North.Holland Publishing Company © Academi~ 1982
REFERENTIAL NETS WITH ATTRIBUTES
Christopher U. Habel *
ODe of the essential problems in natural language production
and understanding is the problem of processing referential
relations. In this paper I describe a model for representing
and processing referential relations: referential nets with
attributes. Both processes (analyzing and generating
referential expressions) are controlled by attributes. There
are two types of attributes, on one hand, the ones to the
internal substitutes of the objects spoken about, on the other hand, the ones to the descriptions of these objects.
I. BASIC NOTIONS: KNOWLEDGE AND SEMANTIC REPRESENTATION
It is a well-known fact that (computational) models of human language production
and understanding have to take the knowledge of the speaker and the listener into
consideration. Thus, one of the main problems is to describe and represent this
knowledge and to distinguish relevant subtypes. A suitable classification which I
described in detail in Habel [7] distinguishes three subtypes:
- knowledge of facts or assertions, e.g. of states of the world, events, etc.,
the "factual knowledge',
- knowledge of rules, e.g. of rule-like relations between objects or states of
the world (or classes of such entities), the "inferential knowledge',
- knowledge of objects, e.g. of persons in the world, the "referential
knowledge'.
(In this paper I give emphasis to the third type: referential knowledge
represented by "Referential Nets').
The common core for representing all these types of knowledge is the "semantic
representation language" SRL. From the formal point of v i e w SRL is a
(propositional) symbolic language as described by Kalish/Montague [8]. Well-formed
SRL-expressions are generated by normal (recursive) formation rules. I
distinguish - as usual - two kinds of meaningful expressions in SRL, terms and
formulas. Thus we define:
TER - the set of well-formed terms FOR - the set of well-formed formulas SRL = <FOR, TER>
For the purpose of knowledge representation SRL contains some specific operators,
e.g.
on the factual level: tenses, beliefs, ...
on the inferential level: entailment relation,...
on the referential level: description operators of various types.
Since SRL is a representation language, I now have to treat the central question
of "representation theory': What kinds of entities are represented by the
different entities of SRL? Here I follow - with minor changes, which I will not
discuss in this paper Miller [I0]: The expressions and entities of SRL
represent cognitive concepts. The operators correspond to different types of
"concepts" in the Miller-approach, e.g. predicative , nominal and modifying
concepts (this third class are the "operators" of Miller.). E.go by means of
description operators it is possible to construct nominal concepts from
102 CH. U. HABEL
p r e d i c a t i v e concepts or other n o m i n a l concepts.
I shall n o w illustrate SRL w i t h an example. In u n d e r s t a n d i n g and r e p r e s e n t i n g the sentence
(i) I w i l l meet B a r b a r a in the u n i v e r s i t y tomorrow.
the act of referring to "I", "Barbara", and "the u n i v e r s i t y " is to be processed, i.e. the ( o b J e c t - ) o r i e n t a t i o n to them is the i n t e r e s t i n g p r o b l e m from the referential point of view. The SRL r e p r e s e n t a t i o n of (i) is
(2) time (meeting (l,Barbara, ETA(x) :university(x) ),tomorrow)
The a r g u m e n t s of "time" and "meeting" p r e s e n t the most relevant types of
referential operators: "I" and "tomorrow" are deictic substitutes (personal or
temporal), B a r b a r a is a p r o p e r n a m e and the e x p r e s s i o n "ETA(x): u n i v e r s i t y ( x ) " is
built up from the n o m i n a l concept "'university" (representing a class of
i n s t i t u t i o n s for e d u c a t i o n and science) and the v a r i a b l e - b i n d i n g d e s c r i p t i o n operator "ETa'. "ETA(x): p(x)" has the m e a n i n g "'an element from the class {x/p(x)>". E T A is an i n d e f i n i t e a n a l o g y to the IOTA-operator of formal logics (Similar o p e r a t o r s are i n t r o d u c e d by Hilbert as E P S I L O N - or E T A - o p e r a t o r . ) . The s e m a n t i c s of the d e s c r i p t i o n operators is given by a set of i n f e r e n c e and
e v a l u a t i o n rules. E.g. there is an inference rule w h i c h relates the ETA-
e x p r e s s i o n s to e x p r e s s i o n s w i t h e x i s t e n t i a l quantification: ( 3 ) p ( E T A ( x ) : q ( x ) ) ~ EX x : p ( x ) & q ( x ) 1"
Some e v a l u a t i o n rules w i l l be d e s c r i b e d in the following chapters, e.g. "creation
of a referential object'. (Both types of rules are d e s c r i b e d in detail in Habel
[5], [6].) Thus, E T A is the formal r e p r e s e n t a t i v e of one of the m e a n i n g s of i n d e f i n i t e articles.
