Slashdot Mirror
Computers Paraphrase English
AhaIndia submits a link to a story discussing computerized paraphrasing of English news articles. This technology, destined to eventually replace most reporters with very small shell scripts, is thankfully still in its infancy.
won't you need someone to write the stuff to be paraphrased in the first place?? explain to me how that replaces reporters with small shell scripts.
So, will there be difference between paraphrasing and copying now in an educational setting? Seems like this could make a report pretty easy...
1) Brainstorm some key points/ideas
2) Have this program data mine for relavent articles online
3) Feed sections of each article into the program and have a finished paper
Granted, the tech isn't quite that powerful yet and probably wouldn't do a whole paper, but it sure looks like it could supply several paragraphs of material per page...
For you to say that this technology will someday replace reporters makes me think that you're clueless about what reporters do. Do you realize that the biggest parts of a reporter's job are gathering facts and making judgments about 1) which stories are worth reporting, 2) which are the relevant facts about a story and 3) who's lying and who's telling the truth about a story? The actual writing that you see is many times almost incidental to most of what a reporter does. You might not like the judgments that a reporter makes (and I could agree with that in many cases), but software can't go out into the world and talk to people and use judgment and intuition to find information to write about.
As an ex-reporter and editor, I find it laughable that anyone might think this technology will replace reporters. It's sort of like suggesting that machines that can read source code and interpret it can somehow figure out what new software people want and then write it. Both possibilities are equally insane.
The poster incorrectly assumes that this could be used to replace reporters. The problem is that computers have a difficult time generating new text. The methods that computers use to evaulate text (as any user of grammar-check would realize) aren't that great.
In fact, most language models cannot generate even a large portion of English text. Those that do have a good range rarely have good accuracy, because there are many things that we "just don't say that way." This is why when you're talking to a non-native speaker, you often cannot explain why something they said was wrong. This is because there is no real grammar rule against speaking in a given way.
So if we rule out syntax-based models, that just leaves statistical-based models. I worked in a NLP lab during the summer of 2002, and my prof there said that syntax and statistics are like the two sides of the force. Statistics are quick and easy but are seductive. They corrupt you and leave you unable to really think about the language itself. You only think in terms of bigrams and HMMs.
So even though these systems are doing well, they are mostly statistical. Thus, it's hard to get incremental improvement. You have to have larger corpora, and larger corpora usually have more errors, thus defeating any advantage you might get by capturing more aspects of a language.
In my opinion, only with well-developed language models that can effectively generate NL can we get anywhere. Which is what Barzilay is working on, but it's still a long, long, long way off.
----------
I am an expert in electricity. My father held the chair of applied electricity at the state prision.
Well, some of us (me, for instance) listen to fox news and NPR...my own personal take on fair and balanced...and see that the party line is alive and well on both major sides of the political fence. That's part of the reason I'll never be a democrat or a republican (or a libertarian, or any other label you want to stick on a like-minded group of people). The news has information in it. Look for it, compare notes, and make up your own mind what's news.
"Murphy was an optimist" - O'Toole's commentary on Murphy's Law
Obviously this is a developing field. The best models seem to use phrases from the original text, anyway the Mac OSX example above shows that it is useful to users willing to take it with a massive grain of salt, even if we are not into full computational sentience yet.
When it works even a little better it will replace all those awful grade school teachers who assign paraphrasing as a homework assignment. The reporters who might have been replaced by it will have already lost their jobs, except for the ones in AhaIndia of course who will paraphrase for the rest of us, usually at a marginally better level than the machine.
The research is interesting - and I'd like to understand Barzilay's notation is that APL or calculus of statement? - in the paper (pdf) I found on google. Also see the papers on her site.
Of course structured text is easier, and news stories are known to have most of the meat in the beginning, but this is great stuff.
One interesting older system is ThoughtTreasure which was built to understand a story and answer questions about it. The author also did work on news analysis ("NewsForms") too. There are tools out there, I've been making a survey myself too. If anyone has information about practical NLP tools for real world tasks please post.
The main problem is that languages, especially English, are so idiomatic that mechanical translators will be a too much of a disadvantage - take the Babelfish translator for instance.
Furthermore, the English language is so flexible that just about any word can arbitrarily substitute for anything else - for instance, take 'bad' meaning 'good'.
It would be impossible to program a machine to be able to understand the full spectrum of idiomatic phrases but the future may lie in employing neural net technologies so that computers can do some limited learning.
cogito ergo sig...
"It might even know the proper use of to/too and your/you're."
Yeah, but can it manage to use "There are" instead of "There is" with a plural subject?
Actually, the long known solution to most of these *oh so difficult* translation problems is to translate everything into a neutral interlanguage like Interlingua and then translate that into other languages, sending the interlingua version along for the ride, thus preventing degradation in further translations. Then all that local linguists have to concentrate on is ONE set of problems: translating their local language into and out of Interlingua, and Interlingua, being tightly defined, is much easier to machine translate into and out of other languages. So...all this lunacy of trying to machine translate Chinese into English, German, Hungarian, Estonian...--you get the picture--is an incredible waste of time and resources and isn't the best way to solve the problem.
Hic iacet Arthurus, rex quondam rexque futurus.