[MUD-Dev] English grammar thoughts

Robert Zubek rob at cs.northwestern.edu
Wed Oct 24 09:40:00 New Zealand Daylight Time 2001


bruce at puremagic.com writes (re link grammar):
>-- begin quote --
>   The Link Grammar Parser is a syntactic parser of English, based
>   on link grammar, an original theory of English syntax.
>--- end quote ---

just as a matter of expanding on this description - from the little
that i've played with link grammar, it looks like an attempt to
focus on context-sensitivity, in contrast to traditional
context-free approaches.  grammatically, this means allowing each
word to impose constraints on the interpretation of their neighbors
- so that an adverb could expect to be preceded by a verb, or an
article to be followed by a noun. and these expectations, rather
than more abstract grammatical constructs, influence how the words
get interpreted. in the authors' own words:

   Rather than thinking in terms of syntactic functions (like
   subject or object) or constituents (like "verb phrase"), one must
   think in terms of relationships between pairs of words.  In the
   sentence below, for example, there is an "S" ("subject") relation
   between "dog" and "has"; a "PP" (past-participle) relationship
   between "has" and "gone"; and a "D" (determiner) relation between
   "the" and "dog".

         +-----Ds-----+
         |     +---A--+-Ss-+-PP-+
         |     |      |    |    |
        the black.a dog.n has gone


>  * How well do these systems deal with the poor grammar, spelling,
>  etc that most people seem to have?

it depends on the type of grammatical problem. it can handle
in-sentence mobility quite well (ie. "wave at the sky with your
sword" vs. "wave with your sword at the sky"), but it seems to be
breaking rather badly on something as mundane as missing articles
(eg. "put sword on box" breaks the parser, while "put the sword on
the box" parses wonderfully). in addition, chances are it won't
recognize even common misspellings, though it nevertheless should be
able to categorize them correctly as nouns, verbs, and other parts
of speech.

all of this is only expected, as the cmu link library had been
trained on a newspaper-style corpus.


>  * How much is the current research directed at handling full
>  sentences?  Will that same research work well in a command-driven
>  environment like a MUD?

my perception is that most research is driven by full sentences,
because it's easy to train probabilistic grammars on them. in regard
to muds, the command-like imperative voice shouldn't be a problem
(eg. "do this", "go there" are fine), but compound sentences and
sloppiness with articles might cause it to hiccup.

>  * What is the performance hit like for a system like Link?

i have a recollection of it being quite fast (tens of milliseconds,
which for a parser would be very fast indeed :) - but i'm afraid i
don't have any details.

fortunately, there's an online parser available for experimentation
at:

  http://www.link.cs.cmu.edu/link/


rob


--
Robert Zubek
rob at cs.northwestern.edu
http://www.cs.northwestern.edu/~rob

_______________________________________________
MUD-Dev mailing list
MUD-Dev at kanga.nu
https://www.kanga.nu/lists/listinfo/mud-dev



More information about the MUD-Dev mailing list