快猫短视频

GOVERNMENTS and big business like to indulge in media spin, and that means knowing what is being said about them. But finding out is becoming ever more difficult, with thousands of news outlets, websites and blogs to monitor.

Now a British company is about to launch a software program that can automatically gauge the tone of any electronic document. It can tell whether a newspaper article is reporting a political party鈥檚 policy in a positive or negative light, for instance, or whether an online review is praising a product or damning it. Welcome to the automation of PR.

Till now, discovering whether the coverage you are getting is good or bad, negative or neutral has usually meant hiring a 鈥渞eputation management鈥� firm. Teams of people employed by the company will read through everything written about a chosen organisation, person, event or issue and report back on how favourable it is.

As well as being expensive, this can be a long, slow process, says Nick Jacobi, director of research for the Surrey-based company, Corpora Software. 鈥淭here鈥檚 a massive information overload.鈥� A single news agency may churn out more than eight articles each hour. That鈥檚 almost 200 stories a day per news outlet.

Previous attempts to automate this kind of analysis have used one of two techniques. In the first, called machine learning, a program is trained by being given thousands of articles already determined by a human reader to be positive or negative in tone.

But learning in this way can lead to mistakes. For example, if a series of the training articles mentions bomb attacks on a mosque in Iraq, the program may incorrectly conclude that all other mentions of mosques are negative too.

The alternative is the lexicon approach, in which certain words are classified as either positive or negative. But plenty of words can be both. 鈥淭he plot was unpredictable鈥� and 鈥渢he steering was unpredictable鈥� differ by just one word. Yet the word 鈥渦npredictable鈥� has a positive connotation in the first example and a negative meaning in the second (see 鈥淲ord play鈥�). And even if that problem is solved, just picking up on positive or negative words can also lead to mistakes, as is demonstrated by the sentence: 鈥淓veryone told me it was terrible, that I would hate it, but in the end it wasn鈥檛 at all bad鈥�.

So Corpora has come up with a program called Sentiment, which uses algorithms to tease out grammatical components, such as nouns, verbs and adjectives, and identify the subjects and objects of verbs. It can even analyse pronouns like 鈥渋t鈥�, 鈥渉e鈥� and 鈥渉er鈥� to work out what words or concepts they are referring to.

Having an understanding of grammatical structure makes it possible to filter out words that are not relevant to the sentiment of the article, Jacobi says. So instead of assuming certain words, such as 鈥渦npredictable鈥� or 鈥渞ubbish鈥�, are positive or negative it allows the structural context to disambiguate them.

It doesn鈥檛 get it right all the time, Jacobi admits, but then neither do people. Three expert readers are likely to agree about an article 85 per cent of the time, and about 90 per cent of non-experts will agree with this consensus. The Sentiment software agrees with the same expert consensus about 80 per cent of the time.

Sentiment was developed principally for Infonic, one of Corpora Software鈥檚 subsidiary companies, which provides clients with online media analysis of websites, chat rooms, bulletin boards and blogs. The company also hopes to use it to analyse the news for its clients.

Sentiment won鈥檛 take the humans out of the equation, says Orlando Plunket Greene of Infonic, because someone is still going to have to evaluate the software鈥檚 report on each article. But because the program will list items in terms of how positive, negative or neutral they are it is possible to skip to the most relevant items. 鈥淚t will allow us to prioritise, and do the job much faster,鈥� he says. While a person might be able to scan 10 articles an hour, Sentiment can zip through 10 a second.

What makes this kind of analysis so challenging is that key words in a text often offer no clues as to what sentiment they carry. Some of the toughest challenges to comprehension, such as identifying irony and rhetoric, are likely to remain unsolved for some time.