F u r t h e r d e s c r i p t i o n s of the objects m e n t i o n e d in (i) or in the SRL e x p r e s s i o n (2)
can be generated by (focussing) t r a n s f o r m a t i o n s similar to the solution of
a l g e b r a i c equations: The u n i v e r s i t y m e n t i o n e d above is also the one described by (4) ETA(x) :time ( m e e t i n g ( l , B a r b a r a , u n i v e r s i t y ( x ) ) , t o m o r r o w ) 2~
or the n a t u r a l language equivalent "a (the) u n i v e r s i t y where Barbara and I will
meet tomorrow". (The uniqueness of the E T A - d e s c r i p t i o n depends on the
s l t u a t l o n a l / t e x t u a l context.)
B e f o r e we pass on to a d e t a i l e d d e s c r i p t i o n of "referential nets', I will make a short r e m a r k on d i s c o u r s e and communication. As m e n t i o n e d above h u m a n language p r o d u c t i o n and u n d e r s t a n d i n g is based on the s p e a k e r - l i s t e n e r ' s k n o w l e d g e of the language, the w o r l d and p a r t i c u l a r l y the p a r t i c i p a n t s in the communication. Thus, I a s s u m e a d i s c o u r s e model 3* which is based Dn t r i p a r t i t e k n o w l e d g e (namely
factual, inferential, and referential), each of these types is r e p r e s e n t e d by
s t r u c t u r e d sets of S R L - e x p r e s s i o n s .
\
2. T H E R E F E R E N T I A L NET
Processing, e.g. s t o r i n g or retrieving, r e f e r e n t i a l relations is executed by a p a r t i c u l a r component of the d i s c o u r s e p r o c e s s i n g system, the "referential procedures" (ReP). T h e RePs w o r k on a m e m o r y s t r u c t u r e w h i c h is a d e q u a t e for the r e p r e s e n t a t i o n of k n o w l e d g e about objects, the "referential net" (RefN) 4*. A RefN consists of entities of three different types:
- referential objects (RefOs), which are the internal s u b s t i t u t e s for the objects s p o k e n about,
- descriptions, i.e. terms of SRL which describe the RefOs,
- attributes, i.e. p r o p e r t i e s of the d e s c r i p t i o n - r e l a t i o n b e t w e e n a R e f O and one of its d e s c r i p t i o n s or p r o p e r t i e s of the R e f O (see chap. 3).
