﻿Speaker Attitudes in Text Planning 
Christine DEFRISE 
IRIDIA 
Universit6 Libre de Bruxelles 
Sergei NIRENBURG 
Center for Machine Translation 
Carnegie Mellon University 
ABSTRACT. Natural language generation needs an in- 
put language whose expressive power is sufficient for 
generating texts with the level of quality desired by var- 
ious NLP applications. In oar generator, DIOGENES(e.g., 
Nirenburg et al., 1989), we use the text meaning repre- 
sentation language TAMERLAN(Nirenburg and Defrise, 
1989 and forthcoming). Expressions in this language 
are used as input by the DIOGENES text planner to pro- 
duce text plan expressions in the text plan language, 
TPL, that in their turn serve as input to syntactic realiza- 
tion. In this paper we describe the treatment of one of 
the several types of knowledge encoded in TAMERLAN, 
namely, speaker attitudes. We aLso illustrate how these 
input components are used in producing text plans. 
I. Introduction 
Our reasons for introducing attitudes as an explicit part of the 
representation of the meaning of a natural language clause 
are manifold. In what follows we will review three (partially 
interconnected) reasons. Representing attitudes a) helps rea- 
soning about speaker goals, b) highlights the argumentative 
structure of a discourse and c) provides a convenient vehicle 
for representing modal meanings, including negation. 
Almost all spoken and written discourse involves the par- 
ticipants' opinions, so much so that producing a perfectly 
'objective' text is an almost impossible task. Within the set 
of possible goals relating to generating text, the introduction 
(explicit or implicit, lexicalized or not) of the producer's 
opinions and points of view serves two goals: 
• modifying the consumer's model of the producer by 
stating facts (including opinions) about self which are 
not in principle observable by the consumer 
• modifying the consumer's opinions by stating pro- 
ducer's opinions about facts of the world (the latter 
can in principle be observed by the consumer) 
The above distinctions only become visible if one decides 
to represent attitudes overtly. Once this decision is made, it 
becomes clear that it brings about better description possibil- 
ities for additional linguistic phenomena, such as the argu- 
mentative structure of discourse. It has been observed (e.g., 
Anscombre and Ducrot, 1983) that texts have a well-defined 
argumentative structure which reflects the producer's cur- 
rent goals and influences such processes as the ordering of 
text components and lexical selection in generation. The 
argumentative structure of a text is realized (or, in text un- 
derstanding, detected) through linguistic means such as the 
use of scalar adverbs ('only', 'even', 'almost', 'hardly', etc.), 
connectives ('but', 'since'), adjectives ('unbearable', 'fasci- 
nating', etc.). Sets of such lexical items may have to be 
considered equivalent from a purely semantic point of view, 
but different in a facet of their pragmatic effect known as 
argumentative orientation. For example, to illustrate the in- 
terplay between semantic content and argumentative orienta- 
tion (i.e. the producer's attitude towards an event), contrast 
(1) and (2), which have opposite truth conditions, but the 
same pragmatic value-- from both (1) and (2) the consumer 
will infer that the producer regards Burma as an inefficient 
sleuth. In this example it is sufficient to retain pragmatic 
information concerning the producer's judgment of Burma 
while the semantic differences (induced by the use of "few" 
versus "none at all") can be disregarded. However, in other 
contexts the semantics will matter much more -- consider, 
for instance, (3) for which there can be no paraphrase with 
"no clues at all." 
(1) Nestor Burma found few clues. Nobody was 
surprised. 
(2) Nestor Burma found no clues at all. Nobody 
was surprised. 
(3) Nestor Burma found few clues. But it was 
still better than having none at all. 
The difference between (4) and (5), whose truth conditions 
are similar, is purely argumentative (or attidudinal) -- (4) 
expresses a positive (optimistic!) attitude, (5) the opposite 
point of view. This example shows how crucial theextraction 
of the argumentative structure is, since it is the only clue for 
the inacceptability of (6). 
(4) Nestor has a little money. 
150 
(5) Nestor has little money. 
(6) ?Nestor has little money. He wouldn't mind 
spending some on chocolate. 
