Selection: Salience, Relevance and the Coupling between 
Domain-Level Tasks and Text Planning 
T. Pattabhiraman 
Nick Cercone 
Centre for Systems Science 
Simon Fraser University 
Burnaby, B.C., Canada V5A 1S6 
patta@cs.sfu.ca and nick@cs.sfu.ca 
Abstract 
In this paper we examine some issues pertaining to the task 
of selection in text planning. We attempt to distinguish 
salience and relevance, and characterize their role as 
important fundamental notions governing selection. We also 
formulate the problem of selection of text content in terms of 
the coupling between domain-level tasks and text planning 
tasks. We describe our research on generating bus route 
descriptions. 
Keywords: Natural Language Generation, Text Planning, 
Selection, Salience, Relevance, Coupling, Route Descriptions 
1. Introduction 
Most models of natural language generation, be they 
computational or psychological, recognize that the task of 
text planning (also called conceptnalization\[14\], \[10\]), 
comprises the following essential subtasks: (1) content 
selection and (2) content organization. The task of content 
selection (hereafter, selection) is concerned with 
computing, or retrieving from a knowledge base, the 
primary content of texts. Selection may also include a 
phase of content expansion, also called additional topic 
inclusion \[8\]. Content expansion consists in inferentially 
generating or selecting additional material for expression in 
text, once the primary material is selected. The task of 
content organization, also called topic organization \[8\], 
content ordering or linearization \[13\], \[14\] deals with 
ordering the chosen content into a sequence of 
propositional (linear) representations, appropriate for 
realization into coherent text. 
Many text planning models concentrate chiefly on 
content organization and porhaps content expansion. Such 
planners accept as input a pre-selected, and to some extent, 
pre-sequenced collection of representations that serve as 
the raw material of text content. The knowledge base from 
which these input elements are selected, as well as the 
selection process itself, are treated to be external to the text 
planner. In such systems, whatever selection is porformed 
by the text planner is confined to the task of choosing a 
subset of the input elements according to control exercised 
by knowledge sources resident within the text planner. 
Other examples of systems in which pre-selected content is 
input to the text planner include natural language 
generation front-ends to export systems, database systems 
or other application programs- in other words, a problem- 
solver or a host system which is devoted to domain-level 
non-linguistic activities. 
In some other models, such as those of Paris \[23\], the 
knowledge base from which much of the text content is 
drawn is resident within the text generation system itself. 
Selection in such systems is totally a responsibility of the 
text planner. In the TAILOR system of \[23\], for instance, 
facts describing objects are stored in a knowledge base, and 
the textual component selects the content of the description 
from the knowledge base under the regulatory influence of 
a user model. The urge to generate text is input to the 
generator in the form of a request for the definition of an object. 
In all cases, the beholder of the natural language output 
sees the text as coming from a single program which can 
porform domain-level tasks as well as text production tasks: 
the speaker and the thinker are one and the same. If the 
beholder were a text planning researcher, she could be 
inclined to pose questions about the origin of text content. 
Our recent research has been motivated and directed by the 
adoption of such a role. The problem of selection in texts 
(in our case, multisentential descriptive texts) led us to 
examine the nature of input to the text planner, and the 
boundary between the domain-level program and the text 
planner. When we regarded the thinker and the speaker as a 
unified whole, we were led to search for very general 
factors that influenced selection in diverse domains of 
discourse production. When we viewed domain-level 
activities and text planning activities as distinct tasks, we 
examined the division of duties between the text planner 
and its underlying program in the task of selection, and 
explored the conditions under which the modular boundary 
between the domain-level and the text-planning level could 
be kept intact and those under which it might break down. 
This papor is devoted to a presentation of our research on 
some issues portinent to selection in text planning. We 
found that the fundamental notions that were crucial to 
understanding selection in text planning were salience and 
relevance. These are not altogether unfamiliar notions. 
Conklin and McDonald \[2\] and Waltz \[33\] describe some 
effects of salience in generating scene descriptions. 
Researchers in natural language understanding have 
studied salience and relevance in some detail and have used 
79 
them profitably in their accounts. However, in the natural 
language generation research community, the terms are 
used in their literal sense, and often interchangeably, 
whereas in fact they are distinct notions. In this paper we 
present the notions of salience and relevance as they pertain 
to natural language generation, in particular, to (content) 
selection. 
Our presentation in section 2 is a synthesis of several 
analyses of salience and relevance in the disciplines of 
language understanding, psycholinguistics and 
communication. In section 3 we consider the coupling 
between domain-level tasks and text planning tasks from 
the point of view of selection. Our work in the domain of 
route description generation is presented in section 4. In 
this domain, interesting questions emer!~e regarding mode 
of knowledge representation, connecUons between text 
planning and domain-level problem solving, and selection. 
In the concluding section we briefly state our current work 
and research plans for the near future. 
2. Salience and Relevance 
Salience and relevance are theoretical notions which are 
influential in accounting for how or why certain objects, 
concepts, properties or actions are highlighted or preferred 
in natural language processing. Of particular interest to the 
subject of this paper is the role of these important notions 
in controlling selection (and omission) in text planning. 
