A LOGIC-ORI~TED ATN 
Grammar Knowledge as Part of the System's Knowledge 
Helmet Gust, Carola Reddig 
Institut of Applied Cemputer Science, Technical 
University of Berlin (West), Ku~f~rstendgmm 202, D-IO00 Berlin 
The eystem BACON (Berlin Autcmatle COnstruction for 
semantic Networks) is an experimental intelligent qaestion- 
-answerlng system with a nataral la~uage interface based on 
single sentence input 1. 
BEAST 1 
SIREN I 
RUECK/ANSGEN 
Figure I: Structure of the BACON-system 
1 This system has been developed in the project "Auto- 
matisehe Erstellung eemantischer Hetze" (Automatic construct- ion of emantic networks) at the Institute of Applied Compuger 
Scienoe at the Technical University of Berlin. The pro~eet was supported by the Ministry for Science and Technology 
(BMFT) of the Federal Republic of German~. 
- 116 - 
Explanations of the s~stems structure: 
BEAST - "BErlin Automatic Semantic oriented Translation" is 
the language understanding component 2 
SIREN - "Semantic REpresentation and Net evaluator" is the 
knowledge base management component 
BACK - performs the retranslation of the semantic represen- 
tation into natural language to check correctness 
of understanding 
ANSGEN - "ANSwer GENerator" translates the system °s react- 
ions into natural language 
MSRL - "Modal Semantic Representation Language" is the 
representation language of the BACON-system and 
therefore the center of the whole system• 
The syntax of MSRL is reou~sivly defined following the 
principles of common logical languages• The basic (finite) 
set of operators is derided into two disjoint subsets: 
- tel~m~ers are operators which describe objects of the 
world of discalrse. They include the term variables. 
- formelmakers describe the relations between these 
objects, between formulas or between objects and 
fol~ulas • 
An operator with its (possibly empty) list of argmnents 
constitutes a well formed MSRL-expression. 
A very important concept of MSRL is the concept of sorts. 
The set of sorts (proper names of sorts) induces a partition 
of the set of terms. A partial ordering called "compatibility 
of sorts" induces a lattice str-cture on the set of sorts• 
By assigning a list of sorts (one sort for each argument 
place) called "sortal pattern" to each operator we get 
additional semantic constraints for MSR L expressions, The 
oonstructlon of the sortal pattern of an operator is based 
2 An earlier implementation of this component has been pre- 
sented at COLING 78: Habel/Rolllnger/Schmidt: "Machine trans- 
lation of natural language into a modal representation 
language" . 
- 117 -- 
on knowledge about the world of discourse. It determines the 
semantic category of the operator's arguments. The process of 
the assignment of meaning by BEAST is based on the corral 
classification of terms and the corral pattern of the operator. 
This process performs a semantically correct assignment be- 
tween the operator constituting the sentence and ~he terms 
being able to become arguments. 
There have been some unaerlying principles/conditions 
which affected the special ATN formalism in BEAST: 
1. BF~ST should emphasize a semantical analysis which con- 
stitutes the relationship between the natural language 
sentence and Its meaning representation. The aim is the 
creation of a MSRL expression that represents the meaning 
of the input sentence. 
2. The Ex~mmar component should follow the "principle of 
minimal information"- The process of translation tries 
to get along With as little information as possible. Only 
if that does not work, more (not only explicate but also • 
implicit) Information contained in the sentence need to 
be discovered. 
3. The ATN had to be pr0Ex~nmed in SXMULA since the rest of 
the system already existed in SXMULA. 
Our AT~ formalism is founded on the formalism developed 
by Woods, uses ideas of the semantic ATSs and takes the des- 
cribed principles into account by some modifications, extent- 
ions and constraints. 
We use an ATN compiling system which tranalates in a 
first step the ATN representation of the g~a~mar into a 
SIMULA program, and in a second step this 3XMULA representat- 
ion together with the other components Is translated into an 
executable machine proE~aa. 
The main feature5 of these AT~s ares 
- they use In addition to syntactic categories semantic 
categories and constituents 
- 118 - 
- that use semantic constraints to select the best pass 
through the network 
- and they create semantic representations. 
The fo~aaliem supports the treatment of alternatives at 
both sides: at the sour0~ . lan~uaRe by the possibility of 
alternative categories for the entries in the word form 
lexicon, and at the target language where alternatives relat- 
ed to a lemma refer to different sortal patterns of operators. 
, R,qLItor I 
~r~lk conditions 
T 
E ........ I @t 
I A~-BaSll L~sten , 
str~ng~l~ 
Figure 2: ATN compiling system 
The focus of our interest in future development is an 
integration of the ATN formali~n at a higher level, namely 
the level of the representation language MSRL itself, instead 
of level of the programming language SY~ULA. For this purpose 
MSRL has to be extended to represent procedural struct~ee 
like A~Ns. The main thing is that these structures are normal 
MSRL expressions, and therefore on the one hand they are 
manageble by the knowledge base (storable and retrievable) 
but on the other hand they are also evaluable. Starting from 
the fact that the aTN represents the grammatical knowledge 
of the system we have the ability to handle this knowledge 
- 119 - 
Just like the knowledge about the world of disooursez to 
ask questions about it and to modify It. 
In the same way we can represent the lexical knowledge 
in terms of the representation language. Together with the 
extension described above there will 'be the possibility to 
handle prooedul'al lexical knowledge, especially in the case 
of ATHs to link a word in the lexicon with a sUbATH which 
has the task to perform the next part of the analysis when 
-triggered by the occu.rance of this word in the natural 
language sentence. 
In ou~ lecture we will carry out the following points: 
- the description of the ATN fomalism in BEAST 
- the projected extensions of MSRL 
- the consequences of these extensions on the system 
o oneept. 
- 120 - 
