The Effect of Text Reuse Length and Fragmentation on Similarity Detection: A Controlled Evaluation of WordBinary
This study demonstrates that while WordBinary's text similarity detection performance systematically declines as reused content is fragmented into shorter passages, the system remains robust and effective, successfully recovering over 90% of deliberately reused words even when dispersed across numerous small fragments.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of writing, the line between borrowing an idea and stealing words is often drawn by software. These programs, known as similarity detectors, scan a document to find text that matches other sources. They are not mind-readers; they cannot know if a writer intended to plagiarize or if they simply forgot to cite a source. Their job is strictly to find overlapping words. But a fundamental question remains: does the software care how those words are arranged? If a writer copies a long paragraph from a book, the software sees a large block of matching text. If that same writer takes the exact same words and scatters them throughout their essay in tiny, disconnected snippets, does the software still see the theft? This is the core puzzle researchers set out to solve. They wanted to know if the physical shape of copied text—whether it is a solid block or a shattered mosaic—changes the ability of a detector to find it.
To answer this, a researcher named Kiah Curan designed a controlled experiment using a specific tool called WordBinary. The goal was to test the system under conditions that mimic how a writer might try to hide copied material. The researcher started with fifty original documents. For each document, they created seven different versions. In every single version, they inserted exactly 300 words taken from a known source. The only thing that changed was how those 300 words were distributed. In the first version, all 300 words appeared as one long, unbroken passage. In the next, they were split into three chunks of 100 words. The researcher continued this process, breaking the text into smaller and smaller pieces: six chunks of 50 words, ten chunks of 30 words, fifteen chunks of 20 words, twenty chunks of 15 words, and finally, thirty tiny chunks of just 10 words each. This created a total of 350 test documents, containing 105,000 words of deliberately copied text and 4,250 separate fragments.
The results showed that the shape of the copied text matters, but not in a way that causes the system to fail completely. When the 300 words appeared as one continuous block, the software found 97.3% of them. As the text was broken into smaller pieces, the detection rate slowly declined. When the words were split into thirty tiny 10-word fragments, the software found 80.3% of them. This means that while breaking up the text makes it harder to find, the system still recovered more than four out of every five words, even in the most difficult scenario. The study also looked at whether the software could spot the individual pieces. It found every single long passage and every 100-word chunk. Even when the text was broken into 10-word snippets, the software successfully identified 78.5% of those individual fragments.
A crucial part of the experiment was ensuring that the software wasn't getting confused by the rest of the document. The researchers checked the "background" text—the parts that were not copied—to see if the fragmentation made the system see matches where none existed. The background similarity remained steady at about 2.2% to 2.4% across all tests, proving that the drop in detection was due to the difficulty of finding the short fragments, not because the system was malfunctioning or becoming erratic. The study also noted that the system performed exceptionally well on fragments of 30 words or longer, finding 96.9% of those passages. This suggests that the software does not rely on finding massive blocks of text to do its job; it remains effective even when the copied material is scattered, provided the pieces are not too small.
The findings clarify that while fragmentation does reduce the ability to detect copied text, it does not render the software useless. The decline in performance was gradual rather than sudden. There was no specific point where the system stopped working; instead, it became slightly less efficient as the pieces got smaller. This is important because it shows that trying to hide plagiarism by breaking text into tiny pieces is not a foolproof strategy. The system continues to recover the vast majority of the reused material, even when it is dispersed throughout a document. The study concludes that while the physical arrangement of text influences detection, the software maintains a high level of robustness, successfully identifying most reused words regardless of whether they appear in a single paragraph or in dozens of small fragments.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.