Could a Computer Write This Story?

An anonymous reader tips an article at CNN about the development of technology that automates the process of writing news articles. It started with simple sports reporting, but now at least one company is setting its sights on more complicated articles. Quoting: "Narrative Science then began branching out into finance and other topics that are driven heavily by data. Soon, Hammond says, large companies came looking for help sorting huge amounts of data themselves. 'I think the place where this technology is absolutely essential is the area that's loosely referred to as big data,' Hammond said. 'So almost every company in the world has decided at one point that in order to do a really good job, they need to meter and monitor everything.' ... Meanwhile, Hammond says Narrative Science is looking to eventually expand into long form news stories. That's an idea that's unsettling to some journalism experts."

Comments

A better question

[Anonymous Coward]

Could a Computer Write Better Stories on Slashdot?

YES.

Re:A better question

[million_monkeys]

Could a Computer Write Better Stories on Slashdot?

Slashdot summaries would be fairly well suited to being done by computer. They are usually taken from existing articles available on the web. They follow a straightforward format that is largely a quote/summary of the article. Occasionally they provide links to previous stories on the same topic. Computers can already do those things. You could even have an algorithm to put in random typos. I'm not sure how successful a computer would be at generating the tag lines like "from the kicking-newspaper-writers-when-they're-down dept.", but the rest seems doable. If slashdot were run by a bunch of geeks with the desire to do so, the story process could probably be automated, including the process of finding and rating interesting stories by by scanning various sites.

Re:A better question

[jd2112]

Perhaps, but it would almost certainly be neuter at checking for dupes.

Re:A better question

[Anonymous Coward]

In fact, I stop reading slashdot summaries as soon as I decided whether I want to read the article or not. Otherwise I keep reading the same material over and over again, which is a waste of my time. Better slashdot summaries would be welcome, yes. The current crop waste time and aren't very strong indicators, as in there's a fairly high "oh the article was crap after all" percentage.

You can automate totals, not faxts.

[LostCluster2.0]

Using baseball as an example, it's possible to automate the box score creation, but only if a user inputs the pitch-by-pitch scoring information for what was thrown, how the batter reacted, and where the ball went among other things.You can't make a computer make the decision whether the play was a hit by the batter or an error by the fielder yet. Bottom line, it's totals that a computer can come up with, but the atomic facts still need to be gathered by a human.

Re:You can automate totals, not faxts.

[martin-boundary]

Then the journalists are fucked. What most of them do is rewriting press releases submitted by companies, copying subscription news stories, and maybe adding a critical sentence or two cribbed from wikipedia.

Re:Editor AI

[LostCluster2.0]

Editors like the kind on Slashdot are hard to automate. You could rely on the +/- buttons in the firehose to pick stories, butt that's not automation, that's crowdsourcing. Aditionally, you can't automate the submitters without Slashdot looking like Digg where every RSS feed that wants in participates. Automation tools can make Slashdot easier to write, but can't fully replace the man-in-the-machine concept.

How news is written

[Dr. Tom]

"${subject} ${verb} ${object}," said a source inside the ${CurrentPresident} ${administration} who spoke on condition of anonymity because they were not authorized to speak to the press.

Easy!

[mustafap]

void main (void) {
    printf("First Post!\n");
}

Re:Easy!

[sco08y]

Wise guy, eh?

#!/bin/sh
# reads post from standard input
grep -i "First Post" > /dev/null
if [ $? -eq 0 ]; then

      exit 255
fi

No, no, no, it's just:

#!/bin/sh
grep -ivq "First Post"

A shell script will always exit with the last return code, so the exit is redundant... -q is "quiet" meaning grep will stop at the first match and just return error or success. -v will negate.

Incidentally, comparing $? is usually redundant:

if do_stuff
then echo it worked
else echo FAILURE
fi

The [ ] (or the preferred [[ ]] in bash) is just a command like any other. Use if ! command to swap clauses.

It's also worth looking in the man pages at how [[ ]] and (( )) work. In (( )), for instance, (( a = x & y ? 5 + 3 * b : c += 2 )) does what you'd expect.

And many if then clauses can be eliminated with && and ||, which sort of work the way you'd expect. (They have identical precedence, so a && b && c && d || failure works, but it's still weird.)

Already Been Done

[Kamel Jockey]

SciGen was out in 2005.

Re:Editor AI

[panda]

In my experience, humans have a tendency to overestimate their intelligence and the intelligence of our species. I specialize in automating "knowledge tasks." The people that I work with are very often surprised at just how much of what they think of as requiring human intelligence can actually be broken down into algorithms that are then applied to the data. Very little of what people actually do on a daily basis is more than the algorithmic application of knowledge.

I agree that there are some things computers cannot do just yet. Those jobs requiring creativity for instance. You could program a computer to imitate Picasso or a certain writer, but I doubt we'll be seeing computers creating truly original works of art or literature in the near future.

Re:Sure you can automate...

[Zocalo]

On the whole, I found this latest work to be somewhat lacking in comparison to earlier works such as "Oh freddled gruntbuggly", but fortunately I don't currently have a poetry appreciation chair available for the full experience.

You mean they aren't already?

[Snaller]

Could have fooled me.

I've developed something similar.

[RyanFenton]

When I worked a bit at EA, as a gameplay programmer on the Tiger Woods PGA Tour 2010 project, one of the things I worked on the scripting/event/audio system that makes the announcers react to the player's actions.

The main task of such an event engine is, working with a finite pool of reactions, it knows what it has said over a given time period, and tuning it so it doesn't repeat a phrase too often, and using it to fill as much 'empty air' as we can while it hasn't reached an annoying threshold.