What is common to salience and relevance is their role as 
determinants of decisions in selection.However, salience is 
connected with speaker-external objects or properties, 
while relevance is related to speaker-internal factors such 
as goals and motivation. 
2.1 Salience 
Salience explains the prioritization or foregrounding of 
objects or information. Such prioritization may arise from 
direct perception of the world or context, or from 
conceptual knowledge, built up by shared perceptual 
experiences of the speaker and the hearer. According to 
Sridhar \[30\] and Levelt \[14\], salience guides the direction 
of the speaker's attention by making certain aspects of a 
situation stand out relative to others. By virtue of being 
ascribed to factors of context, situation or background, 
salience is a relatively fixed notion, invariant with respect 
to specific perceivers. Salience information can therefore 
be lexicalized, stored as an invariant component of the 
knowledge base, incorporated as a structural property of the 
design of the knowledge base, and/or built into the control 
processes that select text content from the knowledge base. 
Perceptual salience arises from the prominence of 
external characteristics such as size (bigness), amplitude of 
sound (loudness), colour (brightness), etc. Osgood and 
Bock \[22\] call this factor vividness. Conceptual salience 
originates in (shared) experiential knowledge. For 
example, the slyness of foxes, the strength of gorillas, and 
other such distinguishing characteristics are conceptually 
salient \[4\]. The head of a queue of people is a concepmaUy 
salient part of the queue with respect to its function \[7\]. 
Conceptual salience also arises from the degree of 
unexpectedness, unusualness or deviation from the norm of 
an object or event, giving the object or event of description 
its information va/ue and thus motivating its selection in 
the planned text. Salience is attributed to properties of 
objects as well as to components of complex objects. 
An important implication of salience for objects that 
possess it is the ease of their availability in the knowledge 
base or knowledge structure \[34\]. The higher the level of 
salience of an object, the higher is its level of activation in 
the speaker's mind, and therefore, the ~-e~ter is its 
probability of being selected in the speaker s description, 
and the earlier is it likely to be mentioned in the 
description. This fact has been recognized by Conklin and 
McDonald \[2\], Osgood \[21\], Sridhar \[30\] and Levelt \[14\]; 
Salience and Text Planning. The significance of visual 
salience in text planning has been pointed out by Conklin 
and McDonald \[2\] in their empirical study involving scene 
descriptions (of suburban houses) and the GENARO 
program that generates similar descriptions. In GENARO, 
the chosen representation for the set of objects is a list, 
ordered according to decreasing levels of salience of the 
objects. Non-salient objects are not represented. This list 
mediates between the deep generation component and the 
domain data base in regulating the order of description of 
objects. 
Other Work on Salience. Sridhar \[30\] amplifies upon 
earlier psycholinguisfic treatments of the role of salience in 
single sentence production. He examines the effects of 
salience on intrasentential syntactic ordering in phenomena 
like passivization, topicalization and constituent order shift. 
His results are of significance to the realization component 
of natural language generation. On the subject of content 
selection in the description of states and events, Sridhar 
hypothesizes that actions are more salient than state 
changes, which in turn are more salient than constant states. 
Ortony \[20\] invokes salience in explaining the 
interpretation of metaphor. Herskovits \[7\] uses salience in 
her account of prepositional semantics and the construction 
of mental models from prepositional phrases. 
2.2 Relevance 
Relevance constrains participants' utterances and 
interpretations in communicative contexts. It connects texts 
and their generators through goals: speakers are assumed 
to generate utterances which they believe to be relevant to 
their goals \[1\]. It connects the context of text generation 
with the generator and the text by the fact that 
communication as well as the goals themselves are situated 
in context, and by the assumption that the conversational 
(or communicative) goals are relevant to the generator's 
personal goals. The speaker's communicative goals are 
characterized by the speaker wanting the goals to be made 
known to the hearer, and intending to reach them by the 
hearer knowing .~d adopting them \[25\]. Relevance is thus 
an important nouon that connects texts, contexts, spewers 
and hearers. Selecting relevant information in a text 
planning scenario is thus an issue of selecting information 
that is pertinent to the generator's communicative goals. 
Leech \[12\] defines relevance as follows: 
An utterance is relevant to a speech situation if it can be 
interpreted as contributing to the conversational goals of 
the speaker or hearer. 
Haslett \[6\] modifies this definition to account for the 
communicative actions and knowledge. Her definition of 
relevance is stated as follows: 
An utterance, action or unit of knowledge is relevant to a 
80 
speech situation ¢" it can be interpreted as contributing to 
the communicative goals of the speaker or listener. 
