GERERALIZED SYNTACTIC RELATIONS AND SUBSTANTIONAL A~TRIBUTES 
Arcady Borkowek¥ 
Computing Center, Academy of Sciences, Moscow, USSR 
The psper presents a conceptual frs~ework for natural 
lan~age analysis within which some experiments were held and 
some ideas had been developed. 
The work concerns the means of translation of natural 
language text into its meaning repreeantation in form of 
semantic network basing on frwnes°fo~ns£1~. The expex~ment~ 
used Russian as 1spur language. 
I. The analysls is essentially vocabulary driven. Semant- 
to lnfox~attor- Is Intensively used; indeed, the formalism 
does not make much difference between grammar and sanant£oe, 
I% could be regarded as a Generalized syntax., The approach 
leads tO distribution of words Into classes quite different 
fx~n usual ~ran-nattoal classes, but having obviously ltnguiBt- 
to. meaning. 
The basic ideas are related with those of /1/, /2/; the 
earlier variant of vocabulary structure is given in /3/; the 
Y~iep implementation of vocabulary and semantic network nses 
property lasts with Inheritance havi~ much in oomaon with 
PaL, 
2. The language description consist of Semantlc and 
Lexle vocabularies. 
2.1, The entries of Semantlo Vocabulary are notlons, 
for~ an abstract sez~utio network for meaning represent- 
ation &u~ented with gremma~ tnfo~natton, 
- 49 - 
The vocabulary article contains the following: 
a) a reference to supernotion! all information from 
supernotion is relevant to actual notion, if it isn't explicit- 
ly euperoeded. The "notion-supernotion" relation imposes 
hierarchical structure on the set of notions. 
b) a list of the notion's attributes with corresponding 
Generalized Syntactic Relations (GSR). The set of all GSRs 
forms the Eralmar used in an~ysis. The GSR attached to attrib- 
ute must hold for the words (or phrases) of the NL text, the 
first (master) referring an instance of the notion, the second 
(slave) referring to the attribute value. E.g. for-Russian, 
the ~RECEI~ENT attribute of ~GIVE would have a GSR demanding 
the slave to be an instance of ~PF~SON and to have the form of 
Dative case. (The dollar sign is used to distin~n~sh notions 
from words. ) 
The attributes inherited from supernotion may also have 
a specification of default or fixed value, which is immediate- 
ly inserted into meaning representation. 
2.2. The Lexio Vocabulary contains words. Words can be 
si~ican% or auxiliary. Significant words are those, which 
name notions, attributes or attributes values. All other words 
are au~liary~ they are treated as components of analytic 
morphological forms and are processed by prescan. Therefore 
only si~ficant words are present in the vocabulary. 
Lexio Vocabulary article contains: 
a) a reference to Semantic Vocabulary with indication 
of the role cl~s (notionPn~e, attribute-n~ne etc., see 
below) and, optionally, a lexical function (see /1/)o 
b) ~ammatical attributes of the word: grammar class, 
morpholo~cal pattern, fixed grammatical values (such as gend- 
er for nouns, aspect for verbs, etc.). 
3. GSR is a logical function of master's and slave's 
attributes" values. These can be grammatical attributes from 
- 50 - 
Lexic Vocabulary or provided by presoan, attributes inherited 
by slave from supernotion or attributes values reflecting the 
meaning of the text. The most usual ce~es involve matohin 8 
slave+s grammatical attributes to some given values or to 
those of master, and demand for slave to refer a subnotion of 
a given notion. However some GSRs are more sophisticated. 
3.1. The grammar part of GSRs has much in common with 
Surface Syntactical Relations (3SR) of /1/. Indeed the GSRs 
had been inspirited by SSRs. 
GSRs differ from SSRs in two aspectsz first, they system- 
atioally use semantic information! second, GSRs usually deal 
with a deeper syntax level! e,g, if the grammar part of GSR 
postulates a "direct-object relation", its description may 
cover active and passive verbal, participial and nominal 
oonstructionsz"t0 write a letter", "a letter is written", 
"the letter written by..." and "writting a letter". 
3.2. However, the GSR teohniq~e allows different ~ye.to 
describe a fra~nent of language~ all depends on the attributes 
tested. Making the 6ran~ar part of GSRs trivial, one recieves 
a pure semantic ~asmar. On the other hand, if the set of 
notions is the set of grammatical classes, GSR~based analysis 
turns into traditional syntax analysis. 
3.3. When NL text has been transformed into a set of 
property lists, fetched from the vocabularies and augmented by 
morphological preeoan, there are many ways to order the applic- 
ation of the relevant GSRs, for, while each GSR is described 
procedurally, the description as a whole is declarative. In 
