Social Representation Analysis via Semantic Network Mapping

Implements the ThemeScope framework for detecting and visualising social representations in large-scale digital text corpora. From raw documents it builds, via 'udpipe' annotation, sentence-level word co-occurrence networks and derives two community-level indicators grounded in Social Representation Theory: the Prototypical Salience Index (PSI) for anchoring and the Concreteness Score (CS) for objectification. Communities are located in a two-dimensional, theoretically grounded representational map. The whole pipeline is usable from the R console; an optional 'shiny' graphical interface calls the same exported functions. The method is described in Misuraca, Spano and D'Aniello (2026) .

themescopeR

Social Representation Analysis via Semantic Network Mapping

themescopeR implements the ThemeScope framework for detecting and visualising social representations in large-scale digital text corpora. From raw documents it builds sentence-level word co-occurrence networks, detects thematic communities, and locates them on a two-dimensional strategic diagram using two indicators grounded in Social Representation Theory (SRT):

• PSI, the Prototypical Salience Index (anchoring): how structurally embedded a theme is in the wider discourse.
• CS, the Concreteness Score (objectification): how far a theme is articulated through concrete, imageable vocabulary.

The whole analysis runs from the R console; an optional Shiny GUI offers the same pipeline point-and-click and is backed by the very same exported functions.

Citation

Using themescopeR or the ThemeScope method requires citing the paper that introduces it. Publishing results without attribution violates the terms of use.

Misuraca, M., Spano, M., and D'Aniello, L. (2026). ThemeScope: A quantitative thematic analysis for depicting social representations in digital arenas. Journal of Information Science. doi:10.1177/01655515261454276

@article{misuraca2026themescope,
  author  = {Misuraca, Michelangelo and Spano, Maria and D'Aniello, Luca},
  title   = {ThemeScope: A quantitative thematic analysis for depicting social representations in digital arenas},
  journal = {Journal of Information Science},
  year    = {2026},
  doi     = {10.1177/01655515261454276},
  url     = {https://doi.org/10.1177/01655515261454276}
}

In R, citation("themescopeR") prints the same reference.

Installation

# install.packages("remotes")
remotes::install_github("lucadaniello/themescopeR")

The optional GUI additionally needs: shiny, bslib, DT, plotly, visNetwork, shinycssloaders. writexl enables the Excel exports.

Quick start (console)

library(themescopeR)

# 1. Read a collection of raw documents (csv / tsv / txt / xlsx / RData / zip)
coll <- read_collection(
  system.file("extdata", "sample_collection.csv", package = "themescopeR"),
  sequential_ids = TRUE          # documents numbered doc_1, doc_2, ...
)

# 2. Annotate with udpipe (the language model is downloaded once and cached)
tokens <- preprocess_texts(coll, model = "english")

# 3. Run the full ThemeScope pipeline
result <- themescope(tokens, vocab_size = 1500, seed = 1)

# 4. Inspect and visualise
print(result)
as.data.frame(result)      # community-level table: PSI, CS, z-scores, quadrant
plot(result, type = "map") # the ThemeScope strategic diagram (ggplot2)
plot(result, type = "network")
top_terms(result, n = 10)  # most representative terms per community

# 5. Save the whole study in one file, and reopen it later
save_themescope(result, "climate_study.themescope", words = tokens, collection = coll)
archive <- read_themescope("climate_study.themescope")
summary(archive)

themescope() can also start directly from raw texts, annotating internally:

result <- themescope(coll, model = "english", vocab_size = 1500, seed = 1)

To skip the annotation step entirely while trying the package out, the bundled demo corpus ships already tokenised, lemmatised and POS tagged (English GUM treebank):

tokens <- readRDS(system.file("extdata", "demo_annotated.rds", package = "themescopeR"))
result <- themescope(tokens, seed = 1)

Language models (downloaded once, cached)

Preprocessing uses the updated Universal Dependencies 2.15 udpipe models maintained for TALL.

ts_list_models()              # available languages / treebanks
ts_download_model("italian")  # download + cache (no-op if already cached)
themescope_cache_dir()        # where models would be stored (creates nothing)
ts_clear_cache()              # delete everything that was cached

The cache lives in tools::R_user_dir("themescopeR", "cache"); override it with options(themescopeR.model_dir = "/your/path"), or pass model_dir to a single call (ts_download_model("italian", model_dir = tempdir())) to keep a download inside the session.

The directory is created only when a model is actually downloaded, that is when you call ts_download_model() or hand a language name to preprocess_texts(). Asking where the cache is creates nothing, and ts_clear_cache() removes the models and the directory again. Nothing else in the package writes outside the session temporary directory.