Finally, we use the attitude markers as a means of ex- 
pressing modality. Traditionally, formal semanticists have 
extended first order logic to modal logic in order to account 
for modals. This places the modals at a purely semantic level, 
and does not allow for a distinction between what is observ- 
able for beth producer and consumer, and what is not-- such 
as opinions, beliefs, etc. We consider that expressions like 
'perhaps,' 'possibly,' 'it is almost certain that' are clues as to 
what the producer's beliefs and attitudes are towards facts of 
the world and help the consumer modify or update his model 
of the producer. It is for the above reasons that we decided 
to include a detailed specification of producer attitudes into 
the input specification for generation. 
1.1. Attitudes in TAMERLAN 
TAMERLAN is a frame-based representation language for rep- 
resenting text meanings. In our approach, treatment of mean- 
ing is agent-centered, that is, all the processes of (and the 
knowledge used for) understanding, representation and real- 
ization of meaning are described with respect to the model 
of an intelligent agent. This agent includes a model of the 
world, a model of language and a model of language un- 
derstanding and generation. 1. The world model includes 
knowledge (beliefs) about other agents in the world, includ- 
ing interlocutors. In understanding language communication 
(text or dialog), an intelligent agent extracts and represents 
a) text meaning; b) the active set of producer (speaker, au- 
thor) goals and plans that led to the production of this text 
and c) a set of active parameters of the speech situation, 
including spatiotemporal characteristics, knowledge about 
participants and a set of pragmatic factors similar to Hovy's 
(1988) rhetorical goals. These three items form what we call 
the supermeaning of a text. 
To represent text meaning proper, TAMERLAN uses the 
following basic entity types: clause, relation and attitude. 
TAMFaLAN clauses delimit the propositional and pragmatic 
content of target language utterances. Relations represent 
links among events, objects, or textual objects (e.g., sen- 
tences, paragraphs, etc.). A definition and detailed descrip- 
tion of TAMF~LAN is given in Nirenburg and Defrise (forth- 
coming). 
Attitudes are represented in TAMFERLAN as a quintuple 
attitudei = typei, valuei, attributed - toi, scopei, timei, 
where typei is the type of the attitude; valuei is the value 
of the attitude, represented as a point or an interval on a 
*A more detailed description of our approach to agent-centered 
processing see in Nirenburg et al., 1986. 
{0,1} scale; attributed- toi points at the intelligent agent 
this attitude is attributed to; scopei takes as its value that part 
of the meaning representation to which the attitude is held; 
and finally timei represents the time at which this attitude is 
held. 
In somewhat greater detail, the definition of the fragment 
of TAMERLAN dealing with attitudes is as follows. 
<attitude> ::= 
<attitude-type> ::= 
<attitude-value> ::= 
<numerical-value> ::= 
<attitude-type> 
<attitude-value> 
<scope> 
<attributed-to> 
<attitude-time> 
epistemic I deontic I 
volition I expectation 
I evaluative I saliency 
<numerical-value> 
<point>* I 
<semi-interval>* I 
<interval>* 
<semi-interval> 
<interval> 
<point> 
<scope> 
::= > <point> I < <point> 
::= <point> <point> 
::= n, 0 <= n <= 1 
::= any Tamerlan expression 
or set of such 
<attributed-to> 
<attitude-time> 
::= 
::= 
any instance of the 
ontological type 
~'intelligent-agent'' 
since <time> until 
I since <time> 
until <time> 
<time> 
<time> ::= <absolute-time> I 
<time-variable> 
<absolute-time> ::= <month>-<date>-<year>- 
<hours>: 
<minutes>:<seconds>. 
<second-fractions> 
<time-valirable> ::= time_<integer> 
The taxonomy of attitude types is an enhancement of Re- 
ichman's treatment of "context spaces" (1985: 56). We use 
the terminology (if not exactly the spirit) of her distinction 
among the epistemic, evaluative and deontic issue-type con- 
text spaces. Context space is Reichman's term for a discourse 
segment. The issue context space corresponds to our atti- 
tude component, while the non-issue context space provides 
a shallow taxonomy for discourse segment types (Reichman 
defines comment, narrative support, and nonnarrative sup- 
port as the non-issue type values). It will be discussed and 
151 
illustrated in the next section. 
