The computer that understood movie reviews

And more recent highlights from the Ideas blog

Analysis of the sentence: “Unlike the surreal but likeable Leon, this movie is weird.”
Analysis of the sentence: “Unlike the surreal but likeable Leon, this movie is weird.”

One of the hottest areas in computer science these days is “deep learning,” or the idea that computers can teach themselves, so that they acquire something like the nimble processing power of the human mind.

Deep learning is a tantalizing idea (and one step towards artificial intelligence), but for nonexperts, it’s hard to identify a machine that’s taken off on its own. Computer scientists at Stanford have recently created a program called NaSent, however, that is both remarkable in its own right, and a satisfyingly clear demonstration of what deep learning looks like in practice.

NaSent, which is short for Neural Analysis of Sentiment, is a program that determines whether movie reviews are positive or negative. There are already programs that do this, largely by counting positive and negative words in a review, but NaSent is more sophisticated: It can extract meaning from whole phrases and sentences, which puts it ever so slightly closer to the realm of a real live reader.


NaSent was created by computer scientists Richard Socher, Christopher Manning, and Andrew Ng, and linguist Christopher Potts, and presented last month at a conference in Seattle. The researchers began by feeding the program 214,000 phrases and sentences from movie reviews that had been coded manually on a scale from “like” to “dislike.” NaSent then draws on that foundation to determine the meaning of unfamiliar sentences.

Using a sentence diagram similar to the ones you might have made in fourth grade, the program is able to tell the difference between two sentences that contain the same words in nearly the same order, but have completely different meanings:

“Unlike the surreal Leon, this movie is weird but likeable.”

“Unlike the surreal but likeable Leon, this movie is weird.”

These tree diagrams are limited compared to how human beings analyze language. They do, though, provide a glimpse of what a more sophisticated program might look like one day, when maybe a computer will be able to tell that a person who raises his eyebrows, reclines his head, lowers his voice, and says, “That was a great movie,” means no such thing.


The Confederate papers

Though it now sounds almost like a myth, the Confederate States of America existed as a real place from 1861 to 1865. Since the Civil War, the Boston Athenaeum has put together one of the largest collections of printed material from the Confederacy, and the collection is now being digitized and placed online through what’s called the Confederate Access Project.

According to Evan Knight, the lead preservationist on the project, particularly notable objects include a broadside printing of a speech by Confederate Lieutenant General Nathan Bedford Forrest in 1864; Forrest praises his troops for the “conspicuous gallantry” they showed during the previous Confederate victory at Fort Pillow, which may have actually involved a massacre of black Union troops. Other notable items include a “Medical and Surgical Journal,” the very idea of which is enough to make you squirm, and a 1864 newspaper advertisement recruiting Rebel soldiers as the Yanks move in. Together, the artifacts document a very specific American civilization in wartime, and stand as a kind of echo of its fall.

Kevin Hartnett is a writer in South Carolina. He can be reached at kshartnett18@gmail.com.