CONTENT SELECTION AND ORGANIZATION 
AS A PROCESS INVOLVING COMPROMISES 
Helmut Horacek 
Universitiit Bielefeld, Fakult/it ffiir Linguistik und Literaturwissenschaft 
Universitiitsstr. 25, 33615 Bielefeld, Deutschland 
ABSTRACT 
Understanding a fairly complex message may demand an 
increasing degree of effort on behalf of the addressee, a fact 
that has been neglected almost completely in automated ap- 
proaches to natural language generation so far. Encountering 
this problam requires making compromises by reducing the 
degree of detail in which information is presented, or by 
explicitly expressing information left implicit otherwise. 
We identify factors that influence the comprehension effort, 
and we develop a model that indicates a rough quantified 
estimate for this effort. We illustrate these ideas by exam- 
ples of explanations comprising a considerable number of 
arguments, and we discuss potential impacts of these 
measurements on the process of composing a message by 
making compromises. 
1. MOTIVATION 
Communication in natural language may occasionally fail if 
the effort associated with mentally captaring the information 
conveyed becomes too'high. In order to avoid this problem, 
generation programs must have moderately accurate estimat- 
es at their disposal, that capture the concepts "degree ofex- 
plicitness/implicimess" and Mrnount of information~ degree 
of detail" to give an indication of the effort associated with 
understanding the messages they create 1. 
We illustrate these considerations by'discussing expla- 
nations that illustrate proposals made by an expert system. 
The system assigns employees to rooms, thereby meeting 
requirements to provide resources needed, to avoid social 
conflicts, etc. These requirements break down into 
constraints, which are relevant for the problem solving 
process, and into justifications, which give the domain- 
specific rationale behind these constraints. When the system 
comes up with a set of kssignments for employees (here: A, 
B, C, and D) to rooms (here: 1, 2, and 3), the user may ask 
questions about hypothetical assignments deviating from 
the proposed solution,l expecting the system to find out a 
relevant part of the problem specifications which are respon- 
sible for the infeasibility of the assignments focussed on. 
Figure 1 represents the text plan of an explanation telling 
why C and D must be ifi room 3. The plan breaks down into 
two subplans addressing rooms 1 and 2, respectively, each 
of which consists of arguments and justifications. 
We believe that these considerations apply to oral as well as to written presentations, since being forced to reread portions of a text 
frequently iz hardly desirable. A similar problem to this issue lies in adequately determining the node size in a hypertext document. 
There are several possibilities to express this text plan (or 
some portion of it) in terms of natural language texts. The 
simplest version (only comprising propositions 1 ands) is 
1) Room I must be assigned to a. and room 2 must be assigned 
roB. 
which is easily comprehensible, but hardly informative. 
Hence, more details should be added, yielding, for instance 
2) Room 1 must be assigned to A because A is a group leader. 
and room 2 mus? be assigned to B because B is a smoker 
(additionally comprising propositions It and ~). 
3) Room 1 must be assigned to a because a as a group leader 
must be in a big single room and room I is the only big 
single room available. Room 2 must be assigned to $ 
because B must not share a room with either C or D, since B 
is" a smoker, and room 2 is" the last single room out of two 
rooms left (comprising propositions I to ~). 
Neither of these explanations seems to be entirely satis- 
factory. Whereas text 2) leaves too much burden on the 
addressee's inferential capabilities, text 3) is simply too 
long. Much is left implicit in text ~2), which might be 
recoverable in principle, provided the addressee is well 
acquainted with the domain regularities (for instance, that 
group leaders must be assigned to big single rooms), but it 
is doubtful whether the addressee will really succeed in doing 
so. In text 3), we suspect, the addressee might easily forget 
some part of the argumentation before the explanation is 
completed. Hence, some compromise is required, like: 
REA~)N 
I room 1 is 
assigned to 
J USTIFICATION 
n~ s 
A must be in 
a big single room 
ELABORATION 
A is a group leader 
4 1 is the onl 
big single room 
~ST 
REASON °/ 
$ room2is \ 
assigned to B\ 
LIST 
~6 room 2 is the last single 
k room out of two rooms left 
JUSTIFICATION 
~ Bisa smoker 
~B "s room must be dtt'lerent 
from C's and D's rooms 
Figure 1: A text structure including arguments in fifll detail 
221 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
4) Room 1 must be assigned to a because A mast be in a big 
single room, and room 2 must be assigned to B because B 
mast not share a room with either C or D, and because room 2 
is the last single room out of tree rooms left (comprising 
propositions 1, 2, $, 6 and Z cutting off the justifications). 
