[4C1-GS-11] AI and Society:

[4C1-GS-11-03] Estimating Political Position Similarities from Twitter texts Based on Multi-Language Embeddings and Graph Neural Network

〇Jinghui Chen^1,2, Takayuki Mizuno^2,1, Shohei Doi³ (1. Graduate University for Advanced Studies, 2. National Institute of Informatics, 3. Hokkaido University)

Keywords:Multilingual NLP, Computational Social Science, Graph Neural Network

Numerous studies have successfully uncovered latent attributes of social media users, such as their gender, age, and political orientations. In the realm of ideology estimation, there are mainly two methodologies: content analysis and network analysis. Content analysis scrutinizes texts and hashtags through algorithms like support vector machines to perform binary or multi-class classification. Conversely, network analysis focuses on examining relationships through reposts, mentions, and followers, using the homogeneity in these networks to classify users. We proposed a novel methodology to detect political position similarities in a multi-language context with content analysis. Starting by establishing high-dimensional political dimensions, we projected politicians' user vectors onto this left-right axis. We showed that there is a clear polarization between US Democratics and Republicans via the analyzation of social media dataset and the sentence-level model outperformed the word-level model. We are also working on a combination methodology to use the position information as attributes to nodes in a retweet graph neural network and to compare the combined methodology and the singular ones for future work.