Click a node to explore · Drag to reposition · Scroll to zoom
Intellectual structure of @Marinelivesorg tweets, 2016-2018 · 4,981 original tweets · 35 concepts
The graph shows concepts as coloured circles and their co-occurrences as connecting lines. Concepts that appear together frequently in tweets are pulled closer by the layout algorithm; concepts that rarely co-occur drift apart. The thickness and opacity of each line reflects the strength of the connection.
In overview mode (no node selected), the side panel lists the strongest connections in the current year view and the year-on-year shifts — which concepts grew or shrank between 2016 and 2018.
In node detail mode (after selecting a concept), it shows the concept's frequency in each year as horizontal bars, plus its strongest connections ranked by co-occurrence weight.
This explorer visualises the intellectual structure of 4,981 original tweets posted by @Marinelivesorg during 2016–2018. Retweets are excluded.
Concept extraction: A dictionary of 35 concepts was defined manually, informed by TF-IDF keyphrase extraction (bigrams and trigrams, min_df=3, max_df=0.3, 3,000 features) applied to the cleaned tweet text. Each concept is defined by a set of trigger words and phrases — for example, the "Textiles" concept is triggered by any of: textile, textiles, silk, wool, linen, cloth, cotton, garment, fabric, dye, dyeing. A tweet is tagged with a concept if any of its trigger terms appear in the cleaned text.
Important limitation: The 35 concepts were chosen by the analyst, not derived statistically. This means the map reflects a particular interpretation of the project's intellectual landscape. Themes not anticipated by the analyst will be absent. The number 35 was not a target — concepts were added until the analyst ran out of recognisable themes — but a different analyst might identify more, fewer, or different categories.
Co-occurrence: Two concepts are "connected" when they appear in the same tweet. The edge weight between any pair is simply the count of tweets containing both. This is computed separately for each year and for the combined 2016–2018 period.
Layout: The graph uses a force-directed layout: nodes repel each other (preventing overlap), edges attract connected nodes (pulling co-occurring concepts together), and a weak central gravity prevents the graph from drifting off-screen. The layout stabilises after approximately 300 simulation steps. Node size reflects tweet frequency; node colour reflects the concept category (see legend).
Coverage: 77% of original tweets (3,824 of 4,981) were tagged with at least one concept. The remaining 23% contain conversational text, acknowledgements, or topics outside the 35-concept dictionary.