Lukasz, Thanks for the information about Word Grammar, which for anyone else interested, is described here,
You asked: (4) I'm interested in how do you handle backtracking: giving up on application of a construction when it leads to inconsistency. Chart-based unification parsing can be optimized to share applications of constructions which are "parallel", and this can be extended to operators which are (like unification) monotonic, e.g. cannot make unsatisfiable/inconsistent state a satisfiable/consistent one. Merging conjuncts new facts to old ones so it is monotonic in monotonic logics. (Default/defeasible logics are nonmonotonic.) My first solution to this problem is to postpone it by employing a controlled English, in which such constructions will be avoided if possible. Secondly, Jerry Ball demonstrated his solution in Double R Grammar at the 2007 AAAI Fall Symposium, Cognitive Approaches to NLP. His slide presentation is here, which I think fully addresses your issues. To summarize Dr. Ball's ideas, which I will ultimately adopt for Texai: Serial processing [word by word parsing] with algorithmic backtracking has no hope for on-line processing in real-time in a large coverage NLP system.The solution is serial processing without backtrackingIf current input is unexpected given the prior context, then accommodate the input by adjusting the representation [parse state] and coerce the input into the representationDr. Ball gives as an example, parsing the utterance "no airspeed or altitude restrictions". Upon processing the word "or", the conjunction is accommodated via function overriding in his grammar, not by backing up. (4a) Does the fact that your parser is incremental mean that you do "early commitment" to constructions? (Double R Grammar seems to support early commitment when there is choice, but backtracking is still needed to get an interpretation when there are only ones without it.) Yes, my parser makes the earliest possible commitment to a construction, but I allow subsequent elaboration of constructions as new constituents are recognized. For example, in my use case sentence "the book is on the table", I recognize a initial Situation Referring Expression construction to cover the partial utterance "the book is", which is elaborated to form the final Situation Referring Expression when the remaining utterance "on the table" is processed. I regret that some aspects of my implementation are difficult to follow because I am using Jerry Ball's Double R Grammar, but not his ACT-R Lisp engine, using instead my own incremental, cognitively plausible, version of Luc Steel's Fluid Construction Grammar engine. I combined these two systems because Jerry Ball's engine is not reversible, Luc Steel's grammar is not a good coverage of English, and the otherwise excellent Fluid Construction Grammar engine is not incremental. -Steve Stephen L. Reed Artificial Intelligence Researcher http://texai.org/blog http://texai.org 3008 Oak Crest Ave. Austin, Texas, USA 78704 512.791.7860 ----- Original Message ---- From: Lukasz Stafiniak <[EMAIL PROTECTED]> To: [email protected] Sent: Sunday, April 13, 2008 3:04:07 PM Subject: Re: [agi] Between logical semantics and linguistic semantics On Wed, Apr 9, 2008 at 6:03 AM, Stephen Reed <[EMAIL PROTECTED]> wrote: > > I would be interested > in your comments on my adoption of Fluid Construction Grammar as a solution > to the NL to semantics mapping problem. > (1) Word Grammar (WG) is a construction-free version of your approach. It is based solely on spreading activation. It doesn't have a sharp separation of syntax and semantics: there's only one net. Nodes representing subgraphs corresponding to constructions can be organized into inheritance hierarchies (extensibility). But "pure WG" makes things very awkward logics-wise, making it work would be a lot of research (the WG book doesn't discuss utterance generation IIRC, but reversing parsing-interpretation seems quite direct: select the most activated word which doesn't have a left landmark, introduce a word-instance node for it, include spreading its activation through a right-landmark (ignoring direction of the landmark) edge). Texai is impure by its very nature, perhaps it could be made more (than just sharing the spreading activation idea) of a mix WG*FCG. (2) FCG is closer to traditional apporaches a la "computational linguistics" than WG. (3) One could give up some FCG features to simplify it, for example by assuming one-to-one correspondence between constructions and atomic predicates. (4) I'm interested in how do you handle backtracking: giving up on application of a construction when it leads to inconsistency. Chart-based unification parsing can be optimized to share applications of constructions which are "parallel", and this can be extended to operators which are (like unification) monotonic, e.g. cannot make unsatisfiable/inconsistent state a satisfiable/consistent one. Merging conjuncts new facts to old ones so it is monotonic in monotonic logics. (Default/defeasible logics are nonmonotonic.) (4a) Does the fact that your parser is incremental mean that you do "early commitment" to constructions? (Double R Grammar seems to support early commitment when there is choice, but backtracking is still needed to get an interpretation when there are only ones without it.) I will get to studying your sources when I'll have some time... ------------------------------------------- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?&; Powered by Listbox: http://www.listbox.com ____________________________________________________________________________________ Be a better friend, newshound, and know-it-all with Yahoo! Mobile. Try it now. http://mobile.yahoo.com/;_ylt=Ahu06i62sR8HDtDypao8Wcj9tAcJ ------------------------------------------- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244&id_secret=98558129-0bdb63 Powered by Listbox: http://www.listbox.com
