Data extraction
A form built from templates, by hand or proposed by the AI; manual, AI pre-fill or AI automatic modes; quotes checked against the study; a compare view for several extractors; tables and quality scores for the paper.
On this page
Extraction turns each included study into structured data: the values your charts, tables and synthesis are built from.
The form
Build the extraction form from:
- Templates, for example the CS techniques (TinyML-style) template or the Quality checklist;
- by hand: text, number and choice fields;
- Ask AI: it proposes a form from your research questions and a sample of studies.
Choice fields (single or multiple) become the categories of the paper's charts and taxonomy. A field can have a parent field, so that its options apply under one parent option (for example sub-techniques under a technique).
Modes
| Mode | Who extracts | What the paper says |
|---|---|---|
| Manual | People fill the form; the AI is not used | Human extraction |
| AI pre-fills | The AI extracts first; each extractor starts from its values and quoted evidence, accepting or correcting each field; a person marks the study done | AI-assisted, human-verified |
| The AI extracts | The AI's values count as the study's data until a person edits or finalises them; fastest | Disclosed as AI extraction |
Quotes are checked
Every value the AI extracts comes with a verbatim quote. The app looks each quote up in the text it read (the PDF, or the abstract when there is no PDF), part by part when it leaves words out, and checks that every number in a value appears in the study. What cannot be found is marked red for the person who verifies the values. It is not necessarily wrong (a PDF's text layer can garble tables and formulas), but check it against the paper before accepting it. The paper's facts report how many quotes were found.
Several extractors
With two or more extractors, the compare view shows every extractor's values side by side (AI values marked), with differences highlighted. Pick a value with Use, or type the agreed value, then save the final version: that is what the tables, charts and manuscript use.
Tables for the paper
- Table A (approach and mechanism) and Table B (results and cost: task and dataset, the headline result as "metric: baseline → proposed", hardware): mark the fields for each table in the form.
- Each study has a main category. Table A lists every study in a category; Table B only the studies whose main category it is, and the writer gives a study its full account only in its main category's section.
- With "Raw numbers stay in the tables" on, the prose explains mechanisms and trade-offs and points to Table B instead of reciting numbers.
The quality checklist
Quality questions are extraction fields answered yes, partly or no (scored 1, 0.5 and 0), so they get the same AI pre-fill, quoted evidence and human verification. The CS techniques template includes seven questions adapted from Kitchenham and Charters (2007), also available as the Quality checklist template. The paper gets a quality table in the appendix (PRISMA 2020 item 18), and the methods report the checklist and the score distribution (item 11).
While extracting, use Note for the paper to send an observation to a manuscript section, tagged with the study's citation key.