- The author shares a personal anecdote about struggling with Australian English despite years of studying English.
- Large language models (LLMs) face similar challenges in detecting sentiment and sarcasm across different English varieties.
- A new tool called BESSTIE evaluates LLMs' ability to detect sentiment and sarcasm in Australian, Indian, and British English.
- LLMs perform better on native English varieties (Australian and British) than non-native ones (Indian English).
- Sarcasm detection is particularly challenging for LLMs, with accuracy rates as low as 57-62%.
- Performance claims by tech companies for LLMs are often inflated compared to real-world performance on non-American English.
- National context is crucial for improving LLM efficacy, as seen in projects targeting Aboriginal English and emergency department use.