Re: Is it time for AI with Accessibility?

Some of this stuff is already in the field. See, for example, Adobe Reader for mobile devices. Its new “liquid mode” feature - intended for reuse rather than accessibility applications - is a hint of the AI-assisted future of live analysis and repurposing of unstructured PDF content.

More info:

https://www.pdfa.org/adobe-announces-liquid-mode-for-acrobat-mobile/

Duff.

> On Dec 8, 2020, at 15:31, Wayne Dick <wayneedick@gmail.com> wrote:
> 
> Thanks Charles,
> Right around 2000 is when I abandoned pattern recognition as a means to rationalize visual pages. At that time, owing to a drop in AI funding, and some hardware limits really signaled a dead end. However, in the last 10 years graphics hardware has improved dramatically, and we have a lot of very regular data... a lot, like a web's worth.
> 
> When we approached this in the 90's to about 2000 we were dealing with digitized pages from articles and books. Just finding the angle of the lines on a page was a process. There was  a lot of noise to filter. 
> 
> With electronic-based literature we have a completely different state of input. Skewed lines are usually decorative, a category of their own. There is a matching between screen regions and runtime code elements.
> 
> I see three approaches at least. Analysis based on generated code, image analysis of displayed text and hybrid analysis based on our knowledge of the content. This could lead to a couple of accommodations that would exceed the impact of anything we have today.
> 
> 1) Very smart screen magnification. This could apply to professional documents in PDF or another printer oriented language. This would make Journal articles accessible.
> 2) Very smart analysis of content based on the image. That is rendering a markup equivalent that could be read and personalized given the user's needs configuration.
> 3) Hybrid analysis resulting in the  output of 2) or 3).
> 
> In the 90s we got bogged down with Post Office examples where the goal was to extract and process address and zip code information from very messy input. 
> 
> It just seems like it may be time to look at this again. We do have dramatically better tools in 2020.
> 
> Thanks all for your feedback,
> Wayne
> 
> 
> On Tue, Dec 8, 2020 at 2:11 AM Charles 'chaals' (McCathie) Nevile <chaals@yandex.ru <mailto:chaals@yandex.ru>> wrote:
> On Tue, 08 Dec 2020 07:20:48 +1100, Wayne Dick <wayneedick@gmail.com <mailto:wayneedick@gmail.com>>  
> wrote:
> 
> > I am interested in any research in this direction. Anybody know about  
> > anything like this in progress?
> 
> Hello Wayne, all.
> 
> I went to a presentation in New Zealand in the early 2000s, at the  
> invitation of Graham Oliver, on a project that had been running for quite  
> some years (if I recall correctly, since the early 90s) to do this.
> 
> I no longer recall enough to easily find it (and I have looked for it  
> before without success).
> 
> The basic idea was to use machine learning systems to look at the  
> interface of a user's computer, and provide a personalised approach to  
> understanding the components. Initially the system used a very expensive  
> high-powered computer to read the interface of a standard desktop PC, but  
> as increasing power became available, it was slowly morphing toward  
> software running directly on the machine.
> 
> I also recall that a large part of the explanation about automatic visual  
> recognition used jet fighter planes as the example object to follow.
> 
> In my mind the project may have been associated with Stanford University,  
> and it may have been called Eureka, although that is widely used as a  
> name, so not a very helpful search term :(
> 
> If this rings a bell with anyone I would love to find more pointers to the  
> work.
> 
> Cheers
> 
> Chaals
> 
> -- 
> Using Opera's mail client: http://www.opera.com/mail/ <http://www.opera.com/mail/>
>