Stereotypical views about social groups (e.g., Muslims, Germans, women, immigrants) play a pervasive role in how we perceive and interact with people. Stereotypes emerge from the way we communicate about categorized people and their behavior both privately and in the media. It is valuable to learn about the exact (linguistic) means through which such (negative) stereotypical views become shared within communities.
The present project aims to merge complementary approaches, and thereby develop a methodology for using NLP1 to automatically identify content and strength of stereotypes about specific groups shared within communities. Read more…

Abstract
The goal of our NIAA project was to develop annotation guidelines to identify different types of implicit bias in natural language. In order to develop these guidelines, we first created an online survey, asking participants to write about gender differences in a controlled context. Social psychological control questions were included to test for biased attitudes. This is one of the first studies to integrate all aspects of implicit bias together in one dataset, creating an important resource. Based on this dataset, we developed an innovative codebook. For instance, we developed a category label model that is the first attempt to classify different types of labels of social categories. We furthermore improved an existing model of abstraction in behavior descriptions. Most insights were gained during the development of these annotation guidelines. Ideally, these guidelines translate insights gained in social and communication psychology to clear, linguistic guidelines that could eventually be used to implement an automatic system. The discussions of these guidelines revealed fundamental differences in perspectives in both fields, which has led to important insights that might not have been discovered in a project with only one of the disciplines.
As the development of the guidelines required substantial discussions, we were only able to annotate a small portion of the collected data set. On the basis of this, two publications are planned:

• Analysis of dataset using annotation guidelines against tested bias (survey)
• A report on the insights gained in developing the guidelines.This project laid the groundworks for further research, such as applying the annotation guidelines different data containing different social categories and implementing an automatic system.