Analyzing political party positions through multi-language twitter text embeddings

Area of Science:

• Computational Social Science
• Natural Language Processing
• Political Science

Background:

• Traditional word embedding models represent word semantics as vectors, useful in social science research.
• Previous work projected words onto antonym-defined dimensions; this study extends this to sentence-level analysis.

Purpose of the Study:

• To develop and evaluate a novel sentence-level methodology for analyzing political positions across languages.
• To compare the efficacy of the new sentence-level approach against traditional word-level embedding models.
• To investigate political positioning and cross-cultural similarities using multilingual Twitter data.

Main Methods:

• Extended word-level cultural dimension architecture to the sentence level.
• Utilized a Language-agnostic BERT model (LaBSE) for multilingual position similarity detection.
• Employed Latent Dirichlet Allocation (LDA) for detailed topic analysis of political tweets.
• Assessed methodology using Twitter data from US and Spanish politicians.

Main Results:

• The proposed sentence-level methodology significantly outperforms traditional word-level embedding models.
• The approach effectively identifies fine-grained political positions, showing variation even within individual politicians.
• Analysis of American and Spanish political data confirmed alignment with common sense, news, and prior research on a liberal-conservative axis.

Conclusions:

• The developed sentence-level architecture offers an improvement over standard word-level methodologies.
• This approach is effective for analyzing political discourse and cross-lingual position similarities.
• The methodology holds promise for future sentence-level applications in social science and NLP research.