Hacker News Reader

Comment by dandare | Hacker News Reader

I wish there was an AI bot that would summarize any marketing speak to such a useful summary like you did.

Also I wonder what would such bot translate my product description to :D

PG's heuristic for this is 'write like you are explaining to a friend.'

The problem is that that kind of clarity is rarely achieved on a corporate website because Conway's Law applies as much to marketing as it does to software design.

For reference: https://en.wikipedia.org/wiki/Conway%27s_law

This is what the NLP community would call 'abstractive summary'. It fundamentally requires generating new sentences - which is an excruciatingly impossible task for current AI tech, at least as far as my understanding goes. The closest tech is the sequence-to-sequence deep networks. I am yet to dig deeper into this subject.

There was recently a Google result here[1] where they were able to summarise a paragraph into one sentence. It looked pretty convincing. How does your "excruciatingly impossible" stand when confronted with that? Were the paragraphs in that result particularly easy to summarise? Does the problem's difficulty grow super-linearly with length of the text? Is there something else in the way?

[1]: https://news.ycombinator.com/item?id=12353955

I would love to see that on some actual text (rather than published "samples"). I expect the more text it has to deal with the more unintelligible the result.

That's intuitive. What isn't is how much harder it becomes with longer text. Imagine there is an abstract measure of "smartness" of your algorithm (that would intuitively have some relations to your model size, teaching sample size and so on). My question could be worded as "what is the asymptotic smartness complexity of the summarisation problem?"

Btw, I'm not aware of existence of such a measure, which I take as a signal it's infeasible to formulate (or I'm just ignorant).

Is language generation (following syntax and grammar and etc.) the hard part? Because personally, I'd be fine with a summary in the form of a series of small concept-token-graphs, as long as it was indeed a series of such graphs (akin to sentences, arranged into paragraphs), instead of one big jumbled graph.

Yep language generation is the hard part. What you have said is a stepping stone in that direction. Any task involving sequences is very hard today. How to generate these sequences ? Tops we can generate limited phrases, what Google smart reply does.

that is a nice idea, but i think we are unfortunately years away from that. i dont have time at the moment, but take this article and run it though this state of the art algorithm and see what comes out:

https://research.googleblog.com/2016/08/text-summarization-w...

A Mechanical Turk-based system could be a good middle ground.

This is more than just a summray

Reply on news.ycombinator.com