Ontological types are concepts in the intelligent agent's 
ontology and domain model. The organization of the ontol- 
ogy used in the DIOGENES project see, e.g., in Nirenburg and 
Levin (1989). Instances of ontological types are actual mean- 
ings, including those comprising a TAMERLAN text. Some 
instances are "remembered instances" (e.g., John Kennedy, 
The Washington Post etc.) and are stored in the agent's 
episodic memory. The absolute time at (or since or until) 
which an attitude has been held is shown, for instance, as 
05-12-90-13:45:11.56. Relative (or unknown) times are lo- 
cally represented as variables and treated with the help of 
temporal relations in TAMERLAN. 
The attributed-to component of the attitude simply binds 
the attitude to a particular cognitive agent (which may be the 
producer of the utterance or some other known or unknown 
agent), who endorses the responsibility of the content of 
the utterance. This is important for understanding reported 
slxech, and more generally the polyphony phenomena, in 
the sense of Ducrot (1984). Ducrot's theory of polyphony, 
an approach to extended reported speech treatment, provides 
a framework for dealing with the interpretation of a number 
of semantic and pragmatic phenomena, e.g. the difference 
in meaning and use between 'since' and 'because', certain 
particularities of negative sentences, etc. 
The scope of the attitude representation pinpoints the entity 
to which this attitude is expressed. The values of the scope 
can be an entire clause, a part of it or even another attitude 
value, with its scope. In understanding the text the text con- 
sumer notes the attitudes of the producer toward the content. 
The attitudes can be expressed toward events (7), objects (8), 
properties (9) or other attitudes (10). 
(7) The train, unfortunately, left at 5 p.m. 
(8) This book is interesting. 
(9) The meeting was reprehensibly unproductive. 
(10) Unfortunately, I ought to leave. 
McKeown and Elhadad (1989) also treat argumentative 
scales and attitudinals in a generation environment. They, 
however, consider these phenomena as part of syntax, thus 
avoiding the need to add a special pragmatic component to 
their system. This decision is appropriate from the point 
of view of minimizing the changes in an existing generator 
due to the inclusion of attitude information. However, if 
compatibility were not an issue, we believe that introducing 
a separate component is a more appropriate choice. 
2. Attitude Types 
The following example illustrates lexical realizations of the 
epistemie attitude (grouped by approximate attitude-value). 
1 Paul left. I know for sure Paul left. I believe without 
doubt that Paul left. It is true that Paul left. 
0.9 Paul must have left. Most probably, Paul left. 
0.8 Paul may have left. I'm prepared to believe that Paul 
left. Perhaps Paul left. I'm almost sure Paul left. 
0.6 It is possible that Paul left. I would think Paul left. 
Chances are Paul left. 
0.5 I don't know whether Paul left (or not). 
0.3 It is unlikely that Paul left. I doubt whether Paul left. 
0 Paul didn't leave. It is impossible for Paul to have left. 
I don't know that Paul left. I don't believe (at all) that 
Paul left. It is not true that Paul left. I know that Paul 
didn't leave. I believe (without a doubt) that Paul didn't 
leave. 
In our representation we do not distinguish what is from 
what the agent knows, believes or is certain about. "Ob- 
jective" reality, thus, doesn't exist in the system. Facts and 
events belong to the "projected reality" (Jackendoff's term), 
i.e., reality as perceived by an intelligent agent. The fact that 
something is or is not, happened or did not happen, bears the 
mark of the agent's perception. Hence the epistemic attitude. 
Degrees of knowledge are identified with degrees of belief 
and degrees of certainty. If an agent knows something, he 
is certain about it and believes it. "Paul left" = "I (the text 
producer) believe that Paul left" = "I know that Paul left." 
