Romanian news from Chișinău and from Bucharest: what still tells them apart

The 2019 study

What 33,564 news stories have in common#

In 2019 Andrei Butnaru and Radu Tudor Ionescu published MOROCO: 33,564 stories from the five most popular news websites in each country, labelled by domain, .md or .ro, with the names of people and places replaced. Their best program guessed the country for 94.13% of the test stories. A simpler program, rebuilt here, reaches 94.5%; without the 453 test stories that also appear word for word among the training or validation ones, 94.1%.

The program on this page is a simple classifier rebuilt after the 2019 study: it counts words and letters and learns which are more frequent on .md or on .ro sites. It is not the MOROCO authors’ and it does not understand the text. The study gives a few examples of clues, such as “sînt” (are) or “tenismen” (tennis player), but does not say how much each one weighs, or whether they hold on other sites. On the right, every word takes the colour of the country the program pushes it towards; most stay grey.

This page compares Romanian-language news from the Republic of Moldova and from Romania. Moldova’s Academy of Sciences established in 1994, and reaffirmed in 1996, that “the state (official) language of the Republic of Moldova is Romanian” (Academy of Sciences of Moldova); the Constitutional Court ruled to the same effect in 2013 (Decision 36/2013), and Law 52 of 2023 replaced “Moldovan language” with “Romanian language” throughout the statute book (the text of the law). The program’s label is the country of the website, not the writer’s dialect, nationality or identity.

First layer

A rule from 2016, followed by the whole press#

Spelling is the best-known clue. Romanian writes one vowel with two letters: â inside a word, î at its edges (with some exceptions). Among the words counted here that can be written either way (“sînt” and “sunt”, “cînd” and “când”), 19.5% use the old form in .md stories and 0.5% in .ro stories. Chișinău was not uniform: 15.6% of .md stories have only old forms, and the rest have only new forms, a mix of both, or none of these words.

Romania went back to â and “sunt” in 1993 (Romanian Academy, 1993). In Chișinău the Academy of Sciences recommended it in July 2016 and asked the press to switch by 1 January 2017 (Academy Resolution 205/2016). In the 2026 articles, read word by word, the old spelling appears once on the Moldovan sites and four times on the Romanian ones; the other matches are correctly spelt compounds such as “supraîncărcat” (overloaded) or “bineînțeles” (of course), or typing errors such as “cunoscutîn”. A program that sees only spelling told 60% of the corpus stories apart (a coin flip gives 50%) and 53% of the 2026 ones.

Second layer

The newsroom’s keyboard#

Some Romanian sites published without diacritics (“si”, “in”, “fara”). Others wrote ş and ţ with a cedilla instead of ș and ț with a comma, a leftover of old fonts (Unicode).

A program that sees only these habits told 66% of the corpus stories apart and 51% of the 2026 ones.

Third layer

The newsroom’s formulas#

Formulas differ too. In the corpus, “transmite” (reports), used to credit an agency (the 2026 articles have, for instance, “transmite IPN”), is 5.7 times more frequent on .md sites; in 2026 it no longer is. Also in the corpus, the .ro sites more often gave the weekday (“sâmbătă”, “vineri”) and the .md sites more often wrote “astăzi” (today); in 2026 only the difference in “astăzi” remains. Some .md sites reprinted stories from Romanian agencies: that is why “scrie Agerpres” (Agerpres writes) pulls towards Chișinău.

A program that sees only the formulas told 70% of the corpus stories apart and 55% of the 2026 ones. Part of what it had learnt were the habits of a few newsrooms.

Fourth layer

Words that name the country#

The corpus’s name filter misses some. Left in are “raionul” (Moldova’s districts), “județul” (Romania’s counties), “moldovenesc”, “românii”, “capitala”: words that have more to do with the subject of a story than with how it is written.

A program that sees only a list of such words tells 83% of the 2026 stories apart, more than any other program tried here with a single kind of clue.

Fifth layer

The topics differ#

The .md half of the corpus has more politics and sport; the .ro half more finance and technology. That is why some words pull one way only through the subject: “partid” (party) was 1.3 times more frequent on .md sites in the corpus, and in 2026 it is more frequent on .ro sites; “companie” (company), the other way round.

The fair comparisons are those within one topic, like the pairs in the chapter “Same story, different words”. With the same number of stories from each country in each topic, the program’s score drops a little, from 91.7% to 91.1%.

What holds

The language of communiqués: what stays, what faded#

Four expressions of reporting on institutions appear on all 11 Moldovan sites and stay several times more frequent there in 2026: “totodată” (at the same time), “în baza” (on the basis of), “în cadrul” (within), “întreprindere” (enterprise). Others are rarer, but still more frequent in Chișinău: “la moment” used for “now” (nine times, on three sites), “întrevedere” (a formal meeting), “peste hotare” (abroad), “formațiune” (for a party). And some, very frequent in Chișinău in the corpus, have almost disappeared in 2026: “cu începere de la” (starting from), “careva” (someone), “cu referire la” (with reference to), “urmează a” (is to), “un șir de” (a number of).

Almost all are in the dictionaries (DEX: întrevedere, DEX: bază, DLRLC: hotar); “cu referire la” is marked as an expression “of administrative language” (DLRLC: referire), and “reieșind din” in the sense “proceeding from” is corrected by the Moldovan linguist Andrei Crijanovschi (Limba Română, 2008). The difference is one of dose: these words appear in the Romanian press too, only less often.

The list of 23 forms was drawn up after a first look at both collections: it is an exploration, not a blind test. A program that sees only the list recognises 42% of the Moldovan stories from 2026 and 87% of the Romanian ones: it mostly finds the Romanian ones, by the absence of these words.

The missing test

On 18 sites read in 2026: 80 in 100#

Studies of MOROCO tested programs on other stories from the same corpus, whose sites are not named. On Twitter posts, the F1 score of a word-based program trained on these stories fell from 0.94 to 0.50 (Ceolin and Zhang, 2020). Here the program trained on the whole corpus read 981 articles published in 2026 on 18 sites, 11 in Moldova and seven in Romania. Whether some of these sites are also in the corpus is not known.

It was right 80.5% of the time, averaged over the two countries (87% of the Romanian stories and 74% of the Moldovan ones), against 94% in the corpus. With the layers above taken out: 77%, that is 91% of the Romanian stories and only 62% of the Moldovan ones. On the right, only the words that still pull the same way in the 2026 articles stay lit.

Where the difference is

When the story is not about Moldova, the difference shrinks#

With the layers taken out, the program recognises 75% of the Moldovan sites’ stories that mention Moldova, Chișinău or its institutions at least twice, but only 46% of the others. Romanian stories stay recognisable without Romania in them: 96% and 88%. On some Moldovan sites (ziarulnational.md, bani.md) the program is wrong more often than right.

That does not mean the difference disappears. Asked to choose between two stories, one from a .md site and one from a .ro site, the program picks the Chișinău one 94 times in 100 when both have at least two markers of their country, and 76 times in 100 when neither has. It learnt on a corpus in which Moldovan news often talked about Moldova, so it recognises those more easily.