How WorldRelat Actually Calculates a Relationship Score
2026-09-03 · by Jordi Fuster
The About page gives the short version of how a WorldRelat score gets made. This is the longer one — what's actually happening under the hood, and how that compares to how other people try to solve the same problem, which turns out to be a genuinely hard one: how do you put a single number on something as messy as how two countries feel about each other?
Start with the text, not a database of treaties
WorldRelat doesn't work from a checklist of alliances, treaties or trade volumes. It starts from written text about a specific pair of countries, run through VADER, an open-source sentiment analysis tool built for exactly this kind of task: reading ordinary prose and scoring how positive, negative or neutral its language is, word by word and then in aggregate. The introduction and the body of the source text are scored separately and then blended, weighted toward the introduction — usually the more careful, representative summary of the two. Short or thin source material pulls the score back toward a neutral middle rather than letting a couple of strongly-worded sentences swing it to an extreme. Shared EU or NATO membership adds a small, deliberate nudge upward, on the theory that formal alliance is real signal even when the prose doesn't spell it out directly.
That's a genuinely different approach from most of the well-known projects trying to measure the same thing.
How that compares to other approaches
The best-known large-scale effort here is probably the GDELT Project, which monitors news media from essentially every country in something close to real time and scores the tone of that coverage using its own content-analysis measures. GDELT operates at a scale WorldRelat doesn't attempt — hundreds of thousands of articles a day — but that scale comes with its own trade-off: it's tracking the tone of current news coverage, which can swing hard around a single event, rather than a more settled, encyclopedic account of a relationship's overall character.
Academic and policy-research projects tend to go the opposite direction: hand-coded indices, where researchers or expert panels directly score things like alliance strength, rivalry, or conflict history — the Correlates of War project is the classic example, and it's been a reference point in international relations research for decades. That approach is rigorous, but it's slow and it doesn't scale to every possible pair of the roughly 200 countries in the world — which is the specific gap WorldRelat's fully-automated approach is built to fill: a score for literally every pairing, refreshed as the underlying text changes, without a research team behind it.
Every approach makes a trade-off. GDELT trades depth for real-time breadth. Correlates of War trades scale for rigor. WorldRelat trades the nuance a human analyst would bring for full, automated coverage of every country pair, refreshed on a schedule rather than compiled by hand once every few years.
What the algorithm doesn't know
It's worth being direct about the limits. The model reads language, not intent — it can't tell you who's right in a dispute, and it has no concept of what "should" happen next. A relationship that's covered in dry, neutral, businesslike language will often score higher than one full of dramatic but ultimately less consequential rhetoric, simply because sentiment analysis reads tone, not stakes. That's a real limitation, and it's one reason the site also carries a separate, vote-adjusted "Combined score" on every relation page — a way for real visitors to push back on a reading they think misses the mark, without letting a handful of votes override what the underlying text actually says.
Put together, that's the honest version of "how it works": one automated, text-based method among several very different ones people use to try to measure the same slippery thing, each with its own strengths and its own blind spots.