our experiments a simplification of the parsing algorithm 
described in /2/ was used. 
4.1. The outlined approach demands a classification of 
words different from one based on grammatical classes. 
Significant words are devided into classes depending on 
the role they play in n~ing corresponding notions. We dieting- 
-51-,. 
uish four main classes - N, A¥, A and SA. 
Class N i8 the largest; it is comprised by words which 
name notionB, instances of notions and venues for some attrib- 
utes. This class covers most nouns, verbs, verbial ad~ectives 
and numerals. It also contains a small but very important 
subclass of pronouns. 
Class AV is formed by words, na~ng attribute together 
with its value. This class covers most adjectives and adverbs. 
Class A consists of words naming attributes. Usually 
they are nounso 
Example: 
The followAng words refer the same notion ~FLIGHT in 
different ways: "to fly", "flight", "flyln~ ~ust name it and 
a~e of the class N; '~peed" is an example of class A, it nsmee 
an attribute of ~FLIGHT! "quick e and wquAckly" refer to the 
same attribute, but provide a v~ue (Magn) for it, these two 
are the me~.ber of class A¥. 
"Simple stories" use mostly N and AV. The words of 
class A are common in NL-aocess to data-basses 
4.2oThe fourth class is formed by Subetantional Attri- 
butes (SA). As a matter of fact, their discussion was the 
main motivation to write this paper. The separation of it 
from other classes makes it possible to process 88 "linguist- 
io e some text r~lations usually treated as "semantic e and 
requiring some deductive system to process them. 
Subetantional Attributes combine the properties of the 
other three classes. They name attribute, provide its value 
but focusize on an object - the attribute *s value. E.K. "ca- 
pital" names an attribute of ~OUNTRY, but the focus i8 nor 
country neither ~he attribute~name but an instance of ~CITY. 
n attribute has, so to say, its own substantion 
'ly separated from the master. 
- 52 - 
Indeed, unlike attributes of class A, Subetantional 
Attributes may be used without explicit reference to its 
master: e,g, "the train goes to the capite£". If the master 
is present, the master-slave relation need not to be express- 
ed syntactically: 
"Peter ~ave the so.._nn an apple." 
Traditional syntax ignores the posessive link between 
subject ("Peter") and object ("son") in this phrase; nor does 
it consider the previous one incomplete. 
Our GSR-grannnar claims the presence of anaphora in 
these phrases. ~he first contains an unresolved reverence to 
some country! in the second one the word "son" includes a 
reference to some "parent", the most probably - Peter (i.e. 
the denotat of "Peter"). 
Defiui tion: 
A word is Substantial Attribute if it refers an object 
by naming its relation to some other object or situation. 
In text Substantional Attribute actfas if it were bound 
by poseseive relation to a "virtual pronoun" of appropriate 
semantic class. For example, these two phrases can be yiewed 
8~ 
and 
"The train goes to the capital Eof-country" 
"Peter gave hi_.ss son an apple" 
In the first phrase "1of-country" stands for such a 
"virtual pronoun". In the second one, "virtual pronoun" occas- 
ionally turned to be a real one; while in English its use is 
quite natural, in Russian the use of posessive pronoun would 
have an emphatic meaning. 
The anaphora resolution for such "virtual pronouns" 
iS done in the same way as for real (lexically expressed) pro- 
nouns. However, it is possible to take benefit of the fact, 
- 53 - 
that "virtual pronoun" refers an instance of specified notion, 
while for real pronoun only grammatical values ere known. 
Another example: 
"President and wife came to capital" 
(Articles and pronouns ere dropped to reflect Russian), 
This phrase is processed as 
"President roof-country with xhis wife came to capital Eof- 
country". 
Indeed the wife is the wife of this president, and they 
came to the city, which is the capital of the country of which 
he is the president. To infer this from the phrase no extra- 
-linguistic deduction is needed. 
4.3, The experiments have shown, that treating SAs as 
two words: one refering an object and another a "virtual pro- 
noun", is helpful in analysis oriented on extraction of mean- 
Ing of the text. In fact It deals with a more general question 
of the limits between "language knowledge" and "knowledge of 
world". 
References

I. "A Theory of Linguistic Models MEANING -- TEXT", Moscow, 
1972 (in Russian). 

2. M.C.McCord "Slot Grammers", Am. Journ. of Computational 
Linguletios, V.6, n.1, 1980, p, 31. 

3. A.B.Borkowsky, G.V.Senln "Lingulstical analysis of Quelrlss 
to Dialogue System", in "Interactive Systems" (the Proceed- 
ings of Soviet-Finnish Conference) Moscow, 1979e 
