AUTOMATIC GENERATION OF SUBWAY DIRECTIONS: 
SALIENCE GRADATION AS A FACTOR FOR \ 
DETERMINING MESSAGE AND FORM 
Lidia Fraczak 
LIMSI-CNRS, BP 133 
F-91403 Orsay Cedex, France 
I idiaQl ims i. fr 
Guy Lapalme 
DIRO, Universit4.de Montr4al 
CP 6128, Succ Centre-Ville 
M0ntrdal Qudbec H3C 3J7 Canada 
lapalme@ iro. umontreal, ca 
" ' Michael Zock 
LIMSI-CNRS, BP 133 
F-91403 Orsay Cedex, France 
zock@l imsi. fr 
Abstract 
A frequently encountered problem in urban life is navigation. In order to get to some place 
we use private means or public transportation, and if we lack clear directions we tend to ask 
for help. We will deal in this paper with the descriptions of subway routes and their automatic 
generation. In particular, we will try to show how the relative importance of a given piece of 
information can effect not only the message but also the form. 
1 Introduction: the problem 
A frequently encountered •problem in urban life is the use of •public transportation: we have to 
get from here to there and quite so often we dori't know how. As it is not always possible to get 
help from a person (be it because nobody is available, or able to speak our language), we might 
appreciate assistance of a machine. 
In order to convey to the user "useful information", we must define what "usefulness" means. 
For example, if we •tell someone how to get from one place to another, we hardly ever specify all 
the intermediate steps, in particular if there is no ambiguity. Also, not all information is of equal 
weight. Yet, as we •will show ' the notion of "relative importance" of information is gradual in nature 
• rather than discrete , that is a simple binary Value (important vs. unimportant) 1. 
• All this reflects, of course, in the content and form of the final text. Relative importance is sig- 
naled by different means at the text level (headers, paragraphs, etc.) and at the sentence level (word 
choice, •syntactic structure: main clause versus subordinate clause, topic-comment •structures). 
• Concerning the prominence status (i.e. relative importance of a piece of information), semi- 
oticians and text linguists have reached a similar conclusion by distinguishing between the 'Yore- 
ground/background" or "primary/secondary level" of a text \[Bar66, vD77, AP89, Com92\]. Ac- 
cording to Combettes \[Com92\], the "primary level" deals with the core meaning, i.e. events and 
facts that make the text progress, while th e "secondary level" deaIs •with descriptions, evaluations, 
comments, and:reformulati0ns. " . -- ~: " i ~. : : : 
The distinction of levels, with information Of varying shades (salience gradation), implies that 
it should be possible to identify corresponding linguistic "markers" for each one of them. Yet, 
as Combettes has pointed out \[Com92\], the means used for marking the relative importance of 
information may vary •from one type of text to another. Nevertheless, certain markers do •hold 
lln this respect we cieviate from:most,. current generation systems. 
• " 58 . ! 
I 
I 
I 
i 
I 
I 
I 
1 
! 
! 
! 
! 
! 
regardless of the text type. This is particularly true for certain syntactic devices such as subordinate 
clauses, appositions, nominalization, all of which are, according to Combettes, markers of the 
secondary level, unlike main clauses which mark the primary level. 
Analyzing a corpus of route descriptions in French we have identified correlations between the 
salience status of specific conceptua ! chunks (landmarks, segment distance, etc.) and linguistic 
structures (independent vs. subordinate clauses). In section 2, we will reveal how the salience 
status of some types of information may affect the content and form of the final text. In section 3 
we will illustrate our use of these data in a generator of subway route descriptions. 
2 A case study: subway route descriptions 
Route descriptions are interesting for at least two reasons: first of all, as navigation aids in general 
they help to solve a real world problem; second, despite their apparent simplicity, especially with 
regard to surface form, they require the solution of a number of non trivial linguistic and discourse 
problems, problems which are intimately rooted in human cognition, 
Our analysis is based on a corpus containing 30 subway route descriptions in French. The data 
were collected from ten subjects via e-mail. Each one of them had to describe three routes in the 
Parisian subway. These routes differ in terms of length and complexity. The first route involves 9 
stops and one transfer. It is the longest. The second one contains 4 stops and no transfer. It is the 
simplest. The third route, though very short (4 stops), is the most complex one as it involves two 
transfers. 