Relevance and Coherence. Haslett\[6\] emphasizes 
relevance over coherence as the essential property of texts 
emerging in communicative contexts. While coherence is 
essential in interpreting texts, it is not sufficient in 
accounting for how texts are planned and generated. Texts 
make sense not only in the general way of being 
interpretable, but also m the particular way of contributing 
to conversational goals. Generators try to make their texts 
both relevant and coherent. A relevant text must also be 
coherent in the sense of being meaningful to the 
conversational goals. However, not all coherent texts need 
be relevant (el. retorts like yes, but that's not relevant to the 
issue). In their work on generating expert system 
explanations, Moore and Swartout \[19\] implicitly aff'mn 
the precedence of relevance over coherence. Although the 
finer details of how they generate explanations are not 
directly relevant to the concerns of this paper, their stress 
on the role of the generator's goals in accounting for 
selected text content ts of significance. A similar view is 
evident in Paris \[24\]. 
Relevance and Selection. I-Iaslett's definition is suited 
to our purposes, since it explicitly recognizes the role of the 
sources (representations) as well as the kn°wledg~nvolved in 
processes selecting text content. Knowledge 
could be represented in any number of ways. However the 
definition of Haslett has implications for the design of 
particular text planners and specific models of text 
planning: a very natural design (for the knowledge base 
that stores much of what is eventually expressed in text, as 
well as for the processes that select from the knowledge 
base) will be one which aids, and is in harmony with, the 
goal structures present in the given setting or domain of 
text generation. Mann and Moore \[15\], in their work on the 
KDS system of text generation, point out the importance of 
relevance in judging whether an item of knowledge should 
be selected for expression in text. Their method, however, 
has numerous deficiencies, a discussion of which is 
presented in \[15\]. Paris \[24\] also recognizes explicitly the 
relation between selection and relevance through goals. 
Other Work on Relevance. Sperber and Wilson 
\[29\] stress the role of relevance in language 
comprehension. Their work is partly a response to the 
shortcomings of the semiotic code model of verbal 
communication. They emphasize the place of intended 
inference in verbal communication, over and above the 
encoding and decoding of messages. The reader is referred 
to \[6\] for a critique of the theory of Sperber and Wilson 
from the point of view of relevance, van Dijk and Kintsch 
\[32\] and Schank \[28\] treat the more local, restricted issue 
of topical relevance in sentences. 
2.3 Relevance and Salience Contrasted 
Salience is a speaker-external, contextual principle, while 
relevance is a speaker-internal principle related to the 
speaker's goals. Waltz \[33\] states that in generating scene 
descriptions, what the speaker notices is a function of 
external factors (scene characteristics) and internal factors 
(like goals). Thus he effectively recognizes the distinction 
between salience and relevance, though he doesn't use the 
term relevance to describe the external factors. Similarly, 
psycholinguists like Sridhar \[30\] use the term salience to 
describe relevance as presented here. Notions derived from 
salience and relevance include focus and attention. 
Not all settings of text production may admit 
characterization of salience, but all settings involve 
relevance, to the extent that text production in the given 
setting is purposive. However, relevance can be a nebulous 
matter in many text production settings. Relevance and 
salience could sometimes reinforce each other and at other 
times, conflict with each other. When they do conflict, 
relevance, given Haslett's definition, would have greater 
weight in its dictates on selection decisions in particular 
and language generation decisions in general. This 
observation is in accord with Sridhar's psycholinguistic 
findings on single sentence production. 
3. Domain-Level Tasks and 
Text Planning Tasks 
We have noted relevance and salience as fundamental 
notions that guide selection of text content. The effects of 
relevance and salience can be brought out in text planning 
systems in several ways, of which one is sketched below. 
We assume that the primary content of texts is drawn from 
a knowledge base, which may consist of several component 
modules in the implementation. 
A large body of surely-irrelevant knowledge is not  p 
resented. This corresponds to a closed worm assumption 
r knowledge bases with respect to relevance and salience. 
The knowledge base only contains potentially relevant and 
salient items for all likely instances of text production. 
Selection is still responsible for choosing only the portion 
of the knowledge base that is pertinent to a particular 
instance of text planning. The tmnciples of relevance and 
salience, as may be embodied in a text planner, are thus a 
matter of representation as well as control. The control 
could be exercised by a variety of knowledge modules, 
such as the user model (as in the ease of Paris \[23\]), the 
hearer model (in effect, the user model) (as in the case of 
Dale \[3\]), etc. What needs to be said is selected; at the 
same time, what should be omitted is skipped, by the search 
process turning a blind eye to irrelevant knowledge, as it 
were. Selection processes mediate between the controlling 
knowledge sources and the knowledge base from which 
content is selected. 
The texts .produced by such systems may be perceived to 
be emerging from communicative goals, though 
communicative goals themselves are not explicitly 
represented. In the text planning models of \[3\] and \[23\], 
communicative goals need not be explicitly represented, 
since relevance information is implicit in the semantics and 
the content of the user model. When do the goals 
themselves require explicit representation and reasoning in 
the selection process? This question directs our analysis 
towards two aspects: 
1. the nature of the text produced 
2. the connection (or, coupling) between 
domain-level tasks and text planning tasks 
3.1 Explanation Generation 
The text could be self-referential, in that (for instance) the 
generator might allude to, justify or explain the prior 
content of text that it produced or the actions that were 
B1 
performed at the non-linguistic, domain-level. (The identity 
of meta-language with object-language is a quintessential 
characteristic of natural languages). An exemplary case in 
point is that of explanation generation in expert systems. 