A f t e r d e t e c t i n g a n e w indefinite d e s c r i p t i o n (as ETA(x) : unlversity(x)) R e P creates a n e w " r e f e r e n t i a l object'" (RefO). D u r i n g t h e d i s c o u r s 6 (after the i d e n t i f i c a t i o n process) further d e s c r i p t i o n s of the same R e f O will be linked to this RefO. (The r e l a t i o n a m o n g the d e s c r i p t i o n s of the same R e f O is the w e l l k n o w n
relation of eoreference,) ~ u s we have, for example, the f o l l o w i n g set of
d e s c r i p t i o n s of the RefO m e n t i o n e d above :
(5) r . l ~ . ~ d. II--ETA(x) : university(x)
For the d e s c r i p t i o n of the r e f e r e n t i a l net I use SRL and in a d d i t i o n a "finite but
extendable" set, REFO, of "referential objects'. The r e f e r e n t i a l objects are
seen as a kind of b a s i c term of SRL. The set R E F O is analogous to the set of
variables. The p h r a s e "finite but extendable" is used to emphasize the dynamic
aspect of the set REFO and the RefN. REFO consists of e x a c t l y those RefOs that are a c t u a l l y needed. It is n e c e s s a r y to d i s t i n g u i s h two types of terms, one type
w h i c h consists of the m e n t a l representatives, i.e. the RefOs, and the other which
contains those terms w h i c h f u n c t i o n as descriptions. N a m i n g this set b y D - T E R we
h a v e T E R - R E F O U D-TER. Thus, referential nets, RefNs, can he defined as
( 6 ) R e f N c D - T E R x REFO •
For <d-ter, r.i> 6 R e f N I also use "d-ter deser r.i'" , "d-ter is a d e s c r i p t i o n of r.i". In this w a y a d e s c r i p t i o n - r e l a t i o n is d e f i n e d on the basis of the RefN. B y m e a n s of referential nets it is possible, e.g. to a r r i v e at one object (i.e. RefO) v i a different aspects, i.e. different d e s c r i p t i o n s , of this object. It is
only by such aspects (cp. Schank's [II] w a y o f m e m o r y discrimination), that a
s y s t e m is able to choose "'a best d e s c r i p t i o n " of the objects w h i c h are the theme
of the d i s c o u r s e (see below). The referential net (5) is s i m p l i f i e d in two
crucial points, first, that the descriptions of r.l o p e r a t e on M S R L ~ p r e s s i o n s and not on RefOs, too. But this can be treated s i m i l a r l y to the focussing process of (4), w h i c h can be f o r m a l l z e d now:
( 7 ) p ( r . i ) 8--- r . i - - ETA(x) : p ( x )
U s i n g ( 7 ) , t h e d e s c r i p t i o n s c a n b e s o l v e d w i t h r e s p e c t t o a l l R e f O s , i . e . a l l a r g u m e n t s a r e f i l l e d b y RefOs o r o p e n t e r m s . Thus we c a n d e r i v e ( 8 ) f r o m ( 5 ) :
(8) r . l ~ - - - d . I I - E T A ( x ) : u n i v e r s i t y ( x )
d . 1 2 - E T A ( x ) : t i m e ( m e e t i n g ( r . 2 , r . 3 , u n l v e r s l t y ( x ) ) , r . 4 ) d . 1 3 - I O T A ( x ) : p r o f e s s o r - a t ( r . 3 , u n i v e r s l t y ( x ) )
r . 2 - - d . 2 1 - " I ' " r . 3 - - d . 3 1 - - B a r b a r a r . 4 - - d . 4 1 - " t o m o r r o w "
Secondly, R e P has tO e x p l i c a t e the r e f e r e n t i a l links of deictlc substitutes, e.g. "r" r e f e r s to the speaker, etc.. " " s t a n d s for the s p e c i f i c ReP w h i c h evaluates d e i c t l c expressions, i.e. w h i c h i n s ~ a n t i a t e s deictic ~xpressions w i t h RefOs, e.g. "I'" - SPEAKER(l) (cp. fn 6.). By the same m e t h o d s n e w RefOs for "we" c a n be created, e.g. in the p r o c e s s i n g of " T h e r e we w i l l discuss some papers." (We = Barbara + I)
3. A T T R I B U T E S IN R E F E R E N T I A L N E T S
I will n o w extend the basic t w o - p l a c e r e l a t i o n b e t w e e n RefOs and d e s c r i p t i o n s to the more adequate concept of a m a n y - p l a c e r e f e r e n t i a l relation. The a d d i t i o n a l p l a c e s w i l l be called "attributes'. In the first step of e x t e n t i o n I give emphasis to "attributes of descriptions'. T h e y r e p r e s e n t p r o p e r t i e s of the d e s c r i p t i o n relation, i.e. of pairs <d, r> E RefN. Thus, the e x t e n t i o n to an a t t r i b u t e d R e f N (ARefN) uses a set A T T of a t t r i b u t e s and changes (6) to
(9) A R e f N c A T T x D - T E R x R E F O .