5) Room 1 mast be assigned to a because a as a group leader 
must be in a big single room and room 1 is the only big 
single room available. And room 2 must be assigned to B 
(comprising propositions I to 4, and elaborating on 1). 
We believe that variants 4) and 5) are fairly comprehensible 
and provide some useful information, but not all that is 
available. However, if the explanation seeking person is 
interested in further detail, he/she may ask for justifications 
to text 4), and for elaborations of text 5~ 
In the following, we briefly review the contributions to the 
concepts "effort required to understand a message" and "degree 
of detail entailed in a message" made in the field so far. We 
describe a first sketch of a formal model that captures these 
two concepts and accounts for the compensative effect 
between them. We discuss examples which illustrate the 
tension between informativeness and comprehensibility 
including possible compromises between these two factors. 
2. PREVIOUS APPROACHES 
There is ample evidence from psychological experiments 
that humans draw causal inferences during reading to dose 
gaps left implicit in narrative texts (\[ 12\]). This reasoning is 
presumably done by building forward-oriented expectations 
and by drawing backward-driven inferences \[5\]. In assessing 
the processing effort associated with understanding the 
relation between two subsequent sentences, \[ 18\] distinguish 
between "direct causes", which can easily be understood, and 
"indirect causes" and "urtrelated facts ~, which both require 
considerable reasoning effort - to infer or, at l~ast, guess a 
relation or to resign in the attempt of doing so. The 
distinction between "direct" and "indirect causes" has been 
made on a statistical basis, which expresses comrnonalities 
about default expectations holding across a set of subjects 2. 
In automated approaches to natural language generation, the 
addressee's comprehension process is anticipated in a few 
aspects only. They comprise: the avoidance of a potential 
ambiguity caused by a particular referring expression \[ 10\], 
the inferability of additional information to convey from 
those propositions uttered explicitly \[7\], and the proper use 
of basic level categories like "dog" and "house" to avoid false 
implicatures \[ 16\]. However, the mechanisms developed are 
mainly motivated by detecting potential sotrrces of misinter- 
pretations and by avoiding the production of redundant text. 
In addition, active checking of the user's understanding has 
been incorporated in a system concerned with providing 
explanations 12\], and a selection mechanism is used in a 
reactive approach by Moore and Swartout \[15\], which incor- 
Within the cases classified as "indirect causes', some expectations 
triggered by the fact presented in f'ust place constituted "direst causeg" of the fact presented in second place, while the "direct 
causes" of some facts presented in second place were too specific to 
be expectations triggered by the facts presented in first place. 
porates decisions about the most promising explanation 
strategy under the context of a given situation. 
Some series of experiments have been carried out to obtain 
empirical evidence about how much information humans 
can typically understand, and what can be considered too 
much to be memorized without problems. Efficiency in 
language processing is then influenced to a great extent by 
finding an optimal mixtt~e of old and new information 
according to the working memory's capacity \[11\], which is 
a capability that may vary strongly among individuals 
hence, it is almost impossible to identify precise qtmntifi- 
cations on general grounds. The only relevant experiments 
we know of have been reported on a long time ago in I131, 
which have resulted in the concept of the magical number 
seven plus~minus two. The major reason for this situation 
is that the variety of knowledge that helps humans in build- 
ing conceptual chunks is enormous, and this knowledge 
causes significant differences in their memory capabilities. 
Very few systems developed in the field have the capability 
of producing messages in significantly varying degrees of 
detail when starting from comparable content specifications. 
Hierarchical organization is exploited for content selection 
(in BLAH \[19\] and in EPICURE \[4\]). The hierarchy is cut 
off at some level, eventually due to the user's assumed 
domain expertise, and only more abstract specifications are 
included in the message to be composed. Some other appro- 
aches addressing this issue are PAULINE 19\] and the appro- 
ach by Bateman and Paris \[1\]. Hence, the motivations 
underlying the texts produced are time constraints and other 
pragmatic parameters, and effective term selection 
techniques based on the addressee's rol~ and commaud of 
domain knowledge. However, there is no evidence in these 
systems about the relation between the pieces of text 
actually generated and the precise communicative effect they 
might achieve (\[1\] constitutes an exception in this sense). 