Explanation systems explicitly represent goals so as to 
explain their domain-level actions, or to exp.lain their own 
prior utterances, for instance in answenng follow-up 
questions in dialogues with the user, as in the works of 
Moore and Paris \[18\], Paris \[24\] and Moore and Swartout 
\[19\]. The explanation system of McKeown \[16\] identifies 
its goals with the user's goals, which are hence represented 
in detail. 
3.2 Coupling between Domain-Level Tasks andText Planning Tasks 
We observed in the introduction that the speaker and the 
thinker are one and the same, in that a single system is 
perceived to perform domain-level activities such as 
problem-solving, action planning, etc, as well as text 
production tasks. In this context, we could speak of loose 
coupling and tight coupling between domain-level and text 
planning tasks. Loose coupling implies modularization 
between the two tasks m process models and 
implementations. On the one hand we could ask whether 
the domain-level tasks are in some sense naturally coupled 
to text planning tasks, and on the other, we could ask what 
the chosen theoretical paradigm, modelling method, 
implementation or application has to do with the coupling. 
Assuredly, the knowledge of relevance and ,salience that 
is necessary in selecting text content is conceptually 
distributed between the non-linguistic, domain-level and 
the text production level. Therefore, the knowledge and the 
processes involved in selecting are, in principle, distributed 
at both levels. In actual implementations or computational 
process models, if the coupling between domain-level tasks 
and text planning tasks is loose, then additional work will 
have to be done to make the coupling appear fight (i.e., to 
make the thinker identical with the speaker). This issue 
arises in the construction of explanation facilities for expert 
systems and front-end natural language generators for 
various application programs. Virtual tight coup.ling is 
achieved by building interfaces, or by augmenting the 
design of the expert system and/or the text planner. 
In the Explainable Expert System architecture described 
in Paris \[24\], this is achieved by designing expert systems 
with explanation in mind, and by using for explanations the 
support knowledge applied in deriving the expert system. 
Problem-solving knowledge is kept insulated in the 
implementation from the knowledge necessary for 
generating explanations. Selection in the explanation task is 
decided by the text planner and mediated by the use of a 
rich text plan language, described in Moore and Paris \[18\]. 
The design of powerful plan languages (for selecting text 
content from the knowledge base) is an interesting research 
problem \[9\]. 
In the EPICURE recipe-generation system of Dale \[3\], 
tight coupling is achieved by modelling discourse plans to 
be isomorphic to domain-level plans. There is no domain- 
level activity proceeding independently of (separately 
from) text planning. In the work of Appelt \[1\], tight 
coupling is ensured in the theoretical paradigm of action 
plannin8 in which both domain-level tasks and utterance 
production (linguistic) tasks are uniformly viewed as 
goal-directed actions. Content selection in Appelt's 
system is distributed in a complex way throughout the 
generation process by commitment to the view that the task 
of what to say is inseparable from the task of how to say. 
What about complex domains of language generation in which 
domain-level activities are naturally coupled to text 
planning activities, irrespective of the theoretical paradigm 
that may be used? In the next section we turn our attention 
to one such domain, viz., that of route communication. 
4. Generating Route Descriptions 
In this section we describe the domain of route description 
generation, and report some aspects of our research on 
generating descriptions of bus route directions from a given 
source to an intended destination within a city. Our work 
includes an implementation in CProlog of a prototype 
system that generates descriptions of bus route directions in 
Vancouver. A representation of the city or region in which 
route description generation occurs is available in the form 
of a cognitive map. The map serves as the knowledge base 
in the domain-level activity of finding a mute, as well as in 
providing the primary content of route description texts. 
4.1 Route Communication 
The verbal protocol that consists of requesting route 
directions and giving them is termed route communication 
by Klein\[ll\]. Klein uses route communications that 
occurred naturally on the streets of Frankfurt in exploring 
the relation between the cognitive (domain-level) task of 
route fmding and the linguistic task of generating a route 
description as manifested in the choice of local deictic 
terms like here, there, left and right. Wunderlich and 
Reinelt \[35\] deal with the discourse structure of route 
communication. Although full route communication occu~ 
as a dialogue, the roles of the participants are not 
symmetric as they would be, say, in a casual conversation. 
All varieties of route communication involve selecting 
from maps. However, the nature of information selected, 
and in particular the kind of spatial information included in 
the route description depend on such factors as the mode of 
transportation used, whether the questioner herself will be 
navigating or whether she will be using a public wansport 
system, etc. 