Similarly, we feel that if someone says "Paul didn't leave," 
it really means (to the text consumer who interprets it) "The 
producer doesn't believe at all that Paul left" = "The producer 
doesn't know that Paul left" = "It is impossible for Paul to 
have left" = "The producer doesn't believe that Paul left" = 
"It's not true that Paul left." Negation can be understood as 
an attitude towards the event "Paul left." Hence our decision 
to collapse the parity of sentence with the epistemic attitudes 
of the agent. Seeing negation as the realization of an agent's 
attitude has further advantages. Some uses of negation (the 
"polemic" use, in denials) as in the following dialog: 
A: Paul came to the party yesterday. 
B: He didn't come. <I saw him downtown with 
his girlfriend. At the time of the party, he was 
o.. 
demand an analysis that take into account more than parity, 
contrasting explicitly different agent's attitudes towards the 
same event (this is similar to Ducrot's (1983) "polyphony"). 
we can provide a good representation of the above dialog 
using the "attributed-to" slot of an epistemic attitude frame. 
This representation will include the representation of the 
152 
meaning of the clause "Paul came to the party yesterday" in 
a TAMERLAN clause, say, clause_l, and two epistemic attitude 
frames, as follows: 
(attitude 1 
(type ep~stemic) 
(value i) 
(attributed-to A) 
(scope clause_l)) 
(attitude 2 
(type ep~stemic) 
(value 0) 
(attributed-to B) 
(scope clause I)) 
In generating spoken text, the fact that the representation 
contains opposite epistemic attitudes with similar scopes will 
be realized through marked intonation. In contrast, a text fea- 
turing a simple negation (not a denial of a previous assertion, 
but a simple negative assertion) will not be represented using 
two opposite-value epistemic attitudes with similar scope. 
Furthermore, representing parity as an attitude gives rise to 
"formulas" that elegantly translate certain semantic relations 
between sentences. For instance the synonymy of the natural 
language sentences "The book is not interesting" and "The 
book is uninteresting" is translated in terms of attitudes as 
The equality will be valid only if the "attributed-to" slots of 
the relevant attitudes have the same fillers. The above means 
that negation is generally understood as having a "lowering 
effect" -- something not interesting is less than interesting. 
When the condition about the "attributed-to" fillers is not 
fulfilled, negation must be understood as polemical, and in 
this case the meaning of "the book is not interesting" could, 
in fact, be as in "the book is not interesting; it is fascinating." 
(Once again, in speech a marked intonation will be used.) 
The realization of the deontic attitude can be illustrated as 
follows: 
1 I must go. I have to go 
0.8-0.2 I ought to go. I'd better go. I should go. You may 
go. 
0 I needn't go. 
Some illustrationsof the realization of the volition attitude: 
1 I wish ... I want to... I will... I will gladly... 
0.8-0.2 I hesitate to... It may be a good idea to... 
reluctant to... 
0 I'm unwilling to... I refuse to... I don't want... 
I'm 
Some lexical realizations of the expectation attitude: 
(attitude 3 
(type ep~stemic) 
(value 0) 
(attributed-to A) 
(scope (clause 2 attitude_4))) 
(attitude 4 
(type evaluative) 
(value I) 
(attributed-to A) 
(scope clause 2)) 
and 
(attitude 5 
(type ep~stemic) 
(value i) 
(attributed-to A) 
(scope (clause 2 attitude_6))) 
(attitude 6 
(type evaluative) 
(value 0) 
(attributed-to A) 
(scope clause 2)) 
respectively (clause_2 represents the meaning glossed as 
"this book," because the entire sentences only express the 
attitude toward the book). Therefore, 
(epistemic 0 (evaluative 1)) = (epistemic 1 (evaluative 0)) 
1 Not surprisingly... As expected... Of course... 
Needless to say... 
0.8-0.2 Even (as in "Even Paul left") 
0 Surprisingly... It couldn't be expected... 
The last two attitudes, evaluative and saliency can have 
in their scope not only clauses, relations or attitudes like the 
previous ones, but also objects and properties. 2 It is therefore 
difficult to give a limited and exhaustive set of examples of 
realizations. 
The evaluative scale goes, like the others, from 1 to 0. 
