ATTITUDE EMERGENCE - AN EFFECTIVE INTERPRETATION SCHEME 
FOR PERSUASIVE DISCOURSE 
ItOR.NG-JYH P. WU and STEVEN L. LYTINEN 
Artificial Intelligence Laboratory 
The University of Michigan 
Ann Arbor, MI 48109. U.S.A. 
ABSTRACT 
Previous approaches have used a reasoning mechanism 
called belief percolation to determine the actual speech 
intent of the speaker (e.g., (Wilks and Bien 1979)). 
In this paper, a similar mechanism, called attitude 
emergence, is proposed as a mechanism for inferring a 
speaker's attitude toward the propositions in a persua- 
sive discourse. It is sbown that in order to adequately 
interpret the statements in advertisements, associa- 
tions of relevant semantic information, through br*dg- 
lug inferences, are to be percolated up through attitude 
model contexts to enhance and calibrate the interpre- 
tation of statements. A system called BUYER is being 
implemented to recognize speech intents through atti- 
tude emergence in the domain of food advertisements 
taken from Reader's Digest. An example of BUYER's 
processing is also presented in the paper. 
Introduction 
One of the most significant characteristics of persuasive 
discourse is that it involves the expression of people's 
beliefs, desires, preferences, etc. These beliefs, desires, 
and preferences constitute a model of mental a~titudes 
(or, an attitude model) which characterizes the mind 
of the speaker engaging in a persuasiw', discourse. An 
attitude model is important for figuring out what the 
speaker means; i.e., his speech iutent. More specifi- 
cally, often when the speaker expresses his beliefs, de- 
sires or preferences in persuasive discourse, he means 
to induce a reaction, in the forms of comparable men- 
tal attitudes, on the part of a hearer. For example, 
an expression of the speaker's belief can be intended 
to induce such an belief in the hearer. However, in 
general, inferring the speech intent through attitudc 
model reasoning is complex. For instance, in our do- 
main of persuasive discourse - advertisements - a ma- 
jor statement may be followed by minor statements, as 
demonstrated in the following passage. 
(1.1) Nabisco is great. 
(1.2) It is nutritious whole wheat, 
(1.3) low in cholesterol and saturated fat, 
(1.4} haz plenty of fiber and vitamin. 
Arranged in this fashion, tbe expression of the 
speaker's preference in (1.1) - Nabisco is great comes 
to be supported by statements (1.2) through (1.4). 
This makes tile acceptance of (1.1) as the heater's 
own much more compelling, intending to induce in 
the hearer the same preference expressed in (1.1). Al- 
though tills explanation sounds intuitively simple and 
correct, the question remains: how do statements such 
as (1.2) to (1.4), which are oil the surface~ "disjoint" 
expressions of the speaker's belief, come to have a real 
psychological impact on the hearer? Some sort of rea- 
soning must have been employed to bridge them with 
(1.1) and to produce the persuasive effects. 
Previously, model-based reasoning has been inves- 
tigate(\[ for many tasks, such as belief ascription and 
metaphor understanding (Ballim et al. 1991), logical 
reasoning (Dinsmore 1987), and natural language un- 
derstanding in general (Faueonnier 1985). One previ- 
ous approach to attitude model reasoning concentrates 
exactly on issues of inferring speech intents. As dis- 
cussed concisely in (Wilks and Bien 1979), the state- 
ment "Frank is coming tomorrow" can be interpreted 
in many ways, depending on the context. For instance, 
if the hearer believes that the speaker believes that 
Frank is hostile to the hearer, and the hearer has uo 
personal knowledge about Frank, then this statement 
might be interpreted as a threat to the hearer. 
To account for different possible interpretations of 
statements like these, Wilks el al. propose an altitude 
percolatiou mechanism, in which a statement is pushed 
down to the frame of the system's belief to create the 
attitude context - the system's belief of the speaker's 
belief. Since in this more specific context, the following 
statements are simultaneously present I : Frank comes 
tomorrow and Frank is hostile to the system. Thus, 
tile system can infer that Frank may harm the sys- 
tem, which in turn, allows the system to interpret the 
original statement ms a threat to itself. 
