Abstract:With the impressive performance in various downstream tasks, large language models (LLMs) have been widely integrated into production pipelines, like recruitment and recommendation systems. A known issue of models trained on natural language data is the presence of human biases, which can impact the fairness of the system. This paper investigates LLMs' behavior with respect to gender stereotypes, in the context of occupation decision making. Our framework is designed to investigate and quantify the presence of gender stereotypes in LLMs' behavior via multi-round question answering. Inspired by prior works, we construct a dataset by leveraging a standard occupation classification knowledge base released by authoritative agencies. We tested three LLMs (RoBERTa-large, GPT-3.5-turbo, and Llama2-70b-chat) and found that all models exhibit gender stereotypes analogous to human biases, but with different preferences. The distinct preferences of GPT-3.5-turbo and Llama2-70b-chat may imply the current alignment methods are insufficient for debiasing and could introduce new biases contradicting the traditional gender stereotypes.
Abstract:Language models have demonstrated strong performance on various natural language understanding tasks. Similar to humans, language models could also have their own bias that is learned from the training data. As more and more downstream tasks integrate language models as part of the pipeline, it is necessary to understand the internal stereotypical representation and the methods to mitigate the negative effects. In this paper, we proposed a simple method to test the internal stereotypical representation in pre-trained language models using counterexamples. We mainly focused on gender bias, but the method can be extended to other types of bias. We evaluated models on 9 different cloze-style prompts consisting of knowledge and base prompts. Our results indicate that pre-trained language models show a certain amount of robustness when using unrelated knowledge, and prefer shallow linguistic cues, such as word position and syntactic structure, to alter the internal stereotypical representation. Such findings shed light on how to manipulate language models in a neutral approach for both finetuning and evaluation.