LKB or TRALE

[Stefan Müller]

Hi Mariana,

> writing and expanding HPSG-structures for the German

Interesting. To bad that the mail adresses do not give anything away
about the affiliations. I would be interested to learn more about the
project ...

> I read the book "Implementing Typed Structure Grammars" and the
> TRALE-Description on Stefan Müller's Homepage.
> 
> Unfortunately I can't decide which is better!

Oh, I failed =;-)

There is a detailed comparison of the two systems in a paper by Nurit
Melnik:

@article{MelnikHandWritten,
author = {Nurit Melnik},
title = {From ``Hand-Written'' to Computationally Implemented {HPSG}
Theories},
journal = {Research on Language and Computation},
Volume = 5,
Number = 2,
pages = {199--236},
year = {2007}
}

A brief summary is this: TRALE can do everything that the LKB does
except default inheritance in the type hierarchy, but the defaults can
be modeled in a different way if you want to.

In addition to what the LKB does you get in TRALE:

1) empty elements (you do not have to use them if you hate them =;-)

Most researchers working on German assume that it is a SOV language and
that there is some trace of the verb when we look at sentences with the
finite verb in initial position.

If you want to express this without traces you are in deep trouble. It
can be done, but it will make you unhappy. You can read about this here:

http://hpsg.fu-berlin.de/~stefan/Pub/surface.html

2) relational constraints (most HPSG papers use them in a way that
cannot be expressed in LKB)

Examples are `append' with several lists, `shuffle', and `last' - a
relation that refers to the last element of a list (of indetermined length).

3) complex antecedents

You will find an example for constraints with complex antecedents here:

http://hpsg.fu-berlin.de/~stefan/Pub/chinese-svc.html

If you have them you can say about a certain configuration that it has
certain properties just in case other properties are instantiated. For
instance: If daughter one has perfective aspect, the whole construct has
a `final' meaning. If daughter two has perfective aspect, the whole
construct has a `causative' meaning.

Without such constraints you have to specifiy all possibilities
explicitely in a hierarchy of constructions and you get
theories/implementations with 350 dominance schemata.

I also used such implicational constraints to determine the clause type
(imperative, question, declarative) lexically instead of using phrasal
constructions to do this. This is described in my HPSG textbook (Section
10.3):

http://hpsg.fu-berlin.de/~stefan/Pub/hpsg-lehrbuch.html

4) cyclic structures

If you want to model verb movement in German in a GBish way (as most of
the German HPSGians do), you need cyclic feature structures. LKB does
not allow them.

Construction grammarians love cyclic structures, since you need them for
the evokes-operator: There is the classical example for the concept
`hypothenuse'.

You can read about this in my grammar theory book in Section 9.6.2.

http://hpsg.fu-berlin.de/~stefan/Pub/grammatiktheorie.html

5) Morphology is fully integrated in TRALE

You can manipulate the phonological contribution of a linguistic object
as you like in the lexical rule component. This can be used to do
morphology of semitic languages, where you insert vowels into consonant
clusters.

6) Gerald probably has to add something here ...

Any of the above points may have changed in the past, but this is what I
know from using LKB till 2003 and from what I heard recently.

I hope this helps. If you need any help with TRALE, we would be glad to
help.

Best wishes

        Stefan

-- 
Stefan Müller       Tel: (+49) (+30) 838 52973
                    Fax: (+49) (030) 838 4 52973
Institut für Deutsche und Niederländische Philologie
Deutsche Grammatik
Habelschwerdter Allee 45
14 195 Berlin

http://hpsg.fu-berlin.de/~stefan/

http://hpsg.fu-berlin.de/~stefan/Babel/Interaktiv/