The endpoints are interpreted as roughly "the best" ("very 
good") and "the worst" ("very bad"). Depending on the 
scope, realizations will greatly vary and will include no lex- 
ical realization at all. If the scope is an event, adverbs like 
fortunately and unfortunately will be used. If the scope is the 
physical appearance of a person, theendpoints of the scale of 
evaluative attitude will be realized as "attractive" and "ugly," 
etc. 
The saliency attitude plays an important role in selecting 
the syntactic structure of the target sentences and in the lexi- 
cal selection. Thus, it will influence the order of elements in a 
2The evaluative attitude to "the book" in the example above 
would, in fact, belong to this latter class. Its scope is a clause 
only because there is no predication in the sentence other than the 
attitudinal one. In a sentence like "John read an interesting book" 
the attitude is clearly toward an object instance. 
153 
conjunction; it will be realized syntactically through topical- 
ization ("It is Paul who ...") and lexically through connective 
expressions such as last but not least or most importantly. 
3. Text Plan Representation 
In a nutshell, the flow of data in DIOGENES can be described 
as follows. The first processing component in DIOGENES 
is its text planner which, taking into account the input "su- 
permeaning" produces a text plan, a structure containing 
information about the order and boundaries of target lan- 
guage sentences; the decisions of reference realization and 
lexical selection (for both open and most closed-class lexi- 
cal items). At the next stage, a set of semantics-to-syntax 
mapping rules are used to produce a set of target-language 
syntactic stn~ctures (we are using the f-structures of LFG 
-- see, e.g., Nirenburg and Levin, 1989). Finally, a syntac- 
tic realizer produces a target language text from the set of 
f-structures. 
The text plan language we use in DIOGENES includes the 
following types of constructs -- the plan-sentence, the plan- 
clause, two kinds of plan-roles and the plan modifier. The 
frames for these constructs are as follows: 
(S_# 
(type plan-sentence) 
(subtype <TYPE>) 
(clauses (<C_#>*)) 
) 
(C # 
(type plan-clause) 
(head <word-sense>) 
(realization {ellipsis I pro I lexical}) 
(features <feature-value>* ) 
(topic {passive I cleft I passive-cleft 
I active}) 
(role <R #>)* 
(modifiers <MOD #>* <R #>*) 
) 
(R_# 
(type plan-role) 
(head <word-sense>) 
(realization {ellipsis I pro 
(features <feature-value>* ) 
(role <R #>)* 
(modifiers <MOD #>* <C #>* <R 
) 
I lexical}) 
#>*) 
(R # 
(type plan-role) 
(head $SET$) 
(elements <R #><R #>*) 
(type {CONJ T DISJ}) 
(realization {ellipsis I pro 
(features <feature-value>* ) 
) 
I lexical}) 
(MOD # 
(type plan-modifier) 
(head <word-sense>) 
(realization {ellipsis I pro 
(features <feature-value>* ) 
(modifiers <MOD #>* ) 
I lexical}) 
Types of plan sentences at present include simple, 
compound-conjunctive and compound-disjunctive. The 
realization property has 3 possible values - lexical, 
ellpisis and pro. topic is used to mark the topical- 
ized/focused elements in the clause; this property is also 
used to specify that the clause will be active or passive and 
whether it will feature an active or passive cleft construction. 
modifiers is a slot in which all the modifiers of a given 
plan concept are listed. 
The text planner will to determine which of the thematic 
roles in the input are to be realized as arguments and which, 
as modifiers. 
3.1. Text Planning Rules for Attitudes. 
Text planning rules in DIOGENES deal with a variety of phe- 
nomena. Some are devoted to text structure proper -- the 
number and order of sentences and clauses to express the 
meanings of input; clause dependency structures, etc. Oth- 
ers deal with treatment of reference -- pronominalization, 
ellipsis, etc. Still others take care of lexical selection, deter- 
mine tense and mood features of the target text, etc. In this 
section we illustrate text planning rules devoted to realization 
of attitudes. 
Rule A1 deals with an attitude of the evaluative type; rules 
A2 through A4 with attitudes of the epistemic type. 
