Why String Matching Isn't Enough
Copy-paste plagiarism is easy to catch. Change a few words and basic tools miss it entirely. "The economy grew by 5% last quarter" becomes "Last quarter saw 5% economic growth" -- same information, different words, invisible to string comparison. Sophisticated plagiarism uses paraphrasing, synonym substitution, and sentence restructuring.
This system uses three detection layers that catch progressively subtler plagiarism:
Layer 1: Document Fingerprinting
MinHash and SimHash algorithms create compact "fingerprints" of each document. Documents with similar fingerprints are flagged for detailed comparison. This layer is fast -- it can compare a document against 100,000 corpus documents in seconds. It catches copy-paste and light rewording.
Layer 2: TF-IDF Similarity
Documents flagged by fingerprinting go through paragraph-level TF-IDF cosine similarity analysis. Each paragraph is vectorized and compared against candidate source paragraphs. This layer catches restructured paragraphs where the same key terms appear in different order.
Layer 3: Semantic Matching
For the highest detection accuracy, the BERT-based semantic similarity model compares meaning, not just words. "The president vetoed the bill" and "The legislation was rejected by the head of state" are semantically similar despite sharing no content words. This layer catches sophisticated paraphrasing that evades keyword-based tools.
Originality Report
The report highlights matched sections in the original document with color coding: red for high similarity (>80%), orange for moderate (50-80%), and yellow for low (30-50%). Each match links to the source document or URL. An overall originality score summarizes the percentage of original content. Reports can be exported as PDF.