In our domain of advertisement persuasive disconrse~ 
much fewer facts arc privately known compared to 
what are mutually known 2. Therefore, it is argued 
~If tile system has personM knowledge otherwise about 
Frank, then this may affect the reasoning proces,q. 
2It is also an essential criteria for "good" advertisements 
to adhere to fashionable viewpoints. 
AC1-ES DE COLING-92, NANTES, 23-28 AOt~r 1992 9 1 1 PROC. OF COLING-92. NANTES, Auo. 23-28. 1992 
that the push dowa operation, which investigates into 
the personal knowledge about one another, will be lim- 
ited. Instead, in this paper, a more relevant mecha. 
nism, called attitude emergence, is investigated. It is a 
more conlprehensive treatmeTit of attitude lnodels~ nti-- 
lizing mutually known semantic and world knowledge 
to convey speech intent, a.s observed in the interpreta- 
tion of our earlier example. More specitically, attitude 
models for statements are eonstnmted, aud a~qsimilated 
through bridging inferences based on mntual knowl- 
edge. The semantic association resulting flora assinfi- 
lation then gets percolated up along attitude model con- 
texts so that the proper interpretation of speech intent 
is recognized. This attitude emergence mechanism is 
being implemented in a system called BUYEI{~, which 
understands food advertisements taken fl'om lgeader's 
Digest. A corpus of 129 advertis(mmnts has been col- 
lected. Part of the ads are used R)r constructing tile 
system, while the rest of them are used to verify the 
generality of BUYI';ll.'s knowledge base. In this papm, 
we present an examl)le of 1111Yl,2d, processing one nf 
these ads. 
The basic framework of atttil;ude 
emergence 
As observed above, the recognition of speech intent in 
persuasive discourse is rather comj)lex. Various sorts 
of mutual knowledge ntay be employed in bridging the 
statements and bringing out the speech intents. In (Wn 
and Lytinen 1991), we proposed a three.step procedure 
of attitude emergence: 
Step 1: Construct tile initial altitude model (or A-model). 
Step 2: Assimilate the successive statements cnherently 
into one A-model (if possible) anti recognize the 
semantic a.qsociation Imtween tile models. 
Step 3: Percolate upwards &tOltg A-model context to effect 
attitude ehauge due to the semantic association 
recognized in Step 2. 
A-models are rceursive structures of infi)rmation with 
attitude contexts, each layer el Milch consists of an 
agent and an attitude he holds toward the deeper hwel 
information. A simple passag;e is analyzed ill (Wu and 
I,ytinen 1991), 
(2.1) Peter loves antique cars. 
(2.2) His favorite model is tile 1887 I)uryca. 
Tire evolution nf A-.models through this three-step pro- 
cedure can be summarized as follows: At tile end of 
Step l (see Fig. 1), a,m A-model is coastructed con- 
sisting of an attitude context (Report i;pcak(~r ...), 
which eml)eds another attitude context • (Love Pe 
ter ...), which contains ~,.*l object antique.ea.k's. Note 
that, these linear R)rmula m'c just, short hands ik)r ulod- 
els. We take thai, ill implmnentatinn, models are ibr- 
mula plus indexing ;rod encapsulation, as tile boxes ill 
Fig. I is intended to capture: the indice,~ on agents are 
called S-boxes and on attitude~q A-boxes. 
l';ach attitude context creates it'~ oegn environnient 
R)r simulative reasoning (Wilks and llavtley 1990), 
 .FaTV _A 
•. \[--V& -~ ....... 
Antique eara \] 
Figure 1: The initial attitude model after (2.1) h&s 
been processed. 
...... ~,. ak-- ...... 
I.,'igure 2: The attitude model and reasoning when (2.2) 
is assimilated. 
which involves only entities with comparable attitude 
,qtatus. In turn, simulative reasoning call produce re- 
suits app\]icable to the attitude context where it takes 
place, and may sometimes affect related contexts, e.g., 
causing re-Evaluation of the attitude in the embedding 
attitude contexts. Thus, ill Step 2, while statement 
(2.2) is being assimilated with statement (2.1), some 
reasoning takes place marked as (A), (B) and tO) in 
Fig. 2. 
