A Neural Network and Embedding Association-Based Approach to Nationalism Dictionary Creation

Studies on nationalism often suffer arbitrariness in measuring core dimensions with widely varying definitions. In this study, we propose a semi-automated approach that inductively creates a dictionary of keywords and quantifies them with varying scales of exclusiveness and inclusiveness. We do so by adopting a neural network feature engineering that can balance the trade-off between representativeness and manual labor cost carefully, creating a nationalistic dictionary from a corpus consisting of 167,174 Weibo posts of a famous nationalistic account in China. We also introduce a Multilevel Embedding Association Test (MLEAT) to compare the performance of the dictionary in this study with two established manual-coding dictionaries in capturing semantic associations between two prominent social groups, state institutions, and the private sectors, and a pair of rhetoric distinctiveness, exclusive and inclusive nationalistic rhetorics in Chinese cyberspace. The results suggest our semi-automated dictionary can efficiently extract key information from large-scale corpora, a task prohibitively expensive for manual methods, while also providing the transparency and interpretability lacking in previous studies. We also find that the limited generalizability of domain-specific dictionaries is an inevitable limitation at the current stage.