Shiny GUI

themescope_app()

The interface follows the pipeline step by step, and every step calls the exported functions, so the GUI and the console give identical results.

1. Import. Choose the data type first: raw text (a single file or a zip of many), a saved .themescope archive, or the bundled demo corpus. Tabular files let you pick which column holds the text. Documents are numbered doc_1, doc_2, ... with the original identifier kept as source_id. A file bigger than the current upload limit opens a window offering to raise it, then imports automatically.
2. Preprocess. Pick the annotation language and the exact treebank; the best model for the language is preselected, with its description underneath. The demo corpus and archives arrive annotated, so this step is skipped.
3. Run. Set the vocabulary, POS filter, co-occurrence window, similarity measure, edge threshold, community algorithm and concreteness lexicon, then run the analysis.
4. Refine. Network appearance in the sidebar controls repulsion between communities, node spacing inside a community (measured in node diameters), edge opacity, label size, node size, the layout seed, term labels and whether unclassified terms are drawn. Refine network redraws the graph without recomputing the analysis.

The results are shown as an interactive map and network (ThemeScope tab), a filterable community table, and top terms as both charts and collapsible per-cluster lists. The Reproduce tab prints the console code equivalent to what the GUI just did, and Export .themescope saves the annotated corpus, the parameters and the term scores in a single file.

Main functions

The package mirrors the ThemeScope pipeline as a chain of small, pure, command-line functions. themescope() runs them all in order, but each stage is exported so you can inspect or customise any step. They are presented below in the order the pipeline calls them.

1. Import: read_collection()

read_collection(path, text_col = NULL, id_col = NULL, sequential_ids = FALSE)

Reads a document collection into a tidy data frame with a doc_id and a text column. It auto-detects the format from the file extension and supports a single .csv/.tsv, .txt, .xlsx/.xls, .RData/.rda, or a .zip containing many such files (unzipped to a temporary directory, read, and row-bound). Text and id columns are guessed but can be forced with text_col / id_col. With sequential_ids = TRUE documents are numbered doc_1, doc_2, ... and the identifier that came with the file is kept as source_id, which is handy when the source ids are long hashes. Malformed input raises a clear, actionable error.

2. Linguistic annotation: preprocess_texts()

preprocess_texts(collection, model = "english", text_col = "text",
                 doc_id_col = "doc_id", verbose = TRUE)

Annotates the raw texts with udpipe: tokenisation, lemmatisation, sentence segmentation and part-of-speech (UPOS) tagging. model accepts a language name, a .udpipe file path, or a loaded model object; language models come from the updated Universal Dependencies 2.15 treebanks and are downloaded once and cached (see ts_list_models(), ts_download_model(), ts_model_path(), themescope_cache_dir()). It returns the full udpipe annotation (one row per token, with doc_id, sentence_id, token, lemma, upos and the rest), which is the input every downstream function expects.

3. Vocabulary: build_vocab()

build_vocab(words_df, unit = "lemma", vocab_size = 1500,
            pos_filter = c("NOUN", "ADJ", "PROPN"))

Filters the annotation to content words (by default nouns, adjectives and proper nouns), counts them on the chosen unit ("lemma" or "token"), and keeps the vocab_size most frequent terms. Returns a data frame of term, freq and upos, the node set of the network.

4. Co-occurrence and association strength: build_cooccurrence_matrix()

build_cooccurrence_matrix(words_df, vocab, unit = "lemma",
                          window = "sentence", normalization = "association")

Counts how often each pair of vocabulary terms appears together within the same window ("sentence" or "document") and rescales the raw co-occurrence counts into a similarity matrix via normalization. The default "association" (Association Strength, AS(t,t') = a_tt' / (a_t * a_t')) is the measure used in the paper; "jaccard", "salton", "inclusion", "equivalence" and "frequency" are also available (normalize_cooccurrence(), compute_association_strength() expose them individually). It returns both the normalised matrix and each term's presence a_t (sentence/document count), which PSI needs later.

5. Network construction: build_cooccurrence_network()

build_cooccurrence_network(as_matrix, threshold_percentile = 0.98)

Turns the similarity matrix into a weighted, undirected igraph network, keeping only the strongest edges (those above the threshold_percentile) and dropping isolated nodes. This sparse backbone is what gets partitioned into themes.

