4 ms·

This is what the NLP community would call 'abstractive summary'. It fundamentally requires generating new sentences - which is an excruciatingly impossible task

by deepGem 10y ago

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.

lrem 10y ago
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 https://news.ycombinator.com/item?id=12353955
- daveguy 10y ago
  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.
  - lrem 10y ago
    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).
derefr 10y ago
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.
- deepGem 10y ago
  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.

by deepGem 10y ago

lrem 10y ago

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

daveguy 10y ago
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.
- lrem 10y ago
  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).

derefr 10y ago

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.

deepGem 10y ago
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.