Hmmm... Googling: "elected as the MP for" "was created using some automation", reveals very little variation. We're not exactly talking AI powered journalism here. More mail merge than ML.
Next week: How computers wrote your last electricity bill.
No 'real' AI/ML journalism is claimed in the article; the meat and potatoes of the article are provided by Mr. Robert McKenzie, editor of BBC News Labs. It is probably best to direct your vitriol towards the right source.
"None of the stories have any quotations in, none of them have any analysis of what happened or what the significance is. It is purely a written version of what has happened based on the data. So that's quite a big downside in terms of quality of journalism."
However, Mr McKenzie said the BBC was still in the "very early stages of understanding what audiences want from the technology".
It's hardly vitriol. I just thought that the headline and opening paragraphs ("a news story for every constituency... all written by a computer", "the BBC's biggest test of machine-generated journalism so far", "Each of nearly 700 articles... was checked by a human editor") did hint at an AI element and more variety of composition in the articles than I found there to be.
> "None of the stories have any quotations in, none of them have any analysis of what happened or what the significance is. It is purely a written version of what has happened based on the data. So that's quite a big downside in terms of quality of journalism."
With the way news works today, are we sure this is really a downside?
I imagine the BBC is not attempting to boast, merely being transparent. It probably says more about us, that the moment we see "machine-generated" we assume it will have an AI component.
About 10 years ago, as an online editor, I built a similar set of scripts -- we called them one-clicks. You'd click a button, the story would be generated, and it would be populated into the CMS.
This works really well for mundane, data-driven stories such as lottery results or surf reports, both of which I built.
There are a bunch of companies, such as Narrative Science, that are doing much more sophisticated versions of this. You give them the data and they can churn out a machine-generated recap of a Little League game that reads much like an MLB story.
But what I think will really be a game changer is personalization for the reader.
Can you imagine the same story written 10 different ways for 10 reader personas? Maybe a reader with an education background might get a different lede than a reader with an interest in politics on a story about a new school district referendum. Or heck, maybe a Republican reader might get a different version of a Trump story than a Democrat. (Not that this would necessarily be a GOOD thing. But imagine the possibilities!)
I worked at a startup in 2015 that did something similar. Using an NLP called Yseop, we wrote sports articles. The team consisted of me on backend, a journalist writing the Yseop code, a technical writer building the story trees, and a data scientist building the DBs from a plethora of sports sources. It's fairly straightforward. Provide enough variation and you don't lose readers.
Yseop is not cheap, though, and the company went under after massive financial fraud. The Tribune Network used a fair chunk of our content at the time. I've been waiting for this to be more common-place.
There used to be an online publisher called Associated Content (bought by Yahoo for $100M in 2010) that used freelancers to write SEO-optimized content.
I realized that I could write some code that would scrape various sources of financial data and generate articles about stock tickers. Managed to make some money before the articles got flagged as too similar to each other, but if I had put a little more effort into differentiating them, I probably could have made a pretty penny.
I think financial data was where the company started. The first vertical I interacted with was real estate listings. We moved to basketball and then baseball. Aside from the articles, we sold "list" ads that would pick stats and throw players in a list. As we built click data, we were able to tune those lists to popular things.
I heard about this several years ago and became attuned to it. I follow a few small cap companies and cryptocurrencies with google news alerts, and most of the articles sound very robotic and just put data in paragraph form. When the headline is “XYZ up 2.3% today, a 43.9% above average increase” I know it’s computer generated.
>Voter turnout was down by 3.5 percentage points since the last general election.
hm, this example they used was my constituency. I walked around for about 15 minutes until I found the entrance because somebody moved the "POLLING STATION" sign against a wall where it was less clearly visible. I wonder if that's related :/
Most "mainstream media" reports on elections solely on the basis of the horserace aspect of it. We hear very little about the actual policies proposed and the arguments for and against them.
> We hear very little about the actual policies proposed and the arguments for and against them.
Really? Most "mainstream media" (in the UK) these days have something called leaders/referendum debates that invite the audience to participate in holding a leader or debater to account and they and the interviewer both scrutinise the debaters' policies in their manifestos in great detail.
There were many debates in this General Election and Local Elections and "that" referendum of 2016 all detailing arguments on both sides. So I believe the media sometimes does help with removing the contradictions in the actual policies rather than simply projecting the whole event like a horserace. (No matter how biased they might be).
