a small experiment with the R&D team, an installable bot that bring election coverage into your Slack room
intended to publish to a small number of friends and techies
well meaning colleague tweeted it out, and…
Nieman Lab found it
before we knew it, in some 500 Slack rooms
large audience made it more daunting
started doing live coverage on primary nights
how did we do that workflow? we used Slack itself as an admin
inspired by other work we’d done for live chats on live blogs
Separate room dedicated to bot posts
Publishing was controlled via emoji responses
Or at least, it was until the incident…
- put an extra layer of security over it
- put an extra layer of security over it
CMS for the Slackbot was actually Slack itself
Separate room dedicated to bot posts
Publishing was controlled via emoji responses
Can see these highlighted posts…
…are then seen in the output room
Names of authors stripped out, posting “as” the bot
But this wasn’t always the case
there was a separate emoji reaction
using that posted as the user…
Paul Volpe - Deputy Washington Bureau Chief
… like so
primary night coverage was OK, but most Slack audience at home
also very busy, difficult to carve our resources
instead, let’s make an event - office hours
Nate Cohn - Reporter for the Upshot
a live chat broadcast on our site
and mirrored by the Slack bot
one problem, though…
there was a lot of it
live for an hour. some users loved it, some hated it
realized we don’t know the context of our bot, what room is it in? how many people?
one success, though…
one success, though: questions
added a slash command, /asknate, piped questions into our Slack room
which we could then select to publish back out again as Nate answered.
Got us thinking about questions - existing command just forwarded to the newsroom
worked again with R&D to try natural language processing of questions
they created a dataset with some candidate positions and election calendar
used wit.ai, an NLP service bought by Facebook
has a mix of entities, intents and expressions
put them all together and you can see how it pulls out various terms from questions
this structured data is then sent back to us to process
worked great
for the first five minutes or so
very quickly realized that people’s questions would be numerous and diverse
some of them are answerable - for example, get donation data, good to go
most are not, kind of questions people asked Nate. Requiring deep analysis - which is why we’re here!
this does. not. scale.
supreme court example: is that an election question? what comes after that? is “I don’t answer that” a valid response?
you would need a full time team answering questions, and even then, would maybe fail
nature of news means dataset is always changing. Apple can afford to tweak Siri endlessly, we’d be starting over and over
within the confines of the experiment we weren’t able to do that justice
so we shut it down for now. Excited for the future, but a chat experience like this has to be very, very focused. Maybe Olympics - data-driven event, lots of people in the office during the day