Gender Prejudice in Large Language Models from the Perspective of Pragmatics-Based on Critical Translation

This paper explores the complex interplay between gender bias in Large Language Models (LLMs) and human empirical perspectives, focusing on pragmatic and critical translation approaches. Considering this pressing issue, this work creates a database of 55 Chinese sentences without explicit gender pronounces and tests translation tasks on ChatGPT, Ernie Bot, and Spark Desk platforms. The study reveals that LLMs exhibit varying levels of gender bias, reflecting cultural differences observed in human translators. In the research, the Gender Stereotype Circle of Large Language Models and Human, a theoretical framework is designed to address these biases through six vital factors and demonstrate the interactions to aid in optimizing how LLMs cope with gender issues, enhancing the output quality, and providing an empirical foundation for the LLMs training.