[Paper Review] Evaluating ChatGPT's Performance for Multilingual and Emoji-based Hate Speech Detection

Mithun Das, Saurabh Kumar Pandey+1 more|arXiv (Cornell University)|May 22, 2023

Hate Speech and Cyberbullying Detection7 citations

TL;DR

The paper assesses ChatGPT’s strengths and weaknesses in detecting hate speech across 11 languages and in emoji-based expressions, using functional tests that reveal granular failures not captured by aggregate metrics.

ABSTRACT

Hate speech is a severe issue that affects many online platforms. So far, several studies have been performed to develop robust hate speech detection systems. Large language models like ChatGPT have recently shown a great promise in performing several tasks, including hate speech detection. However, it is crucial to comprehend the limitations of these models to build robust hate speech detection systems. To bridge this gap, our study aims to evaluate the strengths and weaknesses of the ChatGPT model in detecting hate speech at a granular level across 11 languages. Our evaluation employs a series of functionality tests that reveals various intricate failures of the model which the aggregate metrics like macro F1 or accuracy are not able to unfold. In addition, we investigate the influence of complex emotions, such as the use of emojis in hate speech, on the performance of the ChatGPT model. Our analysis highlights the shortcomings of the generative models in detecting certain types of hate speech and highlighting the need for further research and improvements in the workings of these models.

Motivation & Objective

• Motivate the need for robust hate speech detection systems in the era of powerful LLMs.
• Evaluate ChatGPT’s performance across 11 languages to identify strengths and weaknesses.
• Investigate the influence of emojis and other complex emotions on hate speech detection.
• Show that aggregate metrics may overlook nuanced failure modes in generative models.

Proposed method

• Perform a series of functionality tests to uncover granular failures of ChatGPT in hate speech detection.
• Evaluate ChatGPT across 11 languages to measure multilingual performance.
• Analyze the impact of emoji usage and complex emotions on model performance.
• Compare granular test outcomes to aggregate metrics such as macro F1 or accuracy.
• Highlight shortcomings of generative models in detecting certain hate speech types.

Experimental results

Research questions

• RQ1 How well does ChatGPT detect hate speech across 11 languages?
• RQ2 How does emoji usage influence ChatGPT’s hate speech detection performance?
• RQ3 What granular failures does ChatGPT exhibit that are not captured by macro metrics?
• RQ4 What limitations of generative models emerge in hate speech detection tasks?

Key findings

• ChatGPT exhibits granular failures in hate speech detection not evident from macro metrics.
• The model shows shortcomings in detecting certain types of hate speech.
• Emoji and complex emotional cues can affect performance, revealing limitations of current generative models.
• The study highlights the need for further research and improvements in hate speech detection with LLMs.