Cross-language plagiarism checking: how translated text reuse is detected

Changing the language of a passage can remove nearly every word-level resemblance to its source without changing the underlying argument, explanation or findings. That creates a particular challenge for plagiarism checking: how can software recognise borrowed material when the original and the submitted text use different vocabularies, grammatical structures or writing systems?

Cross-language plagiarism detection, also called cross-lingual plagiarism detection, addresses this problem by looking for relationships between texts written in different languages. Depending on the method, a system may translate material into a common language, compare bilingual word correspondences, identify shared concepts or analyse multilingual representations of meaning. The aim is to locate potentially reused passages and their sources – not simply to find identical wording.

However, detecting a relationship between two texts and establishing plagiarism are different tasks. A translated passage may be properly acknowledged, while two independently written passages may express similar ideas. A useful system therefore needs both technically effective matching and results that a person can meaningfully investigate.

What is cross-language plagiarism?

Cross-language plagiarism occurs when someone uses material from another language without adequately acknowledging its source. It can involve a close translation, a translated paraphrase or the reuse of another author’s argument and distinctive explanation. Translating the words does not, by itself, make the underlying contribution original.

The terminology varies. Academic papers use both cross-language plagiarism detection and cross-lingual plagiarism detection, often abbreviated to CLPD. Translated plagiarism and translation plagiarism are more immediately understandable descriptions for a general audience. Cross-language text reuse is a useful neutral expression when discussing the technical matching task without assuming misconduct. Cross-lingual semantic similarity refers to the related task of assessing how closely texts in different languages correspond in meaning.

Another relevant phrase is multilingual plagiarism checking, but it requires care. A service can support several languages while comparing each document only with sources in the same language. Accepting English, French and Arabic documents does not necessarily mean that an English submission will be checked against French and Arabic originals. Product documentation needs to establish whether comparison actually crosses language boundaries.

These terms overlap, but they should not be treated as interchangeable promises of functionality.

Translation is not inherently plagiarism

Using sources written in other languages is not the problem. The problem is presenting another person’s contribution without appropriate acknowledgement. Academic guidance on plagiarism applies to ideas and arguments as well as copied wording, so replacing words – or changing language – does not remove the need to cite a source.

Consider this invented example:

Spanish source: “La ampliación del horario de la biblioteca facilitó el acceso de los estudiantes que trabajaban por la tarde.”

English rendering: “Students with afternoon jobs found it easier to use the library once it stayed open longer.”

The sentences share little visible wording, but communicate substantially the same proposition.

If the English sentence reports a finding taken from the Spanish source, attribution is needed. Conversely, similarity alone would not establish that the writer actually used that source: context and evidence of dependence still matter.

When quoting a passage translated personally, writers should follow the required referencing style and identify their translation where appropriate. For example, MLA guidance distinguishes a writer’s own translation from an existing published translation. Crediting the original author and, where applicable, the published translator helps readers understand exactly what has been consulted.

How cross-language plagiarism checking works

A complete detection system must solve more than one problem. The established workflow separates candidate retrieval, detailed alignment and post-processing. This distinction matters because success at one stage does not guarantee success at the others.

Finding possible sources

Candidate retrieval identifies documents or passages worth investigating.

The system is not initially trying to make a final judgement about plagiarism. It is trying to reduce a potentially enormous collection to a manageable shortlist. Cross-language retrieval can use translated queries, bilingual resources, shared concepts or multilingual representations to find potentially related material. Research systems have combined relatively lightweight retrieval with more demanding subsequent analysis.

This creates a fundamental limitation: a system cannot align a submission with a source it never retrieves.

Even an excellent passage-comparison model will miss a source excluded during shortlisting. Evaluating the comparison model alone therefore tells us less than evaluating the complete search-and-comparison process.

Aligning the relevant passages

Detailed alignment asks which sections of the submitted document correspond to which sections of a candidate source.

Whole-document similarity is insufficient. Two long articles might discuss the same subject without copying, while an otherwise original dissertation might contain one translated paragraph. The system needs to locate the relevant fragments rather than merely declare the documents broadly similar.

