Document Similarity Checker (Offline)

About the Document Similarity Checker (Offline)

Add 2 to 50 documents — essays, assignments, reports or articles, as Word, PDF, OpenDocument, RTF, HTML, Markdown or plain-text files, or pasted text — and see which of them share passages word for word. Every pair is compared: a similarity matrix shows how much of each document is found in each other one, a ranked list puts the most similar pairs first, and the side-by-side view highlights every shared passage in both documents.

Quotations, reference lists and text every document may contain (the assignment question, a template) can be left out, so the results show the copying that matters. The documents are read and compared on your device and never uploaded. It compares only the documents you add, with each other: it does not search the internet, so it is not a plagiarism check against published sources.

Common uses

• Teachers checking a class set of essays or lab reports for students who copied from each other.
• Comparing several drafts or versions of a document to see what each one kept.
• Editors and content teams checking a batch of articles for duplicated text before publishing.
• Researchers comparing survey answers, interview transcripts or applications for text reused across them.

How it works

Each document is split into words (capitals ignored unless you say otherwise, punctuation never compared). Every run of k consecutive words (5 by default) is turned into a number, a hash, and the document keeps a sample of them as its fingerprints: in every window of hashes it keeps the smallest. This is winnowing, the method of Schleimer, Wilkerson and Aiken (ACM SIGMOD paper); their paper reports that the MOSS copy-detection service, used mostly for programming assignments, runs on it. With a window of t − k + 1 hashes, any passage of at least t words that two documents share is guaranteed to give both of them the same fingerprint, while runs shorter than k words can never match.

Every fingerprint two documents share is then checked word by word and grown to the full length of the shared passage, the longest passages are taken first, and no word counts twice. The share of a document is the number of its compared words inside shared passages divided by all its compared words. Resemblance is Andrei Broder’s measure (“On the resemblance and containment of documents”): of all the different k-word runs in the two documents, the share that occur in both.

What is left out

• Quotations (when you tick Ignore quotations): text between paired quotation marks in one paragraph — “…”, "…", „…“, «…», 「…」 and ‘…’ (a single quote counts only where it cannot be an apostrophe).
• Reference lists (on by default): everything from the last heading such as References, Bibliography, Works Cited, Literaturverzeichnis, Bibliografía or संदर्भ सूची, on a line of its own, to the end of the document or to an appendix after it. A heading near the start, such as a table of contents, is not taken as the start of the references.
• Text to ignore: any run of k words that it shares with a document. Use it for the assignment question, a template or text you handed out.

Left-out text is shown struck through in the side-by-side view, and a passage never runs across it.

Reading the results

A high share is a reason to read the highlighted passages, not proof that anyone copied. Students who answer the same question often repeat its words, set phrases appear in many texts, and correctly quoted and cited text is shared legitimately. Look at where the shared passages are and how long they are: one long passage of ordinary prose is far more telling than several short stock phrases. Choose a longer passage length (12–20 words) for long documents, where short shared phrases happen by chance.

Limitations

• It compares only the documents you add, with each other. It does not search the internet or any database, so it cannot tell whether text was copied from a website, a book or a document you did not add.
• Matching is word for word: reworded passages, translations and synonyms are not found, and a few changed words split a passage into shorter ones.
• Scanned PDFs have no text layer: make them searchable with OCR PDF first. Old .doc files must be saved as .docx; password-protected PDFs must be unlocked first.
• Up to 50 documents, 2 million characters each and 10 million in all. The side-by-side view shows the first 150,000 characters of each document; the passage list and the reports cover everything.
• A reference list is recognised only by its heading on a line of its own, and block quotations without quotation marks are compared like any other text.
• In extremely repetitive text (the same phrase used well over a hundred times in both documents), only the first uses of that phrase are matched, so that the comparison stays fast.

Privacy

Your documents are read and compared in your browser and never uploaded or stored. Only your option choices are remembered in this browser.