2.1 Analysis of the underlying content 
The information contained in subway route descriptions can be divided into two broad categories: 
"global" and "local" information. We describe each one of them below, illustrating particular 
information types with examples taken from the corpus. 
Global information: 
• identification of the route by specifying departure and destination, 
eg. Po~r aller de Saint-Lazare 5 Jussieu... / To go from Saint-Lazare to Jussieu... 
• comments concerning the complexity of the whole route, 
eg. C'est simple et rapide, pas de changement. / It's simple and fast, no transfer. 
• information concerning the distance of the whole trip, 
eg. ~a doit faire 7 ou 8 stations en tout. / This should make 7 or 8 stops for the whole trip. 
Local information: 
• stop of departure, eg. ii partir de Jussieu, tu prends... / Starting from Jussieu, you take..." 
• destination, eg. tu arrives h Gate de Lyon / you arrive at Gate de Lyon 
o 
• lines to take, eg. prendre Ia ligne 5 / take the line number 5 
• transfers, eg. changer h Opdra / change at Opera 
• directions to take, eg. tu prends la direction Gallieni / you take the direction Gallieni 
• partial distances to cover, eg. il y a une seule station /there is only one stop 
59 ~' 
According to Wunderlich and Reinelt \[WR82\], "local information" is the core of route descrip- 
tions, while "global information" is additional as it serves mainly interactional purposes. • In the 
remainder of our analysis we will concentrate on the "local route information" and the way it is 
expressed in the domain of subway route descriptions, the objective being to determine whether 
some information is obligatory or not. Of course, we could have defined on a priori grounds what 
\ . • • should be mentioned explicitly and what not. Yet, we preferred to ground our work on empmcal 
data. 
We assume that "obligatory information" is information that is contained in all descriptions of 
the corpus, whereas "optional information" occurs only occasionally 2. We have also tried to find 
explanations for the omission of optional information. For example, the stations of departure and 
destination could be considered as optional, since they are already known by the "questioner" (either 
because they are a part of the question, or because they are given with the context/situation ). 
Indeed, our data reveal that, while the destination stop is always mentioned, the departure is 
mentioned only in 50%of the cases (eg./i Jussieu, tu prends... / At Jussieu, you take...). In the 
light of these data we conclude that it is useful to make a distinction between given and new, or 
known and unknown information. The problem concerning the "known" information is to decide 
whether to make it explicit or not. This is not a conceptual problem, - the known information must 
already be present at the conceptual level, - the choice is pragmatic in nature (what information 
should be conveyed, because it is really useful?), with possible stylistic side effects. For example, the 
fact that the destination (known information) is mentioned systematically in the corpus seems to be 
based on "stylistic" considerations: if it were not, the description would look like being incomplete. 
On the other hand.: decisions concerning "new" information do involve conceptual choices. They 
consist in determining whether to include a given piece of information in the message or not, and 
in determining its degree of salience. 
The rest of our paper deals only with the analysis of "new" information, since we are mainly 
interested in the choices at the Conceptual level and their consequences on the linguistic form. As 
the data show, information concerning transfer stations and directions of lines is obligatory! both 
types of information systematically occur in the corpus. The corpus also reveals that information 
concerning partial distances (number of stops to travel on a given line) and the names of the lines 
(eg. "line 7" or "orange line")is optional. 
It should'be noted, that partial distance may be represented in two ways in the domain of subway 
route descriptions: either as the length of a route segment (eg. "two stops"), or as the result of the 
number of stops counted (eg. "second stop"), This kind of information is not mentioned at all in 
30% of the cases. We have noticed that the inclusion/exclusion of information concerning partial 
distances depends on contextual factors such as the "value" of the distance itself (one stop vs. 
several) and the position on the route (last route segment or not). A "one-stop distance" is more 
important than a segment containing several stops. Also, the distance of the last segment seems 
to be more important than the distances of the intermediate segments (unless they are equal to 
one stop). Other strategies concerning information on partial distances have been observed: some 
subjects have mentioned all of them in each one of their description, regardless of the number of 
stops and the relative position of the segment, while others did not mention them at all. Another 
kind of optional information are the names of the lines to take. This may vary from place to place, 
but at least in Paris it is the direction (final destination) of the train that tells the user which 
train to take. The names of the lines, represented by numbers, were omitted in one third of the 
descriptions in the corpus. 
