Reasons why I do not care grammar formalism 
Jun-ichi Tsujii 
Dept. of Electrical Engineering, 
Kyoto University, 
Yoshida-honmaehi, Sakyo, Kyoto, 606, JAPAN 
General Comments 
Computational linguistics (CL) has borrowed a lot of ideas from 
Theoretical Linguistics (TZ). We could not have developed even 
a simple parser without the research results in TL. It is obviously 
nonsense to claim that we, computational linguists, do not care 
research results in TL. 
llowever, the researchers in TL, it seems to me, are very fond 
of fighlinq~ especially, those who are called Synlaclicians. They 
always fight with e~h other by asserting that their grammar 
formalisms are superior to the others'. They are oversensitive 
and tend to distinguish people into two groups, the ally and the 
enemy. 
A computational linguist using LFG (or pseudo LFG) as a 
small part in his total system is taken as the ally of LFG, and is 
certainly accused by the other groups. They promptly demon- 
strate that LFG is wrong, by showing a lot of peculiar sentences 
which rarely appear in real texts. 
We are tired of listening to such discussions. 
The Reasons Why 
Formalisms are prepared for accomplishing specific purposes. The 
formalisms in TL have been proposed, roughly speaking, for de- 
scribing the rules of distinguishing grammatical word sequences 
from arbitrary ung~ummaticalsequences, and of relating the gram- 
matical sequences with the other representational levels. 
On the other hand, a formalism we need in CL is for differ- 
eat purposes. That is, we need a formalism for describing the 
rules of distinguishing the most feasible grammatical structures 
from other less feasible but still grammatical ones of the same 
sentences \[Disambiguation\]. We also lined a formalism in which 
we can manage systematically a large amount of knowledge of 
various sorts necessary for NLP. 
Formalisms for different purposes, of course, should be eval- 
uated based on different standards. The current discussions of 
diffhreut formalisms in TL are irrelevant to our standards, though 
they may be important for their fights. The following is a list of 
the reasons why I think so. 
(1)\[Small and Peculiar Examples\]: Linguists Mways argue 
that their formalisms are better than others by using almost the 
same set of peculiar sentences. This implies that the differences 
of the formalisms are revealed only in these types of sentences 
which rarely appear in real texts. 
F~lrthermore, it often happens that all of the proposed formalisms 
can capture the same regularity. They only claim that their for.- 
malisms capture it more elegantly than others, elegantly accord-- 
ing to their standards. 
(2)\[Meaning and Knowledge\]: The elegance of their formalisms 
is obtained partly by ignoring uninteresting phenomena, again 
uninteresling according to their standards. Especially, they ig- 
nore largely phenomena related to meaning, extra-ling~tistic knowl- 
edge, contexts, etc. Or they ignore them by claiming that they 
are related to these factors. 
(3)\[Disambiguation\]: Linguists can ignore the uninteresting 
phenomena, but CL researchers developing actual systems can.. 
not. The systems have to cover wide ranges of phenomena which 
really appear in texts. Furthermore, disambiguation, which is the 
hardest problem in CL but not at all in TL, certainly requires 
considerations in the factors which syntacticians ignore. 
I fear that lhe elegance of their \]ormalisms becomes a serious 
obstacle ~o the introduction of such extra factors and processinga 
on them. 
(4)\[Management of Grammar Descriptions\] : l also fear that 
their elegance becomes an obstacle to the systematic develop. 
meat and management of grammar descriptions. Grammar de- 
scriptions here include the descriptions of both rules and dictio- 
naries. Some formalisms are claimed eleganl in the sense that 
they require only very few rules. But this elegance is obtained 
by very complicated dictionary descriptions. The standards ibr 
being elegant seem different in TL and CL. • 
(5)\[Processing Issues\] : The grammar formalisms ignore rnostly 
the processing issues. Linguists do not care processing issues in 
their formalisms just as we do not care grammar formalism. 
Conclusion 
I have to repeat here that I do not claim that TL research is 
irrelevant to CL. I only claim that grammar formalisms are not 
important. What is important is in their discoveries which are 
described by their formalisms. And what we have to do is to 
describe their discoveries in our own formalisms. 
452 
