I like it .. it seems to be done well. There are issues, though: -- its English language only -- its unclear how it handles questions -- its unclear how it handles speech acts that aren't grammatical -- its not clear how it handles pauses, fumblings, hemming and hawing, shouting, whispering -- It assumes that everything is a dependency tree -- it does not include any way to provide feedback from the reasoning and logic layers back to the parsing layer.
Well .. the same is true for link-grammar, also, at this time, but LG has a clear path on how to handle much or most of the above. It took me a while to understand that there is a lot more to parsing than simply creating dependency trees -- if that was all that there is to it, then most parsers would be "good enough" with any accuracy over, say 90% on some "typical" corpus. But accuracy is not really the issue -- that is not where the problem lies. A large part of the problem is that tweets aren't grammatical, but people still understand them. People abuse "fixed phrases" and "insititutional utterances" all the time, creating new language in that way: "Just do it" "Where's the beef?" "Mission accomplished". Synonymous phrases appear and disappear all the time -- I think Boing Boing had a good exploration of Archie comics from 60 years ago, and pointed out how much that language has changed since then -- teens just don't talk like that any more. English has a way of obscuring some of these issues, because it encodes a lot of meaning in the word order, as opposed to the morphology, and so English speakers are a lot less aware of using institutional phrases, and are less aware of how important they are for semantics, and how much they control the change and usage of language. The upshot is that dependency parsing is really just the tip of the iceberg. The good stuff, the interesting stuff lies elsewhere. Anyway, ... I almost wrote that it would not be hard to slot SyntaxNet in place of Relex. But that would not be quite correct. First of all, relex is inadequate for some of the things that opencog needs, and so opencog falls back to link-grammar in those areas where relex is broken. Another major issue would be how questions get handled -- relex does some funky stuff to try to replace the object of a question by a variable, and R2L uses that. Well, and even if you got past that, its not clear that R2L is doing the right things, since its again a rule-driven system, running on hand-written rules, when it should really be learning rules, via some machine learning mechanism. ... but all that needs to feed back into the parse itself... So perhaps syntaxnet is a step forward, but without a holistic approach, I don't see much utility for it. I can also put it a different way: prior to SyntaxNet, I believe that Link Grammar was the most accurate English language parser out there, and the amount of attention that generated was just miniscule. There is very little demand for parsers, and basically, nobody quite knows what to do with them. For example: with LG, people come by, they try it out, ask a few questions, and go their own way. The information that a parser provides is just not relevant to any software and machine-learning folks. The knowledge-extraction crowd pays lip service, but then doesn't actually use it as a tool. There's no real community, because there are no real applications. No one knows what to do with a parser. --linas On Fri, May 13, 2016 at 12:38 PM, Jeremy Zucker <[email protected]> wrote: > Hi folks, > > I would imagine that Google's release of SyntaxNet in TensorFlow will > provide a valuable boost in OpenCog Natural Languge Understanding efforts. > Although SyntaxNet cannot resolve common sense Winograd-style tests, it > frees up efforts like OpenCog to work on those kinds of problems rather > than the kinds of things that SyntaxNet handles. > > > http://googleresearch.blogspot.com/2016/05/announcing-syntaxnet-worlds-most.html?utm_campaign=Data%2BElixir&utm_medium=email&utm_source=Data_Elixir_79 > > Sincerely, > > Jeremy > > "Any sufficiently complex system is hackable" > http://www.linkedin.com/profile/view?id=4389790 > > -- > You received this message because you are subscribed to the Google Groups > "opencog" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > To post to this group, send email to [email protected]. > Visit this group at https://groups.google.com/group/opencog. > To view this discussion on the web visit > https://groups.google.com/d/msgid/opencog/CA%2BHW2JOL90RHRf4VHeiX_ZbPghg%2B5i7VTLCabjyBknfC%2BBJDpQ%40mail.gmail.com > <https://groups.google.com/d/msgid/opencog/CA%2BHW2JOL90RHRf4VHeiX_ZbPghg%2B5i7VTLCabjyBknfC%2BBJDpQ%40mail.gmail.com?utm_medium=email&utm_source=footer> > . > For more options, visit https://groups.google.com/d/optout. > -- You received this message because you are subscribed to the Google Groups "opencog" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at https://groups.google.com/group/opencog. To view this discussion on the web visit https://groups.google.com/d/msgid/opencog/CAHrUA355OH%3DRReS4vsYQph67o8n81qS4tHTTwwv%2BXzuOwZLSEA%40mail.gmail.com. For more options, visit https://groups.google.com/d/optout.
