Digital Media Buzz > Search Takes a Sentimental Journey

Search Takes a Sentimental Journey

Next Page « 1 2 »

By Ned Smith

The latest heavyweight to grab for the brass ring on the search carousel is Microsoft’s revamped, reinvented and rebranded Bing, which boldly bills itself as “the first ever decision machine.” Its claim to a larger hunk of intellectual real estate is one more example that the search business has gone beyond its humble keyword roots.

Today, sentiment analysis is the new frontier in search, extending the whole text analytics discipline into the thorny thicket of the opinions of people to gauge their likes and dislikes. In his article for Communications of the ACM, “Our Sentiments, Exactly,” author Alex Wright (Glut: Mastering Information Through the Ages) asked analytics strategist Seth Grimes to describe this discipline extension: “Sentiment Analysis is a set of algorithms and tools for identifying and extracting a) features that express attitudes or opinions, b) attributes that indicate sentiment polarity, intensity and other characteristics, c) the topics those sentiments and attributes apply to, and, optimally, d) the opinion holder,” Grimes says.

This is the highly subjective, unstructured text world of “I feel, I think, I want.” What makes this such dicey territory? “Sentiments are very different from conventional facts,” Grimes tells Wright. “Cultural and contextual factors come into play” who’s ‘speaking,’ where is that person speaking, who is she or he speaking to? There’s lots of slang ― phat, cool, bomb, snap ― and ambiguity ― ‘sinful’ is a good thing when applied to chocolate cake ― as well as sarcasm, irony and idiom.”

It’s something that human beings excel at. But it can eat up huge chunks of time. Computers aren’t the answer. Fast though they may be, computers can be pretty thick when it comes to picking up on emotion and tone. The goal is try to find a way to marry human insight and computational speed.

This is not turf for the faint of heart or those discomfited by nuance and ambiguity. Those pitfalls notwithstanding, though, a growing band of start-up hopefuls and veteran vendors are lighting out for this frontier.

One of the promising new ventures in this space is Jodange. Headquartered in a renovated 1900 trolley barn in Yonkers, N.Y. overlooking the Hudson River a few miles north of New York City, Jodange takes its name from a conflation of the names of CEO and co-founder Larry Levy’s children ― Josh, Daniel and Geena. The other co-founder and Jodange’s chief scientist is Claire Cardie, who is also a professor at Cornell University’s Department of Computer Science, where her primary research is in the area of natural language understanding and intelligent text processing.

The company was founded in 2006 because “we wanted to solve a really specific problem,” Levy says. “We saw information online growing at an exponential rate.” And keyword search ― that “blunt instrument,” as he describes it ― wasn’t hacking it. It’s good at finding things, but woefully inadequate when it comes to extracting meaning.

Jodange uses linguistic analysis to extract opinions from documents and identify the opinion holder and topic each opinion expresses. It ignores any sentence or phrase deemed to be factual, and instead focuses on phrases that are subjective in order to determine the sentiment ― positive, neutral or negative ― of the subjective statement. “We look for well-constructed language,” Levy says. There is a pipeline of up to 30 algorithms, he says, that serves as an information filter to weed out the dross.

“We’re the only company that can do that kind of analysis,” Levy says. Competitors, he adds, “can’t pick up who the real opinion holder is. That nuance is lost on 99 percent of the people doing this.”

“On a daily basis,” he says, “there are 600,000 to 650,000 pieces of information from 435,000 sources that are worth processing.” Jodange’s information sources include its own Web crawlers, Factiva and moreover. “We’re trying to get at the information sources of that content,” Levy says. “Where is the meaningful stuff?”

Sentiment analysis, he says, lives or dies on its degree of accuracy. The key test for accuracy for Jodange is inter-annotation agreement ― a metric based on two people reading the same document and scoring it the same way. Jodange claims a 77-79 percent agreement rate. “You’ll never get a system that is 100-percent accurate,” he says.

Though Jodange is a newcomer, its underlying technology has been gestating for a number of years at Cornell. “Claire has been worrying about this problem for more than 10 years,” Levy says. “It’s fantastic having Cornell as a partner. I know I have the best people on the planet to do what we do. We have the world’s leading experts in this field. The key issue for our next wave of things is scalability and what we call disambiguation.”

Jodange’s offering now comes in two flavors: Opinion Utility for Publishers, which matches and resurfaces opinions extracted from a publisher’s network of sites, including reader comments, and Top of Mind, an application that provides media monitoring across social and traditional media. The latter is targeted at the corporate market with a starting price of $2,000 a month for up to 10 users. The Opinion Utility for Publishers starts at $8,000 a month for top-tier publishers.

Anyone who has spent much time around people in the text mining and sentiment analysis field is familiar with concepts such as “entity extraction” and “tokenization.” Simply put, entity extraction and tokenization describe the processes that translate text into a form that computers can understand.

Next Page « 1 2 »