Bus Route Descriptions. In our research on generating 
bus route directions we emulated the methods of Klein 
\[11\] and Wunderlich and Reinelt\[35\] to gather and 
analyze natural text. Our primary linguistic d_am consist of 
over 40 bus route descriptions to various destinations in 
Vancouver, and include written (printed) as well as spoken 
descriptions. We concentrate on the route description phase 
of the dialogue, which emerges more or less as monological 
text Route descriptions involve knowledge of different 
granularities. For instance, when the full journey involves 
taking two or more buses, the connecting buses may be 
available at the point of disembarkation (from the previous 
bus) itself, or just across the street. But often, one may have 
to walk a fair distance before transferring to the next bus. In 
such cases, the bus route description also includes walk- 
route information, as illustrated by the italicized portion of 
the following text: 
from Klngsway and Edmonds you want to catch bus 
number 106 called Metrotown ... take it down to the 
82 
Edmonds skytrain station ... lake the skytrain as far as 
Burrard station ... and at Burrard station you want to 
walk north on Burrard one block ... as far as West 
Pender ... and on West Pender going westbound it's bus 
number 19 and it's called Stanley Park 
Route Descriptions vs Route Sketches. Route 
descriptions are also interesting in that they can also be 
communicated non-verbally (graphically) in the form of a 
route sketch. The following spoken text describes how to 
get from Lougheed Mall to Grouse Mountain in 
Vancouver. The "..."s correspond to pauses or confirmatory 
expressions like ok or yeah by the questioner. 
from Lougheed Mall.9 ... ok, you could catch a 151 or a 
152 called Vancouver ... get off at Hastings and 
Kootenay right by the Kootenay Loop ... on Hastings in 
front of the loop transfer to bus number 28 called Phibbs 
Exchange ... will cross Second Narrows Bridge to Phibbs 
Exchange on the other side ... and that's where you 
could catch the 232 Grouse Mountain bus ... and it goes 
fight up to Grouse Mountain. 
The information conveyed by this text could also be 
expressed graphically as illustrated in figure 1. There are 
interesting parallels between natural language route 
descriptions and route sketches as bearers of information. A 
one-to-one correspondence may be seen between certain 
aspects of route sketches and natural language route 
descriptions. However, route descriptions are linear, and 
the use of certain mechanisms such as connectives, 
pronominals and deictic terms is exclusively associated 
with the use of natural language as the medium of 
description. 
4.2 Planning and Discourse Units in 
Route Description Generation 
A journey consists of a series of connected route sections. 
However, not all route sections need be described in a route 
description. For instance, when a public transportation 
system is used, all the details of the turns taken by the 
conveyance used are not relevant for the description. 
However, when one has to drive or walk, information on 
each route section and turn should be communicated. 
Thus, route planning units at the domain level are in 
general more detailed than route description units at the 
discourse level. 
Route descriptions consist of a sequence of units, each 
specifying, at the very least, a source, a destination and a 
route label. In the case of bus route descriptions, such a unit 
consists of a specification of a boarding point, a 
disembarkation point and the bus label for a single bus ride. 
Such units provide the skeletal plans for route description 
texts. In figure 1, the skeletal plans correspond to the 
connected sequence of arrows marked with the bus labels. 
The skeletal plans are computed by the route finder using 
the knowledge base (map), and are input to the text planner. 
The text planner may augment the skeletal plans by 
selecting from the knowledge base additional descriptions 
of landmarks, location, orientation and so on. 
4.3 Knowledge Base and Route Finder 
We have described the design of the knowledge base and 
the route finder in \[26\]. For discussions of issues on the 
form, content, function and structure of spatial knowledge 
in the domain of route description generation, the reader is 
referred to Habel \[5\], and to experimental works such as 
those of McNamara \[17\] and Thorndyke \[31\]. 
As noted earlier, the cognitive (problem-solving) task of 
route finding and the linguistic task of description 
generation are simultaneously manifested in route 
communication. We do not make specific cognitive claims 
regarding how the route finding and description generation 
tasks are interleaved. In instances of human route 
communication, one may become aware of the final 
sections of the route even before the initial sections are 
realized into descriptions. However, a route section will 
have to be determined before it is realized. Accordingly, 
we separate route finding and route description generation, 
and feed the text planner with the sequence of skeletal 
plans (described in section 4.2) furnished by the route 
finder. 
The route finder is specialized for the description 
generation task, in the sense that its output units correspond 
to the discourse units (skeletal plans) of the descriptions. 
The sequence of skeletal plans processed by the text 
planner preserves the spatiotemporal connectivity of the 
journey described. The route finder thus contributes to the 
discourse structure, coherence and gross linearization of the 
route description. Therefore the text planner need not 
maintain any explicit paragraph-level representation for the 
overall text in its discourse model. 
The input to the route finder is a top-level goal 
represented in Prolog as in 
.9_ rfind(pats_house,sfu). 
It outputs the following skeletal plans in sequence. We 
use the Prolog list notation to indicate that the three 
elements of the skeletal plan have no positional 
significance in the representation. 
