Paraphrase

A paraphrase (/ˈpærəˌfreɪz/) or rephrase is the rendering of a text through the use of different words without altering the text's original meaning. Most of the time, a paraphrased text can convey its meaning more effectively than the original words. In other words, paraphrasing uses different vocabulary than the original text but maintains the same concept. For example, when someone tells a story they have heard, in their own words, they paraphrase it, with the meaning being the same. The phrase itself is derived via Latin paraphrasis, from Ancient Greek παράφρασις (paráphrasis) 'additional manner of expression'. The action of paraphrasing is also called paraphrasis.

Although paraphrases likely abounded in oral traditions, paraphrasing as a specific educational exercise dates back to at least Roman times, when the author Quintilian recommended it for students to develop dexterity in language. In the Middle Ages, this tradition continued, with authors such as Geoffrey of Vinsauf developing schoolroom exercises that included both rhetorical manipulations and paraphrasing as a way of generating poems and speeches. There is interest in the study of paraphrases relating to concerns around plagiarism and original authorship.

For the purposes of education, Fred Inglis identifies five levels of paraphrase:

Paraphrasing with synonyms is considered by some to be an acceptable stage in teaching paraphrase, but it is necessary that it is ultimately combined with techniques for altering sentence structure to avoid the appearance of plagiarism. Studies of English language students have found that ESL learners tend to rely on using synonyms rather than changing sentence structure when paraphrasing. Participants in a study of some Vietnamese ESL learners expressed that they preferred using synonyms out of a fear that using the wrong sentence structure would lead to the sentence having a different meaning. Na and Mai suggest that ESL teachers should provide varied activities including tasks requiring changes in syntax, and that ESL students should be given source texts to paraphrase whose meaning they are already readily able to understand.

Natural Language Processing researchers have defined various (atomic) paraphrase types to better understand how paraphrasing occurs in humans. These types fall into six broad categories, each reflecting different ways in which a text can be altered to convey the same or similar meaning:

Morphology-based changes involve alterations at the level of word formation, such as changing the tense of verbs or the number of nouns. For instance, converting "walks" to "walked" represents a morphological change by altering the verb's tense. Lexicon-based changes include changes made by substituting words with their synonyms or closely related words without significantly altering the sentence structure. An example could be changing "quick" to "fast" in a sentence, where both adjectives convey a similar speed attribute. Lexico-syntactic-based changes contain both lexical alterations and modifications in the sentence structure. An example might be transforming an active voice sentence like "The cat chased the mouse" into a passive voice "The mouse was chased by the cat," where both the sentence structure and some words are altered. Syntax-based changes are primarily focused on the structure of the sentence rather than the words themselves. For example, changing a complex sentence into two simpler sentences while maintaining the overall meaning falls into this category. Discourse-based changes are alterations that affect the larger discourse or text structure, such as reordering points in a paragraph or changing the way arguments are presented without altering the factual content. Extremes include changes that significantly alter the text, possibly introducing new information or omitting crucial details, thus pushing the boundaries of what might typically be considered a paraphrase.

Machine learning models have been trained to generate paraphrases with specific properties, such as high semantic similarity and syntactic diversity, or to generate specific paraphrase types. Models that have high capacity in paraphrasing can be used for various applications. For example, the granular understanding of the linguistic changes involved in paraphrase generation could be directly applied to support language learners. A model can provide simpler paraphrases considering specific linguistic variations (e.g., syntax) to support students in learning new words and concepts. Universities could create a linguistic profile of their students based on their assignments and better assess their thesis with content similarity detection for potential plagiarism cases. Different types of paraphrases such as syntax and lexicon changes have also been used for prompt engineering to adjust prompts in specific linguistic aspects to achieve better model outputs.

A paraphrase typically explains or clarifies the text that is being paraphrased. For example, "The signal was red" might be paraphrased as "The train was not allowed to pass as the red signal light was illuminated". A paraphrase can be introduced with verbum dicendi—a declaratory expression to signal the transition to the paraphrase. For example, in "The author states 'The signal was red,' that is, the train was not allowed to proceed," the that is signals the paraphrase that follows.