Source-linked AI summary
Is ChatGPT better than Human Annotators? Potential and Limitations of ChatGPT in Explaining Implicit Hate Speech
Fan Huang, Haewoon Kwak, Jisun An
TL;DR
Implicit hate speech is subtle, difficult to detect, and challenging to explain, motivating an empirical study of ChatGPT’s classification and natural-language explanations. Using LatentHatred, repeated prompted responses, and human evaluations against human-written explanations, the paper finds strong classification agreement and clearer ChatGPT explanations, while cautioning that convincing errors may mislead lay people.
Problem
Implicit hate speech is difficult to detect and explain, while subjective social tasks also have low human agreement and costly annotation demands.
Method
The study evaluates ChatGPT on LatentHatred using a concise prompt, repeated responses, and human comparisons of classification and generated explanations.
Results
ChatGPT correctly identifies 80% of implicit hateful tweets, and its generated explanations are perceived as clearer than human-written explanations.
Takeaways & Limitations
ChatGPT shows potential as a data annotation tool and for generating explanations, but its convincing responses require careful use when decisions are wrong.
Takeaways & Limitations
When ChatGPT is wrong, its explanations may reinforce human perceptions and potentially lead lay people toward incorrect labels.
Abstract
from arXiv · showhide
Recent studies have alarmed that many online hate speeches are implicit. With its subtle nature, the explainability of the detection of such hateful speech has been a challenging problem. In this work, we examine whether ChatGPT can be used for providing natural language explanations (NLEs) for implicit hateful speech detection. We design our prompt to elicit concise ChatGPT-generated NLEs and conduct user studies to evaluate their qualities by comparison with human-written NLEs. We discuss the potential and limitations of ChatGPT in the context of implicit hateful speech research.
1 INTRODUCTION
The paper investigates ChatGPT for subjective social tasks through implicit hate speech detection and explanation, addressing challenges in human agreement, annotation cost, detection, and explainability.
- Subjective social tasks require social judgment and decision-making, yet are difficult even for humans and costly to annotate at scale.
- Implicit hate speech is difficult to detect because of its subtlety, and explaining model decisions with natural language explanations remains challenging.
- The study evaluates ChatGPT on two tasks: classifying implicit hate speech and generating explanations for implicit hate speech.
- Using LatentHatred, the authors compare ChatGPT responses with human-written data through human evaluations to identify its potential and limitations.
2 RELATED WORK
Prior work has used generative language models for hate-speech explanations and evaluated ChatGPT across several domains; this paper applies ChatGPT to concise explanations for implicit hateful speech.
- Explicit hate speech is easier to detect than implicit hate speech, whose nuanced and context-dependent nature complicates explanation generation.
- Earlier studies applied GPT-2, BART, OPT, and T5 to generate natural language explanations for hateful text.
- Existing ChatGPT research examined question answering, medical report simplification, and bug fixing across varied evaluation settings.
- This work extends that line of research to concise natural language explanations for implicit hateful speech, where toxicity boundaries are unclear.
3 CHATGPT-BASED EXPLANATIONS
The study uses LatentHatred tweets, a concise prompt, repeated ChatGPT responses, aggregation rules, and human evaluations to assess classification and explanation quality against human-written explanations.
- 3.1 Case Study: LatentHatred Dataset: The LatentHatred dataset contains 6,358 implicit hateful tweets with human annotations of target groups and implied statements; the study samples 795 instances.
- 3.2 ChatGPT Response Collection: ChatGPT receives each tweet and is prompted to answer yes or no about implicit hatefulness and explain its decision in one sentence.
- 3.2 ChatGPT Response Collection: The authors collect three responses per tweet in separate chat sessions to avoid influence from conversation history.
- 3.2 ChatGPT Response Collection: Responses are converted to scores of +1, -1, or 0 and aggregated into Hateful, Non-Hateful, or Uncertain labels.
- 3.3 Human Evaluation: Human evaluations use Mechanical Turk annotations and compare posts alone, posts with human-written explanations, and posts with ChatGPT-generated explanations.
- 3.3 Human Evaluation: Explanation quality is evaluated with Informativeness and Clarity ratings on randomly sampled tweet-explanation pairs, with reviewed annotations for disputed cases.
4 RESULT
ChatGPT identified 80% of implicit hateful tweets, and human evaluations examined disagreement cases and explanation quality against human-written NLEs. Its generated explanations were clearer than human-written ones, while informativeness did not differ significantly.
- ChatGPT classified 636 of 795 instances as implicitly hateful (80%), with 146 non-hateful (18.4%) and 13 uncertain (1.6%).
- In disagreement cases, laypeople rated posts with ChatGPT explanations less hateful than posts alone, with means of -0.52 and -0.41, respectively.The difference was statistically significant: t(316) = -13.75, p<0.001.
- Posts paired with human-written explanations received a positive average hatefulness score of 0.29, lower in absolute value than the other conditions.The authors interpret this as human-written NLEs being less convincing than the post-only and ChatGPT-NLE conditions.
- ChatGPT-generated NLEs had higher Clarity scores than human-written NLEs, 5.39 versus 4.68, in evaluations of 100 tweets.The difference was statistically significant: t(197) = -3.73, p < 0.001.
- Informativeness did not differ significantly between ChatGPT-generated and human-written NLEs (t(198) = -0.69, p = 0.49).
- The authors conclude that ChatGPT generates quality NLEs comparable to those of human annotators for implicit hate speech.
5 DISCUSSIONS AND CONCLUSION
The discussion presents ChatGPT as promising for implicit hate-speech classification and explanation, while warning that convincing errors could mislead laypeople and produce incorrect labels.
- ChatGPT correctly identified 80% of implicit hateful tweets and its generated NLEs were perceived as clearer than human-written NLEs.
- For the 20% disagreement cases, ChatGPT’s classifications were more likely to align with laypeople’s perceptions.
- ChatGPT-generated explanations can reinforce human perception, creating a risk of misleading laypeople when ChatGPT’s classification is wrong.The authors therefore call for extra caution when using ChatGPT to assist data annotation.
6 ETHICAL CONSIDERATIONS
The study describes ethical safeguards for Mechanical Turk data collection, including warnings about possible hatefulness, permission to stop, and access to consultant hotlines.
- The Mechanical Turk data-collection design received approval from Singapore Management University (Approval No.: IRB-22-076-A043(622)).
- Task instructions highlighted possible hatefulness, allowed annotators to stop if uncomfortable, and provided consultant hotlines.