🤖 machine learning

Detecting LGBTQ+ Instances of Cyberbullying

This study evaluates the effectiveness of various transformer models in accurately detecting and distinguishing cyberbullying incidents specifically targeting the LGBTQ+ community using real social media data.

Original authors: Arslan Bisharat, Manuel Sandoval Madrigal, Mohammed Abuhamad, Deborah L. Hall, Yasin N. Silva

Published 2026-06-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Arslan Bisharat, Manuel Sandoval Madrigal, Mohammed Abuhamad, Deborah L. Hall, Yasin N. Silva

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine social media as a giant, bustling town square. In the past, if someone wanted to be mean, they had to stand right next to you to shout insults. Today, the "town square" has a keyboard, and people can throw digital rocks from miles away. This is cyberbullying.

While everyone can get hurt by these digital rocks, the paper explains that a specific group of people—the LGBTQ+ community—is getting hit much harder and more often than others. It's like they are standing in a target zone where the bullies are aiming specifically at their identity.

Here is what the researchers did, explained simply:

The Mission: Building a Digital Security Guard

The researchers wanted to build a "smart security guard" (a computer program) that could look at comments on Instagram and instantly say, "Hey, this is bullying, and it's specifically targeting LGBTQ+ people."

They knew that regular security guards often make mistakes. They might think a comment is just a joke when it's actually hate speech, or they might get confused by slang and hidden insults. So, the team tested three different "brains" (computer models) to see which one was the best guard:

RoBERTa
BERT
GPT-2

The Training Ground: A Small, Tricky Class

To teach these guards, the researchers used a specific set of 1,083 Instagram comments.

The Problem: Most of the comments were not about LGBTQ+ bullying. Only about 200 of them were. It's like trying to teach a dog to find a specific rare flower in a field full of dandelions. The dog might just learn to ignore the rare flower because there are so many dandelions.
The Fix: To help the guards learn, the researchers used a technique called "oversampling" (SMOTE and ADASYN). Imagine this as taking the few rare flower examples and making photocopies of them so the guards have more practice spotting them.

The Results: Who Was the Best Guard?

After training, they put the guards to the test. Here is what happened:

RoBERTa was the champion: It was the smartest guard. It got the highest overall score and was the best at spotting the bullying.
The others struggled: BERT and GPT-2 were good at spotting general bullying, but they got confused when the bullying was specifically about LGBTQ+ issues.

The Big Catch: The "Missed Spot" Problem

Even though RoBERTa was the best, it still had a major blind spot.

The Good News: It was excellent at saying, "This is not LGBTQ+ bullying." (It rarely cried wolf).
The Bad News: It was still bad at saying, "This is LGBTQ+ bullying." It missed a lot of the actual attacks.

Why did it miss them?
The paper explains that bullying against LGBTQ+ people is often like a secret code.

Sometimes it uses sarcasm.
Sometimes it uses slang that changes meaning depending on the situation.
Sometimes it hides the insult inside a sentence that looks innocent.

The computer models are like students who memorized the dictionary but don't understand the joke or the tone. They see the words but miss the sting. For example, the paper found that the model sometimes missed the word "dyke" because it didn't have enough examples of that specific slur in its training, or it missed a comment because it didn't understand the context of the conversation.

The Takeaway

The researchers concluded that while we have built some very smart "security guards" (AI models), they aren't perfect yet.

RoBERTa is the strongest guard we have right now.
Oversampling (making more practice examples) helped a little, but it didn't fix the problem completely.
The Challenge: The "bullying" is too subtle and tricky for the current technology to catch every time. The models are still missing too many of the attacks (False Negatives).

The paper suggests that to make these guards truly effective, we need to teach them with more diverse examples, show them pictures and videos (not just text), and help them understand the deeper context of human conversation. Until then, the digital town square still has blind spots where the LGBTQ+ community remains vulnerable.

