TL;DR
Counterhate tweets, intended to combat hate speech, can sometimes backfire and escalate conflict. This paper introduces a novel dataset and analysis of replies to counterhate tweets, revealing that nearly half of these replies disagree and a significant portion actually support the original hateful tweet. The research also demonstrates that it's possible to predict the type of reply a counterhate tweet will elicit based on its linguistic features.
What Problem Does This Paper Address?
The internet, particularly social media, is rife with hate speech. While counterhate arguments are a promising strategy to combat this without resorting to censorship, they can inadvertently worsen the situation. Some counterhate tweets might provoke aggressive responses or even bolster the original hateful message, leading to more toxic discourse. Existing research often treats counterhate as a simple binary (present/absent) or focuses on whether a reply agrees or disagrees. This paper aims to go deeper, analyzing the nature of replies to counterhate tweets with a fine-grained approach, and exploring whether the characteristics of counterhate tweets themselves can predict the type of reply they will receive.
Methods
The authors collected a dataset of over 2,600 "triples," each consisting of a hateful tweet, a counterhate tweet responding to it, and all replies to that counterhate tweet. These replies were then meticulously annotated with four fine-grained characteristics:
- Agree?: Does the reply agree with the counterhate tweet?
- Support?: If the reply disagrees, does it support the original hateful tweet?
- Attacks Author?: If the reply disagrees, does it attack the author of the counterhate tweet?
- Addtl. Counterhate?: If the reply agrees, does it add new counterhate arguments?
They then performed a linguistic analysis of the counterhate tweets, correlating various features (token count, sentiment, profanity, etc.) with the types of replies elicited. Finally, they built binary classifiers to predict these four reply characteristics using BERTweet, a transformer model pre-trained on Twitter data. They experimented with different input combinations (hateful tweet, counterhate tweet, reply) and employed data augmentation and blending with related datasets to improve performance.
Datasets
The primary dataset introduced in this paper consists of 2,621 (hateful tweet, counterhate tweet, reply) triples. This dataset was collected using a multi-stage process involving existing hate speech corpora and Twitter API calls, followed by manual validation and annotation. The authors also leveraged existing datasets for:
- Stance detection: To help determine if a reply agrees with the counterhate tweet.
- Offensive and hateful tweet-reply pair datasets: To assist in classifying whether a reply supports the hateful tweet, attacks the counterhate author, or adds additional counterhate.
Key Insights and Ideas
- Counterhate isn't always effective: A significant portion of replies (51%) disagree with the counterhate tweet, and a substantial number of these (40%) actually support the original hateful tweet. This highlights the potential for counterhate to backfire.
- Replies are complex: Beyond simple agreement/disagreement, replies can be nuanced, either supporting the original hate or attacking the counterhate author when disagreeing, or adding further arguments when agreeing.
- Linguistic cues matter: The characteristics of the counterhate tweet itself (e.g., length, use of question marks, sentiment) can predict the type of reply it will receive. This opens the door for more targeted counterhate strategies.
- Predicting reply types is feasible: The research demonstrates that machine learning models can effectively predict the fine-grained characteristics of replies to counterhate tweets.
Key Results
- Label Distribution: Approximately half of the replies (51%) disagree with the counterhate tweet. Of those that disagree, 40% support the hateful tweet, and 24% attack the counterhate author. Of the replies that agree, 35% add additional counterhate arguments.
- Linguistic Analysis: Counterhate tweets with more tokens or question marks tend to elicit replies that don't attack the author or add additional counterhate. Positive words in counterhate tweets correlate with replies that don't attack the author or add counterhate. Profanity in counterhate tweets is associated with replies that agree or don't support the hateful tweet.
- Model Performance: The best performing models, using a combination of reply, counterhate tweet, and hateful tweet as input, achieved F1-scores around 0.80-0.85 for most tasks, with the "support hateful tweet" task showing slightly lower performance. Blending complementary corpora significantly improved results.
- Error Analysis: The models struggled most with intricate text (sarcasm, implicit meaning), general knowledge, short texts, misspellings/abbreviations, and rhetorical questions.
Limitations and Future Work
The authors acknowledge several limitations:
- Data Availability: The data collection process was affected by tweets being unavailable and not all counterhate tweets having replies, leading to fewer triples than initially hoped.
- Classifier Accuracy: The use of classifiers to filter hateful/counterhate tweets might have incorrectly discarded or included some tweets.
- Text-Only Analysis: The study only considered tweet text, excluding images or multimedia content, which could provide additional context.
Future work suggested includes:
- Generating Customized Counterhate: Developing counterhate arguments that are tailored to specific contexts and characteristics found to be more effective at stopping hate.
- Causal Claims: Moving beyond correlation to investigate causal relationships between counterhate characteristics and reply types.
- Incorporating Multimedia: Analyzing tweets that include images or videos to gain a more comprehensive understanding.
Summary updated