In the next section we describe the results of our linguistic analysis. We will show what spe- 
2Of course, this poses the problem of completeness and representativeness of the corpus. 
60 
I 
I 
I 
I 
I 
!1 
,I 
Li 
I 
I 
cific linguistic resources (independent clauses vs. sufiordinate structures) are used for expressing 
obligatory or optional parts of information. 
2.2 Correspondence between •conceptual saliency and linguistic resources 
It comes as no surprise that independent • clauses are the major syntactic structure used. Their 
function is to convey information of primary importance. Our analysis of the corpus Shows that 
independent clauses are mostly used in order to convey "obligatory" information, namely informa- 
tion specifying the names of the stations where to get off and directions to take. This is the case 
in example 1 below 3, where only these two chunks of information are contained in the independent 
clauses. 
Ex. 1 ,~ Saint-Lazare, prendre la direction GaIlieni. Descendre ~ Opdra (deux stations plus loin). Ps'endre 
alors la direction Mairie d'lvry/Villejuif jusqu'~ Jussieu (7-~me station). 
At Saint-Lazare, take the direction Gallieni (two stops ahead). Then take the direction Mairie d'lvry/Villejuif until 
Jussieu (Tth stop). 
However, independent clauses may also convey optional information. Ill this case, we consider 
it as a way of signaling prominence. For example, in our corpus there are cases where a "one-stop 
distance" (distance being optional information) is expressed by an independent clause: 
Ex. 2 A Bastille, prendre le m~tro n 1 direction Chhteau de Vincennes et descendre & la prochaine station 
qui est la Gate de Lyon. 
At Bastille, take the line number 1, direction Chateau de Vincennes, and get off at the next stop which is Gare 
de Lyon, 
Tile names of the lines (optional information), together with information concerning tile di- 
rection (obligatory information), are also quite frequently mentioned in independent clauses (see 
example 3). Again, we consider this as a sign for signaling high prominence: 
Ex. 3 A Saint-Lazare , prendre la ligne 3 direction Gallieni et changer h Opera. Prendre ensuite la iigne 7 
direction Mairie d'Ivry et descendre h Jussieu. 
At Saint kazare, take the line 3 direction Gallieni and change at Opera. Then, take the line 7 direction Mairie 
d'lvry and get off at Jussieu. 
Subordinate structures are generally used to convey optional information or information of minor 
importance. This is in our case information concerning partial distances and names of lines. In 
example 4 below, the information concerning partial distance is included only for the last segment, 
which is expressed by an "anaphoric clause". Example 5illustrates a strategy whereby prominence 
of the names of the lines is decreased: they are expressed in bracketed appositions. 
Ex. 4 A Saint-Lazare prendre le m~tro n 3 direction Gallieni, changer b. OPera et prendre le m~tro n 7 
direction Mairie d'Ivry/Villejuif et descendre ~t Jussieu (c'est la 7-~me station). 
At Saint-Lazare take the number 3, direction Gallienl, change at Opera and take the number 7 direction Mairie 
d'lvry/Villejuif anti get off at Jussieu (it's the 7th stop). 
Ex. 5 Prendre direction Gallieni (ligne 3). Sortir b. Opera (2 stations). Prendre direction Mairie d'Ivry 
(ligne 7). Descendre h Jussieu (7-~me station). 
Take direction Gallieni (line 3) and change at Opera (2 stops). Take the direction Mairie d'lvry (line 7) and get 
off at Jussieu (7th stop). 
aExamples from the corpus are followed by their English equivalents. 