Research on fine-grained cross-lingual alignment has therefore examined smaller units, dictionaries, term proximity and dynamic alignment techniques. These approaches aim to connect specific passages while filtering out weaker candidate matches.

Consolidating the findings

Post-processing turns individual detections into a coherent result. Adjacent matches may need combining, overlapping detections may need resolving, and the boundaries of the suspected reuse must be identified.

The PAN text-alignment evaluation framework makes this localisation requirement explicit: systems identify matching spans in both documents, including their positions and lengths. A useful result is therefore more informative than a single document-level percentage.

For a reviewer, the practical question is: Can I see the submitted passage, the corresponding source passage and the basis for the proposed match?

The main detection methods

The principal approaches include translation-based comparison, statistical alignment, character matching, conceptual representations, knowledge graphs and neural models. These are complementary families of techniques, not a simple sequence in which every new approach makes the earlier ones redundant.

Machine translation followed by monolingual analysis

The most intuitive approach translates one or both texts into a common language and then applies same-language comparison. Researchers often call this translation plus monolingual analysis, or T+MA.

For example, a system might translate a submitted French passage into English before comparing it with English sources. The subsequent analysis may examine wording, word sequences or meaning.

In their 2013 comparison, Barrón-Cedeño, Gupta and Rosso found that T+MA produced the best overall results among the approaches tested, with an alignment-based method close behind. That finding supports translation as a useful strategy, but not a universal claim that it is always the most accurate approach.

Its central vulnerability is the translation step. Where the translation changes a technical meaning, resolves an ambiguity incorrectly or renders an idiom differently, later comparison must work with that altered representation. Translation followed by paraphrasing introduces further variation. The supplied research identifies both resource availability and non-literal modification as important constraints on these workflows.

Cross-language character n-grams

Cross-language character n-grams, commonly abbreviated to CL-CNG, compare short sequences of characters rather than complete words.

Some languages retain enough shared written features for this apparently simple method to be useful. Potthast and colleagues found strong performance for related language pairs in their foundational evaluation. However, this does not mean that character matching can bridge arbitrary languages or writing systems: its usefulness depends on the surface similarities that remain available.

The practical appeal is efficiency. The limitation is equally clear: when translation removes the relevant character overlap, the method has little evidence left to compare.

Bilingual dictionaries and statistical alignment

Cross-language alignment-based similarity analysis, or CL-ASA, uses bilingual word relationships and translation probabilities to assess correspondence between passages.

Rather than producing a complete, readable translation, it asks whether words in one passage plausibly align with words in another language. Bilingual dictionaries derived from parallel texts can supply these relationships. The supplied research identifies this approach as particularly useful for direct translations.

Barrón-Cedeño and colleagues found that CL-ASA achieved higher precision than the other methods in their comparison, despite T+MA’s stronger overall results. This illustrates an important trade-off: one approach may find more reuse, while another produces fewer incorrect matches.

Cross-language explicit semantic analysis

Cross-language explicit semantic analysis, or CL-ESA, moves away from direct word correspondence towards shared concepts.

A passage is represented through its relationship to reference concepts, such as corresponding Wikipedia articles in different languages. Two passages can then be compared through those concept-based representations even when their wording differs.

Potthast and colleagues examined CL-ESA alongside character-based and alignment-based methods, demonstrating the value of conceptual comparison across language pairs.

The interpretative limitation is important: sharing concepts can establish topical relatedness without establishing reuse. Two independently written explanations of the same subject may be conceptually similar.

Multilingual embeddings and neural models

An embedding represents text as a numerical vector. In a suitably trained multilingual system, passages expressing similar meanings in different languages should have comparable representations.

This is not simply a matter of feeding two languages into any neural model. The representations must be trained or adapted so that cross-language comparison is meaningful. Reimers and Gurevych’s work on multilingual sentence embeddings, for example, trained a multilingual model to reproduce meaningful sentence representations across translated examples.

