Text Network Analysis — Mapping Meaning as a Graph
- Introduced by
- Dmitry Paranyushkin; Jana Diesner & Kathleen M. Carley
- Year
- 2011 (Paranyushkin); 2005 (Diesner & Carley)
- Type
- Text-mining network method
What does it do?
Text network analysis models the words or concepts in a text as nodes and their co-occurrences as edges, then uses network metrics to reveal the structure of meaning.
The approach was advanced by Diesner and Carley (2005) for communication networks and by Paranyushkin (2011) for tracing the pathways of meaning circulation in text.
The intuition behind the method
A document is more than a bag of words — meaning lives in how words appear together. If two terms keep showing up near each other, they are connected; if a term sits between many others, it acts as a bridge that holds the discourse together. By turning text into a graph of co-occurring terms, text network analysis lets you see which concepts are central, which clusters of ideas form, and which words connect otherwise separate themes, rather than reading the corpus word by word.
History
Text network analysis grew from the study of communication and semantic networks. Diesner and Carley (2005) explored extracting communication networks from text corpora, while Paranyushkin (2011) framed text as a network whose structure reveals the pathways through which meaning circulates, popularising co-occurrence-based concept graphs and centrality metrics for textual analysis.
When to use it
Use text network analysis when you have text data and want to explore the relational structure of concepts or describe how themes connect, rather than just count words. It suits exploratory and relational research questions on cross-sectional or longitudinal text. A co-occurrence window must be defined and the text must be preprocessed first; a corpus of at least roughly 20 documents is recommended for stable structure.
Limitations and considerations
- Results depend heavily on the chosen co-occurrence window; a different window can change which terms appear connected.
- Requires careful text preprocessing; inconsistent normalisation fragments nodes and distorts edge weights.
- Co-occurrence captures association, not directed syntactic or causal relations between terms.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Reveals the relational structure of meaning — central concepts, bridges, and clusters — that simple frequency counts miss.
- Works on any preprocessed text corpus and produces an interpretable graph plus quantitative network metrics.
- Bridges qualitative reading and quantitative analysis, supporting both exploratory and relational research questions.
Applications
- Mapping the central concepts and themes in a corpus of documents or interviews.
- Tracing how ideas connect and circulate across political, media, or social discourse.
- Identifying bridging terms that link otherwise separate topic clusters.
- Exploring the conceptual structure of academic literature or organisational communication.
How it works
How it works
The raw text is cleaned and standardised before any network is built: it is tokenised, noise is removed, and terms are normalised so that the same concept is counted consistently. Without this preprocessing the resulting nodes would be fragmented and the co-occurrence counts unreliable.
A co-occurrence window is defined — the span (for example a sentence, a paragraph, or a sliding window of words) within which two terms are treated as appearing together. Choosing this window is a key decision: it determines which word pairs become connected and how dense the resulting network will be.
Words or concepts become nodes and their co-occurrences become weighted edges, where the weight reflects how often the pair appears together within the chosen window. The result is a network that encodes the relational structure of the text rather than just its vocabulary.
Network metrics are computed over the graph — for example betweenness centrality to find the terms that bridge different parts of the discourse, alongside other centrality and clustering measures. These metrics surface the most influential concepts and the community structure of meaning in the corpus.
Common pitfalls
- Skipping or rushing preprocessing leaves noisy, fragmented nodes that make the network metrics meaningless.
- Picking the co-occurrence window arbitrarily — too wide inflates spurious links, too narrow misses real associations.
- Running on too small a corpus (well under ~20 documents), where the network structure is unstable and over-interpreted.
- Reading co-occurrence edges as causal or grammatical relations rather than mere association.
Frequently asked questions
How is text network analysis different from word frequency analysis?
Frequency analysis counts how often each word appears in isolation. Text network analysis instead models how words co-occur, building a graph whose edges capture relationships. It tells you not just which terms are common but which are central, which bridge different themes, and how concepts cluster together.
How do I choose the co-occurrence window?
The window defines the span within which two terms count as appearing together — for example a sentence, a paragraph, or a sliding window of a few words. There is no universal value: a wider window creates a denser network with more links, a narrower one keeps only the closest associations. Choose it deliberately based on your unit of meaning and check that the resulting structure is stable.
What do the network metrics tell me?
Centrality metrics such as betweenness highlight the terms that bridge different parts of the discourse and the concepts that hold the network together. Combined with clustering measures, they reveal the most influential ideas and the community structure of meaning across the corpus.
How much text do I need?
A corpus of at least roughly 20 documents is recommended so the network structure is stable enough to interpret. With far fewer documents the graph is dominated by noise and easily over-interpreted.
Related methods
Sources
Paranyushkin, D. (2011). Identifying the Pathways for Meaning Circulation Using Text Network Analysis. Nodus Labs.
Open publication referenceCopied.Diesner, J. & Carley, K. M. (2005). Exploration of Communication Networks from the Enron Email Corpus. SIAM International Conference on Data Mining, Workshop on Link Analysis, Counterterrorism and Security.
Open publication referenceCopied.