:) it's grammatically correct because @gist uses an extraction based strategy. (ie, an existing representative sentence). If you want an abstract summary (ie, a new representative sentence), best to ask a human for now.
goto one of the following sites:
www.bbc.com/news, washingtonpost.com, huffintonpost.com, reuters.com, theglobeandmail.com
On those sites, find any text article that's medium+ size in length, and submit the article (not the home page of the news site) to gist.
You should see that it's not just reiterating the title.
On many sites it's quite buggy, and what you're seeing is likely the result of gist either a.) attempting to gist something when it shouldn't (like say bbc news home page), or b.) attempting to gist when the scrape/parse failed and there was too few or poorly parsed sentences.
We do extract the html title and stick it in at the top of the page as a reference, but that title, like each gist, is editable by the community, and serves only to act as a jumping off point.
In the distant future, but as of right now we're focused more on text and getting humans to start participating. The machine learning approaches we're using are getting better every day, and the latest deep learning techniques are showing promise with pictures, so that is certainly a possibility down the road.
Ya, it returns about 1/6 of the sentences in the documents as gists. For an item with 12 sentences, 2 is too few, but for something with 100 sentences, 17 is too many.
We'll probably change the ratio to scale in a smarter way rather than a constant fraction.
Go to any of those as a start and submit one of the articles you find and everything suggested by @gist was generated by an algorithm.
Our hope is that will help get the ball rolling. We are certainly very subject to network effects, so focusing on a niche and providing tremendous value to that one vertical is a good strategy.
Ah, your landing page should make that abundantly clear. That's a very cool feature.
I tried a couple of articles on Washington Post and the algorithm did a decent job. I then read the gist first before reading an article, and while some sentences were a little hard to understand w/o the surrounding context, I felt I still got a decent summary. I can see myself skimming the article through this service instead of relying solely on my eye balls when scanning through.
Couple of suggestions:
- Linking the extracted sentences back to the original page if I want the surrounding context
- A tool/browser plug-in which can allow me to select a representative sentence from the story and submit to gist
Seems like a useful stand alone service as it is; having people vote on and submit gists themselves would be cherry on top.
Thanks! There is a browser plugin planned that will show you the gists of any url you navigate to, and additionally could show the gists of links on a page before you click on them...
Ya, the website shown here doesn't have privileged access to the api. So, short of some basic rate limiting, we can release the same api the website uses to the public as a start.
It's a bit buggy still. To get a sense of it, stick to BBC articles.