Separating questions and answers without losing their connection
Numbering, sections and source positions connect each answer to an inspectable question and its original evidence.
Separate the sources before aligning them
Questions and answers may appear on different pages, and printed question numbers are not always unique. Vispo preserves their sources and page locations before mapping them together, including restarted numbering and parent–child question structure.
The current upload flow first tries the answer PDF’s text layer, using OCR results where needed. Alignment draws on numbering, hierarchy and available source context rather than list position alone.
Keep uncertain matches inspectable
Completeness checks identify items requiring review. When the structure does not support a clear match, source crops and review signals remain available; vision-based matching can provide additional support where applicable.
Teachers can move from an extracted item back to its source to check numbering, answers and crop locations before accepting it. Document understanding becomes an editing workflow they can inspect and correct.
Research context: answers need location evidence
An earlier answer-localization module explicitly cites LaAP-Net’s evidence-localization idea. The COLING 2020 paper pairs answer prediction with bounding-box evidence in text visual question answering.
Vispo adapts the idea of locating answer evidence through its own exam fields, section structure and matching checks. This does not imply reproduction or training of the paper’s complete neural network.