SexWEs: Domain-Aware Word Embeddings via Cross-Lingual Semantic Specialisation for Chinese Sexism Detection in Social Media

Jiang, Aiqi; Zubiaga, Arkaitz

doi:10.1609/icwsm.v17i1.22159

Cited by 2 publications

(1 citation statement)

References 44 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Even though the current literature on hate speech detection, diffusion, and interventions is increasingly trying to tackle the problem [14], [23], an important factor of online abusive events has remained unexplored: the characteristics of online hate targets. Despite the volume and diversity of the existing datasets [24], there is a dearth of research providing a holistic view of the abusive events, which significantly inhibits its investigation and our work aims to progress on.…”

Section: Introductionmentioning

confidence: 99%

Target-Oriented Investigation of Online Abusive Attacks: A Dataset and Analysis

Alharthi

Shekhar

et al. 2023

IEEE Access

Self Cite

View full text Add to dashboard Cite

Despite a body of research revolving around online abusive language, aiming at different objectives such as detection, diffusion prediction, and mitigation, existing research has seldom looked at factors motivating this behaviour. To further research in this direction, we investigate the motivations behind online abuse by looking at the characteristics of the targets of such abuse, i.e. is the abuse more prominent for specific characteristics of the targets? To enable target-oriented research into online abuse, we introduce the Online Abusive Attacks (OAA) dataset, the first benchmark dataset providing a holistic view of online abusive attacks, including social media profile data and metadata for both targets and perpetrators, in addition to context. The dataset contains 2.3K Twitter accounts, 5M tweets, and 106.9K categorised conversations. Further, we conduct an in-depth statistical analysis of online abuse centred around the targets' characteristics. We identify two types of abusive attacks: those motivated by characteristics of the targets (identity-based attacks) and others (behavioural attacks). We find that online abusive attacks are predominantly motivated by the targets' identities (97%), behavioural attacks accounting for a much smaller proportion (3%). Abuse is also more likely to target users who are popular and have a verified status. Interestingly, an analysis of the user bios shows no clear indication that keywords used in the bios are likely to trigger abuse. Additionally, we also look at the frequency with which perpetrators perform online abusive attacks. Our analysis shows a large number of infrequent perpetrators, with only a few recurrent perpetrators. Findings from our study have important implications for the development of abusive language detection models that incorporate an awareness of the targets to improve their potential for prediction.

show abstract

Section: Introductionmentioning

confidence: 99%

Target-Oriented Investigation of Online Abusive Attacks: A Dataset and Analysis

Alharthi

Shekhar

et al. 2023

IEEE Access

Self Cite

View full text Add to dashboard Cite

show abstract

Learning like human annotators: Cyberbullying detection in lengthy social media sessions

Zubiaga

2023

Proceedings of the ACM Web Conference 2023

Self Cite

View full text Add to dashboard Cite

The inherent characteristic of cyberbullying of being a recurrent attitude calls for the investigation of the problem by looking at social media sessions as a whole, beyond just isolated social media posts. However, the lengthy nature of social media sessions challenges the applicability and performance of session-based cyberbullying detection models. This is especially true when one aims to use state-of-the-art Transformer-based pre-trained language models, which only take inputs of a limited length. In this paper, we address this limitation of transformer models by proposing a conceptually intuitive framework called LS-CB, which enables cyberbullying detection from lengthy social media sessions. LS-CB relies on the intuition that we can efectively aggregate the predictions made by transformer models on smaller sliding windows extracted from lengthy social media sessions, leading to an overall improved performance. Our extensive experiments with six transformer models on two session-based datasets show that LS-CB consistently outperforms three types of competitive baselines including state-of-the-art cyberbullying detection models. In addition, we conduct a set of qualitative analyses to validate the hypotheses that cyberbullying incidents can be detected through aggregated analysis of smaller chunks derived from lengthy social media sessions (H1), and that cyberbullying incidents can occur at diferent points of the session (H2), hence positing that frequently used text truncation strategies are suboptimal compared to relying on holistic views of sessions. Our research in turn opens an avenue for fne-grained cyberbullying detection within sessions in future work. CCS CONCEPTS• Artifcial intelligence → Natural language processing; • Humancentered computing → Social media.

show abstract

SexWEs: Domain-Aware Word Embeddings via Cross-Lingual Semantic Specialisation for Chinese Sexism Detection in Social Media

Cited by 2 publications

References 44 publications

Target-Oriented Investigation of Online Abusive Attacks: A Dataset and Analysis

Target-Oriented Investigation of Online Abusive Attacks: A Dataset and Analysis

Learning like human annotators: Cyberbullying detection in lengthy social media sessions

Contact Info

Product

Resources

About