That's something that should be spelt out in the article.
Also, if one has the resources to proofread all articles, how much more expensive would it have been to have humans write such trivial texts? The data is there, writing an article like those takes maybe two minutes?
Bloomberg has been running automated stories for a while - all flagged as automated so you know. Makes sense for company / eco releases but noticed they are getting more 'editorial' as well with some of newer automated stories.
Good piece by NYT earlier this year diving into some of the details and tech used.
The information in the example story could have been communicated much more succinctly via some type of (generated) infographic. Or, heck, just a table.
IMO, paragraphs of text are the worst way to communicate this type of data.
I read stories about stocks sometimes that I think are computer generated. They sound like XYZ reported profits up 10% while margins downs 2%. Typical margins in XYZ corporation's sector are 22%, while XY maintains a margin of 24%.." etc.
Sometimes I think, I could get this information from reading the 10q. Still I like getting the info in paragraph form. So yay computers.
For stories that are basically just regurgitating easily parsable statistics, this is going to become more and more common. The LA Times also has an algorithm that produces a story every time there's a large enough earthquake reported by the USGS. The byline gives the credit to "Quakebot", although the disclaimer that the article was auto-generated also credits the person who wrote the bot's code
25 comments
[ 2.3 ms ] story [ 49.4 ms ] threadHmmm... Googling: "elected as the MP for" "was created using some automation", reveals very little variation. We're not exactly talking AI powered journalism here. More mail merge than ML.
Next week: How computers wrote your last electricity bill.
"None of the stories have any quotations in, none of them have any analysis of what happened or what the significance is. It is purely a written version of what has happened based on the data. So that's quite a big downside in terms of quality of journalism."
However, Mr McKenzie said the BBC was still in the "very early stages of understanding what audiences want from the technology".
https://twitter.com/nicerbloke
With the way news works today, are we sure this is really a downside?
This works really well for mundane, data-driven stories such as lottery results or surf reports, both of which I built.
There are a bunch of companies, such as Narrative Science, that are doing much more sophisticated versions of this. You give them the data and they can churn out a machine-generated recap of a Little League game that reads much like an MLB story.
But what I think will really be a game changer is personalization for the reader.
Can you imagine the same story written 10 different ways for 10 reader personas? Maybe a reader with an education background might get a different lede than a reader with an interest in politics on a story about a new school district referendum. Or heck, maybe a Republican reader might get a different version of a Trump story than a Democrat. (Not that this would necessarily be a GOOD thing. But imagine the possibilities!)
Yseop is not cheap, though, and the company went under after massive financial fraud. The Tribune Network used a fair chunk of our content at the time. I've been waiting for this to be more common-place.
I realized that I could write some code that would scrape various sources of financial data and generate articles about stock tickers. Managed to make some money before the articles got flagged as too similar to each other, but if I had put a little more effort into differentiating them, I probably could have made a pretty penny.
hm, this example they used was my constituency. I walked around for about 15 minutes until I found the entrance because somebody moved the "POLLING STATION" sign against a wall where it was less clearly visible. I wonder if that's related :/
I wonder why?
Really? Most "mainstream media" (in the UK) these days have something called leaders/referendum debates that invite the audience to participate in holding a leader or debater to account and they and the interviewer both scrutinise the debaters' policies in their manifestos in great detail.
There were many debates in this General Election and Local Elections and "that" referendum of 2016 all detailing arguments on both sides. So I believe the media sometimes does help with removing the contradictions in the actual policies rather than simply projecting the whole event like a horserace. (No matter how biased they might be).
That would have been more expensive using humans.
Factual reporting is fairly trivial to automate. Add some language generation and you can even make it sound less dry.
More difficult is adding local contextualization and the impact of e.g. specific local factors that might have had an influence on the vote outcome.
Implying 'the computer' is doing anything here that a dedicated human could not is just bad communication at best and utter lies at worst.
The BBC has a fixed income, if something is too expensive it's functionally impossible for them to do.
Also, if one has the resources to proofread all articles, how much more expensive would it have been to have humans write such trivial texts? The data is there, writing an article like those takes maybe two minutes?
Good piece by NYT earlier this year diving into some of the details and tech used.
https://www.nytimes.com/2019/02/05/business/media/artificial...
IMO, paragraphs of text are the worst way to communicate this type of data.
Sometimes I think, I could get this information from reading the 10q. Still I like getting the info in paragraph form. So yay computers.