Latent semantic analysis (LSA) for plagiarism detection

Plagiarism detection is the task of identifying instances where content has been inappropriately copied or imitated without proper attribution. Traditional plagiarism detection methods often rely on string matching or fingerprinting techniques to catch verbatim copying. However, plagiarism has evolved beyond simple copy-paste to include paraphrasing, translation, and idea-based plagiarism, among other tactics. These sophisticated forms of plagiarism – sometimes termed intelligent plagiarism, as they preserve the original meaning while altering wording and structure – pose a significant challenge to conventional detectors (Kakkonen & Mozgovoy, 2010). In response, researchers have turned to semantic analysis techniques. These methods can recognise textual similarity in terms of meaning rather than exact wording. One such technique is Latent Semantic Analysis (LSA). It has emerged as a powerful approach for detecting paraphrased or contextually similar plagiarised content. LSA works by mapping documents into a semantic vector space.

This article provides an in-depth examination of how LSA is applied to plagiarism detection. We focus on its methodology, its effectiveness in catching paraphrased plagiarism, and its strengths and weaknesses. We discuss LSA’s ability to capture the underlying meaning of documents. We highlight examples of its use in detecting plagiarism in both text and source code. We also consider its performance in scenarios such as cross-language plagiarism. Furthermore, we analyse the advantages that LSA offers over traditional methods. Additionally, we outline the limitations and challenges associated with this approach.

Understanding latent semantic analysis

Latent semantic analysis is a mathematical and linguistic technique that represents text documents as vectors in a high-dimensional semantic space. The core idea is to capture latent meaning relationships between words and documents. This is achieved by analysing patterns of word co-occurrence in a large corpus (Landauer et al., 1998). Rather than treating words as independent features (as in a simple bag-of-words model), LSA leverages the context in which words appear. It constructs a large term–document matrix in which each cell reflects the weight of a word in a document. Often, a TF–IDF weighting scheme is used for this purpose. LSA then applies a dimensionality reduction—typically Singular Value Decomposition (SVD)—to this matrix to uncover latent semantic dimensions (Dumais, 2004). The SVD factorisation produces a lower-dimensional approximation of the matrix. Essentially, words and documents that frequently co-occur are grouped into the same latent semantic space (Dumais, 2004).

This process reveals hidden conceptual associations. Documents that share topics or meaning will end up near each other in the truncated vector space. Even if they do not share many exact words, they remain close in semantic terms. A document’s vector in the LSA space is often thought of as capturing the “topic” structure or meaning. LSA’s representation is derived from large-scale patterns of usage. For example, LSA can learn that “car” and “automobile” (or “research” and “study”) occur in similar contexts. It therefore treats such words as semantically related. The technique effectively sidesteps the limitations of exact keyword matching. It does so by abstracting textual content into a concept-based vector form (Landauer et al., 1998). LSA can therefore judge documents by meaning rather than exact phrasing. It can find two passages to be similar in content even if they have few or no identical words.

It is important to note that LSA is an unsupervised, corpus-driven method. The quality of the semantic space depends on the representativeness of the training corpus. It also depends on key parameter choices (such as the number of dimensions to retain). Researchers describe LSA as a highly parameterised approach. Its effectiveness is contingent on careful tuning of several factors. These include preprocessing steps (e.g., stop-word removal and stemming) and the term weighting scheme. Another critical factor is the number of dimensions used (Cosma & Joy, 2012). When properly configured, LSA yields a robust semantic representation of documents. This representation can be exploited for various information retrieval tasks. In plagiarism detection, the goal is to capture meaning rather than surface similarity, which LSA supports well.

Detecting paraphrased plagiarism with LSA

One major advantage of LSA in plagiarism detection is its ability to detect paraphrasing and semantic equivalence between texts. Plagiarists often try to mask copying by rewriting sentences, substituting synonyms, or altering the structure of the text while preserving the original ideas. These obfuscation tactics make plagiarism harder to detect through naive lexical checks. Traditional detection systems based on exact matching or simple lexical similarity struggle with such disguised copying. In contrast, LSA is inherently designed to capture the contextual similarity of documents. LSA maps texts into a space defined by concepts and word co-occurrence patterns. Two passages conveying the same idea will occupy nearby positions in that space, even if the actual words differ (Kundu & Karthik, 2017). In other words, LSA focuses on what the text means, not just the literal words used.

