Elide

Anonymize research interviews — privately, in your browser, no account needed.

Elide runs entirely on your device. Nothing is sent to any server.

What it does

• Import fieldwork documents: interviews, field notes, meeting notes or other (.txt, .md, .docx, .rescribe.json, .corpus.json)
• Automatically detect names, places, organizations, and dates
• Tag them with colored labels — or do it by hand
• Replace names with pseudonyms
• Variants (case, accents, hyphens) and homonyms are flagged for review (⚑) instead of being guessed
• Export the anonymized text as .txt, .pdf, or .docx
• Work on spreadsheets of quotes (.xlsx, .xls, .csv): choose which columns to anonymize, tag cells in a table view (search, sort, edit the other columns, delete rows), export back to .xlsx, .xls or .csv
• Export validated documents for analysis as a .corpus.json bundle

How to use

Just open index.html in your browser. No installation, no setup.

You can also use the hosted version on GitHub Pages.

Pipeline

Rescribe  ──.rescribe.json / .corpus.json (raw)──▶  Elide  ──.corpus.json (anonymized)──▶  analysis tool
transcript cleanup                                  anonymization + human validation        coding / analysis

Elide reads and writes the shared format described in corpus-schema/corpus.schema.json.

Import. A .rescribe.json carrying a corpus block, a single corpus-doc or a whole corpus-bundle is read directly: document, speaker and segment ids are kept, as are the metadata and the provenance chain. Older .rescribe.json files without corpus are still read (the transcript HTML is parsed). Plain .txt / .md / .docx files get stable segment ids (seg_0001…) once, at import. The re-editable Rescribe project embedded in a bundle (ext.rescribe) is ignored and never stored.

Export for analysis (project view → For analysis (.corpus.json)): only documents marked as validated, with no tag left to review, are included. Each document is sensitivity: "anonymized" and contains only:

• the anonymized text of each segment, with its id, speaker id, time and hash;
• speakers as tags ([ENQ], [ENQ·E_1], [PERS_2]…), never their names;
• the list of tags present (type and count), never their real value;
• the review status and the provenance chain (file names removed);
• the study title, date and mode — the interview title, transcriber name and free metadata fields stay on the device.

Before writing the file, Elide scans every value of the bundle for any real value it knows (entity bank, speaker names), ignoring case, accents, ligatures, apostrophe and dash variants, and invisible characters (zero-width spaces, soft hyphens). If one is found, the export stops and lists the segment and expected tag: fix it in the editor, or confirm that the occurrence is not identifying and use Export anyway. Words you deliberately kept as is must be confirmed one by one.

The scan has its own test suite, run against the shipped index.html (Node 18+):

node --test tests/scan.test.mjs tests/parser.test.mjs tests/table.test.mjs

What leaves the device: only the files you export yourself. What never leaves it: the raw text, the pseudonym key (real value ↔ tag), the entity bank, corrections and original file names. The private .elide.json project file still contains everything — it is a working file, keep it private.

Spreadsheets

At import, each column gets a role: text to anonymize (quotes, comments — every cell is tagged like interview text), participant / speaker (replaced by its tag, e.g. [ENQ·E_1]), keep as is (theme, code, page — editable in the table view) or do not import. Elide suggests the roles from the column names and contents. Each sheet of a workbook becomes a document; .csv files are read as UTF-8 or Windows-1252, with ,, ; or tab as separator.

Quotes themselves can't be edited in Elide (tags stay reliable): fix a typo in the source file and import it again. The spreadsheet library (SheetJS, Apache 2.0) is downloaded on first use only, like the .docx reader; your files never leave the device.

A workbook with several sheets can be exported back as a whole (Export → Whole workbook (.xlsx)): one anonymized sheet per imported sheet, in the original order, named after the titles chosen at import. Sheets that are not validated are listed first, with an Export anyway option. A participant gets the same tag in every sheet of the workbook.

In the export for analysis, one quote cell is one segment; the other columns travel in ext.elide.table and are scanned for real values like everything else.

Automatic detection

Elide includes a multilingual name-detection model (FR, EN, DE) that runs locally in your browser — no internet connection required after the first load.

Optional: connect a local spaCy or CamemBERT server for higher accuracy, or use Ollama for LLM-based anonymization.

Projects can be marked multilingual when interviews mix languages: the local servers then use a multilingual model (xx_ent_wiki_sm for spaCy, bert-base-multilingual-cased-ner-hrl for the HF server).

Languages

French · English · German

Changelog

0.3.4

• Security: imported .rescribe.json HTML is parsed in an inert document; segment text is escaped in the HTML/PDF exports; titles, project names and tags are no longer interpolated into inline handlers.

0.3.3

• Spreadsheet import: coding tables without a header row are recognised (column names guessed from content: No., Time, Code, Quote), theme rows become a Section column, column names can be edited in the import dialog, which is now readable with long names. New document types Coding table and Quote list (exported as kind other with meta.genre).

This is the beginning of the README. Read the rest at the original. Read the full README ↗