1. \[source(pats house),dest(lougheezlmall), 
rlabel(busid(lm 134))\] 
2.\[source(lougheeclmall),dest(sfu), 
rlabel(busid(sfu 145))\] 
4.4 Selection 
Input to the text planner is a sequence of skeletal plans 
generated by the route finder. The planner consists of two 
major submodules that correspond to two ordered stages of 
processing: (1) selector and (2) realization-specification (r- 
spec) synthesizer. The details of the r-spec synthesizer and 
other aspects of the route description generation system 
will be presented in \[27\]. The selector augments the 
skeletal plan with additional information retrieved (or 
computed) from the knowledge base (map), possibly under 
the control of various knowledge sources. The output of 
the selector is an expanded plan. The expanded plan is 
input to the r-spec synthesizer, which accommodates the 
information in the plan into one or more r-specs. It attends 
to such tasks as forming predicates, choosing the utterance 
type and certain language-oriented tasks like topicalization, 
choice of verb modality and tense. The r-specs are input to 
the realization module which generates surface sentences. 
The text planner thus deals with two distinct sorts of 
representations, one of which is closer to the domain 
(knowledge-base) and one, to the language. 
83 
Phibbs Eli 232 Grouse Mountain 
Narrows 
Bridge 
28 Phibbs Ex. t.-I 
D Grouse Mountain 
151/152 Vancouver 
K HASTINGS 
O 
O 
T 
E 
N 
A 
Y 
El Lougheed Mall 
Figure 
What kind of extra information is selected? The 
following are some of the kinds of information selected at 
this stage. 
• Direction or orientation, as in the 
descriptions On West Pender going 
westbound it's bus number 19. Walk north 
on Renfrew. 
• Distance information, for realizing such 
descriptions as You walk north on Burrard 
one block as far as West Pender. 
• Landmarks and location descriptions for 
sources and destinations of route sections, as 
included in the descriptions Get off at 
Hastings and Kootenay right by the 
Kootenay Loop. On Hastings in front of the 
loop transfer to bus number 28 called Phibbs 
Exchange. 
• Additional descriptions of route sections 
involving landmarks, as in Will cross Second 
Narrows Bridge to Phibbs Exchange on the 
other side. 
• Straightforward matters like retrieving names 
corresponding to route labels in the skeletal 
plan (bus number, bus name, etc). 
How is the information derived? Computation of 
direction (orientation) information requires availability of a 
co-ordinate system and additional inference modules. The 
text planner (selector) uses additional procedural domain 
knowledge for retrieving or computing orientation 
information. Information so computed is expressed 
linguistically as north, left, right, and so on, and 
incorporated into the text plan. The search and 
computation of orientation from a map using qualitative or 
1: FrornLougheed Mall to Grouse Mountain 
quantitative information forms a separate subject matter of 
study by itself. 
Distance is expressed in route descriptions in terms of 
various units, as, for example, in one block, 200 yards or 
three stoplights, but may not be represented in the map in 
the same terms. In our system we simplify the task of 
distance estimation/evaluation by directly representing 
distances in terms required by the text planner. 
Retrieving landmarks for describing locations of sources 
and destinations requires examining finer-grainer spatial 
layout information at the neighbourhood of these points. 
Salience plays a key role in the choice of landmarks chosen 
for location description. Along with the task of choosing 
the landmark comes the task of computing the locative 
relation between the two objects (for example, adjacence), 
and expressing the relation in language (as in right by). 
Salience also influences the choice of landmarks chosen 
for describing long route sections (as in will cross Second 
Narrows Bridge...). In our implementation we have 
represented certain landmarks along the bus route stages. 
Given the end-points of a route section, the landmarks 
along the way can be selected for description. It is to be 
noted that landmarks like the Second Narrows Bridge 
aren't quite points, but are idealized as points and 
represented in the knowledge base as such. Our 
implementation at present doesn't include the detailed 
representations and procedures necessary for deriving 
orientation, distance and location information. We sidestep 
the task by incorporating such information in the form 
required by the text planner. 
What is the rationale for selecting such information? 
The kinds of information described in this section are 
common in route descriptions, but are nevertheless 
secondary to the essential information contained in the 
skeletal plans. When a route section should be covered by a 
walk or a drive, some kind of distance information is 
required. Location information is crucial at transfer points 
(intermediate destinations). Additional descriptions of 
84 
routes in terms of landmarks crossed give the prospective 
traveller a feel for how long she needs to be on the bus, 
train, etc, and assure her that she is still on the right track. 
Orientation information is given in bus route descriptions 
when the listener has to take a bus after a spell of walk 
(self-navigation), as evident in the description 
...you want to walk north on Burrard one block as far as 
West Pender... and on West Pender going westbound it's 
bus number 19... 
But one cannot always attribute motives for including 
additional material in route descriptions. At present we 
include in the text plan all additional relevant material that 
can be extracted from the knowledge base. 