Consider a simple example. One author writes, “the apparatus was small and compact.” Another author writes, “the device was tiny and had a minimalist design.” A straightforward keyword comparison might find little overlap between these sentences (“apparatus” vs “device”, “small and compact” vs “tiny”). However, an LSA-based system would recognise that “apparatus” and “device” appear in similar contexts. It would also recognise that both sentences describe a small device (i.e., they have the same meaning). Thus, LSA would place these sentences close together in semantic space. It would flag them as similar in meaning despite the lack of any verbatim overlap.

This capability makes LSA particularly effective for detecting paraphrased plagiarism. Even “intelligent plagiarism” – where the plagiarist heavily rewords the source text but keeps its meaning – can be uncovered by LSA (Ali & Taqa, 2022). Such intelligent plagiarism (including paraphrasing, synonym substitution, translation, or idea theft) is notoriously difficult to catch with purely lexical algorithms. LSA offers a solution by matching documents based on semantics rather than surface text. Research studies have demonstrated LSA’s strength in this regard. For instance, Kundu and Karthik (2017) categorise plagiarism cases as either text-based similarity (direct overlap) or context-based similarity (same meaning). Their work uses LSA to perform context-based plagiarism detection. It effectively finds a “match between two pieces of text with respect to the implemented meaning or ideas implied” (Kundu & Karthik, 2017). Similarly, other researchers have found that LSA is highly sensitive to semantic similarity. It can detect when one text is a rephrased version of another because the two share a high semantic correlation. This holds true even if the wording differs.

For example, a study by Kakkonen and Mozgovoy (2010) noted that semantic methods like LSA can identify disguised plagiarism instances that involve extensive paraphrasing. In contrast, simpler keyword matching often fails on those cases. Moreover, studies have shown that LSA often outperforms traditional vector space models in identifying paraphrases. The dimensionality reduction in LSA gives it an edge over basic bag-of-words approaches. Empirical evaluations confirm that a well-tuned LSA model can achieve higher recall on paraphrased plagiarism than simple VSM or n-gram matching approaches (Cosma & Joy, 2012; Alzahrani et al., 2012). In other words, LSA finds more paraphrased plagiarism cases that basic lexical methods would miss. By focusing on meaning, LSA improves the detector’s ability to identify when two works are substantively the same. This holds true even if the wording has been cleverly altered to mask the similarity.

Of course, adopting an LSA-based approach does not eliminate all challenges. For example, extremely subtle forms of plagiarism that involve stealing ideas without any shared terminology can still be hard to detect. Nonetheless, LSA significantly expands the detection scope beyond simple copy-paste cases. Therefore, LSA serves as a critical component in modern plagiarism detection systems. It is especially useful for recognising paraphrase, summarisation, and other advanced plagiarism tactics that undermine academic integrity.

LSA in text plagiarism detection

Latent semantic analysis has been applied successfully to detect plagiarism in natural-language texts such as essays, articles, and reports. In extrinsic plagiarism detection (where a submitted document is compared to a collection of sources), an LSA-based system typically works as follows. It first builds a semantic index from a large corpus of documents (which may include known sources or a background corpus like Wikipedia). It then represents each document as a vector in the reduced semantic space. The submitted (suspect) document is similarly transformed into the LSA space using the same preprocessing and weighting. Plagiarism detection then boils down to comparing the query document’s vector to all source document vectors, usually by computing cosine similarity. A high cosine similarity indicates that two documents share a lot of semantic content.

