skip to main content

Title: Abusive Language Detection in Heterogeneous Contexts: Dataset Collection and the Role of Supervised Attention
Abusive language is a massive problem in online social platforms. Existing abusive language detection techniques are particularly ill-suited to comments containing heterogeneous abusive language patterns, i.e., both abusive and non-abusive parts. This is due in part to the lack of datasets that explicitly annotate heterogeneity in abusive language. We tackle this challenge by providing an annotated dataset of abusive language in over 11,000 comments from YouTube. We account for heterogeneity in this dataset by separately annotating both the comment as a whole and the individual sentences that comprise each comment. We then propose an algorithm that uses a supervised attention mechanism to detect and categorize abusive content using multi-task learning. We empirically demonstrate the challenges of using traditional techniques on heterogeneous content and the comparative gains in performance of the proposed approach over state-of-the-art methods.
; ; ; ; ;
Award ID(s):
Publication Date:
Journal Name:
Proceedings of the AAAI Conference on Artificial Intelligence
Page Range or eLocation-ID:
Sponsoring Org:
National Science Foundation
More Like this
  1. Budak, Ceren ; Cha, Meeyoung ; Quercia, Daniele ; Xie, Lexing (Ed.)
    We present the first large-scale measurement study of cross-partisan discussions between liberals and conservatives on YouTube, based on a dataset of 274,241 political videos from 973 channels of US partisan media and 134M comments from 9.3M users over eight months in 2020. Contrary to a simple narrative of echo chambers, we find a surprising amount of cross-talk: most users with at least 10 comments posted at least once on both left-leaning and right-leaning YouTube channels. Cross-talk, however, was not symmetric. Based on the user leaning predicted by a hierarchical attention model, we find that conservatives were much more likely to comment on left-leaning videos than liberals on right-leaning videos. Secondly, YouTube's comment sorting algorithm made cross-partisan comments modestly less visible; for example, comments from conservatives made up 26.3% of all comments on left-leaning videos but just over 20% of the comments were in the top 20 positions. Lastly, using Perspective API's toxicity score as a measure of quality, we find that conservatives were not significantly more toxic than liberals when users directly commented on the content of videos. However, when users replied to comments from other users, we find that cross-partisan replies were more toxic than co-partisan replies on bothmore »left-leaning and right-leaning videos, with cross-partisan replies being especially toxic on the replier's home turf.« less
  2. We licensed a dataset from a mental health peer support platform catering mainly to teens and young adults. We anonymized the name of this platform to protect the individuals on our dataset. On this platform, users can post content and comment on others’ posts. Interactions are semi-anonymous: users share a photo and screen name with others. They have the option to post with their username visible or anonymously. The platform is moderated, but the ratio of moderators to posters is low (0.00007). The original dataset included over 5 million posts and 15 million comments from 2011- 2017. It was scaled to a feasible size for qualitative analysis by running a query to identify posts by a) adolescents aged 13-17 that were seeking support for b) online sexual experiences (not offline) with people they know (not strangers).
  3. Ad hominem attacks are those that target some feature of a person’s character instead of the position the person is maintaining. These attacks are harmful because they propagate implicit biases and diminish a person’s credibility. Since dialogue systems respond directly to user input, it is important to study ad hominems in dialogue responses. To this end, we propose categories of ad hominems, compose an annotated dataset, and build a classifier to analyze human and dialogue system responses to English Twitter posts. We specifically compare responses to Twitter topics about marginalized communities (#BlackLivesMatter, #MeToo) versus other topics (#Vegan, #WFH), because the abusive language of ad hominems could further amplify the skew of power away from marginalized populations. Furthermore, we propose a constrained decoding technique that uses salient n-gram similarity as a soft constraint for top-k sampling to reduce the amount of ad hominems generated. Our results indicate that 1) responses from both humans and DialoGPT contain more ad hominems for discussions around marginalized communities, 2) different quantities of ad hominems in the training data can influence the likelihood of generating ad hominems, and 3) we can use constrained decoding techniques to reduce ad hominems in generated dialogue responses.
  4. Cyberbullying, identified as intended and repeated online bullying behavior, has become increasingly prevalent in the past few decades. Despite the significant progress made thus far, the focus of most existing work on cyberbullying detection lies in the independent content analysis of different comments within a social media session. We argue that such leading notions of analysis suffer from three key limitations: they overlook the temporal correlations among different comments; they only consider the content within a single comment rather than the topic coherence across comments; they remain generic and exploit limited interactions between social media users. In this work, we observe that user comments in the same session may be inherently related, e.g., discussing similar topics, and their interaction may evolve over time. We also show that modeling such topic coherence and temporal interaction are critical to capture the repetitive characteristics of bullying behavior, thus leading to better predicting performance. To achieve the goal, we first construct a unified temporal graph for each social media session. Drawing on recent advances in graph neural network, we then propose a principled graph-based approach for modeling the temporal dynamics and topic coherence throughout user interactions. We empirically evaluate the effectiveness of our approachmore »with the tasks of session-level bullying detection and comment-level case study. Our code is released to public.« less
  5. Prior work and perception theory suggests that when exposed to discussion related to a particular piece of crowdsourced text content, readers generally perceive that content to be of lower quality than readers who do not see those comments, and that the effect is stronger if the comments display conflict. This paper presents a controlled experiment with over 1000 participants testing to see if this effect carries over to other documents from the same platform, including those with similar content or by the same author. Although we do generally find that perceived quality of the commented-on document is affected, effects do not carry over to the second item and readers are able to judge the second in isolation from the comment on the first. We confirm a prior finding about the negative effects conflict can have on perceived quality but note that readers report learning more from constructive conflict comments.