For <att, d, r > C A R e f N I also say "'att is an a t t r i b u t e of d w i t h regard to r". Some p o s s i b l e a t t r i b u t e s are (Note t h e fact that this list is not c o m p l e t e and not fully adequate):
- syntactic and semantic features of the description, e.g. gender, number, sexus,
etc.
- n u m e r i c a l values, e.g. "grades of relevance of a description", "degrees of "being the TOPIC / b e i n g in the FOCUS" ", "recency". 5* A simple s t r a t e g y for d e - / c o d i n g is: '~se/try the d e s c r i p t i o n or RefO with h i g h e s t degrees !'" - names of persons, e.g. p a r t i c i p a n t s of earlier discourses. By this a t t r i b u t e it
is p o s s i b l e to find a d e s c r i p t i o n relevant to speaker and listener (Cp. Clark / M a r s h a l l ' s [I] co-presence triples).
- situations in which the d e s c r i p t i o n / R e f O Is relevant or was i n t r o d u c e d (cp. W e b b e r ' s [12] " e v o k e ° - p r e d i c a t e ) .
104 CH. U. HABEL
Before I explain the concept of "attributes to descriptions" by some examples, I will new sketch some aspects of function and statue of the attributes. Firstly, in
both processes, analysis as well as generation of referential expressions, it is
necessary to take into consideration a set of alternatives. All of them are
possible candidates to be the referent (in analyzing) or to be a good description (in producing) respectively. The goal of the ReP is to choose exactly one from the
alternatives, namely "the best'. It is obvious that trying to choose the best can
fail. Then the decision has to be revised afterwards. (But this is a total
different problem.) In other words, the function of the attributes is to point to
the most appropriate (or being more careful: to that which seems to be the most
appropriate) of the alternatives. Secondly, I think that, on one hand, knowing and processing attributes is part of human natural language processing, and therefore,
these attributes have to be part of the discourse processing model. On the other
hand, knowledge of attributes is different (cp. the list of possible attributes,
above) from the factual or referential knowledge. Thus, I propose a strict
separation between descriptions and their attributes. This strict distinction
contrasts to Webber's [12] solution: her use of the "evoke'-predieate mixes the
concepts of descriptions a n d their attributes in a cognitive inadequate manner.
Now I shall continue with examples of the use of attributes • Let us suppose a
situation in which a questioner Q asks:
(I0) Where can I study Computational Linguistics?
and the answerer A knows that the university which is (internally) represented by r.l is a good university to study CL. The selection of the best description of r.l
depends on Q and A (and the situational context). I assume the following
attributes to the descriptions of r.l (from (5), (8)):
(ii) descr, general relevant to: time of relevance:
-d.ll "unspecified"
-d.12 SPEAKER(I), LISTENER(1), Barbara NEIGHBOR(r.4) 6*
-d.13 SPEAKER(1), LISTENER(I), KNOW(Barbara) always
(II) describes a situation, in which the name / location of the university in
question is not represented, e.g. since the information about the possibility to
study C L w a s given by an earlier utterance of Barbara "We have some courses in CL" but A (whose RefN is represented by (ii)) does not know where Barbara teaches. The best description can be chosen by comparing the attributes of Q with those of the
descriptions of r.l. If Q knows Barbara, .i.e. <KNOW(Barbara)> ATT (Q) 7*, then
the best description will be:
(12) d.13 - "The university where Barbara teaches (as a professor)."
Ranking d.13 higher / better t h a n d.12 depends on the dimension "time of
relevance'. Furthermore, if Q knows more about r.l it is possible that Q will
react with,
(13) "Well, at Amherst.'"
and now A is able to extend hls/her RefN with a further description of r.l (14) r.l - d.14 - Amherst - - "general high relevance" "always"
(This attribute depends on the fact that proper names have high relevance in
general.)
Anagously we can use attributes in the decoding process. Look at the following
pair of sentences (similar to (I)).