As \[14\] have pointed out, the assumption underlying many 
of today's text planners 3 that discourse is composed of hier- 
archically structured segments in a completely recatrsive way 
is inadequate for extended explanations. Hence, if saying "X" 
is assumed to be comprehensible, and saying "Y" as well. 
then it is concluded that saying "X and Y" is also a suitable 
message. Within a certain range, this assumption is reason- 
able (for instance, mentioning the track in addition to a 
train's departure time is perfect), but this is clearly not 
without limitations (for instance, the description of a non- 
trivial route may easily become too complex). 
Hence, generationsystems never have developed a quantified 
model of the communicative effort involved in. under- 
standing their messages - but this has not been necessary so 
far: The situation description from which a communicative 
act is to be produced is hardly detailed and accurate enough 
to allow tor the rich variety of reactions humans are able to 
produce under comparable circumstances, and the amount of 
information to convey is hardly ever so large that techniques 
to mamtam the addressee's attention are a serious concern. 
3 This assumption is based on Cohen and Levesquc's model of 
rational action and interaction 13\]. 
222 
1 
1 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
3. TOWARDS A FORMAL MODEL 
3.1. Choosing Degrees of Explicitness 
In general terms, the goal in selecting a suitable degree of 
explicitness and implicitness lies in following the coneept 
identified in the course of the experiments carried out by 
\[18\], namely leaving: "direct causes" to be inferred by the 
reader, and expressing "indirect causes" explicitly. This 
oi • • strategy tends to achieve a balance between pieces of refor- 
mation expressed exlSlieitly and others left implicit, in the 
spirit of the discussion of explanations given in section 1. 
In the approach adopted in \[8\], we have envisioned this goal 
by exploiting inference patterns based on the relations 
between generic regularities and indivuals involved in 
principle-based explanations, thereby also taking scalar 
implicature \[6\] nto account. This reasoning is based on the 
relevance that each oflthese pieces of information bears in an 
explanation context for understanding the rationale behind 
the decisions made (see also the examples in sections I and 
3.2). As argued in \[8.1, the mechanism developed is extend- 
able to capture at least the following types of inferences: 
• Inferring plausible sequences of actions; by exploiting 
expectations about intermediate steps indescriptions of 
action seqtfences can and should be omitted to produce 
less verbose and more natural text (see \[12\] and \[18\]). 
- Inferring causal relations between action and their 
results; reference to a problem-solving step to be 
accomplished can be established by the action to be per- 
formed or by the State to be achieved, according to the 
linguistic repertoire available (see \[ I7\]). 
In our application, the distinction between "direct" and 
"indirect causes" is implemented by leaving the results of 
chained inference steps implicit under particular circum- 
stances only: the results of scalar implicature (and logical 
deduction) are left to be uncovered by the addressee's infer- 
ence capability, even if combined with other inferences to be 
made. Eventually, this strategy may be refined if the effect 
of default expectations in the domain at hand is better 
understood and can be captured more systematically. 
We conceive that it is also possible to pursue other 
strategies, either by producing more concise utterances and 
thereby increasing the burden on the addressee's inferential 
capability, or by expressing more inference steps explicitly. 
However, we do not igo into potential variations of this 
aspect here, we simply adopt the approach of selecting the 
"best" balance between expressing information explicitly or 
implicitly, on the basis of the heuristics incorporated. 
3.2. Deciding Upon Degrees of Detail 
In order to decide whether the complexity of a message is 
beyond what the addressee can reasonably be assumed to 
understand with convenience, we propose the computation 
of three measures that serve as partial indications for this 
purpose: (1) The number of entities entailed in the proposi- 
tional representation of the message; (2) The number of pre- 
dicates used to build assertions about these entities; (3) The 
• number of relations holding between entities and predicates. 
In order to define entities and predicates for this purpose, we 
rely on the domain ontology of the underlying world model. 
The motivation behind lies in the assumption that, if the 
communicative purpose of the message is recognized, the 
conceptualization of the addressee should mirror to a large 
extent this categorization. In the office planniug domain, for 
instance, "employees', "rooms', "resources" constitute 
entities, while "next-to', "assigned-to" constitute predicates. 
