Recipes / Monitor a market whose language nobody on the team reads
Monitor a market whose language nobody on the team reads
/stories in semantic mode. The parameters matter less
than the reasoning behind them, so that is most of what is below.
The request
GET /stories
?q=<topic, in English>
&mode=semantic
&language=<iso639-3>
&from=<now - 3d>Why these parameters
Semantic mode lets you query in one language and match coverage in another, which is the only workable approach when nobody on the team can compose a query in the target language. Scope by language explicitly rather than relying on the topic to imply geography — a topic query without a language filter returns whatever the largest streams are saying about that market, which is not the same as what the market is saying about itself.
What comes back
Results in the target language with headlines in their original script. Set text direction per result when rendering, or right-to-left headlines will display with punctuation on the wrong side.
How this goes wrong
Assuming English-language coverage of a country substitutes for that country's own press. It is thinner, later, and selected for what interests an outside audience — which is precisely the wrong filter if you are monitoring for risk.
Adapting it
To cover a region rather than a country, query several languages over the same window and merge — for the Balkans that means Serbian, Croatian and Bosnian together, since mutually intelligible coverage is split across separate language facets and any one of them alone gives a partial picture. Serbian additionally needs care because it is written in two scripts and a single-script keyword query silently returns roughly half the coverage.
Running it for real
Have someone who reads the language spot-check the results monthly. Semantic matching across languages is good but not transparent, and the failure mode is quiet: plausible results that systematically miss a subject area. A thirty-minute review each month catches drift that no automated metric will, and it is the only real safeguard when nobody on the team can read the output.
When this is the wrong tool
This is not a substitute for someone who reads the language. Semantic matching across languages is good enough to surface the right material and not good enough to tell you what it says, and a translated headline loses the register, hedging and attribution that often carry the actual meaning. Teams that treat this as full coverage of a market tend to discover the gap during the one week it matters. The recipe is also weaker for languages with small streams: where a language has few outlets, a bounded query returns little and the temptation is to widen it until noise dominates. Widen the window instead, and if that still returns nothing, accept that the coverage is not there rather than loosening until something appears.