Text Embeddings to Measure Text Topics

This is the authors' abstract. We don't add key points for this paper.

Automated content analysis for measuring specific topics has become increasingly popular in social science research. This article demonstrates that text embedding, combined with cosine similarity, can accurately measure specific topics and explore the relationship between text topics and meta-information. Unlike probabilistic topic models (PTMs) such as the Structural Topic Model (STM) and Keyword-Assisted Topic Model (keyATM), the text embedding method does not require the selection of hyperparameters, keywords, or a training process. This significantly reduces computational resources and makes the method more convenient and faster to use. Additionally, the text embedding method can better capture the meaning of short texts. The findings suggest that text embedding is a superior alternative to commonly used PTMs.