Cross-language plagiarism research also uses Siamese architectures: paired processing branches learn to compare two inputs. Bouaine and Benabbou’s 2024 study combined multilingual mBART features with a Siamese long short-term memory model to classify document pairs.

These methods offer a way to recognise correspondence beyond identical vocabulary. However, they still require a distinction between equivalent meaning, broad topic similarity and actual dependence on a source.

A numerical similarity score is not automatically a probability of plagiarism. A score of 0.92, for example, should not be interpreted as “a 92% probability that the author plagiarised” unless a separately validated probabilistic model supports that interpretation.

The underlying lesson is that semantic matching supplies evidence about relationships between texts, not a self-contained judgement about attribution.

Knowledge graphs and ontology-based methods

Knowledge-based approaches represent concepts, entities and their relationships rather than relying only on words.

The Cross-Language Ontology-Based Similarity Analysis method, or CL-OSA, introduced by Stegmüller and colleagues, represents documents using entities from Wikidata. Its design does not require full machine translation or pre-training on parallel corpora, and the authors reported particular benefits for detecting sense-for-sense translations in their evaluations.

A simple way to understand the approach is to imagine two passages mentioning the same institution, process and scientific concept under different language-specific names. Connecting those mentions to shared entities provides another basis for comparison.

More recent research combines meaning and grammatical structure. Bouaine and colleagues’ 2026 study combines WordNet semantic information with Universal Dependencies syntax in knowledge graphs, using graph-learning methods to compare multilingual material. This represents a hybrid direction: relationships between concepts and relationships between words both contribute to the analysis.

These methods still depend on the resources they use. Representing a passage through a knowledge base does not make missing, ambiguous or poorly represented concepts disappear as a problem.

Citation-based comparison

Scholarly writing contains another possible signal: its references.

Citation-based plagiarism detection examines patterns such as the order and proximity of citations. Those patterns can survive translation even when the surrounding wording changes substantially. Gipp’s work on citation-pattern analysis presents this as a complementary approach to detecting disguised and cross-language reuse.

The idea is particularly relevant to research literature, but its interpretation requires restraint. Two papers may legitimately cite the same standard works. An unusual sequence supporting closely corresponding passages warrants investigation; a shared bibliography alone does not establish plagiarism.

Hybrid systems

A practical architecture can combine several of these methods.

For example, Avetisyan and colleagues’ 2023 preprint uses multilingual thesauri for candidate retrieval and multilingual BERT-based models for detailed analysis. Its approach avoids machine translation during use, although that should not be confused with an absence of translated resources anywhere in development or training.

Roostaee and colleagues similarly proposed a two-level alignment approach: an initial stage selects promising fragment pairs, followed by more precise sentence-level analysis using word relationships.

The design principle is economical: use faster methods to narrow the search, then spend more computation on the strongest candidates.

Why translated plagiarism remains difficult to detect

Language coverage is uneven

Supporting a language is not the same as performing equally well in it.

The availability of parallel corpora, bilingual dictionaries, translation systems and other linguistic resources varies. A model’s performance on English–French material does not establish its performance on English–Hindi or Arabic–English material. The supplied research identifies resource scarcity and linguistic differences as continuing obstacles, particularly for distant and less-resourced language pairs.

English–Arabic research by Alotaibi and Joy illustrates why multiple signals may be useful: their sentence-level approach combines semantic and syntactic features, including word order, embeddings and alignment. The language pair is not treated as a simple word-substitution problem.

Translation can be combined with paraphrasing

A translated passage does not necessarily preserve the original sentence boundaries, word order or level of detail.

Sense-for-sense translation, restructuring and paraphrasing make the correspondence less explicit. Research comparing continuous-space and knowledge-graph representations has therefore examined different lengths and types of disguised reuse, rather than treating all translated passages as equivalent cases.

There is also a distinction between finding a close translation and recognising a compressed summary. A paragraph may draw on a much longer source section, making one-to-one sentence comparison a poor fit.

For evaluation purposes, literal translation should therefore be only one category—not a substitute for testing more demanding transformations.

