Low Quality Label

It is the system that recognises content of low information value by the measurable patterns it carries, and marks it.

The label reads the form of a headline, not its intent: listicle and service headlines (‘7 practical methods’), horoscope, fortune-telling and dream-interpretation content, and question patterns such as ‘who is’, ‘when is’. Content that receives the label instantly loses -25 penalty points from its score. The machine learning model has an F1: 0.75 in low quality TR detection, while the rule engine has an F1: 0.41 success by looking at specific phrases (‘ne zaman’, ‘kimdir’ etc.) or question mark patterns (2 or more ‘?’).

Headlines that carry no such pattern and merely leave a curiosity gap mostly do not get this label. A headline like ‘Here is that detail’ or ‘It shocked those who saw it’ steals the reader’s time just the same, but leaves no measurable trace; the system cannot see it.

Example: The headline ‘What does it mean to dream about water?’ receiving the low quality label — while ‘Here is that development concerning millions!’ does not.

Go deeper

The main goal of the low quality label is to push the content trying to steal the reader’s time to the bottom of the feed. In English low quality detection, the ML model’s success (F1: 0.55) is behind the Turkish model. The rule-based backup engine achieves an F1: 0.60 success by looking at lists starting with numbers or patterns like ‘how to’ / ‘guide to’ in English. While the recall rate provided by the ML model for low quality TR here is at a reasonable level like 0.61, the essential thing is keeping the precision (P: 0.99) very high in order to prevent the reader’s time from being wasted.