The problem in that case, of course is that we only got to record so many responses with a professional voice actor, and only so much room on the disc.

With a news response engine, you wouldn't have it respond to everything - you'd have a very specific class of stories used to patch holes, the kind that is already nearly automatic already. Grabbing retweets, say "this person said this about this person", send it to an editor for review, then use it to fill gaps in a web page layout.

But then you'd still have to balance the rate of repetition of such types of news stories - which is a game of novelty and adaptive tuning.

It's certainly possible - but given the company, I expect it to be used for a while with lots of embarrassing things the editors miss showing up, until the marketing crew discovers they can use it to inject advertising messages into news stream. This input from several sources gaming the system will lead to it becoming useless over time, leading to it eventually being reinvented independently several times.

Meanwhile, Fox news will become a 24/7 lottery news channel - you too can become rich! They'll put parts of a lottery number in each commercial, then have the exact same news hosts as now tell people about how much you have to gain, using traditional conservative talking points to bolster the appeal.

MSNBC? They'll just keep selling airtime to infomercials when they can - they've already become the costs-nothing-to-produce-prison-shows channel.

Ryan Fenton

Re:Editor AI

[yelvington]

No, I'm serious. there are somethings computers can do and some they can't. You can't tell a computer to watch news come in and output a newscast.

Actually ... you can,sort of.

Voice recognition -> extracting facts from text -> story generation. All three are currently functional processes (varying degrees of quality). Having C3PO observe an arbitrary event, "understand" by inferring meaning (mathematics, probability, context database) and generate a report is not nearly so far out of reach as we might imagine. It is currently out of reach because each step introduces error rates that would result in hilarious crap, so the short-term R&D focus tends to be on domains of information where data is already encoded (such as sports and business information).

In the near term, I think the interesting opportunity is likely to be machine intelligence aiding humans in the process of reporting and analyzing. Some of this is already going on in lab situations; I've seen a system at Northwestern University "read" a brief political story and quickly connect the actors and actions with data about political contributions and connections.

Since all of this is based on machine learning, the interaction with human journalists has the potential to make the AI smarter over time, sort of how Google Translate has mutated from hilarity to utility in just a few years.

Newspeak makes it easier for computers to write...

[mrroot]

Automated story writing will be easier once the next updated version of the newspeak dictionary is released. Unfortunately I am a doubleplusungood newspeaker, but at least computer written stories will help me avoid crimethink.

Re:Editor AI

[TaoPhoenix]

"In my experience, humans have a tendency to overestimate their intelligence and the intelligence of our species. I specialize in automating "knowledge tasks." The people that I work with are very often surprised at just how much of what they think of as requiring human intelligence can actually be broken down into algorithms that are then applied to the data. Very little of what people actually do on a daily basis is more than the algorithmic application of knowledge.

I agree that there are some things computers cannot do just yet. Those jobs requiring creativity for instance. You could program a computer to imitate Picasso or a certain writer, but I doubt we'll be seeing computers creating truly original works of art or literature in the near future."

Unfortunately, your first paragraph slides is Insightful and your second one is Underrated as "Damn I wish this were true but watch out for my first paragraph".

Taking the writing example (because I know almost zero about painting), in one sense it is Not So Tough to create a computer that can create "truly original works of literature". We're currently playing a No True Intelligence (Scotsman) game against AI because of our crushing need not to get vaporized by automation and end up like the Matrix pods. As I've ranted elsewhere, we've purposely stunted funding into AI because of this zenophobic fear of what happens if we're not on top. The second fallacy is "zero human interaction". That gives us the comfort of watching the threatening AI crash and burn the minute it cannot get a little help. Instead, a "90-10" pattern is the devastating mix of automation.

Let's do a thought experiment. This will sound a little stilted for reasons of easy conceptual value, but then all you need is a second program to "polish" it. Here we go: (Parentheses used for pseudocode structure instead of other characters, in an attempt to avoid filters)

-------------------
http://falkvinge.net/2012/05/12/dutch-judge-who-ordered-pirate-bay-links-censored-found-to-be-corrupt/
Thought Experiment of Machine Assisted Article
(Find Topic I Like)
      (Topic I Like = Modules Loaded As Expertise)
(IntegrityOfLaw (IsOneOfModulesLoaded)
(Story Matching TopicILike-IntegrityOfLaw)

(Pick ConversationalPhraseOpener)
You may have heard of
(Introduce DeepSubject1)
the Pirate Bay.
(Beginner Mode On. EasyDescription Follows).
That is a site where users post music. However, in most places it is only legal to post music that you own. Some or many people post music that they do not own there. This is illegal. Also, the people that own the music do not like this. Subject to differences in law by countries, if the owner of a song asks for it to be removed, it is supposed to be removed. However, this takes time and time costs money. The music owners do not like this either.
(Intermediate Warmup to Story Follows)
Suppose someone uploaded a song. By itself maybe no one will see it. Then no one downloads any copies. The music owner "feels he has not lost any sales". But if someone posts the exact file name or other method, then many people will see it. Many people will download copies. The music owner "feels he has lost sales". The music owner does not like this. The music owner asked a court to tell people to stop posting file names to songs they do not own. The court agreed.
(Actual Story Begins Here)
The entire concept of the pure theory of law requires a fair judge. When the judge is unfair, this creates a social problem point. Generally, a fair judge has nothing to gain for himself. However, today's news is that the judge made money with the plaintiff. Therefore this is not a fair judge. Rulings by an unfair judge can sometimes still be fair. However, more likely the ruling from an unfair judge is unfair.
(Insert Poll Here)
75% of people polled say this ruling was unfair.
The question now for a higher court is whether the judge is fair.

(End of Pseudo-Generated Story.)

--------------------