As a total measure for complexity we propose the formula 
max (e+p, r, 2*0, where.p, e, r, i represent the number of 
entities, predicates, relations, and inferences, respectively. In 
the texts we are dealing with, these factors tend to be rough- 
ly the same; by the consideration of all factors, we intend to 
treat also other types of texts adequately: in enumerations, 
for instance, the number of relations decreases in compar- 
ison to the number of entities and predicates involved, while 
it increases if network:like relations between a set of entities 
and a set of predicates are to be expressed. 
Returning back to the examples discussed in section 1, we 
are able to justify our assessments in some formal sense. 
Figure 2 illustrates the complexity measures computed for 
each of these sentences, confirming the intuitive assessment 
made in section 1: the measurements introduced indicate a 
fair correspondence with Miller's magical number seven. 
The categorization into what constitutes an entity and what 
constitutes a predicate is problematical, since there are 
several sources of inaccuracies in this approach: for instance, 
complications due to vague expressions and due to qlmntifi- 
cation relations, and conceptualizations made by the 
addressee, which may differ significantly from those embo- 
died in the system's conception consisting of entities and 
relations. We may encounter this problem by incrementing 
the calculations in case terminological transformations or 
lexical operations involve significant changes in the repre- 
sentation of the message. Eventually, the calculation 
schema must be augmented in case involved quantification 
relations occur more frequently than in our application. 
In cases where the complexity assessments indicate potential 
comprehension problems on behalf of the addressee, several 
strategies are possible: a simple, but risky one, in which the 
complex utterance is fully elaborated, and two more sensible 
strategies, One of them consists in reducing the degree of 
detail in a coherent way (i.e., not leaving out details 
arbitrarily, but orienting this reduction on structural 
regularities), especially if demanded by time and/or space 
linfitations, and the other one consists in re, organizing the 
text so that additional structuring hints in the text produced 
enable the addressee to perform intermediate conceptuali- 
zation already with parts of the information' presented. 
sentence number 
number of entities (e) 
number of predicates (p) 
number of relations (r) 
number of inferences (i) 
total = max (e+p, r, 2"i) 
1) 2) 
4 6 
1 1 
4 6 
o 
4 
3) 4) 5) 
10 7 6 
3 3 1 
13 9 7 
110 1 3 7 
Figure 2: Complexity assessments for some example sentences 
223 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
The choice among alternatives should prefereably be 
interest-based. For instance, the choice between alternatives 
4) and 5) in section I may be made in favor of text 4) if the 
user wants to vary the relative significance of constraints in 
defning problem specifications (therefore, he/she wants to 
know which ones prove relevant for the aspect in question). 
Alternative 5) is preferable for a user who wants to learn the 
rationale behind system decisions, which particularly 
includes associating justifications with constraints. 
4, CONCLUSION 
A mechanism assessing the comprehension effort of the 
addressee of a text or an utterance certainly seems to be of 
interest for theoretical models of natural language generation 
processes. While the proposal entailed in our method can 
hardly be considered a psychologically motivated approach, 
we believe that it contributes to the understanding of what 
ingredients a performance-oriented model of natural language 
generation should consist of, and how interaction between 
these ingredients can suitably be organized. 
Apart from a purely theoretical perspective, we believe that 
these considerations are also relevant in practice, even if an 
assessment of the comprehension effort of the addressee does 
not manifest itself in the majority of generator programs 
themselves - which also may not even prove necessary in a 
good deal of applications. Hence, for simpler, eventually 
application-oriented systems, taking these considerations 
into account will help in making explicit the assumptions 
underlying the simplifications embodied, so that conditions 
for an eventual transfer to other domains become more 
evident - hence, the assumptions must hold across domains. 
However, it will be important to know about a system's 
limitations, where they manifest themselves, and when they 
are likely to weaken the system's usefulness. 
Conditions under which assessing the complexity of a 
message to be generated may prove benefieial, hold in at 
least the following types of application: 
• Presenting a significant quantity of records selected from 
a database; when envisioning an ambitious, flexible 
presentation, summary facilities play a crucial role based 
on suitable assessments in the spirit of our method. 