Using cosine similarity in the LSA vector space is a common technique. Cosine measures the angle between two document vectors, effectively capturing how closely related the topics of the documents are. In practice, if a student paper and a published article have a cosine similarity above a chosen threshold in LSA space, it strongly suggests the student’s paper may be derived from the article. This holds true even if they do not share many exact words, since a high semantic similarity implies overlapping content. The choice of similarity threshold is important: one typically determines through experiments what cosine similarity value constitutes likely plagiarism, to balance catching plagiarism against false positives. In practice, systems might rank potential source matches by cosine similarity and present the top candidates for manual inspection.

Numerous research prototypes and tools have implemented LSA for text plagiarism detection. For instance, Gupta and Gupta (2015) describe a system that integrates LSA to go beyond simple string matching. They reported significantly better detection of reworded plagiarism cases when using LSA. Other authors have combined LSA with complementary techniques. For example, some systems use a thesaurus or WordNet to explicitly account for synonyms, further boosting semantic coverage (Stamatatos, 2011). The consensus in the literature is that incorporating LSA enhances recall for disguised plagiarism. Cases that would be missed by strict string matching are often uncovered when documents are compared in the latent semantic domain.

LSA is largely domain-independent and language-agnostic. One can train the LSA model on any subject matter or language, and it will capture that domain’s or language’s semantic structure. This means LSA can be used to detect plagiarism in diverse fields (from scientific articles to student essays) by training on the relevant corpus. Furthermore, because LSA is based on statistical co-occurrence rather than language-specific rules, it can be applied to any language given a suitable corpus. This flexibility has been demonstrated by studies applying LSA to languages beyond English, and even to cross-language plagiarism detection (Ratna et al., 2017).

Another benefit of LSA is the informative output it can provide to investigators. Its continuous similarity scores can be visualised or analysed to show which parts of two documents align semantically. For example, Cosma (2008) showed that LSA can highlight corresponding code fragments between two programs, allowing instructors to see exactly where plagiarism likely occurred. Such evidence is valuable for demonstrating the overlap to students or evaluators. Traditional tools highlight overlaps only when the text is identical. LSA goes further by indicating conceptual overlap even if phrased differently, thus providing deeper insight into the plagiarism.

LSA in source code plagiarism detection

Beyond natural language, latent semantic analysis has also been applied to plagiarism detection in programming source code. Source code plagiarism (e.g., students copying programming assignments) is a well-known problem in computer science education. Traditional code plagiarism detectors often use techniques like syntax tree matching, tokenisation, or string edit distance. LSA offers a different perspective by treating code as text and attempting to identify semantic similarities between programs.

Source code does not have synonyms in the usual sense. However, different variable names, altered comments, or reordering of code blocks can make plagiarised code look superficially different. To apply LSA, code files can be converted into a text-like form (for example, using only identifiers and keywords). We then treat each program as a “document” composed of those tokens. After building a term–document matrix for a set of code submissions, LSA is applied to reduce this matrix. Programs that implement the same functionality (or that derive from the same base code) tend to appear close together in the semantic space. This holds true even if the actual code has been altered significantly (Cosma & Joy, 2012).

Georgina Cosma and Mike Joy (2012) provided a foundational example of LSA applied to source-code plagiarism detection. They found that, with careful parameter tuning, LSA could effectively detect plagiarised Java programs. In their study, source code files were preprocessed (comments removed and identifiers normalised) and then indexed using LSA. The best retrieval performance was achieved after such preprocessing and by using a combined term weighting scheme, with cosine similarity to rank potential matches. Their experiments showed that an LSA-based approach, properly optimised, outperformed a standard vector space approach for detecting similarity in code. It was able to highlight segments of code with high semantic similarity even when those segments were not exact copies.

One innovative aspect of using LSA in code plagiarism is its ability to assist in the investigation phase. Cosma (2008) developed a tool named PlaGate that employs LSA to cluster suspicious source-code files and visualise their similarities. The system provides graphical evidence indicating which code fragments contribute most to the similarity between files. In practice, an instructor can see a visual map of the code alignment. Certain functions or blocks are highlighted if LSA deems them important to the match. This greatly aids in making a case for plagiarism. Rather than just providing a similarity score, the system pinpoints likely copied segments and shows how strongly they are related (Cosma, 2008). Traditional exact-match tools also highlight overlaps, but they are limited to identical strings. LSA goes further by suggesting conceptual overlaps that may not be word-for-word, thus broadening the scope of evidence available to the investigator.

