This is the oldest receipt I have for the claim that I do not stop at advice. The report was published in April 2019 by ANTHRO.AI, a division of Marain Pte Ltd, a company registered in Singapore, and commissioned by Xynteo for Europe Delivers. It runs to 211 pages, and behind it sits a piece of software.
The problem with asking what a continent thinks
The brief was to study discourse and attitudes across Europe on four themes: the future of work, a green and resilient economy, a new social contract, and a global Europe. The obvious approaches both fail. A panel survey tells you what people say when asked a question somebody else wrote. A keyword scrape tells you about the words you already thought of, which means it can only confirm the brief.
So the collection started from concepts rather than keywords. Analysts identified seed subjects per theme using a subject exploration tool, and an engine trained on Wikipedia's ontology expanded each seed into a family of related concepts. Articles were then pulled against that family, across twenty-nine European countries including Russia, covering news, opinion, essays, blogs and conversations published between September 15 and October 15, 2018. Seven countries were analyzed in depth: Belgium, France, Germany, Italy, Poland, Sweden and the United Kingdom. A Russia supplement was added later at the client's request.
Where the machines stopped and the people started
Each article was scored for the concepts it contained and the weight of each one. Sentiment came from Google Cloud's Natural Language service. Headlines were machine-translated so an analyst could scan a corpus in a language they did not read, and open the original when it mattered. A t-SNE dimension reduction turned the concept relationships into a map per country, where distance means relatedness, so a person could see which ideas were traveling together in Poland and not in Sweden.
Then analysts read the maps. That is the whole design. The machine could hold twenty-nine countries in view at once and could not tell you what any of it meant. The analysts could tell you what it meant and could not read twenty-nine countries in a month. Neither half is the product.
Two numbers instead of one
Sentiment was reported as both a score and a magnitude, separately, and the report explains why. Score is normalized, so a long article containing furious agreement and furious disagreement returns something close to neutral. Magnitude is not normalized: every expression of emotion in the text adds to it. A country can look calm on score and be extremely loud on magnitude, and since the study was about conflict, reporting the single normalized number would have hidden the exact thing being measured.
I still do this. The public search on this site reports the number of results returned and the number that matched separately, so the interface can never claim eleven results over a list of eight. Same reflex, seven years apart.
No data transport, in 2019
The maps shipped as a browser interface. The client loaded the provided data packs and everything ran locally. The report states it plainly: all processing happens in the browser and there is no data transport to a server, so a powerful computer may be needed, and the maps may take a while to load on slower machines.
That is a product decision with a cost, taken deliberately and disclosed to the reader in the same sentence. The ethics section says the same thing in a different register: study aggregates, identify no individuals, store no personal data to build profiles. Private where it matters was already how I built things seven years ago, before the market started caring, and I was willing to make the user experience worse for it.
What I would do differently
I would version the ontology. The concept expansion depended on Wikipedia, which changes underneath you, so the same seed run six months apart is not the same study and nothing in the pipeline recorded that. I would also log which analyst read which map. The emotional analysis ran programmatically for every country except Poland, which was handled differently, and that kind of exception belongs in machine-readable provenance next to the finding rather than in a footnote a reader has to notice.