CONNOTATION AS A FORM OF I ~qFERENC E 
JB BERTHELIN, University of Essex (language and linguistics), Wivenhoe Pk, 
COLCHESTER C043Sq, GREAT BRITAIN, 
and INSTITUT DE PROGRAMMATION, Universit~ Paris 6, 75230 PARIS CEDEX 05, FRANCE. 
Abstract 
Story-processing systems have to deal, or avoid 
dealing, with INFERENCE CONTROL ''6' Ii, 13. 
when designing one such system 2, we were great- 
ly helped by the "(ERC)RC" expression for con- 
notation I. Our system is specialised in multi - 
faceted descriptions of characters (not, how- 
ever, in the most difficult problems of beliefs 
about beliefs 4 and 8): here we present another 
aspect of the story character processing, name- 
ly the recursive EXPLANATION of inconsistencies 
appearing in the description of a character. 
We give a very schematic system overview, then 
some details about the CONNOTATION rules and an 
example of their application to a story. 
posed to ("fromage" RC("cheese")) RC("french"), 
but also situation-specific ones, as ("cheese" 
RC("cheese as geological stuff") RC("bizarre"), 
in the processing of a robot who asked what the 
moon is made of. As Barthes then remarks, con- 
notations may be based on a set of initial ex- 
pressions rather than on a single one, which is 
expressed by (EiRCl, E2RC 2 ..... EnRCn) R C 
This seems to be an essential feature of conno- 
tations, since it allows emphasis on 'connoted' 
contents by means of an accumulation of signi- 
fiers. Another feature, elegantly illustrated 
in 'l'envers des signes 'i0 by the two steps 
(("voile" RC("navire"))RC("po~sie"))RC("rh~to- 
rique"), is their recursivity. 
Introduction 
The allusiveness of human languages, in addition 
to being quite convenient in social life, justi- 
fies the use of variable amounts of i n t e 1 - 
1 i g e n c e in processing a sentencer accor- 
ding to the number of reasoning steps leading 
to a position where a satisfactory reaction be- 
,14 comes possible, like "to redefine 'substance 
when reading Spinoza". Let us represent a first 
step by "ERC" i.e. "an expression is related 
to a content ''1, which Barthes calls the denota- 
tion: the second reasoning step will be repre- 
sented in the connotation formula: " (ERC) RC". 
It does not determine exactly where the reaso- 
ning step leads to: if we write E's content as 
C(E), we may have quite general connotations 
like ("cheese"RC("cheese"))RC("english") as op- 
As a beginning we made an attempt to express 
this in AI terms by writing a program, BAQUIL, 
which finds the connotations with structure(E I 
RCi, E2RC2)RC in a recursive way. The interest 
of such connotations can be shown by15: 
.a doctor asks:"how is he feeling ?" 
.the nurse answers:"he is groaning." 
where the nurse means, by connotation,that the 
patient is suffering: ("groaning" RC("groaning") 
RC("suffering)), but our understanding is direc- 
ted by the previous interrogation, so that the 
definitive result we expect from our system will 
be ("feeling"RC("feeling"), "groaning"RC("groa- 
ning") )RC("suffering"). This result shall be 
reached by consulting a semantic network and ob- 
serving that 'groaning' is not exactly a case of 
'feeling', but an expression of it. Thus " the 
nurse means he is suffering" is both a conno- 
tation and an inference, and seemingly a useful 
one: compare with Charniak's 6 "demon-demon in- 
228 
teraction, 'want to buy candies' and 'shake 
piggy-bank' together trigger 'need money'; but 
BAQUIL has no such extended world-knowledge. It 
is specialised in DISCOURSE ATTITUDES like 'er- 
rors', 'lies', 'jokes' etc. This also means it 
can run without an elaborate model 5 of persona- 
lities, and so deal with fables and folktales 
about foxes and sparrows, which would perhaps 
fail to have goals 'common to most people', al- 
though they are meant as human in a way. 
System overview 
Before explaining how it works, we give an idea 
of BAQUIL's construction (i.e. its hierarchy). 
/ "., 
/ 
INPUT / 
HAlt 14, ~,. j,~ ,, ,. ,, 
o u'rP~JT 
¢~/Al~Acr~l~,V'-" 
\ 
~ ~-scmrr'lo~ \ ou~vr 
--229-- 
Schema.tic explanation of how the system runs 
i) the dictionary input procedure builds, from a 
file whose structure is shown in the ' sample 
session ', a classical semantic network 7'9"'' 
2) the story specialist submits sentences to a 
parser 12 and the resulting case structures 
to the story character specialist, whose ac- 
tions include the c o m p a r i s o n s and 
i n f e r e n c e s detailed here. 
3) output procedures express the inferences in 
English and, on request, detail the repre- 
sentation of each character in the story. 
Metarules of BAQUIL 
(this page and the following three will develop 
what rules are applied by the story character 
specialist: this includes METARULES, RULES OF 
COMPARISON and INFERENCE RULES.) 
MI BAQUIL starts a connotation or inference on- 
ly when a comparison rule has been applied 
to a pair of predicates which are related to 
the set of descriptors of one character: the 
predicates are versions of the semantic case 
structure in terms of the current character 
description, and are called 'NOTATIONS' 
M2 Except when specified otherwise (inference 
rules R6, R7), the inference is expressed 
by a 'notation' whose verb is a subcatego- 
ry of either CHANGE or INCONSISTENCY. 
consistency. The metarules are represented in BA- 
QUIL by instructions: Mi in the NOTATION proce- 
dure, M2 and M3 in the CONNOTATION procedure. 
The latter also contain the instruction corres- 