In source code applications, LSA typically complements other analyses. Since programming languages have strict syntax, structural matching or token alignment is often used alongside LSA’s semantic comparison to improve accuracy. However, LSA adds value by being language-agnostic and noise-tolerant. Changes in identifier names or comments (a frequent trick by plagiarists) will have minimal effect on an LSA similarity, since those changes do not alter the underlying logic of the code. For example, renaming a variable from total to sum will change one token. However, two code files solving the same problem will still share a similar distribution of relevant terms (e.g., key API calls and operators). LSA can capture those similarities. As a result, LSA has been included in hybrid plagiarism detection frameworks for source code. Studies have found that it improves the detection of modified plagiarism cases (Zhang et al., 2014). Overall, LSA is not a standalone solution for code plagiarism (it ignores code structure). Nevertheless, it serves as a powerful filter to identify program pairs that feel similar in their use of programming elements. It can catch plagiarism that might be missed by purely syntactic analyses when the plagiarist has applied superficial disguises.

LSA in cross-language plagiarism detection

Plagiarism can also cross language boundaries. For instance, a student might take an article in one language and translate it into another without attribution. Detecting such cross-language plagiarism is inherently challenging, because the texts share no common wording if the translation is done well. However, LSA can be extended to cross-language contexts to find documents that are semantically similar across different languages.

One approach is to construct a bilingual latent semantic space. Researchers have experimented with feeding a parallel or comparable corpus (i.e., texts that are translations of each other or share topics across two languages) into LSA. This effectively binds two languages into one semantic space. Documents from both languages are then represented as language-independent concept vectors. In theory, a Spanish document about renewable energy and an English document on the same topic would end up close together in this bilingual semantic space. Their underlying concepts align, even though the actual words in Spanish and English are different. In practice, building a robust cross-language LSA model requires a significant bilingual corpus and careful alignment.

For example, Ratna et al. (2017) developed a cross-language plagiarism detector focusing on Indonesian and English. Their system translates the Indonesian documents into English at the word level (using a bilingual dictionary or translation tool). It then applies LSA on the translated texts. Because LSA ignores grammar and word order, a simple word-by-word translation was sufficient to capture the semantics in their approach. The authors reported that this LSA-based system achieved promising performance. In their experiments, it attained up to 87% detection accuracy on test data when optimally configured (Ratna et al., 2017). This indicates that even with basic translation, LSA can identify cross-language plagiarism by focusing on shared semantic content.

Another noteworthy aspect of Ratna et al.’s approach is the use of a classifier (Learning Vector Quantization, a neural network) on top of LSA to decide if a pair of documents is plagiarised. LSA provides features (such as similarity scores or norms), and the classifier learns to distinguish plagiarism based on those features. This reflects a general trend: LSA can generate semantic similarity metrics that are then fed into machine learning models for the final decision. Regardless of the classifier, the key point is that LSA enables the quantification of semantic overlap across languages.

Cross-language plagiarism detection via LSA is still an emerging area, and it faces challenges such as translation quality and the need for multilingual corpora. However, early studies have shown it to be feasible. For instance, experiments on English–German and English–Arabic plagiarism detection have shown that LSA (and similar semantic approaches) can achieve high recall. This is especially true when these methods are combined with other cross-lingual resources or translation aids (Franco-Salvador et al., 2016). The ability to detect translated plagiarism means that journals and universities can guard against a cunning form of academic dishonesty – one that exploits language barriers to avoid detection. LSA’s language-agnostic mechanism makes it naturally suited to be a backbone of such cross-language detection systems, given appropriate training data or translation support.

