Plagiarism and Text Reuse: Detection Methods and Evidence

It helps to start with a clear map before going near detection algorithms, because each type of reuse has a different evidentiary weight and calls for a different expert response.

A key concept across all these categories is the distinctiveness of the shared material. Finding the phrase 'the data show' in two papers is trivial. Finding the same unusual three-sentence construction and the same idiosyncratic spelling error in both is not. Forensic linguists focus on what is distinctive, rare, and unlikely to arise independently, rather than on the total percentage of matched text.

Detection software typically works in three generations of increasing sophistication. Understanding the mechanics explains why a report from any one tool is an incomplete picture.

1. First generation: fingerprinting
   Rabin-Karp and related rolling-hash methods break both texts into overlapping windows of characters or words (n-grams), hash each window to a fingerprint, and store those fingerprints in an index. Matching fingerprints across documents identify shared passages. The method is fast and scales to enormous corpora, but only catches verbatim or near-verbatim copies. A single word change in every sentence defeats it.
2. Second generation: vector-space models
   TF-IDF weighted bag-of-words vectors represent each document as a point in a high-dimensional space, and cosine similarity measures the angle between two points. Documents using the same vocabulary at similar frequencies cluster together. This catches paraphrase that preserves vocabulary but not order, and handles synonym substitution poorly.
3. Third generation: neural sentence embeddings
   Models like BERT, SBERT, and LaBSE encode sentences as dense semantic vectors. Two sentences meaning the same thing in different words land close together; two sentences using the same words with opposite meanings land apart. Cross-lingual models like LaBSE align languages in the same space, making translation plagiarism detectable. These methods are the current best practice for semantic and cross-language similarity, but they require more computational resource and produce scores that are harder to explain to a lay audience.

Cross-language plagiarism is more common than most institutions realise. A researcher reads a Spanish paper, translates the argument into English, and submits it as original work. Because no English string in the output matches the Spanish source, standard detectors give a clean result. The same logic applies to Japanese-to-Chinese copying in scientific publishing, or French-to-Portuguese in academic theses.

The detection approach requires mapping both texts into a shared meaning space. Multilingual sentence encoders, trained on parallel translation corpora, assign similar vectors to sentences that express the same content regardless of language. Once both documents are embedded, cosine similarity across sentence pairs flags candidate matches, which a bilingual reviewer then verifies.

• Loose machine translation before comparison is a cheaper alternative, but introduces errors that lower detection precision.
• Reference databases for cross-language detection are far smaller than English-only corpora, limiting recall in less-indexed language pairs.
• Expert bilingual analysis is still required to distinguish coincidental conceptual overlap from genuine copying, particularly in technical fields where vocabulary is constrained.

The factual question, 'did this person copy from this source?', looks similar in both academic and legal contexts. The standard of proof and the definition of actionable similarity are different in each, and a forensic linguist must calibrate their analysis accordingly.

The practical implication: a passage paraphrased from a textbook may constitute academic plagiarism even though no copyrightable expression was reproduced. Conversely, a verbatim copy of a factual database entry may not infringe copyright if facts lack originality under the applicable jurisdiction's law. An expert who applies the wrong framework to the wrong forum will confuse rather than assist the decision-maker.

Courts routinely ask why they need an expert if a software report already shows 42% similarity. The answer is that the number by itself means almost nothing. Consider two scenarios. In the first, a 42% match comes from an entirely shared legal disclaimer and references section, with the body of the text showing no overlap. In the second, a 12% match comes from a single 300-word block, verbatim, that contains the core original argument of the source. The second scenario is the more serious case, but the software scores it lower.

• Identifying which matched material is distinctive versus common: boilerplate, citations, and genre formulas match across texts innocently.
• Assessing directionality: which text came first, based on external evidence and internal features such as errors that were copied forward.
• Evaluating the uniqueness of matching passages using corpus frequency data: a rare three-word phrase appearing in both texts has far more weight than a common one.
• Ruling out independent creation from a common source: if both parties drew on the same publicly available government report, their texts may overlap without either copying the other.
• Explaining what the similarity score means in plain terms that allow the court to weigh it proportionately.

The challenge of explaining a similarity score to a lay audience is not merely pedagogical. Courts have excluded or heavily discounted expert testimony that rested on software output without adequate explanation of the method's assumptions, its validated accuracy, and the gap between the raw score and the conclusion.

Under Daubert and Kumho Tire in the United States, a court may ask whether the detection method has a known error rate, has been subjected to peer review, and is generally accepted in the relevant scientific community. Neural sentence embedding for cross-language plagiarism is relatively recent and its validation for forensic purposes is still developing, which an expert must disclose. Older fingerprinting methods have far more documented use and published accuracy data.

• Disclose the tool, its version, and the comparison corpus used.
• Separate matched passages into meaningful categories: distinctive expression, common genre language, and citations.
• Present findings visually where possible: a side-by-side comparison of key passages is often more persuasive and more honest than a number.
• State the limits of the analysis explicitly: what the similarity shows, and what it cannot rule out.