Sentence Embedding
Definition
A neural representation of a sentence as a dense numeric vector, trained so that semantically similar sentences land near each other in vector space regardless of wording. Used for semantic plagiarism and cross-language comparison.
- Output type
- Dense numeric vector per sentence
- Training goal
- Semantically similar sentences land near each other
- Independent of
- Exact wording or word order
- Forensic use
- Semantic plagiarism and cross-language comparison
- Field
- Text reuse and forensic linguistics
Common questions
How does sentence embedding catch plagiarism that word-matching tools miss?+
A paraphrased passage can share almost no exact words with its source, but if the meaning is preserved, the two sentences' embedding vectors still land close together in vector space, letting the method flag reuse that surface string comparison would miss entirely.
Why is sentence embedding useful for cross-language plagiarism detection?+
Some embedding models are trained on multilingual corpora so that a sentence and its translation map to nearby vectors, which allows detection of a passage copied from one language and translated into another, a case exact or fuzzy string matching cannot handle.
What is a limitation of relying on sentence embeddings as evidence?+
Embedding similarity is probabilistic and model-dependent, so two sentences on the same narrow topic can score as similar without actual copying, meaning a high similarity score needs corroborating evidence, such as structural or stylistic overlap, before being treated as proof of reuse.
Related terms
- Mosaic Plagiarism
- Borrowing phrases and expressions from a source and embedding them across new sentences, so no single passage is a direct copy but...
- Paraphrase Plagiarism
- Reproducing the ideas and structure of a source while replacing most of the wording, making character-level detection ineffective. Requires semantic rather than...
- Substantial Similarity
- The legal standard in copyright infringement: whether the protected expression in one work is reproduced in another to a degree that a...
- Translation Plagiarism
- Copying from a source in another language and rendering it in the target language. Defeats monolingual detection entirely and requires cross-lingual semantic...
- Verbatim Plagiarism
- Copying text word-for-word from a source without attribution. The simplest form to detect computationally because character-level matching directly finds the copied string.