Strengths of LSA in plagiarism detection

Latent semantic analysis offers several compelling strengths when applied to plagiarism detection:

• Captures semantic similarity: The foremost advantage is LSA’s ability to identify similarity in meaning rather than just in wording. This dramatically improves detection of paraphrased or obfuscated plagiarism. It helps catch instances where the plagiarist used different vocabulary or structure but retained the original ideas. By focusing on content rather than form, LSA can reveal connections between texts that would evade purely lexical comparison (Landauer et al., 1998).
• Robust to minor edits and noise: LSA tends to be tolerant of small changes that plagiarists often introduce to throw off detectors. Changing a few words to synonyms, adding extraneous phrases, or slightly reordering sentences usually does not fool an LSA-based system, because these minor edits have little impact on the document’s semantic vector. This robustness leads to higher recall when detecting the common tricks of plagiarism. For example, if two documents differ in wording but are semantically the same, LSA will still recognise them as similar, whereas a simple string-matching approach might fail once too many words are changed.
• Domain and language flexibility: LSA is largely domain-independent and language-agnostic. An LSA model can be trained on any subject matter or language, and it will capture that domain’s or language’s semantic structure. Thus, LSA can be used to detect plagiarism in diverse fields (scientific articles, literature, programming code, etc.) as well as in multiple languages, simply by providing an appropriate training corpus. Studies have successfully applied LSA to languages beyond English and even to cross-language plagiarism detection (Ratna et al., 2017). This flexibility means LSA-based methods can be deployed in many different educational and professional contexts.
• Informative evidence for investigators: LSA can provide richer output to help understand and demonstrate plagiarism cases. Because it yields continuous similarity scores, those can be visualised or further analysed to highlight which parts of a document pair contribute most to the similarity. As discussed, systems like PlaGate (Cosma, 2008) use LSA to present graphical evidence of source-code plagiarism, highlighting corresponding fragments. Even in text plagiarism, an LSA-based system can underline or list semantically matching passages (not necessarily identical words) between a student’s essay and a source. Traditional detectors show exact overlaps, but LSA can show conceptual overlaps, giving instructors or editors deeper insight into how the content was copied. This can be very useful for presenting evidence to students or for legal/disciplinary proceedings, as it points out the theft of ideas, not just words.
• Complementary to other methods: LSA can be integrated with other detection techniques to create a more robust plagiarism detection system. Its outputs can serve as features in machine learning classifiers or as a filtering step before more fine-grained analysis. Because it addresses semantic similarity, LSA catches cases that others miss. Conversely, exact string matching can catch cases (like nearly verbatim quotes) that LSA might not flag as strongly because they use unusual terminology not well-represented in the semantic space. Combining methods ensures comprehensive coverage. In evaluations, semantic methods including LSA have been ranked as essential components for detecting advanced plagiarism (Jambi et al., 2022). In summary, LSA strengthens the overall detection pipeline by adding a semantic perspective, and it works best in concert with lexical and structural analyses.
• Proven and well-studied approach: LSA is a mature technique with a long track record in research. It has been widely used in plagiarism detection experiments, and its behavior is well understood. This means practitioners have guidance on how to tune LSA (e.g., choosing an appropriate number of dimensions for a given corpus size) and how to interpret its results. The widespread adoption of LSA in the plagiarism detection literature – it has been called one of the most studied semantic approaches for this task – attests to its reliability and impact (Zhang et al., 2022). Newer embedding models now exist, but LSA remains a solid, explainable choice with decades of supporting research.

Limitations and challenges of LSA

While LSA brings powerful capabilities to plagiarism detection, it is not without limitations. It is important to be aware of these challenges when deploying LSA-based solutions:

• Dependence on a representative corpus: LSA’s semantic space is only as good as its training data. If the corpus used to create the LSA model is too small or not representative of the domain of the documents being checked, the model may miss important nuances of meaning. For example, detecting plagiarism in biomedical research papers requires an LSA model trained on biomedical literature – a generic corpus might not effectively capture terms like “CRISPR” or “beta-amyloid.” Building and maintaining a proper corpus for each domain can be resource-intensive. Moreover, if a plagiarised text contains very unusual or technical terminology not seen in the training data, LSA might not map those terms well, reducing its effectiveness. In short, an LSA-based system must be used with a suitable corpus that covers the subject matter of the documents under scrutiny.
• Computational complexity: Constructing an LSA model involves computing an SVD on a large term–document matrix, which can be computationally expensive for very large corpora. While there are optimisations and incremental algorithms, the dimensionality reduction step can become a bottleneck, especially if the plagiarism detection system’s index needs frequent updating with new documents. In high-scale scenarios (e.g., scanning the entire internet or a vast publication database), a naive LSA implementation might be impractical. One can mitigate this issue by restricting the scope of LSA (for example, using it within a course’s submissions or a predefined repository) or by using more efficient updating techniques. Nonetheless, computational cost and scalability are considerations when using LSA for very large-scale or real-time plagiarism detection.
• Parameter tuning and sensitivity: LSA requires choosing a number of latent dimensions to retain, and its performance is quite sensitive to this choice. Too few dimensions may oversimplify the semantic space (merging distinct concepts together), whereas too many dimensions may retain too much noise (undermining the benefit of abstraction). Selecting the optimal number often requires experimentation or cross-validation. Similarly, decisions about preprocessing (e.g., whether to stem words, how to handle stop words) and term weighting schemes (TF–IDF vs. entropy weighting, etc.) can significantly affect results. There is no one-size-fits-all setting for LSA. A model that works well for one dataset or language might need re-tuning for another. This sensitivity means deploying LSA effectively demands expertise and careful calibration. A poorly tuned LSA model may perform no better than simpler methods, so effort must be put into parameter selection and validation.
• Interpretability and false positives: LSA produces dense vectors and similarity scores that are not as immediately transparent as exact matches. If an LSA-based system flags two documents as similar, it is not always obvious which parts of the text are responsible, without additional analysis. Investigators may need to use tools to extract the most contributing terms or segments (as some systems do by highlighting likely matching segments). This lack of direct interpretability can make it harder to explain the findings to authors or students compared to a side-by-side text comparison. Additionally, there is a risk of false positives: two documents discussing very similar topics might naturally end up close in semantic space even if one did not plagiarise the other. For example, different students writing about the same well-defined topic might use similar terminology and arguments, and LSA could incorrectly flag their work as too similar. To mitigate this, one must set similarity thresholds carefully and often corroborate LSA-based findings with other evidence (or a manual check). In practice, LSA is best used as a guide to potential issues, rather than an absolute judge – human review is still important to distinguish between coincidental similarity and actual plagiarism.
• Not a standalone solution: As powerful as LSA is, relying on it exclusively would leave certain gaps in plagiarism detection. LSA might not effectively detect very short plagiarised snippets, because a single sentence may not significantly influence a document’s overall vector. In such cases, exact substring matching might be more sensitive. Likewise, if plagiarised content is heavily interwoven from multiple sources (“mosaic” plagiarism), the semantic signature of any one source might be diluted, making LSA less confident even though copying occurred. Modern plagiarism detectors therefore typically use a multi-faceted approach: they combine LSA or other semantic methods with direct text matching, stylometry, citation analysis, and so on (Foltynek et al., 2020). LSA addresses the semantic aspect of plagiarism, but it should be complemented with techniques that address verbatim copying and writing style anomalies. In summary, LSA is a valuable component of an advanced plagiarism detection toolkit, but it should not be the only component.
• Competition from newer models: It’s worth noting that LSA, developed in the 1990s, has been joined (and in some respects surpassed) by newer language representation models. Techniques such as word embeddings (e.g., Word2Vec, GloVe) and contextual language models (e.g., BERT, GPT) also capture semantic similarity, often with greater nuance. These models have started to be applied to plagiarism detection research in recent years. For instance, BERT-based approaches can understand synonyms and context on a sentence level and might catch certain paraphrases that LSA could miss. One limitation of LSA is that it gives each word a single vector in the semantic space, regardless of context (so it struggles with polysemy – words with multiple meanings). Contextual models can overcome that by adjusting word representations based on surrounding text. While a full comparison is beyond our scope, it’s important to acknowledge that LSA is not the newest or most sophisticated semantic tool available. However, LSA remains popular due to its simplicity, interpretability, and the fact that it still performs well for many tasks. In practice, one might use LSA alongside newer methods or as a baseline. The concepts behind LSA have also paved the way for these advanced techniques. Going forward, it’s conceivable that LSA could be gradually replaced or augmented by deep learning models in plagiarism detection, but for now it continues to be a foundational and trusted approach.