How is the selected information incorporated in the 
plan? Additional information selected at this stage is 
incorporated in a skeletal plan like 
\[source(lougheed mall),dest(hastings@ kootenay), 
flabel(busid(van 151 ))\] 
by adding to the list or by argument-adjunction and 
replacement. For example, the skeletal plan given above 
may be expanded as: 
\[source(lougheed_mall),dest(d(hastings@kootenay), 
loc(next(kootenay_loop))), 
rlabel(busid(busnumber(151),busname(vancouver))), 
via(brentwood mall)\] 
Discussion. We have used the skeletal plan as a substrate 
on which further information is deposited by the selector in 
the text planner. At the same time, the information 
available in the skeletal plan enables fast access to the 
information needed by the selector, by providing the route- 
label, source and destination names. This corresponds to 
our intuition regarding route description generation using 
visual search over maps: once we spot the end-points of a 
route section, we can easily spot the additional relevant 
objects around the end-points and over the route-section 
connecting them. The selector derives the information 
from the same knowledge base (map) consulted by the 
route finder. 
The text plan thus has a dual function: (1) for selection 
from the knowledge base, in its skeletal form and (2) for 
conversion into text eventually, in its expanded form. The 
expanded plan also has a third important function: a copy 
of the expanded plan is retained in the discourse record for 
consultation when the corresponding r-spec and sentences, 
and the next expanded plan are generated. It thus provides a 
record of the objects that have been introduced in the 
discourse, and forms the basis for decisions on 
pronominalization, referential expression generation and 
other factors that govern text coherence. 
It is to be noted that the domain of route description 
generation involves the use of a knowledge base which is 
primarily object-oriented. The knowledge base consists of 
points, landmarks, bus routes, streets and so on. The route 
finder works with the knowledge base and gives some 
objects the status of source, desanaaon and route-label in 
particular instances of route finding. The selector in the text 
planner picks up some more objects, endowing them with 
such attributes as landmarks, distance, direction. As the 
expanded plan stands now, it can just as well be expressed 
pictorially as a route-sketch. 
5. Work in Progress and Future Work 
We have examined the roles of salience and relevance as 
determinants of selection decisions, and presented their 
distinction mostly in intuitive terms. The notions should be 
formulated in more precise terms to be usable in specific 
models or implementations. We are continuing our 
investigation on ,salience and relevance by attempting to 
cast them in concrete terms in the domain we are currently 
examining, viz., route description generation. As in the 
case of Grice's maxims and many other pragmatic 
principles, the formulation of formal principles appears to 
be a difficult task. Herskovits \[7\] has encountered similar 
difficulties in attempting to formalize salience and 
relevance rigorously (in her case, for the purpose of 
preposition understanding). 
We are also continuing our research on the coupling 
between the thinker and the speaker. Examining the kinds 
and degrees of coupling will help us understand the nature 
of text planning tasks, design better text planners and 
compare various models of natural language generation. 
Our system is intended to generate bus route descriptions. 
However, there are other route descriptions that are more 
demanding in the kind of knowledge representation and 
processing that they require, and more complex in their 
syntax. The repreesentations for text plans and r-specs we 
use are also limited in variety and are tailored to the 
domain under examination. We are at present completing 
the implementation of the system. 
Route description generation is in itself a very complex 
process, and this research can proceed along several lines 
of further work: introduction of various knowledge sources 
to regulate selection from the knowledge base, 
representation of more detailed spatial knowledge at 
several levels, and so on. Upon completion of the current 
implementation effort, our primary plans, however, are to 
examine the implications of the system as a process model 
for speech-like monological multisentential text generation, 
refine its details and examine its applicability to other 
domains. 
Acknowledgments 
Our thanks are due to Ed Hovy, Fred Popowich and Dan 
Fass for help with some of the literature, and to the three 
anonymous reviewers for their comments. This work has 
been supported in part by a Simon Fraser University 
Graduate Research Fellowship and a British Columbia 
Advanced Systems Institute Graduate Student Scholarship. 

References 
I. Appelt, D.E. Planning English Sentences. Cambridge 
University Press, 1985. 
2. Conklin, EJ. and McDonald, D.D. Salience: The Key 
to the Selection Problem in Natural Language Generation. 
Proceedings of the 20th ACL Conference, Toronto, 1982, 
pp. 129-135. 
3. Dale, R. Generating Recipes: An Overview of 
EPICURE. The Second European Workshop on Natural 
Language Generation, Edinburgh, 1989, Extended 
abstract. 
4. Fass, D. Lexical Semantic Constraints. CSS/LCCR TR 
89-11, Centre for Systems Science, Simon Fraser 
University,1989. 
5. Habel, C. Cognitive Linguistics: The Processing of 
Spatial Concepts. Lilog-Report 45, IBM Deutschland 
GmbH, 1988. 
6. Haslett, BJ. Communication: Strategic Action in 
Context. Lawrence Erlbaum Associates, 1987. 
7. Herskovits, A. Language and Spatial Cognition: An 
Interdisciplinary Study of the Prepositions in English. 
Cambridge University Press, 1986. 
8. Hovy, E.H. Planning Coherent Multisentential Text. 
Proceedings,26th Conference of the ACL, Buffalo, 1988. 
9. Hovy, E.H. Unresolved Issues in Paragraph Planning. 
Selected Papers from the Second European Workshop on 
Language Generation, Edinburgh, 1989. Mellish, C. and 
Dale, R., Editors, to appear, 1990. Preprint obtained from 
the author. 