• • 61 
We distinguish between two cases of subordinate structures: subordinate clauses and apposi- 
tions. The former include relative clauses (eg. "descends £ Opera qui est la 2-~me station" / "get 
off at Opera, which is the second stop") and anaphoric clauses (eg. "tu prends la direction Mairie 
d'Ivry, c'est la ligne 7" / "you take direction Mairie d'Ivry, it's the line 7"). We divide apposi- 
tions into nominal and prepositional appositions: Nominal appositions occur after an independent 
clause and may be used with various punctuation devices such as comma, colon, or brackets. In 
our corpus, they generally occur in brackets, for example; "descendre ~ Gare de Lyon (station 
suivante)" / "get of at Gare de Lyon (the following stop)". Prepositional appositions occur before 
.an independent clause. They are used to mention "known" information like "get-on stations" (the 
"departure station" or a "get-on station" that has been mentioned before as a "get-off" or "transfer 
station"), for example: "Descendre ~ Bastille. De Bastille, prendre.,."/"Get off at Bastille. From 
Bastille take...". 
: In order to be able to automatically generate route descriptions in line with these linguistic 
data, we have defined a set of rules that map the relative salience of a given piece of information 
onto one or several syntactic structures (cf: section 3 below, table 1 and table 2). 
3 A subway route description generator based on empirical data 
As we have shown, when people give directions the)' tend to use specific strategies for signaling 
the relative importance of a given piece of information. These strategies have been encoded in a 
program, written in GNU Emacs Lisp, that generates subway route descriptions. The generator is, 
at its present state,operational for two subway networks: Paris and Montreal. The examples used 
to illustrate our approach deal with a route in the subway of Montreal. 
The generator is composed of two main modules: a referential module and a discourse module. 
The former computes the fastest route between two stations, while the latter generates the corre- 
sponding text. The discourse module is divided into two Components. Th e first one divides the 
route given• by the referential module into conceptual clusters, whereas the second one plan's the 
textual structure of the description. 
The conceptual structuring consists in dividing a route into segments and landmarks \[Maa93\]. 
Landmarks are defined in terms of attribute-value pa!rs of relevant "places" and "paths" which, 
inour application, correspond respectively tO "stops" (stations) and to •"portions of subway lines 
in a given direction". The output of the conceptualizer takes the form Of a list of attribute-value 
pairs, representing landmarks for route segments and for stops. Examples of such representations 
and their corresponding texts are given in section 3.1. 
Information concerning partial distances is represented either as an attribute of path landmarks 
(dimension) or as an attribute of stop landmarks (order) • . The relative importance of this infor- 
mation is represented by the attributes degree-dim and degree-order. The names of the lines 
are represented by the path-landmark's •attribute name, and their importance is specified by the 
attribute degree-name. 
The fun ctiort creating the conceptual representation (the message) takes as input two "prag- 
matic" parameters. The first onespecifies the relative importance (with values ranging from 0 to 
3) of partial distances (represented by the attributes dimension and order), while the second one 
specifies the relative importance (possible values: 0=-2) of the names of the lines (attribute name). 
The values of the attributes degree-dim and degree-order, which express the "local importance" 
of partial distances, are computed on the basis of the value of the "global importance", as specified 
by the input parameters, and by looking at a given segment's distance (one stop vs. several) and 
checking the segment of the route currently processed (the last one or not). The obtained values of 
62 
! 
! 
"local importance" (i.e. values of the attributes degree-dim and degree-order) allow for making 
later on certain choices at the text level. Table 1 shows the rules for making these choices: the most 
important information (value 3) is expressed by an independent clause, the next highest (value 2) 
is expressed by a relative or anaphoric clause, while the least important information (value 1) is 
expressed by a bracketed apposition. 
\ 
Global import. Conditions of inclusion Local import: Textual realization 
0 never included none none 
1 if only one stop 1 apposition between brackets 
2 if only one stop 
if last route segment 
if only one stop 
if last route segment 
in all Other cases 
2 
1 
t 3 
2 
1 
subordinate or anaphoric clause 
apposition between brackets 
independent clause 
subordinate or anaphoric clause 
apposition between brackets 
Table 1: Correspondences between the saliency of partial distances and linguistic forms. 
