Szeged, 2003 december 10-11 115
K n o w le d g e -B a se d N a tu ra l-L a n g u a g e P ro c essin g
László K álm án1, László Balázs2, and Miklós Erdélyi Szabó3 . 1 Applied Logic Láb, Budapest
Research Institute for Linguistics, Hungarian Academy of Sciences, Budapest Theoretical Linguistics Programme, Loránd Eötvös University, Budapest
kalmanQnytud.hu 2 Applied Logic Lab, Budapest
bazsiQall.hu 3 Applied Logic Lab, Budapest
Alfréd Rényi Institute for Mathematics, Hungarian Academy of Sciences, Budapest mszaboQrenyi.hu
K eyw ords: computational linguistics, syntax, semantics, construction grammar, knowledge-based approaches
C om putational linguistics’ enthusiasm for the formal models of syntax and se
m antics introduced by Chomsky and Montague in the 70’s declined during the 80’s and 90’s. I t turned ou t th a t an autonomous syntax, without access to se
m antics, or an autonom ous semantics, w ithout access to pragmatics and world knowledge, are hopelessely untractable computationally and not suitable for ap
plications.
T he alternative th a t we propose is th a t modules axe needed, bu t they m ust correspond to different types of linguistic knowledge and different modes of op
eration (such as statistical, associative and logical modes) rather than different linguistic ‘levels’. T h e statistical modules perform jobs like speech recognition and synthesis, associative modules take care of invoking forms, structures and meanings from memory, and logical modules perform reasoning in understanding and planning tasks.
The m ain ingredient of our system is a b d u c tio n : b o th understanding and producing utterances requires us to complete an array of pieces of information about th e formal an d semantic aspects of an utterance. All sources of informa
tion, including linguistic, pragm atic and cognitive sources can be used in order to achieve this goal. This is th e key procedure th a t our system performs.
Linguistic knowledge itself is stored in a form at th a t does not draw a sharp line between linguistic, pragm atic and cognitive aspects of signs. The framework th a t makes this possible is called c o n s tru c tio n g ra m m a r. Construction gram
m ar is abo u t system atic (conventional) associations of form types and mean
ing types, and inform ation about the probabilities of their co-occurrence. The starting-point of th e abduction process in our system is a loosely connected network of ‘activized’ constructions.