ponding to the inference rules, while the com- 
parison rules form the comparison procedure. 
Discussion of the metarules 
M1expresses the hypothesis that many interes- 
ting antinomies can be detected during the 
pairwise matching of predicates concerning 
one story character." 
M2 aims to expres a more or less syntactic 
finding about the description of a charac- 
ter (co-presence of two antinomic 'notations') 
in terms of the 'notations' themselves. 
M3 means "use the lexical taxonomy when trying 
to recognize a situation"; as a result it in- 
troduce a distinction between natural languages, 
for the subcategories of (for instance) "incon- 
sistency" are not the same in different dic- 
tionariesl Consider French and English, "error" 
and "mistake" vs. "erreur" and "m@rise", or 
worse: the two cases of "to tell a lie" in Rus- 
sian, i.e. "vrat'" vs. "l'gat'". Still it is 
perhaps acceptable to allow for important 
pragmatic divergences between languages (and, 
indeed, dialects or sociolects.) Moreover, we 
did not represent the vocabulary of other lan- 
guages than French and English in our system, 
so we lack precisions about how "whorfian" it 
would turn out. 
M3 Those subcategories are examined in the 
dictionary order and the first one which 
permits the application of an inference 
rule is selected. 
To sum up: (M1) comparison and then inference 
about (M3) subcat(gories of (M2) change or in- 
Although a similar discussion of the comparison 
and inference rules would be necessary, we will 
simply present them here along with some exam- 
ples. The examples are taken from a set of 20 
stories (4 to 200 sentences)which were dealt 
with by the system at Essex in 1979-1980. 
--230-- 
Compari son rules Exagpl e s. 
Their object is to tell whether an inference 
must be started, or not. 
C1 the predicates differ only by a negation in 
one of them (i.e. same environment, verb etc. 
and none was inferred.) 
C1 Confucius is handsome, Confucius is not hand- 
some (in which case 'handsome' need not be 
a priori present in the lexicon.). 
C2 similar to C1 but there is a hyponymy be- 
tween the verbs. 
C2 Confucius is horrible, Confucius is not ugly. 
C3 lexical exclusion between the verbs of two 
affirmative predicates. 
C4 transgression of lexical interdiction, or 
lexical necessity ignored. 
C5 C1, C2 or C3 applies and the first predicate 
expresses an inference. 
C3 Confucius is rich, Confucius is broke. 
C4 Confucius is human and flies away. 
(interdiction) 
Confucius flies away, he does not exist. 
(necessity) 
C5 the systeminferred Confucius is lying, the 
story reveals he is joking (like C3). 
C6 a predicate confirms an inference. 
C7 a predicate confirm a discarded inference. 
(this occurs after a C5 situation led to 
the application of the relevant inference 
rule. ) 
C6 in the situation above, Lao-tsu says that 
Confucius is lying BEFORE the story discon- 
firms it. 
C7 Lao-tsu's remark comes AFTER the revision 
of the inference. 
The predicate or 'notation' structure, which 
permits the comparison, is 
(affirmat.-or-neg., case frame, char. descr.) 
and the connotation has the same structure aug- 
mented by reference to premises ( 2, 3 or a list 
of pairs if an inference has been confirmed.) 
The lexicon element has following slots in 
its structure: 
(list of subcategories, are-they-mutually-ex- 
clusive-or-not, supercategory, property list) 
and the property list contain references to 
other lexical elements. 
--231- 
Connotation rules: exam~es 
As the rules by themselves do not suggest the 
situations which make them useful, let's have 
some examples first. 
RI ...the servant said: "that cow is not going 
to eat you." (...) The next morning, she 
sees one of them is missing and says: "Oh 
my God ! the grey cow ate one of these men". 
.INFERENCE ABOUT THE SERVANT'S VIEW OF THE 
COW. 
R2 (same example as RI) 
.ACCORDING TO THE SERVANT, THE COW UNDERWENT 
A CHANGE. 
R3 A teacher quoting Krylov said that God sent 
a cheese to a raven; a child objects that 
there is no God. 
.INFERENCE ABOUT AN INCONSISTENCY; USING THE 
LEXICAL CONNEXION BETWEEN THE TOPIC 'God' AND 
'religious discourse', BAQUIL SELECTS THAT 
KIND OF INCONSISTENCY. 
R4 the peasant believes that the tree will be 
hit by other rabbits.. But it is not. 
.INFERENCE: ERROR OF THE PEASANT. 
tween levels of discourse independently of 
which character the inference is about, R3 and 
R5 make use of some knowledge associated either, 
as in the examples, with the name of the charac- 
ter, or with previous sentences about it, e.g. 
'Krylov is an author' could be part of the be- 
ginning of the story. 
R3 and R5, which connect the inference with a 
previously observed detail, are "causality" 
rules: in addition to references to the two 
premises whose comparison started it, the in- 
ference has one reference to the predicate 
that justified the choice of a more precise 
verb like 'fiction' or 'joke'. 
A difficulty (which we provisionnally avoided) 
is that some lexical connections could repre- 
sent 'necessary truths' that are just sometimes 
true, as "teachers are sometimes igorant"; what 
R5 should do with them is not clear. However, if 
one considers folktales and fables, the diction 2 
nary connection one uses are almost always 
"foxes are sly, full stop" 
so the question does not arise. 
R5 (continuation of example for R3) 
the teacher answers that there is no cheese 
either. 
.INFERENCE ABOUT AN INCONSISTENCY; USING THE 