The attribute degree-name, which represents the importance of•information concerning the 
names of the lines ("local importance"), takes the Value of the input parameter ("global impor- 
tance"). This value (between 0 and 2) determines whether and how to expressthe information in 
the final text. The rules presented in table 2 are simpler than the ones given for partial distances 
(table 1). This is so because the names of the lines are either included or not in a description, 
and this is valid for all the segments described. The relative importance of this information shows 
up at the text level via its linguistic form: independent clauses signaling higher prominence than 
appositions. 
Global import. Conditions of inclusion Local import. 
0 not included none 
1 included 1 
2 included 2 
Textual realization 
none 
apposition between brackets' 
independent clause 
Table 2:' Correspondences between the saliency of names of the lines and linguistic forms. 
The text module relies on schemata which, for a given conceptual input, specify the possible 
linguistic forms on a local and global level 4 It should be noted though that, even if the global and 
local choices depend fairly much on one another and on the conceptual input, there is still quite 
some freedom for "stylistic" variation: the way the schemata are defined allow us to generate a 
whole class of texts for a given conceptual input. 
3.1 Examples and analysis of results 
In this section, •~ve illustrate through automatically generated examples how the relative importance 
of partial distances and line names lead to quite different texts. 
The examples given below refer to the route from "Charlevoix" to "Acadie" in the subway .of 
Montreal (see figure 1). As the route contains two transfers (one at "Lionel-Groulx" and another one 
4The reason why we believe in the virtue of a "schema-driven approach" is based on the observation that people 
operate on larger "chunks" rather than on atomic units such as words or concepts \[Zoc96\]. 
53 
at "Snowdon"), we divide it into three segments, the distances of each segment being respectively 
one, four and five stops. 
Ligne : bleue 
Direction : Saint-Michel 
Ligne : orange 
Direction : C6te-Vertu 
C6te-des-Neiges 
Villa-Maria 
Vend6me 
• " - 
Place-Saint-Henri 
Ligne : verte 
Direction : Honor~-Beaugrand 
LIONEL-GROULX 
Charlevoix 
Figure 1" The route from "Charlevoix" to "Acadie" in the Montreal subway. 
If the "global importance" of partial distances and names of the lines is 0, no reference will be 
made to them, neither in the conceptual output represe,~tation, nor in the final text. This is the 
case in our first example.• 
Example 1. Parameters of "global importance": partial distances = 0, names of the lines 0. 
(((entity . line) (direction . "Honor~-Beaugrand")) 
((entity ., station) (name . "Lionel-Groulx")) 
((entity. line) (direction . "CSte-Vertu")) 
((entity . station) (name . "Sno"don")) 
((entity . line) (direction. "Saint-Michel")) 
((entity . station) (name . ,Acadie,))) " 
D'abord, Prendre la direction Honor~-Beaugrand et Changer ~ Lionel-Groulx. 
Ensuite, prendre la direction CSte-Yertu et changer & Snowdon. Enfin, prendre 
la direction Saint-Michel et descendre ~ Acadie. 
In the following two examples, the "global importance" concerning partial distances remains 
0 while the importance of the names of the lines changes respectively to 1 (example 2) and to 2 
(example 3). In consequence, the information concerning partial distances is neither represented 
nor expressed, while the information concerning the names of the lines, including their •"local 
importance" values (attributes degree-narae), is represented and expressed accordingly either by 
a subordinate structure (apposition) or by an independent clause. 
64 
I 
I 
I 
Example 
(((entity 
((entity 
((entity 
((entity 
((entity 
((entity 
2. Parameters of "globalimportance": partial distances = 0, names of the lines=i. 
line) (direction . "Honor~-Beaugrand") (name . "verte") (degree-name • 1)) 
station) (name . "Lionel-Groulx")) 
line) (direction . "CSte-Yertu °') (name . "orange") (degree-name . I)) 
station) (name . "Snowdon")) 
line) •(direction. "Saint-Michel") (name . "bleue") (degree-name . i)) 
station) (name . "Acadie"))) 
A Charlevoix, prendre la direction Honor4-Beaugrand (ligne verZe) et changer ~ Lionel- 
Groulx. A Lionel-Groulx, prendre la direction CSte-Vertu (ligne orange) e~s'arr~ter ~ 
Snowdon. Enfin, prendre la direction Saint-Michel (ligne bleue) et sortir ~ Acadie. 