At point (A), an IS-A semantic relation is recog- 
nized between "antique_cars" and "1887_Duryea" in 
the semantic space whicb, ,as depicted as orthogonal to 
attitude space in Fig. 2, stores attitude-independent 
semantic information. Ill order to trigger emergence 
of more attitude related information, the recognition 
of the LqoA relation is percolated up along layers of 
e~ttit.ude contexts to reach tile (Report Speaker ...) con- text, 
where (B) and (C) occur. At points (B) and (C), 
~'~ta~,emeuts (2.2) and (2.1) are found to be related; in 
particular, tile (Report Speaker ...) context of (2.2) is 
calibrated to be "evidential" (to (2.1)), wbile, that of 
(2.1) to be "snl)l)orted" (by (2.2)). "lhat (B) and (C) 
occur is due to tile following world knowledge (WKI): 
,'\~:li s ) IJ()I.INi 1-97, I<IAitii~s, 23 28 Ao()r 1992. 9 I 2 Pltoc. OF COLING.92, NANTES, AUG. 23-28, 1992 
(Ru=alng ~ ) 
(R~lng =} 
(84) ~ p4arllti~d In AIB lud~= apm~e 
Figure 3: The two attitude ruodels for the readings of 
$I. 
If the speaker provides more detailed information 
about a clMm (Y) in a statement (X) 
Then Y is "supported" by "evidence" X and becomes 
more believable 
As demonstrated by (A) - (C), semantic association 
emerges from embedded attitude contexts to calibrate 
the attitude in higher contexts - i.e., how attitude 
emergence happens. 
In this simple example, the attitude emergence 
mechanism has involved, nonetheless, a large amount 
of knowledge. This knowledge is briefly reviewed be- 
low. Yet, more sophisticated reasoning is required to 
process real world ads (see below). First, there is 
knowledge concerning A-model construction based on 
the following mapping rules: (1) Sentence types to A- 
boxes, e.g, a declarative sentence type maps into a 
belief; (2) Attitude verbs to A-boxes, e.g., "loves" into 
a preference; (3) Evaluative predicates to A-boxes, e.g., 
"favorite" into a preference; (4) Adverbs to A-models, 
e.g., "certainly" into a belief; (5) Cue phrases to A- 
models, "it is time that" into a desire (recommenda- 
tion). 
Secondly, there is knowledge concerning A-model as- 
similation and bridging inferences. For example, in 
passage (2), the expression "his favorite model" is re- 
solved through the following three separate bridging 
inference steps: 
1. That "his" refers to tldngs pertinent to Peter. 
2. That "favorite" is an evMuative predicate translated into 
a "prefer" attitude box. 
3. That cars have models. 
For step (1), a focusing mechanisnr is required to locate 
the S-box - Peter, since a pronominal expression usu- 
ally (but not always) refers to some object in the focus. 
For step (2), the A-model helps to guide the resolution 
filrther into the most relevant A-box; in other words, 
an A-model itself carl serve as a marker to find the most 
relevant A-model earlier in tile discourse. For step (3), 
a basic semantic association occurs to recognize pos~ 
sitde semantic relations. In summary, the resolution 
of the expression "his favorite model," and similarly 
for all other expressions, is achieved by bridging in- 
ferences which synthesize many knowledge sources, in- 
cluding focusing, A-models themselves, and semantic 
association. 
A real world example from BUYER. 
In this section, a real world advertisement is presented 
to demonstrate the application of attitude emergence 
in processing persuasive discourse. As discussed above, 
a simple version of A-model construction and assimila- 
tion has to be extended to include more general world 
knowledge. The following ad, which BUYER has pro- 
cessed, demonstrates this. 
The Folgers ad. 
SI. l.s your decaffeinated ~s dark as ours? 
$2. Star~ with one teaspoon of both. 
$3. Hut just bec&tlse the anmunts are cquaJ 
doesn'~ t:lean the results w~ll he. 
$4. Mount,~t~ Grown ~blgers dark, sparkling 
Crystals are the dilt~reJtce. 
S6. So dark a3~d rich, shouldn't you switch? 
