PR-Embed is an interactive system that combines semantic embeddings and topic modeling (STM) to analyze conceptual relationships in public relations research. It is built from 4,128 full-text articles published between 2004 and 2024 in six leading public relations journals. The system helps you explore how meanings, topics, and conceptual boundaries shift over time and across the literature.
What Are Embeddings?
Embeddings are numerical representations of words, phrases, or documents that capture meaning from their surrounding context. Words that frequently appear together in similar contexts have closer vectors in a multi-dimensional semantic space.
For example, coffee, tea, and espresso cluster closely together, while coffee and bicycle lie much farther away from each other.
This toolkit uses Word2Vec (Skip-gram) embeddings trained on overlapping five-word windows (5-grams) to capture phrase-level semantics. The cosine similarity between two word vectors indicates how semantically related they are (high = identical meaning, low = unrelated).
Learn more about Word2Vec embeddings →
Developed by Mikolov et al. (2013, Google Research).
Corpus and Model Training
PR-Embed includes two sets of Word2Vec models trained on peer-reviewed public relations scholarship:
Full-Corpus Model — trained on all full texts from six leading PR journals between 2004 and 2024:
- Public Relations Review (PRR)
- Journal of Public Relations Research (JPRR)
- Public Relations Inquiry (PRI)
- Corporate Communications: An International Journal (CC)
- Journal of Communication Management (JCM)
- International Journal of Strategic Communication (IJSC)
This model captures long-term, cumulative meanings across the field.
Yearly Models — trained separately for each publication year. These allow you to trace semantic drift (how relationships between concepts evolve over time). For example, you can examine whether AI becomes more closely associated with ethics or automation in later years.
Each model uses identical preprocessing: lowercasing, punctuation removal, token filtering (≥ 25 occurrences for full-corpus, ≥ 10 for yearly), and a context window size of 5.
What Is Structural Topic Modeling (STM)?
STM identifies latent topics in large text collections by grouping words that frequently co-occur. Each document is represented as a mixture of topics, and each topic is represented as a distribution of words.
- Prob words: Most frequent words within a topic.
- FREX words: Words most exclusive to that topic.
- Topic proportions: How much of a document belongs to each topic.
STM answers “What are people writing about?” while embeddings answer “How are those concepts related linguistically?”
How PR-Embed Combines STM and Embeddings
STM provides topic-level insights, where the unit of analysis is a topic or a paper. Embeddings provide word-level insights, focusing on how individual terms and concepts relate to one another in semantic space. PR-Embed embeds both topics (from STM) and words/papers (from Word2Vec) into a shared space, enabling comparison of conceptual alignment and distance.
This integration connects what scholars write about (topics and documents) with how those ideas relate linguistically (word meanings and associations).
Exploration Tabs
Word-Level: Explore nearest neighbors, semantic drift between two words, and word equations that reveal derived or contextualized meanings.
Paper-Level: Search papers by title or author, inspect paper information, view topic composition, and find semantically similar papers.
Topic-Level: Browse the 93 STM topics, inspect Prob and FREX keywords, and view representative papers aligned with each topic.
Topic Projection: Project topics along a semantic dimension between two anchor words and compare how topics align across the continuum.
Technical Summary
- Embedding architecture: Skip-gram Word2Vec (Mikolov et al., 2013).
- Corpora: Six leading PR journals (2004–2024).
- Models: One full-corpus model + yearly models for semantic drift analysis.
- Topic model: Structural Topic Model (Roberts et al., 2014) with 93 topics.
- Integration: Each topic and paper represented by an embedding vector for alignment.
- Similarity metric: Cosine similarity.
- Interface: Interactive Streamlit + Plotly visualization.
Refer to the paper cited below for more technical detail on the corpus, modeling choices, validation, and interpretation.
Citation:
Huang, J., Hurst, B., Capizzo, L. W., & Zhou, A. (In Press). Corporate Social Responsibility (CSR) versus Corporate Social Advocacy (CSA): Conceptual Boundaries, Distinct Roots, and Growing Convergence Revealed by Topic Modeling and Word Embedding. Public Relations Review.