(15) a. I will meet the head of the linguistic department tomorrow. b. She is a specialist for Montague grammar.
To solve this referential problem ( and not to be puzzled by the pronoun "she" )
we need further background knowledge, e.g. "Barbara is the head of the linguistic
department.". With this prior knowledge and the additional (theoretical) concept
of "attributes to RefOs" 8* as described in (16) "female" "local TOPIC'- r.3-- d.31- Barbara
d.32- IOTA(x) : head-of(x ,linguist. dept.)
the referential problem of (15) is easy to solve. A system (human listener or
machine) has to match the pattern corresponding to the pronoun "she" ~ I c h is
induced by the context (15.5) against the descriptions, their attributes and the
attributes of the RefO's. The crucial attribute is here "female" for both "she"
and "r.3".
between attributes of descriptions and those of the RefOs, is very important in
German. In spite of the syntaetlc ineongruence (neutr. vs. fem.) it is possible
(in colloquial German) to say:
(17) Ich traf gestern ein Maedchen. Sle hatte rote Haare.
I m e t yesterday a glrl . She had red hair .
neutr. ~ fem.
4. CONCLUDING REMARKS
By means of an extended referential net (w~th attributes both of descriptions and
RefOs) it is possible to describe how to store and process referential entities
(and expressions) in a cognitive adequate and powerful way.
Following the principles and concepts described above a system for anaphora
resolution has been implemented (Guenther [3]), which will be a component of the
next version of the BACON-system (cp. fn 3).
FOOTNOTES:
* Parts of the first draft of this paper were prepared during my work with the
project "Automatic Construction of Semantic Nets" at the Technical University
of Berlin. My current research on "referential nets" has been supported by the
DFG (German Research Foundation). Author's address: Blelbtreustr. 36 a, D-1000
Berlin 15; Fed. Rep. Germany.
I* From a formal point of v i e w "inference rules" are transformation rules on SRL.
Their cognitive and philosphlcal status, which is analogous to Miller's [I0]
"conceptual entailment', is described in }{abe], [7].
Here, i,e. in (3), and in the focussing rule (7) "p" stands for all types of
concepts which are represented by open formulas.
2* The focussing transformation mentioned above is based on the concept of
referential nets and therefore its description will be postponed to the
following chapter. In (4) a special "bar-convention" is used, which I do not
want to explain in detail here. In case a term is necessary as argument, e.g.
of "meeting", the "bar" will change the open formula "p(x)'" to an "open term"
3* Cp. }{abel [7] and Gust / Hahel / Rollinger [4]. This model is the theoretical
basis of the natural language QAS BACON (Berlin Automatic COnstruction of
semantic Nets), developed at the Technical University of Berlin, which was
supported by a grant from the Federal Government since 1978. It should not be
confused with the BACON-system of Carnegie-Mellon University (we chose the same name at the same time).
4* The concept of referential nets is similar to discourse referents (Karttunen
[9]), reference diaries (Clark/Marshall [ 1 ] ) and discourse models (Webber
[12]). The most relevant difference between referential nets and the other
approaches and the main advantage of referential nets consist in the further
attributes described below.
Note the fact that the acrony=s ReP, RefN, etc. can stand for the respective
concept, e.g. "referential net', as well as for a specific instantlation, e.g.
the RefN (5) or (8). Furthermore ReP concerns the system of referential
procedures as well as the individual procedures. In all these cases the meaning of the acronym is obvious through context.
5* In this paper TOPIC / FOCUS are used in an informal and intuitive way. (For a
detailed investigation of these concepts cp, Grosz. [ 2 ] ) Note that the
106 CH.U. HABEL
6* SPEAKER(l) stands for "the speaker of u t t e r a n c e ( I ) ' , i.e. this a t t r i b u t e gives more i n f o r m a t i o n than W e b b e r ' s [12] "evoke(l)'. "KNOW(_)" represents "people knowing _ ' . "NEIGHBOR(--)" stands for "points / periods of time near to the point of time b e i n g the argument'. I do not want to explain the u n d e r l y i n g "logic of time" here.