In tile process of understanding this ad, tim system 
has to tigure out many tbings, for example: ls S1 a 
question or a prompt fur suggesting an action (for tile 
hearer to perform)? Is $2, given tile proper interpre- 
tation of $1, an order or a recommendation; Does $3 
affect the status of tile attii.ude expressed in 817 and 
so on. In order to answer these questions concerning 
speech intcuts, tile attltude model of each statement 
h~ to go through more involved calibration and en- 
hancement tban the simple version presented in the 
previous section. First, the reading of statement 5'I 
is ambiguous. It could mean that the speaker wants 
the hearer to inform him as to whetber the bearer's 
decaffeinated is as dark a~s ours. Or, it can have an- 
other reading: the speaker wants the hearer to know 
whether the hearer's decaffeinatcd is ms dark as ours. 
The latter reading is inferred by the following world 
knowledge (WK2)~: 
If the advertiser already knows everything about his 
products (which is reasonable to assume), 
Then a questio~ concerning the product is actually 
all intention to iu.\[orm. 
The two possible attitude models for the two readings 
of $1 are depicted in Fig. 3. Note that the two vari- 
ables el and el' stand for tile different events speci- 
fied by (Inform-whether Hearer ...) and (Know-whether 
Hearer ...), respectively. The two events are indexed by 
their event types, as demonstrated by the ovals in the 
boxes in Fig. 3. Then, when $2 is processed, from its 
sentence type (imperative), it is inferred that its atti- 
tude context is compatible witll both readiugs of $1, 
3For simplicity and readability, we represent BUYER's 
knowledge in this and following examples as English-like 
rules. 
AcI~:S DE COLING-92, Nntcn~s, 23-28 ^o\[rr 1992 9 1 3 I)rtoc. or COLING-92, NANTES, AU~. 23-28, 1992 
(83) aa a conditional In Attitude Sp~ce 
\[P1 ); The arrlour~t~ are equal, 
{Q1 ): Tim resutbs are riot equal. 
Figure 4: The attitude models of tim conditional $3. 
i.e., (Want Speaker (P Hearer ...)). Hence, the semantic 
association begins to emerge when the events e2, speci- 
fied as "start witb one teaspoon of both," and el or el' 
(see Fig. 3) are being assimilated. The coherence rea- 
soning component of BUYER. (Wu and Lytinen 1990) 
is able to recognize the following Enable coherence re- 
lation: 
That tile hearer starts with one temspoon of both kinds 
of coffee can enable that hc knows which coffee is bet- 
ter. 
That is, the action e2suggested in $2 is an e~:perirnent 
to find out something. The choice between el and 
el ~ is now clear, due to the following world knowledge 
(WE3): 
If an agent has a goal to find out something about X, 
Then he can perform an experiment with X. 
Since eL' - (Know-whetl)er ttearer ...) -- is acquired as 
a goal for the hearer according to Reading 2, el' and 
hence, Reading 2 is determined as the speech intent. 
Then, when $3 is processed, it logically means: 
It is )lot the case that Pl implies Q1. 
where P1 stands for - the amounts are equal; Ql the 
results are equal. The cormnon sense logical reason- 
ing employed in constructing the attitude model of $3 
proceeds ms follows: assume P1 is a fact/observation, 
should we assmne QI or ~ Ql according PFI? If Q1 is 
assumed, then it produces: P1 impliesQt, which wonkl 
be contradictory. So, the only alternative is to choose 
Q1. The result is then the attitude model shown 
in Fig. 4. Following attitude percolation, the attitude 
model of Fig. 4 is pushed down into the one in Fig. 3 
for t?~eading 2, creating the attitude context of (Want 
Speaker (Believe Hearer ...)). At this point, BUYER is 
able to reason that Pl - the amounts are equal - is a 
fact, by resolving "the amounts" to "the amounts of 
coffee used in e27' Given that the conditional PF1 has 
a satisfying antecedent, the conseqnent ~ QI (or "the 
results are not eqnal") is derived as a fact. Note that, 
similar to how "tile amounts" is resolved, "the results" 
would be resolved ms "the results in the experiments 
with the two coffees." 
Next, when $4 is processed, by pushing down and 
resolving "the difference" to "the difference between 
the resnlts in the experimellt," it is recognized tl~at $4 
is supl)orting the implicit speech intent made in $3 -. 
53: The speaker claims that the results 
of Y ar~florent. , 
$5: Tile speaker molivates the use 
S~: t~Ot feooxpepa0kOr~na~ to~eKrl~0~. ~ fcofrFl~ cl~mo0ffeo, dnawing =ppert 
$4: The apeaker gives evidence 
supporting the claim. 