Technical Summary: Detecting LGBTQ+ Instances of Cyberbullying

Problem Definition
Cyberbullying poses a significant threat to global adolescent mental health, with evidence linking exposure to peer aggression and increased suicidality. While general cyberbullying detection is an established field, there is a critical gap in developing models specifically sensitive to the LGBTQ+ community. Research indicates that LGBTQ+ individuals face disproportionately higher rates of online harassment compared to their heterosexual and cisgender peers. Furthermore, existing "one-size-fits-all" detection models often fail to capture the unique, subtle, and context-dependent nature of harassment targeting sexual and gender minorities. This study addresses the need for machine learning models capable of accurately discerning cyberbullying incidents specifically targeting LGBTQ+ individuals in real-time social media environments.

Methodology
The study implements and evaluates three pre-trained transformer language models—RoBERTa, BERT, and GPT-2—to classify Instagram comments as either containing LGBTQ+-related cyberbullying or not.

Dataset: The research utilizes a dataset of 1,083 Instagram comments, of which 217 are specifically annotated as targeting the LGBTQ+ community. The data was labeled by a diverse team with expertise in psychology and computer science, capturing nuances in content type, purpose, and directionality.
Model Architecture:
- RoBERTa and BERT: Adapted with sequence classification heads and trained using binary cross-entropy loss and the AdamW optimizer.
- GPT-2: Modified using GPT2ForSequenceClassification to add a linear layer on top of the pooled output, enabling sequence classification tasks despite its original design for text generation.
- Hyperparameters: Models were trained for 3 epochs with a batch size of 16 and a learning rate of $5 \times 10^{-5}$ .
Handling Class Imbalance: To address the imbalance between the minority class (LGBTQ+ bullying) and the majority class (Non-LGBTQ+ bullying), the authors experimented with three configurations: the original dataset, oversampling using SMOTE (Synthetic Minority Over-sampling Technique), and oversampling using ADASYN (Adaptive Synthetic Sampling).
Evaluation: Performance was assessed using 5-fold cross-validation, measuring Accuracy, Precision, Recall, F1 score, and Area Under the Receiver Operating Characteristic curve (AUROC).

Key Results
The experimental results indicate that RoBERTa consistently outperformed BERT and GPT-2 across most configurations.

Best Performance: RoBERTa achieved the highest accuracy (0.9456) and F1 score (0.733) on the original dataset without oversampling. The highest AUROC score (0.8926) was achieved by RoBERTa using ADASYN oversampling.
Class-Specific Disparities: A per-class analysis reveals a significant performance gap. While models demonstrated high precision and recall for identifying Non-LGBTQ+ bullying (True Negatives), they struggled significantly with LGBTQ+ bullying (True Positives).
- For RoBERTa (Original), the recall for the LGBTQ+ class was only 0.6372, compared to 0.986 for the Non-LGBTQ+ class.
- This trend persisted across all oversampling techniques, indicating that the models frequently misclassify LGBTQ+ bullying comments as Non-LGBTQ+ (False Negatives).
Oversampling Impact: While SMOTE and ADASYN generally increased the number of True Positives compared to the original dataset, they did not fully resolve the issue of low recall for the minority class.

Limitations and Challenges
The authors identify specific challenges in the models' performance:

Context and Nuance: The models struggle with implicit language, sarcasm, and coded expressions common in LGBTQ+ harassment. For instance, the model failed to identify the homophobic slur "dyke" in a false negative example, likely due to insufficient training data for specific slurs.
False Negatives: The high rate of False Negatives suggests that current models are biased toward classifying ambiguous comments as non-bullying, potentially missing severe harassment.
Dataset Constraints: The study is limited by the size and diversity of the dataset and its exclusive focus on Instagram data.

Significance and Future Directions
The paper claims that while state-of-the-art transformer models show promise in distinguishing between general and LGBTQ+-specific bullying, they are not yet robust enough to handle the complex, subtle nature of such harassment without significant false negatives. The study underscores the necessity for tailored online safety tools rather than generic detection systems.

The authors suggest that future work should focus on:

Integrating multi-modal data (images, videos, social network metrics) to enrich context.
Developing datasets that reflect a wider range of bullying scenarios.
Creating models that account for the sequence and timing of interactions.
Exploring fairness-aware training methodologies to enhance detection capabilities for the LGBTQ+ community.

In conclusion, the study demonstrates that while RoBERTa is the most effective model among those tested, current technology requires further refinement to accurately detect the nuanced and often implicit cyberbullying faced by LGBTQ+ individuals.