Related meaning can produce misleading matches

A powerful semantic model may correctly recognise that two passages mean similar things, yet still provide weak evidence of copying.

Consider these invented sentences:

“Poor ventilation can allow airborne contaminants to accumulate indoors.”

“Indoor pollutants may build up when air exchange is inadequate.”

Even across languages, these statements could arise independently. They express a general relationship rather than a distinctive contribution.

As a practical interpretation principle, a match deserves more attention when it preserves an unusual combination of details, a distinctive explanation or the same sequence of specific claims. Broad topic overlap deserves less weight.

This is why semantic similarity should be treated as the beginning of investigation rather than its conclusion.

The original source may be unavailable

Cross-language checking has a coverage problem as well as a language problem.

The logical consequence of a retrieval-based system is that unavailable sources cannot contribute to its matches. If the relevant material is absent from the searchable collection, even a strong alignment method has nothing to compare against. Candidate retrieval and detailed analysis are separate tasks precisely because locating the right material cannot be assumed.

For that reason, “no match found” is a statement about the search performed—not a guarantee that no source exists.

A meaningful evaluation should distinguish failure to retrieve an available source from the absence of that source altogether.

Short passages and long documents create different problems

Short passages may contain too little distinctive information to support confident comparison. Long documents create the opposite challenge: a small reused section can be obscured by a large amount of unrelated material.

This helps explain the research emphasis on fine-grained alignment and multiple comparison levels. Roostaee and colleagues’ two-stage method, for example, narrows candidate fragments before examining their sentence-level relationships more closely.

A useful test should therefore include short insertions inside long documents, not only pairs of passages already known to correspond.

Accurate comparison must also be computationally affordable

Comparing every passage with every possible source using an expensive model is difficult to scale.

The supplied research identifies hybrid pipelines as a response: lightweight candidate filtering through keywords, concepts or entity indexes, followed by more expressive comparison of selected segments.

This creates a balancing problem. Aggressive filtering saves computation but can discard genuine sources. A broader shortlist preserves more possibilities but increases processing and review costs.

The most useful system is therefore not necessarily the model with the highest score on a small comparison task. It is a complete workflow that maintains worthwhile detection performance at the required scale.

What accuracy figures actually tell us

Research on cross-language plagiarism contains impressive results, but headline percentages can conceal substantial differences in what was tested.

Pair classification is not the same as finding an unknown source

Bouaine and Benabbou reported 98.83% accuracy for English–French document-pair classification using their mBART and Siamese model. Their reported accuracy differed across the tested language pairs.

That is evidence about the evaluated model and dataset. It is not evidence that a deployed system will discover 98.83% of all translated plagiarism across the internet.

A model given two documents has a narrower task than a system required to discover the relevant source, locate the reused section and present interpretable evidence.

Difficult cases can produce very different results

Omar, Hamouda and Altohami investigated Arabic literary material translated from an English short story. Their evaluated approaches—Rabin–Karp, KNN and Word2Vec—achieved overall accuracies of approximately 23–26%. The material included direct translation, paraphrasing and original content.

The scope matters: this was a study based on one literary text and three particular approaches. It should not be generalised into a claim that every Arabic plagiarism checker performs at that level.

Equally, the high results in the mBART study should not be generalised to all translated writing. The two studies are not directly comparable league-table entries. They illustrate why task, genre, dataset, language pair and model must accompany any accuracy claim.

Precision, recall and localisation matter

Accuracy alone is inadequate.

As a hypothetical example, suppose a dataset contains 9,900 non-plagiarised items and 100 plagiarised items. A system that labels every item “non-plagiarised” achieves 99% accuracy while finding none of the plagiarism.

Evaluation therefore needs more informative measures:

The PAN evaluation framework assesses precision, recall and granularity, combining them in the PlagDet measure. Such measures address aspects of passage detection that a document-level accuracy figure can miss.

At the retrieval stage, a separate question remains: did the correct sources survive shortlisting?