10. Kempen, G. and Hoenkamp, E. "An Incremental 
Procedural Grammar for Sentence Formulation". Cognitive 
Science 11 (1987), 201-258. 
U. Klein, W. Local Deixis in Route Directions. In 
Speech, Place and Action, Jarvella, RJ. and Klein, W., 
Editors, John Wiley & Sons, Ltd., 1982, pp. 161-182. 
12. Leech, G. Principles ofPragmatics. Longman, 
London, 1983. 
13. Levelt, W. J. M. Linearization in Describing Spatial 
Networks. In Processes, Beliefs, and Questions, Peters, 
S. and Saarinen, E., Editors, Reidel, Dordrecht, 1982, pp. 
199-220. 
14. Levelt, WJ.M. Speaking: From Intention to 
Articulation. M.I.T.Press, 1989. 
15. Mann W. C. and Moore, J. A. Computer as Author- 
Results and Prospects. IS!/RR-79-82, Information Sciences 
Institute,1980. 
16. McKeown, K.R. "Generating Goal-Oriented 
Explanations". International Journal of Expert Systems 1, 
4 (1988), 377-395. 
17. McNamara, T.P. "Mental Representation of Spatial 
Relations". Cognitive Psychology 18 (1986), 87-121. 
18. Moore, J.D. and Paris, C.L. Planning Text for 
Advisory Dialogues. Proceedings of the 27th Annual 
Meeting of the ACL, Vancouver, 1989, pp. 203-211. 
19. Moore, J.D. and Swartout, W.R. A Reactive Approach 
to Explanation. Presented at the Fourth International 
Workshop on Natural Language Generation, Preprint, 
1988. 
20. Ortony, A. "Beyond Literal Similarity". 
Psychological Review 86, 3 (1979), 161-179. 
21. Osgood, C.E. Where do sentences come from? In 
Semantics: An Interdisciplinary Reader in Philosophy, 
Linguistics and Psychology, Steinberg, D. D. and 
Jakobovits, L. A., Editors, Cambridge University Press, 
1971. 
22. Osgood, C.E. and Book, J.K. Salience and Sentencing: 
Some Production Principles. In Sentence Production: 
Developments in Research and Theory, Rosenberg, S., 
Editor, Erlbaum, Hilldale, NJ., 1977. 
23. Paris, C.L. Responding to a User's Domain 
Knowledge: Incorporating the User's Level of Expertise 
into a Generation System. Proceedings of the International 
Computer Science Conference, Hong Kong, 1988, pp. 
301-309. 
24. Paris, C.L. Generation and Explanation: Building an 
Explanation Facility for the Explainable Expert Systems 
Framework. Submitted to the Fourth International 
Workshop on Natural Language Generation, draft version, 
1988. 
25. Parisi, D. and Castelfranchi, C. A Goal Analysis of 
Some Pragmatic Aspects of Language. In Possibilities and 
Limitations of Pragmatics, Parrett, H., Sbisa, M. and 
Verscheuren, J., Editors, John Benjamins B.V., 
Amsterdam, 1981, pp. 551-567. 
26. Pattabhiraman, T. and Cercone, N. Representing and 
Using Protosemantic Information in Generating Bus Route 
Descriptions. In Knowledge Based Computer Systems, 
Ramani, S. et al., Editors, Narosa Publishing House, 1990, 
pp. 341-352. To appear in the Springer-Verlag LNCS 
Series. 
27. Pattabhiraman, T. and Cercone, N. Generation of 
Speech-Like Monologue: The Case of Bus Route 
Descriptions. CSS, Simon Fraser University, 1990. 
Unpublished manuscript. 
28. Schank, R. "Rules and Topics in Conversation". 
Cognitive Science I (1977), 421 A.A.A.. 
29. Sperber, D. and Wilson, D. Relevance: 
Communication and Cognition. Basil Bhackwell, 1986. 
30. Sridhar, S.N. Cognition and Sentence Production. 
Springer-Verlag, 1988. Springer Series in Language and 
Communication, 22. 
31. Thorndyke, P.W. "Distance Estimation from 
Cognitive Maps". Cognitive Psychology13 (1981), 
526-550. 
32. van Dijk, T.A. and Kintsch, W. Strategies of 
Discourse Comprehension. Academic Press, 1983. 
33. Waltz, D. Generating and Understanding Scene 
Descriptions. In Elements of Discourse Understanding, 
Joshi, A. et al., Editors, Cambridge University Press, 1982, 
pp. 266-282. 
34. Weiner, EJ. "A Knowledge Representation Approach 
to Understanding Metaphors". American Journal of 
Computational Linguistics 10 (1984), 1-14. 
35. Wunderlich, D. and Reinelt, R. How to Get There 
From Here. In Speech, Place and Action, Jarvella, R.J. and 
Klein, W., Editors, John Wiley & Sons, Ltd., 1982, pp. 
183-201. 
