Web Concepts

Posts Tagged ‘language’

If a picture is worth 1,000 words, how many words make up a thought?

Researchers at Princeton University looked at fMRI scans to identify brain regions that were activated when participants thought about certain objects, like a carrot or a house. Then the team generated a list of topics that were also associated with those words. They looked at the same fMRI scans to determine the brain activity that was shared by the words within each topic, as a Princeton news release puts it. For instance, thoughts about the idea of “furniture” shared similar patterns with words like “table,” “desk” and “chair.”

Once they could tell the fMRI activity that would be sparked by a particular topic, the researchers were able to look at the brain activity alone and extrapolate what the person was thinking about. If a scientist studying brain scans spotted the neural patterns corresponding to “chair,” he could tell that the person was thinking about furniture.

“Whatever subject matter is on someone's mind — not just topics or concepts, but also emotions, plans or socially oriented thoughts — is ultimately reflected in the pattern of activity across all areas of his or her brain," said the team’s senior researcher, Matthew Botvinick, an associate professor in Princeton's Department of Psychology and in the Princeton Neuroscience Institute.

The researchers started out with brain scans from a 2008 word association study, in which participants were shown a picture and a word of five objects in 12 categories. The participants had been asked to visualize the object for three seconds, and the fMRI recorded their neural activity. Then the Princeton researchers came up with a list of their own topics with which to characterize this fMRI data. They used a computer program to condense 3,500 Wikipedia articles about objects — like an airplane, heroin, birds and manual transmission. The program came up with 40 topics to which these things could relate — i.e. aviation, drugs, animals or machinery. (Their full paper is available online for those interested in the specific methods.)

They arranged the fMRI scans by subject matter, and were ultimately able to tell the general topic on a person’s mind. It was harder to pick out an individual object, however, the Princeton news office explains. The eventual goal is to translate brain activity patterns into the correct words to fully describe thoughts, the researchers say.

This could have applications for helping people with disabilities, for whom brain scans might be able to elucidate their thinking more effectively than pictures. The research appears in the journal Frontiers in Human Neuroscience.

[Princeton News]

NELL, the self-teaching artificial intelligence at Carnegie Mellon, has a Twitter account and knows how to use it

When she learns a new fact, NELL adds it to her online database, and tweets her discoveries, so you can follow her progress. For example, today she learned that the phrase “dem franchise boys” fits in the category of “music artist.” NELL’s knowledge base currently consists of around 440,000 facts, with around 87 percent accuracy. Many of her tweeted lessons demonstrate amazing analytical capabilities – she was able to get “tool” from “Anchor Fixing Self Tapping Screw Bolts.” The New York Times recently published an in-depth look at exactly how NELL learns these things.

NELL isn’t perfect though. And thank goodness, because that means we get some pretty adorable not-quite-there-yet tweets, such as: “I think "chicken recipe time" is a #condiment” and “I think "anonymously" is a #fish.” NELL’s followers can tweet corrections to her and help her improve her associations. Check out her feed or browse her whole knowledge database to see more.

By drawing on the ever-evolving organism that is the Internet, Engkoo (loosely meaning “English vault” in Chinese) should be able to stay abreast of changes in colloquialisms and idioms in both the source language and the one it is translating to. In theory, it should also be able to catch errors or mistranslations easier, since an error is unlikely to be prevalent across the entire Web.

When a user searches for a word or sentence in either language – Microsoft plans to adapt the system for other languages but this initial phase is focused on Chinese-to-English translation – the software driving Engkoo searches through the database for the relevant data and draws upon statistics to translate as accurately as possible. Where possible it links to the sources where it drew the initial data from and often can provide example sentences for a word or phrase.

Engkoo is also a multimedia experience. Computer generated audio translations exist for many English words and sentences to help Chinese speakers with their pronunciation, and researchers are cultivating a video dictation library so users can see the way native speakers’ lips move as they enunciate.

Next up? Ultrasound images that show the movement of the tongue inside the mouth, a critical step in learning pronunciation but one that is often hidden from plain view. Researchers are already gathering ultrasound data for the library, but those of you who find that kind of imagery less-than-savory, worry not; the black-and-white ultrasounds will be converted into cartoon animation to make them a bit more – how do you say? – palatable.

There’s also a mobile app in the works that will run on Windows phones – other mobile OS types are being considered – that allows for translation on the go. Which means perhaps we’re seeing the first real baby steps toward the universal translator you can keep in your pocket for real-time translation of any language into your own.

[WSJ]