6. Community detection: detect_communities()

detect_communities(graph, algorithm = "walktrap", min_size = 10, seed = NULL)

Partitions the network into thematic communities (candidate social representations) with "walktrap" (default), "louvain" or "leiden". Communities smaller than min_size are dropped; pass a seed for reproducible results. get_community_subgraphs() splits the graph into one subgraph per community.

7. The two SRT indicators: compute_psi() and compute_cs()

compute_psi(graph, communities, presence)
compute_cs(graph, communities, concreteness_lexicon = brysbaert)

• compute_psi(), the Prototypical Salience Index (anchoring): how structurally salient and frequent a community's terms are within the wider discourse, normalised by community density.
• compute_cs(), the Concreteness Score (objectification): the association-weighted mean concreteness of a community's terms, using the bundled brysbaert lexicon (lexicon_coverage() / match_concreteness() report and apply the matching).

Both return one value per community; themescope() z-scores them and places each community in a quadrant (assign_quadrant(), zscore()).

8. Representative terms: term_relevance() and top_terms()

term_relevance(graph, membership, presence)
top_terms(x, n = 10, by = "relevance")   # or by = "frequency" / "degree"

Rank the terms inside each community so each theme can be labelled by its most characteristic words. term_relevance() implements the case-study relevance measure

$$R_t(g_i) = \log(1 + a_t)\cdot\frac{s_t^{\text{in}}(g_i)}{s_t^{\text{in}}(g_i) + s_t^{\text{out}}(g_i)},$$

where a_t is the term presence, and s_t^in / s_t^out are the total association strength linking t to terms inside / outside its own community. top_terms() ranks by "relevance" (the default R_t, which downweights generic terms shared across clusters), by "frequency" (term presence a_t alone), or by "degree". Relevance is the labelling method used throughout ThemeScope; frequency is kept as a simpler alternative.

9. Visualisation: plot_themescope() and plot_network()

plot_themescope(psi, cs, ...)   # the strategic diagram (ggplot2)
plot_network(graph, ...)        # the coloured co-occurrence network

plot_themescope() draws the z-scored PSI by CS strategic diagram with the four labelled quadrants; plot_network() draws the community-coloured network. Both share themescope_colours(), so a community keeps the same colour on the map and on the network. On a themescope object simply call plot(result, type = "map" | "network"); plot(result, label = "terms") labels each community with its top terms (ranked by relevance by default; pass label_by = "frequency" to label by term frequency instead).

10. Saving a study: save_themescope() and read_themescope()

save_themescope(result, "study.themescope", words = tokens, collection = coll,
                meta = list(language = "english", lexicon = "Brysbaert"))
archive <- read_themescope("study.themescope")
summary(archive)          # corpus, parameters, vocabulary, network, lexicon
plot(archive$result, type = "map", label = "terms")

A .themescope file is a compressed archive holding the themescope object, the annotated corpus it was computed from, the raw texts, the term-level table with the scores behind the maps, the community table and the parameters of the run. Reopening it restores maps, communities and top terms with nothing recomputed, in the console or in the GUI.

11. Orchestrator and GUI: themescope() and themescope_app()

themescope() runs stages 3 to 9 (and stages 1 to 2 if you hand it a raw collection plus a model) and returns a structured themescope object with print(), summary(), plot() and as.data.frame() methods. themescope_app() launches the Shiny GUI, which calls these same exported functions.

Summary table

Key options of themescope() / build_cooccurrence_matrix(): unit ("lemma"/"token"), window ("sentence"/"document"), normalization ("association", "jaccard", "salton", "inclusion", "equivalence", "frequency"), and community_algorithm ("walktrap", "louvain", "leiden").

Methodology

For communities in the Association-Strength network (AS(t,t') = a_tt' / (a_t * a_t'), with a_t the sentence presence of term t):

where s_t is the mean Association Strength of t's in-community neighbours, delta(g_i) = 2 W_i / (|V_i|(|V_i|-1)) the weighted community density, and c(.) the Brysbaert concreteness rating. Communities are placed on the z-scored PSI by CS plane:

Data and lexicon

• Bundled lexicon brysbaert: 39,954 English concreteness norms (Brysbaert, Warriner and Kuperman, 2014, Behavior Research Methods, doi:10.3758/s13428-013-0403-5).
• Example data inst/extdata/sample_collection.csv: a 1000-document random sample of the publicly available Public Opinion on Climate Change Reddit dataset on Kaggle, plus inst/extdata/demo_annotated.rds, the same 1000 documents already annotated with udpipe.

Authors

Michelangelo Misuraca, Maria Spano, Luca D'Aniello (maintainer).

License

MIT, 2026. See LICENSE.