7* I w i l l not explain the concept of "attributes of a p e r s o n / participant" here. I use it in a s t r a i g h t f o r w a r d and intuitive manner.
8* The extention to a t t r i b u t e s of RefOs is a n a l o g o u s to that f r o m (6) to (9): A A R e f N c ATT. I x D - T E R x (REDO x ATT.2) •
ATT.I stands for the attributes of descriptions, A T T . 2 for the a t t r i b u t e s /
p r o p e r t i e s of RefOs. A A R e f N is a "double attributed RefN'. N o t e the fact that s a m e of the a t t r i b u t e s of the RefOs w i l l be c o m p u t e d f r o m those of descriptions~ e.g. "female" from a a t t r i b u t e of "Barbara'.
REFERENCES:
[ I ] C l a r k , H . H . / M a r s h a l l , C . R . ( 1 9 8 1 ) : " D e f i n i t e r e f e r e n c e a n d m u t u a l k n o w l e d g e ' " • i n : A . K . J o s h l / I . Sag / B. W e b b e r ( e d s ° ) : E l e m e n t s o f D i s c o u r s e U n d e r s t a n d i n g . C a m b r i d g e U n i v . P r e s s : C a m b r i d g e .
[2] G r o s z , B. ( 1 9 7 8 ) : " D i s c o u r s e k n o w l e d g e " i n : D. W a l k e r ( e d . ) : U n d e r s t a n d i n g Spoken L a n g u a g e • North-Holland: N e w York.
[3] Guenther, S. (1981) : V e r f a h r e n der R e f e r e n z a n a l y s e fuer das n a t u e r l i c h - s p r a c h l i c h e System BACON . Diplomarbeit (Thesis) T U B e r l i n , FB-Informatik.
[4] Gust, H. / Habel, Ch. / Rollinger, C.-R. (1982) : " L a n g u a g e - W o r l d Systems : A c o m m u n i c a t i o n m o d e l " in : Proceedings of the w o r k s h o p "Models of d i a l o g u e : theory and a p p l i c a t i o n " Informatlcs L a b o r a t o r y of L i n k o e p i n g University. Jan.1981. ed. by U. H e i n / S. Haegglund. to a p p e a r at H a n s e r / M o M i l l a n : M u n i c h / London.
[5] Mabel, Ch. (1980) : " R e f e r e n z e n Inferenzen". in : Rollinger,
C . - R . / S c h n e l d e r , H.-J. (Hrsg.) : I n f e r e n z e n in n a t u e r l i c h - s p r a c h l i c h e n S y s t e m e n • Einhorn: Berlin.
[6] Habel, Ch. (1981) : "Zur G e s c h i c h t e y o n R e f e r e n z o b J e k t e n - Tell I: D i s k u r s m o d e l l e " . ms. , Berlln.
[7] Habel, Ch. (1982) : "Inferences - the b a s e ot ~emantics?" to a p p e a r in the P r o c e e d i n g s of the K ~ n s t a n z Conference on "Maaning~ Use, and I n t e r p r e t a t i o n of L a n g u a g e " (Sept. 1981), de Gruyter: Berlin.
[8] Kallsh, D. / M o n t a g u e , R. (1964) : L o g i c - Techniques of formal reasoning. Harcourtt Brace &World : N e w Y o r k •
[9] Karttunen, L. (1976) : "'Discourse referents" in. J.D. M c C a w l e y (ed.) : Notes f r o m the linguistic underground. Syntax and Semantics 7.
[10] Miller, G. (1978) : "Semantic relatlons among words" , in : Halle, M. /
Bresnan~ J. / Miller, G. (eds.) : L i n g u i s t i c theory and p s y c h o l o g i c a l reality. Cambridge, M a s s . .
[11] Schank~ R.: (1979) : R e m i n d i n g and M e m o r y Organization: An I n t r o d u c t i o n to M O P s • Yale Univ. D e p t . C o m p . S c i e n c e . R e s e a r c h Report 170.
