I know the sum total of nothing about computational linguistics, but I do agree with your assessment that too many approaches go top-down and hence are pretty much useless for a language like Hebrew. Just for a single example,
S -> NP VP clearly can't handle a Hebrew verbless clause, nor can it handle a clause with a purely inflectional subject like the typical wayyiqtol clause. So as I see it, we need to extract the basal phrase structure rules from the data we have rather than trying to impose something from another language onto Hebrew. I would suggest as a starting point CL -> (P) That is, a clause is a single predicated unit, whether the predicate (verb, for most purposes) is explicit or not. On 19 Jun 2010 at 20:12, James Christian wrote: > > Hi Eric, traditionally in computational linguistics literature a > distinction is made between deep > and shallow methods. In general, deep methods are considered to be > rule based while shallow > methods are typical driven by statistical information. Obviously, > shallow methods of generation > can be used as well as deep methods and, no surprises, shallow > methods are performing better > than deep methods as in just about every other area of computational > linguistics. But please > don't fall into the classic trap of assuming that shallow methods > have no rule system. They have > a rule system, a very complex one. It just isn't made explicit. > As you note 'the number of rules could get out of hand'. However, I > would venture that the very > reason shallow methods are working better than deep methods is > precisely because they have > many more implicit rules that give a wider coverage of real > linguistic high frequency > phenomenon. And so I would like to address the following part of > your statement 'out of hand'. > Now I'm not sure exactly what you mean by out of hand. If you are > assuming that a large amount > of rules is a bad thing then I don't necessarily agree. In fact, > precisely the opposite. The very > reason that most modern attempts to define a generative grammar > aren't as wide coverage as > they set out to be is because they are designed to be compact and, > as a result, over generate. > That is not to say I disagree with you entirely. If many rules are > defined with no control of how > they affect each other then unpredictable results will naturally > arise. My personal experience > suggests to me that starting from the specific and working up to the > generic is a wiser approach > to defining a generative grammar. Most attempts to build a > generative grammar start with a > generic sentence rule like: > S --> NP VP > It comes as no surprise that such attempts generally fail. They are > starting with a linguistic > assumption about the data that they have not extracted from the data > but rather imposed upon it. > This is what I would call a top down approach starting with a > linguistic assumption about what > the top is. It would seem to make more sense to me to adopt a bottom > up approach with no > linguistic assumptions other than letting the data speak for itself. > That is to say to start by > defining rules for the smallest frequently observable phenomenon and > working your way up to > sentence structure, paragraph structure etc. and thus letting the > data speak for itself what the > definition of S is. It may be S --> NP VP, it may be something > entirely different or it may even be > a whole collection of statements. The point is to let the data speak > for itself. > And so, Eric, what I would like you to consider is that it is not > generative grammar in itself that is > the problem. It the method of defining one that may be at fault. > Please also bear in mind that > statistical methods of generative grammar are also rule based. They > only differ in that their rules > system is superior and machine learned from the data. > James Christian Dave Washburn http://www.nyx.net/~dwashbur _______________________________________________ b-hebrew mailing list [email protected] http://lists.ibiblio.org/mailman/listinfo/b-hebrew