Conclusion

Latent semantic analysis has significantly advanced the field of plagiarism detection by enabling comparisons of documents based on meaning rather than just matching strings. By representing texts in a latent semantic space, LSA captures the conceptual fingerprints of documents. This allows plagiarism detectors to identify paraphrased or otherwise disguised copying that would elude traditional methods. We have seen that LSA-based approaches can detect “intelligent” plagiarism – cases where content is heavily rewritten yet semantically equivalent – with a high degree of success, as documented in various studies. Whether it is applied to student essays, academic articles, or even computer source code, LSA has demonstrated an ability to find overlaps in ideas that are not obvious from a surface-level reading.

The strengths of LSA in plagiarism detection are clear. It greatly improves the detection of subtle plagiarism by focusing on synonyms and context, it is relatively language-independent, and it can provide investigators with helpful insights (such as highlighting the most similar passages between documents). At the same time, using LSA requires careful attention to its limitations. It must be trained on appropriate corpora, tuned properly, and ideally combined with other techniques to ensure high precision and recall. As plagiarism techniques evolve – and as new text generation tools (including AI) introduce further complexities – the importance of semantic analysis in detection will only grow. Indeed, LSA was a forerunner in this semantic approach, and it continues to be relevant in the era of advanced language models.

In conclusion, LSA enhances our ability to detect not just verbatim copying but also more subtle forms of plagiarism that rely on paraphrasing and idea re-use. It plays a key role in modern plagiarism detection systems, providing a layer of understanding that goes beyond exact words to the level of meaning. By focusing on what is being said rather than how it is said, LSA exemplifies a more comprehensive approach to identifying academic dishonesty. As educators and publishers strive to uphold integrity, tools like LSA – built on solid theoretical foundations and refined through years of research – remain essential in the ongoing effort to detect and deter plagiarism.

References

Ratna, A. A. P., Purnamasari, P. D., & Adhi, B. (2017).Cross-Language Plagiarism Detection System Using Latent Semantic Analysis and Learning Vector Quantization. Algorithms. 10(2), 69. DOI: 10.3390/a10020069.

Alzahrani, S. M., Salim, N., & Abraham, A. (2012). Understanding plagiarism linguistic patterns, textual features, and detection methods. IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews, 42(2), 133–149.

Cosma, G. (2012).An Approach to Source-Code Plagiarism Detection Using Latent Semantic Analysis. IEEE Transactions on Computers.

Cosma, G., & Joy, M. (2012). Evaluating the performance of LSA for source-code plagiarism detection. Informatica (Slovenia), 36(4), 409–424.

Kakkonen, T., & Mozgovoy, M. (2010). Hermetic and web plagiarism detection systems for student essays: An evaluation of the state-of-the-art. Journal of Educational Computing Research, 42(2), 135–158.

Kundu, R., & Karthik, K. (2017). Contextual plagiarism detection using latent semantic analysis. International Research Journal of Advanced Engineering and Science, 2(1), 214–217.

Landauer, T. K., Foltz, P. W., & Laham, D. (1998). An introduction to latent semantic analysis. Discourse Processes, 25(2–3), 259–284.

Further reading:

1. Wagh, V., Laddha, S., & Kadam, P. (2024).Detecting plagiarism using Latent Semantic Analysis. 2024 IEEE International Conference on Blockchain and Distributed Systems Security.
2. Ullah, F., Jabbar, S., & Mostarda, L. (2021).An intelligent decision support system for software plagiarism detection. International Journal of Intelligent Systems.
3. Soleman, S., & Purwarianti, A. (2014).Experiments on the Indonesian plagiarism detection system using latent semantic analysis. 2014 2nd International Conference on Information and Communication Technology (ICOICT).