LEXICAL CONNECTION BETWEEN 'Krylov' AND 'au- 
thor', AND ONE FROM 'author' TO 'fiction', 
BAQUIL SELECTS THE LATTER SUBCATEGORY OF 
INCONSISTENCY. 
A difficulty for R3 is that a 'topic' could be 
connected both to 'religious discourse' and 
'joke', so in the present state of the system 
one of the connections would be systematically 
ignored: but for the time being, there are very 
few lexical connections, the reason being that 
we cannot decide which necessary truths really 
belong in a lexicon. 
While the other rules exploit comparisons be- 
232 - 
BAQUIL's connotation rules 
'references' are the two initial premises of the 
inference; 'hypotheses come from the list of 
subcategories of the initial verb, i.e. the 
different cases of 'change' or 'inconsistency' 
as represented in the lexicon. 
R1 if both references are about the same charac- 
ter descriptor, the inference is also about 
that descriptor. (this allows for stories 
inside the story). 
R2 if the inference's verb is not yet specified 
and the inference is about the descriptor of 
the second reference, BAQUIL tries the cases 
of 'change'. 
Rules R6 and R7 are illustrated by the story 
of the vegetarian wolf: given a dictionary 
which connects 'wolf' to 'bad action', hence 
to 'lie', the system first infers from the be- 
ginning of the story, where the wolf is con- 
tradicted by someone, that it lies; later, 
the story-telling warns the reader about it, 
telling explicitly that it does not. The in- 
ference is NOT erased, but earmarked as falla- 
cious; so R7 applies when a character says: 
this wolf is lying; but, if the story had 
not denied the possibility of a lie, R6 would 
have applied. One can see that R6, R7 and R8 
are not quite satisfying. I suspect them of 
being in need of some refinement. The set of 
situations 'clever character or lucky guess 
or etc.' is not clearly defined, and I do not 
know what to do with the Liar's Paradox. 
R3 if the verb of a hypothesis is connected to 
the name of the character in a reference by 
a 'be about' link, the hypothesis is selec- 
ted. 
On the contrary, R3 and R5 give less trouble. 
This is shown by the 'sample session' next 
page, representing approximately 1.8 second 
CPU (using about 25K core) on a PDP-iO. 
R4 if one reference comes from a 'belief' and 
the other from the 'story-telling', the verb 
of the inference should be 'error'. 
R5 if the verb of a hypothesis is connected to 
its character's name by a 'tendency' link, 
the hypothesis is selected. 
RULES APPLIED: C4 "lexical necessity" 
R3 "topic" 
R6 if a character's discourse matches an infe- 
rence, BAQUIL checks whether the character 
is clever or has made a lucky guess etc. 
R5 "tendency" 
C3 "lexical exclusion" 
R71if the story-telling contradicts an inferen- 
and~ce, the inference is 'discarded'; if the 
R8Ydiscarded inference is matched by a discour- 
se, one looks for an 'error' or 'trick' etc. 
- 233-- 
SamPle session 
(DICTIONARY INPUT) 
WORDS 
FOOD ISA SUBSTANCE 
SHORT ISA STATE 
TEACHER ISA MAN 
QUOTE ISA DISCOURSE 
AUTHOR ISA MAN 
KRYLOV ISA AUTHOR 
WRITE ISA DISCOURSE 
SAY ISA DISCOURSE 
GOD ISA PUTATIVE BEING 
CHEESE ISA FOOD 
RAVEN ISA BIRD 
CHILD ISA MAN 
PRODUCE INCONSISTENCY ISA DISCOURSE 
(comment: abbreviated PRODINC) 
LIE ISA PRODINC 
MISTAKE ISA PRODINC 
JOKE ISA PRODINC 
ILLUSION ISA PRODINC 
RELIGIOUS DISCOURSE ISA PRODINC 
FICTION ISA PRODINC 
BE ABOUT ISA DISCOURSE 
MEAN ISA DISCOURSE 
IGNORANT ISA STATE 
END WORDS 
LINKS 
USUAL(AUTHOR, FICTION) 
USUAL(CHILD, IGNORANT) 
USUAL(IGNORANT, MISTAKE) 
BE ABOUT(RELIGIOUS DISCOURSE, GOD) 
FORBIDE(RELIGIOUS DISCOURSE, QUOTE) 
comment: that is to account for the location 
being Soviet Russia. 
NECESSITY(ACTION, EXIST) 
NECESSITY(ACTION, AVAILABLE) 
(STORY INPUT) 
Food was short. 
A teacher quoted Krylov. 
He said that Krylov wrote: 
'God sent a cheese to the raven.' 
A child said: 
'There is no God ! ' 
but the teacher replied 
'there is no cheese, 
either ...' 
THE END 
(INFERENCES) 
(God sent a cheese 
1(there is no God 
C4 in description of God and R3/God: 
"the child means Krylov has a religious 
discourse". 
(Krylov has a religious discourse 
2(the teacher quotes Krylov 
C4 in description of Krylov and default: 
"the child means the teacher has an inconsis- 
tent discourse". 
(God sent a cheese 
3(there is no cheese 
C4 in description of cheese and R5/Krylov: 
"the teacher means Krylov's discourse is fic- 
tion". 
(Krylov has a religious discourse 
4 (Krylov's discourse is fiction 
C3 in description of Krylov and R5/child: 
"the teacher means the child is mistaken" 
END OF INFERENCES 
END LINKS END OF SESSION 
-234- 
Conclusion 
While not achieving much by itself, Baquil is 
2 an important component of the larger system 
currently built at Paris 6, and could probably 
also be integrated in a large AI MT system as 
an expert of indirect descriptions of charac- 
ters, for instance it could recognize the use 
of a nationality adjective suggesting a charac- 
ter trait (Cretan for liar, etc.), which is 
helpful in many cases: when the nationality 
adjective is not familiar in the target lan- 
guage (Chines texts about the ancient king- 
doms), but also when there are several possible 
interpretations for the nationality adjective 
in terms of personality ( 'qua' ambiguity). 
The story in the sample session is adapted from 
"The Big Red Joke Book" by Benton and Loomes, 
Pluto Press, London, 1979. 
(i0~ GENETTE: Figures I, Paris, 1966. 