Ai. IF (and (= clause i.attitude.type 
evaluative) 
(= clause i.attitude.value 
(< 0.3~) 
(= clause i.attitude.scope 
clause--i.proposition)) 
THEN (add-unit-filler C i 
~attitude ~unfortunately) 
A2. IF (and (= clause i.attitude.type 
epistemic) 
(= clause i.attitude.value 
(= clause i.attitude.scope 
clause_i.proposition)) 
THEN (add-unit-facet-filler C i 
~features 'mood 'declarative) 
i) 
A3. IF (and 
THEN 
(= clause i.attitude.type 
epistemic) 
(= clause i.attitude.value 
(= clause i.attitude.scope 
clause i.proposition)) 
(add-unit-facet-filler C i 
0) 
154 
'features 'mood 'negative) 
A4. IF (and (= clause i.attitude.type 
epistemic) 
(= clause i.attitude.value 
(= clause-i.attitude.scope 
clause i.proposition)) 
THEN (add-unit-filler C i 
~attitude ~perhap~ 
0.5) 
Attitudes get realized either lexically, through the inclu- 
sion of a lexical unit or through grammatical features. In the 
sample rules, the if clauses check the values in TAMERLAN 
and, depending on the actual match, either add features to 
the text plan or add a lexical realization for the attitudinal 
meaning (as in Rule A4). 
4. Status and Future Work 
In the DIOGENES project we adopt the methodological attitude 
of developing the generator functionalities in a breadth-first 
fashion. That is to say that, unlike many other projects, 
we do not tend to describe exhaustively a specific linguis- 
tic phenomenon (e.g., negation, anaphora, aspect, scope of 
quantifiers) or type of processing (e.g., text planning, lexical 
selection, syntactic realization) before proceeding to the next 
one (this approach can be considered depth-first). We prefer 
to go for a complete functioning system which contains all 
(or, in practice, most) of the above components and covers 
all (or most) of the above phenomena. It is clear that, at the 
beginning, the treatment of each (or most) of these compo- 
nents is incomplete, and not every phenomenon is described 
in sufficient detail. However, this methodology allows us to 
benefit from a complete experimentation envkonment and 
an open-ended architecture that facilitates the addition of 
knowledge to the system and its testing and debugging. At 
present we have a working prototype text planning and gen- 
eration system with narrow coverage. Our current work is 
devoted to expanding the knowledge needed for achieving a 
deeper level of analysis of each of the linguistic phenomena 
covered in the system. 

Bibliography
Anscombre, J.-C. and O. Ducrot. 1983. L'argumentation dans la 
langue. Brussels: Mardaga. 
Defrise, C. and S. Nirenburg (in preparation). Aspects of Text 
Meaning. Center for Machine Translation, Carnegie Mellon Uni- 
versity. 
Duerot, O. 1984. Polyphonie. In La//es, 4. 
Hovy, E. 1987. Generating Natural Language under Pragmatic 
Constraints. Yale University Ph.D. Dissertation. 
McKeown, K. and M. Elhadad. 1989. A Comparison of Surface 
Language Generators: A Case Study in Choice of Connectives. 
MS. Columbia University. 
Nirenburg, S. and V. Raskin. 1987. The Subworld Concept Lexicon 
and the Lexicon Management System. Computational Linguistics, 
Volume 13, Issue 3-4. 
Nirenburg, S., E. Nyberg, R. McCardell, S. Huffman, E. Kenschaft 
and I. Nirenburg. 1988. Diogenes-88. Technical Report CMU- 
CMT-88-107. Carnegie-Mellon University. June. 
Nirenburg, S. and L. Levin. 1989. Knowledge Representation 
Support. Machine Translation, 4, pp. 25 - 52. 
Nirenburg, S., E. Nyberg and C. Defrise. 1989. Text Planning 
with Opportunistic Control. Technical Report CMU-CMT-88-113. 
Carnegie-Mellon University. June. 
Reichman, R. 1985. Getting Computers to Talk Like You and 
Me. Cambridge, MA: MIT Press. 