What a convincing evaluation should include

A practical evaluation should include literal translations, substantial paraphrases, short reused fragments, unrelated passages and independently written material on the same topic.

It should also test the languages, document types and source collections relevant to the intended use. Mixed-language documents and imperfect text extraction are worth testing explicitly rather than assuming that success on clean, single-language inputs will transfer.

These are evaluation recommendations, not claims that every published study omits such cases. Their purpose is to make the test resemble the task the system will actually perform.

What to check in a multilingual plagiarism checker

Product claims need to be read at the level of individual features, language directions and settings.

Check the direction of comparison

Turnitin’s documentation describes Translated Matching as an optional feature that translates supported non-English submissions into English and compares the translated text with English-language content. Its standard and translated reports are distinct.

That documented workflow should not be interpreted as a promise of every possible language-to-language comparison. Translating a French submission into English to find English sources is a different operation from checking an English submission against French originals.

Turnitin’s enhanced-report guidance explicitly notes that an English translated report may be unavailable when the submitted text is already English.

Distinguish submission languages from result languages

Copyleaks’ API documentation makes a different distinction: it lists nine supported languages for uploaded documents and more than 30 result languages for cross-language detection. It also describes cross-language checking as a configurable feature that uses additional credits for selected languages.

The practical lesson is broader than either vendor: a language count does not describe the full capability.

Documentation should make clear which language the submission can use, which source languages can be searched, whether the feature must be enabled and whether additional processing is charged. These examples describe published functionality, not an independent comparison of the vendors’ detection accuracy.

Look for evidence that can be inspected

A useful purchasing or evaluation question is not merely “What percentage does it report?” but “What does the report allow a reviewer to verify?”

Prefer outputs that identify the source, locate both passages and distinguish original text from any machine-generated translation used to explain the match. Check whether exclusions and report settings are visible.

For unpublished or confidential work, also examine the service’s retention, indexing and sharing terms before uploading. A technically capable checker may still be unsuitable for a particular document-handling requirement.

Cross-language plagiarism detection is not AI detection

Cross-language plagiarism checking and AI-writing detection address different questions.

The first looks for relationships between a submitted passage and source material in another language. The second attempts to identify characteristics associated with AI-generated or AI-modified writing. Turnitin explicitly states that its AI-writing percentage is separate from and independent of its similarity score, and warns against treating AI-detection output as the sole basis for adverse action.

The distinction produces several straightforward consequences. A person can translate and plagiarise a source without using AI. A source match does not establish that an AI tool was involved. Conversely, the absence of a source match does not establish that the submitted wording was written without AI assistance.

Whether translation assistance is permitted is therefore a separate policy question from whether the underlying material has been properly acknowledged.

How to review a suspected translated match fairly

A sensible review should begin with the passages, not the headline score.

Read enough of both documents to understand what is actually shared. Check whether the source is cited and whether the apparent correspondence concerns a distinctive contribution or a general statement. Where the translation is important to the allegation, involve someone capable of assessing the relevant languages rather than relying solely on an automatically produced rendering.

Consider alternative explanations, including a common earlier source or a legitimate translated version. Similarity between two documents does not, on its own, establish which author relied on which text.

University of Georgia guidance explains that similarity results require interpretation and can be affected by quotations, references, templates and other legitimate matches. A score should not replace examination of the underlying evidence.

A defensible review record should therefore distinguish the observed match, the proposed explanation and the eventual judgement. Those are related, but not identical, findings.

Conclusion

Cross-language plagiarism checking is not simply ordinary plagiarism checking with extra languages added. It requires methods that can connect meaning, structure or other evidence across linguistic boundaries – and a retrieval process capable of finding the relevant sources.

Translation-based comparison, bilingual alignment, multilingual embeddings, knowledge graphs and citation analysis each offer useful signals. Their value depends on the language pair, the kind of reuse, the available sources and the way the complete system is evaluated.

The strongest result is not an unexplained percentage. It is a traceable match that helps a reviewer determine whether another author’s contribution has been used without adequate acknowledgement.