Computer text analysis of a huge database of scientific papers shows a large amount of 'text overlap,' where authors use text from previous papers of their own and others, not always with attribution. According to Cornell University Library researchers, this is not necessarily good or bad.
The analysis was conducted on thousands of papers in the arXiv. Paul Ginsparg, professor of physics and information science and founder of the online arXiv collection of scientific papers, now maintained by Cornell University Library, and Cornell graduate student Daniel Citron reported their study in the Dec. 8 online edition of the Proceedings of National Academy of Sciences.
Sometimes the copied text is a description of experimental apparatus or procedures. Mathematicians often repeat well-known theorems that underlie their arguments. Many authors reuse material from their own previous papers about the same line of research, and a Ph.D. candidate's final thesis may pull in material from papers the student had published along the way. An amusing sidelight is the copying of acknowledgements with new names inserted, possibly because authors think what they are using is a standard format.
Sometimes the roots of the practice may be cultural. In China and some other Eastern cultures it is considered disrespectful to change someone else's words. And if English is not their native language, authors may not feel comfortable rewriting their source material. Some may just have a different view of what is acceptable and may have been trained that way by their mentors.
In what may seem like the Law of Karma in action, the analysis showed that papers with the most text overlap were the least cited. Partly, the researchers said, this is because many papers with large overlap come from a group of countries in Eastern Europe and the Middle East, and papers from these countries are less often cited anyway.
The study compared text in 757,000 articles deposited from 1991 to 2012. Each text was broken into overlapping phrases seven words long, and other texts were then scanned to see how many matching phrases they had in common. The system skips common usages like "The remainder of this article is organized as follows," as well as direct quotations with proper attribution.
The entire database was preprocessed to create an index of all the seven-word patterns in the arXiv. The analysis only became possible, Ginsparg said, with the low-cost availability of computers with enough random-access memory (RAM) to hold the entire 12.5 GB index. Given that, a single paper can be checked in less than a second, he said.
Since June 2011, every new article submitted to the arXiv has been scanned against the entire database, and papers with significant text overlap have been flagged with a notice to that effect. The current study grew partly from the need to establish a threshold for aberrant practice, permitting a response to indignant authors saying in effect, 'Everybody does it.'
The work was partially supported by the National Science Foundation.
