← All method guides

Annotated reading

How Much Annotation Is Enough for Language Learning?

The tempting version of language annotation is also the least readable: label everything. Every noun gets a feature, every verb gets three tags, arrows cross the sentence and the learner is left with a technically rich object that no longer behaves like text. Metkagram takes the opposite position for learner-facing reading. The underlying data may be rich; the visible layer should show only what earns attention on this pass.

Annotation competes for attention with the sentence

A visual cue works because it becomes salient. That is precisely why twenty cues can become a problem: they all ask to be looked at. Research on textual enhancement studies this general problem of drawing attention to target forms. The findings justify experimenting with salience, but they also give no basis for the simplistic rule that maximum salience is maximum learning.

For a learner-facing page, the sentence should still win the competition. If the first thing you notice is a forest of badges and only later discover there was a sentence underneath, the interface has inverted the priority.

Choose annotation from the learning question

Start with the question, not the available labels. “Where is the main verb?” may need one cue. “How is this German clause organised?” may need two or three. “What is the reusable core of this expression?” may be better served by a pattern span than by traditional grammatical roles.

This rule also prevents annotation from becoming an encyclopaedia. A cue is shown because it helps the current decision. Facts that are correct but irrelevant can stay in the data and remain hidden.

Use layers rather than one permanent maximal view

The same sentence can support several passes. One view may show the main verbal structure; another may show a reusable construction; a richer research view may expose morphology or provenance. These layers should not all be forced into one learner screenshot.

Metkagram’s annotation architecture separates the original text from spans, which makes selective rendering possible. The learner can receive a small signal while the system keeps a more inspectable record underneath.

A practical test: remove every Mark that cannot justify itself

For each visible cue, ask what the learner can do after seeing it that was harder before. Locate a verb? Compare word order? Extract a Frame? Avoid a predictable mistake? If the answer is merely “it is linguistically true,” the information may belong in an inspector or reference view rather than the reading layer.

Minimal does not mean vague. The remaining Marks must still be explicit enough to understand without relying on colour alone. Selectivity is about reducing noise, not hiding meaning.