View in EDS

When comments aren't what they seem: The social media comment toxicity detector for understanding contextual comments.

Saved in:
Bibliographic Details
Title: When comments aren't what they seem: The social media comment toxicity detector for understanding contextual comments.
Authors: Song, Zichen1 (AUTHOR) songzch21@lzu.edu.cn, Fan, Xiaopeng1 (AUTHOR) fanxp2023@lzu.edu.cn, Wang, Yutong1 (AUTHOR) yutongwang2023@lzu.edu.cn, Yan, Feixuan1 (AUTHOR) yanfx2023@lzu.edu.cn, Wu, Zijin1 (AUTHOR) wuzj2023@lzu.edu.cn, Kang, Zhongfeng1 (AUTHOR) kangzf@lzu.edu.cn
Source: Expert Systems with Applications. Apr2026, Vol. 305, pN.PAG-N.PAG. 1p.
Subjects: Context-aware computing, Generative pre-trained transformers, Machine learning, Sentiment analysis, Electronic data processing
Abstract: Social media comments often contain toxic, sarcastic, or ambiguous language, challenging traditional toxicity detection models. However, most existing methods treat comments as isolated units and rely solely on surface-level textual cues, making them ineffective in identifying implicit toxicity such as sarcasm or context-dependent insults. These limitations hinder their performance in real-world scenarios, where toxicity often emerges across multi-turn interactions. This paper proposes a context-aware interactive framework using GPT models to simulate human-like interactions, transforming single comments into multi-level threads. This approach improves toxicity detection by leveraging contextual dialogue. We evaluate our method on 5018 comments from platforms such as Zhihu, Bilibili, Weibo, YouTube, and RedNote. Our experiments show that our method outperforms traditional models, especially on YouTube, where sensitive word frequency is high. Additionally, our approach demonstrates strong multilingual support, successfully detecting toxicity in six languages. This paper also compares the performance of GPT-4O, GPT-4, GPT-3.5 and DeepSeek-V3, highlighting the advantage of multimodal processing in the former for more robust detection. The dataset we have collected is publicly available at https://github.com/kanglzu/When-Comments-Aren-t-What-They-Seem. [ABSTRACT FROM AUTHOR]
Copyright of Expert Systems with Applications is the property of Pergamon Press - An Imprint of Elsevier Science and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:Social media comments often contain toxic, sarcastic, or ambiguous language, challenging traditional toxicity detection models. However, most existing methods treat comments as isolated units and rely solely on surface-level textual cues, making them ineffective in identifying implicit toxicity such as sarcasm or context-dependent insults. These limitations hinder their performance in real-world scenarios, where toxicity often emerges across multi-turn interactions. This paper proposes a context-aware interactive framework using GPT models to simulate human-like interactions, transforming single comments into multi-level threads. This approach improves toxicity detection by leveraging contextual dialogue. We evaluate our method on 5018 comments from platforms such as Zhihu, Bilibili, Weibo, YouTube, and RedNote. Our experiments show that our method outperforms traditional models, especially on YouTube, where sensitive word frequency is high. Additionally, our approach demonstrates strong multilingual support, successfully detecting toxicity in six languages. This paper also compares the performance of GPT-4O, GPT-4, GPT-3.5 and DeepSeek-V3, highlighting the advantage of multimodal processing in the former for more robust detection. The dataset we have collected is publicly available at https://github.com/kanglzu/When-Comments-Aren-t-What-They-Seem. [ABSTRACT FROM AUTHOR]
ISSN:09574174
DOI:10.1016/j.eswa.2025.130902