Proceedings of the Annual Meeting of the Cognitive Science Society

Our implicit biases are often reflected in our utterances. Existing AI research works are limited by their narrow focus on predefined, explicit biases. Thus, we propose a more generalized approach to bias analysis by implementing Linguistic Intergroup Bias (LIB) theory, which suggests people tend to use abstract terms to describe in-group positive and out-group negative behaviors, and concrete terms for the opposite. To leverage LIB, we propose a novel task named language abstraction span extraction, and finetuned a large language model for this task. With the model, we analyze the intergroup bias between the political left and right on Twitter and news datasets. Results show that our method can be used for bias analysis both quantitatively and qualitatively. More importantly, our paper provides the first large-sample-based empirical evidence to support LIB, affirming the correlation between abstraction and intergroup bias in social media and news domains.