References 

(1) BARTHES: Elements of semiology, Hill and 
Wang, New York. 

(2) BERTHELIN and SABAH: Le traitement des 
personnages, congr~s AFCET-IRIA, Toulouse 
1979. 

(11) HAYES: The logic of frames, in METZING, edo 
Frames, conceptions and text understanding, 
De Gruyter, 1980. 

(12) SABAH: A contribution to story understan- 
ding, thesis at Universit~ Paris 6, 1978. 

(3) BIEN: Toward a multiple environments model 
of NL, IJCAI-4. 

(13) SCHANK and ABELSON: Scripts, plans, goals 
and understanding, LEA, 1977. 

(4) BIEN and WILKS: Speech acts and multiple 
environments, IJCAI-6. 

(14) WILKS: Grammar, meaning and the computer 
analysis of language, RKP, 1972 

(5) CARBONNELL: Computer model of human per- 
sonality traits, IJCAI-6. 

(15) WITTGENSTEIN: Philosophical Investigations, 
Blackwell, 1953. 

(6) CHARNIAK: Toward a model of children story 
understanding, thesis at MIT, 1972. 

(7) CHOURAQUI: Sur un r~seau s~mantique actif, 
congr~s AFCET-IRIA, Toulouse 1979. 
IJCAI-4 and IJCAI-6 refer to the Proceedings of 
the international Joint Conferences on Artifi- 
cial Intelligence. 

(8) CLARK and MARSHALL: Mutual knowledge and 
definite reference, mimeo, Stanford Univ. 

(9) FAHLMAN: Representing everyday knowledge, 
MIT, 1977. 
