Turkish Path
How AIVEX Feed scores Turkish financial news with an English financial model — safely.
The Problem
Financial sentiment models are trained on English text. Applied directly to Turkish market news, they produce noise: articles are scored incorrectly or skipped entirely.
The Mechanism
AIVEX Feed scores Turkish news in three gated steps:
- Language detection — a language identifier detects the article's language. Below the language-confidence gate, the article is flagged and never enters scoring.
- Translation — Turkish articles are machine-translated to English. If translation fails, the article is skipped — not guessed at.
- Sentiment scoring — a financial-English model scores the title first; below the title-confidence gate, it is re-scored against the title and body together.
Skip Rather Than Guess
An article with uncertain language or a failed translation is marked neutral and passed over. It never pollutes the numbers. The system's failure mode is silence, not invention.
Three Gates, All Configurable
Three confidence thresholds gate the pipeline — sentiment confidence, title confidence, and language confidence. Each is set from an environment variable, so the gate can be tightened or loosened without touching code.
Sentiment accuracy figures for the Turkish path have not yet been published; no accuracy claim is made until they are measured against a labelled evaluation set.