• Generating business reports including a considerable 
amount of (heterogeneous) data, where summaries also 
play an important role; in that genre, it is the contextual 
inferability of information from some key propositions 
conveyed, which constitute the main difficulty. Hence, a 
quality assessment function has to take this particular 
aspect into account, whereas other influences on the ease 
of comprehension can be widely neglected. 
• Explaining an issue of at least moderate complexity (for 
instance, presenting a chain of rules or a set of 
constraints with underlying justifications) in explanation 
generation; all aspects neglected in report generation 
(degree of detail, slructure, acquaintance with terms) bear 
a significant amount of relevance here, and they should 
be incorporated in a suitable quality function. 

REFERENCES 
\[1\] J. Bateman, C. Paris: Phrasing a Text in Terms the User 
can Understand. In Proc. of IJCAI-89, pp. 1511-1517, 
Detroit, 1989. 
\[2\] A. Cawsey: Generating Explanatory Discourse. In Current 
Issues in Natural Lanzuage Generation, R. Dale, C. 
Mellish, M. Zock (eds.), pp. 75-102, Academic Press, 
1990. 
\[3l P. Cohen, H. Levesque: Speech Acts and Rationality. In 
Proe. of ACL-85, pp. 49-60, 1985. 
\[4\] R. Dale: Cooking Up Referring Expressions. In Proe. of 
A CL-89, pp. 68-75, Vancouver, 1989. 
\[5\] A. Gamham: Testing Psychological Theories about Infer- 
ence Making. In Memory and Cognition 10(4), pp. 341- 
349, 1982. 
\[6\] J. Hirsehberg: A Theory of Scalar bnplicature. Garland 
Press, New York, 1991. 
\[7\] H. Horaeek: Exploiting Conversational lmplicature for 
Generating Concise Explanations. In Proc. of EACL-9!, 
Vol. 1, pp. 191-193, Berlin, 1991. 
\[8\] H. Horacek: How to Avoid Explaning Obvious Things 
(Without Omitting Central Information), to appear in 
ECAI-94. 
\[91 E. Hovy: Pragmatics and Natural Language Generation. In 
Artificial Intelligence, 43, pp. 153-197, 1990. 
\[10\] T. Jameson, W. Wahlster: User Modeling in Anaphora 
Generation." Ellipsis and Definite Description. In Proe. of 
ECAI-82, pp. 222-227, Orsay, 1982. 
\[11\] M. Just, P. Carpenter: A Capacity Theory of Compre- 
hension: Individual Differences in Working Memory. In 
Psychological Review 99, pp. 122-149, 1992. 
\[12\] W. Kintseh, J. Keenan, G. McKoon: Memory for Infor- 
mation Inferred During Reading. In The Representation of 
Meaning in Memory, W. Kintsch (ed.), Earlbaum, 
HiUsdale, New Jersey, 1974. 
\[13J G. Miller: The Magical Number Seven Plus or Minus Two: 
Some Limits on Our Capacity for Processing Information. 
In Psychological Review 63, pp. 81-97, 1956. 
\[141 D. Mooney, S. Carberry, K. McCoy: The Generation of 
High-Level Structure for Extended Explanations. In Proc. 
of COLING-90, pp. 276~281, Helsinki, 1990. 
\[15\] J. Moore, W. Swartout: A Reactive Approach to Eapla- 
nation. In Proe. of IJCAI-89, pp. 1504-1510, Detroit, 
1989. 
\[16\] E. Reiter: Generating Descriptions that Exploit a User's 
Domain Knowledge. In Current Issues in Natural Languaee 
Generation, R. Dale, C. MeUish, M. Zoek (eds.), pp. 257- 
285, Academic Press, New York, 1990. 
\[17\] D. R6sner: Intentions, Rhetoric, or Discourse Relations? 
A Case .from Multilingual Document Generation. In 
Intentionalitv and Structure in Discourse .Relations, 
Workshop held at the ACL-93, O. Rambow (ed.), pp. 106- 
109, Columbus, Ohio, 1993. 
\[181 M. Thfiring, K. Wender: Ober kausale inferenzen beim 
Lesen. In Spraehe und Ko~nition 2. pp. 76-86, 1985. 
\[191 J. Weiner: BLAH: A System which Explains its 
Reasoning. In Artificial Intelligence, 15(1), pp. 19-48. 
1980. 