$1 : The speaker wants the hearer 
to Know X. 
Figure 5: The statements of the Folgers ad in view of 
their attitude models. 
the speaker wants the hearer to believe that tbe results 
of the experiments are not equal. This is due to the 
following world knowledge (WK4): 
If the speaker gives the physical cause of a conse- 
quence statement, 
Then the consequence statement becomes more be- 
lievable. 
Thus, the intended belief that tim results are not equal 
is further enhanced. Finally, $5, like 83, is not a literal 
conditional statement. It logically means: 
P~ implies should?(-, Q2) 
where P7 is that Folgers is so dark and rich and Q2, 
switch to Folgers. However, the intended meaning is, 
obviously - switch to Folgers. The derivation hinges 
on the following "dogma" about "abnormal vs. nor- 
real states": (1) Unfamiliar external states may be ab- 
normal; (2) If unknown external states are indeed ab- 
normal, people query about them using "should X?'; 
(3) Abnormal states are to be corrected. Due to (1) - 
(3), the intended meaning can be derived, since ~ Q2 
is abnormal and to be corrected; in addition, Pz is also 
a fact. Fig. 5 summarizes tim speech intents reasoned 
by attitude emergence for the Folgers ad: 
Related work 
Work on belief percolation ((Wilks and Bien 1979), 
(Wilk~ and Bien 1983), (Wilks and Ballim 1987), (Hal- 
lira et al. 1991)) has strongly inspired our work. flow- 
ever, n~lost of this work concentrates on one single oper- 
ation of attitude/belief percolation .- the "push down" 
operation. Although this operation is important for in- 
vestigative reasoning and assimilating attitude models, 
effects on attitude models themselves due to attitude 
percolation are more important for our domain of per- 
snasive disconrse. Thus, attitude emergence stands as 
a more relevant mechanism to reason about persuasive 
speech intent. Moreover, the proposed thretr-step pr~ 
cedure for computing attitude emergence proves to be 
a general framework for recognizing speech intents. 
The mapping rules proposed in (Hinkelman and 
Allen 1989), as well as those in (Gerlach and Sprenger 
1988), are similar to the attitude model construction 
rules, while deeper reasoning may underlie some of 
AclT.s t)E COL1NG-92, NA/VIq!S, 23-28 AOt';r 1992 9 l 4 I'ROC. Ol: COLING-92. NANTES. AUG. 23-28, 1992 
their rules, e.g., the interpretation of should?('~Q~) 
as Q2, as we have done to ours. Ill (Hiukelman and 
Allen 1989), they also proposed plan-recognition to 
further identify the speech intent. However, we be- 
lieve that in persuasive discourse, identifying speech 
intent through A-models can be done more locally us- 
ing coherence and bridging inferences. In this sense, 
our approach is closer to those proposed in (Cohen 
1987) and (Mann and Thompson 1988), while (Cohen 
1987) considered only support relations and (Mann and 
Thompson 1988) remained as a descriptive theory, and 
both did not consider A-models as essential lot infer- 
ring speech intent. 
A-models also arc somewhat similar to model- 
theoretical approaches to semantics. While theories 
of the more formal kind -- e.g, Discourse Represen- 
talional Theory (DRT) (Kamps 1981) -- and of the 
nrore cognitive - e.g., Mental Spaces (Fauconnier 1985) 
- emphasize the fundamental issues of reference and 
presupposition, attitude emergence sees application of 
model reasoning to recognize speech intents. These 
also demonstrate that mental attitudes serve as only 
one (though important) way to organize information. 
There are other ways information should be organized. 
I,br example, our formulation of conditionals (for $3) 
is organized not according to attitudes, but principles 
studied in DRT and mental spaces. 
Conclusion and future work 
In this paper, a mechanism called attitude emergence 
is discussed. The basic framework of attitude emer- 
gence, which consists of attitude model construction, 
assimilation, and effects propagation, was first pro- 
posed in (Wu and Lytinen 1991) with limited oper- 
ations. The mechanism is filrther inrproved aud ex- 
tended to recognize more indirect speech in persua- 
sive discourse, by adding other common sense and log- 
ical reasoning. The generality of attitude emergence 
is demonstrated by a real world ad which BUYER has 
proce~ed. BUYER is the computer implementation of 
attitude emergence and is implentented as a rule-based 
system which currently has 348 rules organized in 1O 
problem-solving modules. These problem-solving mod- 
ules are organized in a way that rules are both forward- 
and backward-chained, depending tim deductive and 
abduetive nature of the rules, respectively, and allows 
efficient backtracking. 
