Skip to content / 跳至主要內容
Technology & research

Classify how a student needs to respond

Selection, text entry and drawing call for different interactions. Classification connects the document to the right response interface.

Classification serves the interaction

A worksheet can contain multiple choice, written responses, drawing tasks and headings. Vispo's response classification asks a concrete question: what does the student need to do here?

The current pipeline distinguishes select, input, draw and header, with single- or multiple-selection modes and option labels. This is distinct from concept tagging, difficulty estimation or inferring student ability.

Connect structured output to the product

The inspected pipeline classifies questions in page-level batches using Claude Haiku 4.5, producing response types and related fields. A usable result needs a checkable structure, not just a model-generated label.

A confidence score is not accuracy

Model-reported confidence is a processing signal, not a calibrated probability of correctness. Ambiguous layouts, mixed response types and incomplete options require conservative handling and teacher review.

The practical test is whether students receive the right way to respond, not whether a classification score looks impressive.

References

  1. Guo et al. · On Calibration of Modern Neural Networks
Read more technical notes
Vispo|智加雲 AI 教學助手