Linguistic Biases in LLM Recommendations Studied
A research study examines the linguistic biases present in recommendations generated by LLMs for Southern American English, Indian English, and Code-Switched Hindi-English dialects. Utilizing the Yelp Open dataset (Yelp Inc., 2023) alongside Walmart product reviews (PromptCloud, 2020), the researchers incorporated balanced lists of restaurant and product names into their prompts. They employed zero-shot prompting in a cold-start scenario to identify the top 20 recommendations for each dialect. Aggregate response counts by cuisine type and product category were analyzed using mixed-effects regression models across 20 seeds.
Key facts
- Study investigates linguistic biases in LLM-based recommendations.
- Dialects tested: Southern American English, Indian English, Code-Switched Hindi-English.
- Datasets used: Yelp Open dataset (Yelp Inc., 2023) and Walmart product reviews (PromptCloud, 2020).
- LLMs zero-shot prompted to select top-20 recommendations from balanced lists.
- Mixed-effects regression models used for analysis across 20 seeds.
Entities
Institutions
- Yelp Inc.
- PromptCloud