One future work of attitude emergence lies on tile 
further systematization on the dynamics of attitude 
models. Only the force aspect of a statement is present 
in the current formulation. A tidier formulation of at- 
titude dynamics should include both force and counter 
force, reflecting the enforcement and the resistance to- 
ward an expressed attitude. For example, in Folgers 
ad, the "counter force" induced by $2 -- the inertia of 
the hearer not to be told what to do - can he over- 
corned by the (attracting) force expressed in $1 - the 
speaker wants the hearer to know something important 
- and the force due to the common knowledge that the 
experiment urged ill $2 can enable the attainment of 
such knowledge. Relating statements to psychological 
forces arc steps toward explaining the psychological re- 
ality of persuasive force and pressure. Along this line, 
it is found that the work on "force dynamics" proposed 
in (Tahny 1988) is highly relevant. We are currently 
looking into the relation between the two. 
References 
tlallim, A; Wilks, Y.; and Barnden, J. 1991. Belief ascrip- 
tion, metaphor, and inteitsional identification. Cognitive 
Science 15:133 171. 
Cohen, R. 1987. Analyzing the structure of argumentative 
discourse. Computational Liguistlcs 13:11-23. 
l)insmore, J. 1987. Mental spaces from a functional per~ 
spective. Cognitive Science 11:1-12. 
Fauconnier, G. 1985. Mental spaces. The MIT Press, 
Cambridge, MA. 
Gerlach, M. and Sprengerp M. 1988. Semantic interpre- 
tation of pragmatic clues: Connective, modal verbs, and 
indirect speech acts. In Proceddings of the Eleventh Inter. 
national Conference on Computational Linguistics. 191 
195. 
ltinkelman, E. and Allen, J. 1989. Two constraints on 
speech act ambiguity. In Proceedings of the Twenty See. 
enth Annual Meeting of the Association for Computational 
Linguistics. 212-219. 
Kamps, H. 1981. A theory of truth and semantic repre- 
sentation. In Groenendijk, J.; Janssen, T.; and Stokhof, 
M., editors 1981, Formal methods in the study o\] language. 
Mathematish Centrum, Amsterdam. 277-322. 
Mann, W. and Thompson, S. 1988. Rhetorical structure 
theory: qbward a functional theory of text organization. 
Text 8:243-281. 
q'almy, L. 1988. Force dynamics in language and cognition. 
Cognitive Science 12:49 100. 
Wilks, Y. and Ballim, A. 1987. Multiple agents and heuris- 
tic ascription of belief, ht Proceedings of the Tenth hater- 
natior~al Joint Conference on Artificial Intelligence. 118- 
124. 
Wilks, Y. and Bien, J. 1979. Speech acts and multiple 
environments. In Proceedings of the Sith International 
Joint Conference on Artificial Intelligence. 
Wilks, Y. and Bien, J. 1983. Beliefs, points of view and 
multiple environments. Cognitive Sclence 8:120-146. 
Wilks, Y. and Hartley, A. 1990. Belief ascription and 
model generative reasoning: joining two paradigms to a 
robust parser of messages. In The 1990 DARPA Work- 
shop. 219-239. 
Wu, H-J. and Lytinen, S. 1990. Coherence reasoning in 
persuasive discourse. In Proceedings o\[ the Twelfth Con- 
ference of the Cognitive Science Society, Cambridge, Mas- 
sachusetts. 503-510. 
Wu, It-J. and Lytinen, S. 1991. Attitude and coherence 
reazoning in persuasive discourse. In Proceedings of the 
1991 A A A 1 Spring Symposium on A tyumentation and Be- 
lie\], Stanford, California. 
AL'rF.S DE COLING..92, NANTES, 23-28 ^ot"rr 1992 9 1 5 PROC. OF COLING-92, NANTES, Aua. 23-28, 1992 
