Organising the web
The science of science
Apr 28th 2011
COMPUTER scientists have long tried to foist order on the explosion of data that is the internet. One obvious way is to group information by topic, but tagging it all comprehensively by hand is impossible. David Blei, of Princeton University, has therefore been trying to teach machines to do the job.
He starts with defining topics as sets of words that tend to crop up in the same document. For example, “Big Bang” and “black hole” often will co-occur, but not as often as each does with “galaxy”. Neither, however, would be expected to pop up next to “genome”. This captures the intuition that the first three terms, but not the fourth, are part of a single topic. Of course, much depends on how narrow you want a topic to be. But Dr Blei’s model, which he developed with John Lafferty, of Carnegie Mellon University, allows for that.
The user decides how fine-grained he wants the analysis to be by picking the number of topics. The computer then creates a virtual bin for each topic and begins to read the documents to be analysed. After removing common words that it finds evenly spread through the original documents, it assigns each of the remaining ones, at random, to a bin. The computer then selects pairs of words in a bin to see if they co-occur more often than they would by chance in the original documents. If so, the association is preserved. If not, the words (together with others to which they have already been tied) are dropped at random into another bin. Repeat this process and networks of linked words will emerge. Repeat it enough and each network will correspond with a single bin.
And it works. When Dr Blei and Dr Lafferty asked their software to find 50 topics in papers published in Science between 1980 and 2002, the words it threw up as belonging together were instantly recognisable as being related. One topic included “orbit”, “dust”, “Jupiter”, “line”, “system”, “solar”, “gas”, “atmospheric”, “Mars” and “field”. Another contained “computer”, “methods”, “number”, “two”, “principle”, “design”, “access” and “processing”.
All of which is interesting as a way of dealing with information overload, and tagging papers so that they can be searched in a more useful way. But Dr Blei found himself wondering if his method could yield any truly novel insights into the scientific method. And he thinks it can. In tandem with Sean Gerrish, a doctoral student at Princeton, he has now produced a version that not only peruses text for topics, but also tracks how these topics evolve, by looking at how the patterns in each topic bin change from year to year.
The new version is able to trace a topic over time. For example, a 1903 paper with the evocative title “The Brain of Professor Laborde” was correctly assigned to the same topic bin as “Reshaping the Cortical Motor Map by Unmasking Latent Intracortical Connections”, published in 1991. This allows important shifts in terminology to be tracked down to their origins, which offers a way to identify truly ground-breaking work—the sort of stuff that introduces new concepts, or mixes old ones in novel and useful ways that are picked up and replicated in subsequent texts. So a paper’s impact can be determined by looking at how big a shift it creates in the structure of the relevant topic.
In effect, Dr Blei and Mr Gerrish have devised an alternative to the citation indices beloved of scientific publishers. These reflect how often a particular publication or author is cited as a source by others. High scores are treated as a proxy for high impact. But a proxy is all they are.
Dr Blei and Mr Gerrish are not claiming their method is necessarily a better proxy. But it can cast its net more widely, depending on the set of documents fed into it at the beginning. Citation indices, which work only where publications refer to their sources explicitly, form a tiny nebula in the digital universe. News articles, blog posts and e-mails often lack a systematic reference list that could be used to make a citation index. Yet they, too, are part of what makes an idea influential.
Besides, despite academia’s pretensions to objectivity, it is as subject to political considerations as any area of human endeavour. Many authors cite colleagues, bosses and mentors out of courtesy or supplication rather than because such citations are strictly required. More rarely, an author may undercite. Albert Einstein’s original paper on special relativity, for example, had no references at all, even though it drew heavily on previous work. The upshot is that the Blei-Gerrish method may get closer to the real ebb and flow of scientific ideas and thus, in its way, offer a more scientific approach to science.
The Economist
He starts with defining topics as sets of words that tend to crop up in the same document. For example, “Big Bang” and “black hole” often will co-occur, but not as often as each does with “galaxy”. Neither, however, would be expected to pop up next to “genome”. This captures the intuition that the first three terms, but not the fourth, are part of a single topic. Of course, much depends on how narrow you want a topic to be. But Dr Blei’s model, which he developed with John Lafferty, of Carnegie Mellon University, allows for that.
The user decides how fine-grained he wants the analysis to be by picking the number of topics. The computer then creates a virtual bin for each topic and begins to read the documents to be analysed. After removing common words that it finds evenly spread through the original documents, it assigns each of the remaining ones, at random, to a bin. The computer then selects pairs of words in a bin to see if they co-occur more often than they would by chance in the original documents. If so, the association is preserved. If not, the words (together with others to which they have already been tied) are dropped at random into another bin. Repeat this process and networks of linked words will emerge. Repeat it enough and each network will correspond with a single bin.
And it works. When Dr Blei and Dr Lafferty asked their software to find 50 topics in papers published in Science between 1980 and 2002, the words it threw up as belonging together were instantly recognisable as being related. One topic included “orbit”, “dust”, “Jupiter”, “line”, “system”, “solar”, “gas”, “atmospheric”, “Mars” and “field”. Another contained “computer”, “methods”, “number”, “two”, “principle”, “design”, “access” and “processing”.
All of which is interesting as a way of dealing with information overload, and tagging papers so that they can be searched in a more useful way. But Dr Blei found himself wondering if his method could yield any truly novel insights into the scientific method. And he thinks it can. In tandem with Sean Gerrish, a doctoral student at Princeton, he has now produced a version that not only peruses text for topics, but also tracks how these topics evolve, by looking at how the patterns in each topic bin change from year to year.
The new version is able to trace a topic over time. For example, a 1903 paper with the evocative title “The Brain of Professor Laborde” was correctly assigned to the same topic bin as “Reshaping the Cortical Motor Map by Unmasking Latent Intracortical Connections”, published in 1991. This allows important shifts in terminology to be tracked down to their origins, which offers a way to identify truly ground-breaking work—the sort of stuff that introduces new concepts, or mixes old ones in novel and useful ways that are picked up and replicated in subsequent texts. So a paper’s impact can be determined by looking at how big a shift it creates in the structure of the relevant topic.
In effect, Dr Blei and Mr Gerrish have devised an alternative to the citation indices beloved of scientific publishers. These reflect how often a particular publication or author is cited as a source by others. High scores are treated as a proxy for high impact. But a proxy is all they are.
Dr Blei and Mr Gerrish are not claiming their method is necessarily a better proxy. But it can cast its net more widely, depending on the set of documents fed into it at the beginning. Citation indices, which work only where publications refer to their sources explicitly, form a tiny nebula in the digital universe. News articles, blog posts and e-mails often lack a systematic reference list that could be used to make a citation index. Yet they, too, are part of what makes an idea influential.
Besides, despite academia’s pretensions to objectivity, it is as subject to political considerations as any area of human endeavour. Many authors cite colleagues, bosses and mentors out of courtesy or supplication rather than because such citations are strictly required. More rarely, an author may undercite. Albert Einstein’s original paper on special relativity, for example, had no references at all, even though it drew heavily on previous work. The upshot is that the Blei-Gerrish method may get closer to the real ebb and flow of scientific ideas and thus, in its way, offer a more scientific approach to science.
The Economist
Nenhum comentário:
Postar um comentário