while reducing the need for (and possible harms to) human moderators.

Graduation Semester

2022-08

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Wanzheng Zhu

Owning Collections