Skip to content
Back to search results

Semantic Network Analysis

Introduced by
George Barnett, Marya Doerfel, Steven Corman (communication applications)
Year
1999
Type
Network representation of concepts and their co-occurrence in text
Family
Text-as-network analysis
Font size

What does it do?

Semantic network analysis represents the meaning of a text or corpus as a network of concepts connected by their co-occurrence or grammatical proximity, then uses network-analytic measures to reveal which ideas are central, how concepts cluster, and how shared meaning is structured.

In communication research it is the standard way to map the conceptual architecture of media coverage, organizational discourse, and public conversation at scale.

The intuition behind the method

Reading a thousand articles to find the 'big ideas' and how they hang together is impossible by hand. Semantic network analysis automates the insight that words appearing together carry related meaning: if 'climate' and 'economy' keep showing up in the same sentences, they are linked in the discourse. Turning a corpus into a network of co-occurring concepts lets us see the conversation's skeleton — the hub concepts everyone routes through, the tight clusters that form sub-topics, and the bridges that connect otherwise separate themes — using the same centrality and community tools social scientists use on networks of people.

History

Network representations of meaning originate in cognitive science and in James Danowski's WORDij co-occurrence approach, but communication scholars adapted them for discourse and organizational research in the 1990s. George Barnett and Marya Doerfel's 1999 semantic network analysis of the International Communication Association demonstrated the comparative, structural payoff, and Steven Corman and colleagues' 2002 centering resonance analysis grounded concept extraction in linguistic theory. The method now anchors computational framing, network agenda-setting, and large-scale social-media discourse studies.

When to use it

Use semantic network analysis when you want to map the conceptual structure of a body of text — the central ideas, their clustering, and the bridges between themes — especially across a corpus too large to read closely. It suits questions about shared meaning, discourse organization, and how conceptual structures differ across outlets, communities, or periods, and it underpins network agenda-setting and computational framing work. It assumes that co-occurrence is a reasonable proxy for semantic association and that the extracted concepts capture the meaningful vocabulary. It is less appropriate when meaning hinges on sentence-level argument, negation, or narrative sequence that co-occurrence flattens, or when the corpus is small enough that close interpretive reading would be richer.

Limitations and considerations

  • Co-occurrence ignores negation, irony, and argument structure, so 'linked' concepts may not be meaningfully associated.
  • Results are sensitive to preprocessing choices — concept extraction, window size, thresholds — that can change the network substantially.
  • Large networks become visually and analytically cluttered, requiring thresholds that discard information.
  • It describes structure, not the propositional content of relations; an edge says two concepts co-occur, not what is asserted about them.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Scales interpretation to large corpora, surfacing the conceptual skeleton no analyst could read by hand.
  • Imports the full toolkit of network measures — centrality, clustering, cohesion — into the analysis of meaning.
  • Produces intuitive visual maps that communicate discourse structure to non-specialists.
  • Supports rigorous comparison of conceptual structures across sources, time, languages, or communities.

Applications

  • Mapping how news outlets conceptually structure coverage of an issue and comparing those structures.
  • Analyzing organizational and strategic communication to find the central themes of mission statements or crisis responses.
  • Tracking how public discourse around a topic on social media clusters and evolves over time.
  • Supporting network agenda-setting by building the media and public concept networks to be compared.
  • Studying scientific or political fields by mapping the conceptual co-citation or term networks of their literature.

How it works

How it works

1

The corpus is cleaned and tokenized, stop-words removed, and concepts extracted — often the nouns and noun phrases that carry topical meaning. Methods like centering resonance analysis select discursively central words using linguistic theory, while simpler approaches use frequency-filtered terms.

2

A link weight w_ij between concepts i and j is created by counting how often they co-occur within a defined window — the same sentence, a sliding window of words, or the same document — so that frequently associated concepts are strongly tied.

Formula
w_{ij} = \sum_{u} \mathbb{1}[i,j \text{ co-occur within window in } u]
3

The weighted co-occurrences populate a network in which nodes are concepts and edges are their associations. The result can be a single corpus-level network or a set of networks compared across outlets, time periods, or speakers.

4

Network-analytic measures characterize the structure: degree and betweenness centrality identify hub and bridge concepts, community detection finds thematic clusters, and density or cohesion summarize how tightly the discourse is organized.

5

The structural findings are interpreted substantively — central concepts indicate what the discourse is organized around, clusters indicate sub-frames or topics, and comparisons across networks reveal how meaning differs between sources or shifts over time.

Common pitfalls

  • Treating co-occurrence edges as semantic assertions, e.g., reading a 'vaccine–autism' link as endorsement rather than mention.
  • Letting inconsistent stemming, synonym handling, or stop-word lists fragment or merge concepts in ways that distort the network.
  • Choosing a co-occurrence window or edge threshold arbitrarily and not reporting sensitivity to that choice.
  • Over-interpreting centrality of high-frequency but semantically empty terms that survived preprocessing.

Frequently asked questions

How is semantic network analysis different from topic modeling?

Topic modeling (e.g., LDA) discovers latent topics as probability distributions over words and assigns documents to topics. Semantic network analysis instead represents concepts as nodes and their co-occurrence as edges, then uses network measures to study structure — which concepts are central, how they cluster, what bridges themes. Topic models answer 'what themes exist,' while semantic networks answer 'how are concepts related and organized,' and the two are often used together.

What is centering resonance analysis?

Centering resonance analysis (CRA), developed by Corman and colleagues, is a principled way to build semantic networks. Grounded in centering theory from linguistics, it selects the noun phrases that are discursively central within sentences and links them, then uses network influence measures to index each word's importance. CRA gives a theory-based alternative to ad hoc keyword selection and produces networks whose central nodes correspond to the text's organizing concepts.

Does the choice of co-occurrence window matter?

Yes, substantially. Linking concepts only within the same sentence yields sparser, more precise networks; using a wide sliding window or whole-document co-occurrence yields denser networks that may connect unrelated ideas. Because the window directly shapes which edges exist, best practice is to justify the window from the research question and to report whether key findings are robust to reasonable alternative window sizes and thresholds.

Related methods

Sources

  1. Corman, S. R., Kuhn, T., McPhee, R. D., & Dooley, K. J. (2002). Studying complex discursive systems: Centering resonance analysis of communication. Human Communication Research, 28(2), 157–206.

    Open DOI reference
    Copied.
  2. Doerfel, M. L., & Barnett, G. A. (1999). A semantic network analysis of the International Communication Association. Human Communication Research, 25(4), 589–603.

    Open DOI reference
    Copied.