Example 
(((entity 
((entity 
((entity 
((entity 
((entity 
((entity 
3. Parameters of "global importance": partial distances = 0, names of the lines = 2. 
.line) (direction . "Honor$-Beaugrand") (name . "verte'!) (degree-name . 2)) 
• station) (name . "Lionel-Groulx")) 
• line) (direction . "CSte-Vertu")(name . "orange") (degree-ham e • 2)) 
• station) (name . "Snowdon")) 
• line) (direction . "Saint-Michel") (name . "bleue") (degree-name . 211 
• station) (name . "Acadie"))) 
De Charlevoix, prendre la ligne verte en direction de Honor~:Beaugrand et changer 
Lionel-Groulx. A Lionel-Groulx, prendre la ligne orange en direction de C6te-Vertu 
jusqu'~ Snowdon. A Snowdon, changer pour prendre la ligne bleue en direction de 
Saint-Michel et descendre ~ Acadie. 
The following three examples illustrate how the relative importai~ce of partial distances may 
influence the surface form. (In order to produce more variations we have changed simultaneously 
the importance of the line names.) In example 4, the importance of partial distances is 1. Hence, 
according to the rules given in table 1, only the information concerning "one-stop distance" (first 
segment of the route here) is included (attributes dimension and degree-dim, and attributes order 
and degree-order) and expressed by" a bracketed apposition. 
Example 4, Parameters of"globalimportance': partial distances= 1, names ofthe lines = 2. 
(((entity . line) (direction . "Honor$-Beaugrand") (name . "verte") (degree-name . 2) 
(dimension . i) (degree-dim • i)) 
((entity . station) (name . "Lionel-Groulx") (order . i) (degree-order . I)1 
((entity . line) (direction . "C6te-Vertu") (name . "orange") (degree-name . 2)) 
((entity . station) (name . "Snowdon")) 
((entity . line) (direction . "Saint-Michel") (name . "bleue") (degree-name . 2)) 
((entity . station) (name . "Acadie"))) 
Charlevoix, prendre la ligne verte en direction de Honor@-Beaugrand et s'arr~ter 
Lionel-Groulx (station suivante). Puis, changer pour prendre la ligne orange en•direction 
de C6te-Vertu @t sortir ~ Snowdon. Enfin, prendre la ligne bleue en direction de 
Saint-Michel jusqu'& Acadie. 
65 
In the example 5 below, the "global importance" of partial distances is set to 2. This yields 
including the information concerning the "one-stop distance" and expressing it by a subordinate 
clause. This also yields including the information concern\]ng the distance of the last segment and 
expressing it in a bracketed apposition. 
Example 5. Parameters of "global importance": partial distances = 2, names of the lines = 2 
(((entity . +line) (direction . "Honor~-Beaugrand") (name . "verte") (degree-name \. 2) 
(dimension . 1) (degree-dim . 2)) 
((entity . station) (name . "Lionel-Groulx") (order . 1) (degree-order . 2)) 
((entity . line) (direction . "C5te-Vertu") (name . "orange") (degree-name. 2)) 
((entity . station) (name . "Snowdon")) 
((entity . line) (direction "Saint-Michel") (name . "bleue") (degree-name . 2) 
(dimension . 5) (degree,dim . 1)) 
((entity . station) (name . "Acadie")(order. 5)+(degree-order . I))) 
De Charlevoix, prendre laligne verte en direction de Honor@-Beaugrand et sortir & 
+ionel-GrouIx qui est la station suivante. ALionel-Groulx, prendre la ligne orange 
en direction de CGte-Vertu et s'arr~ter ~ Snowdon. Changerpour prendre la ligne bleue 
en direction de Saint-Michel et descendre ~ Acadie (5-~me station). 
In our last example, the importance of partial distances is set to its maximum value: 3, which 
requires including the relevant information for all three segments. The distances of the first segment 
(1 stop), the second one (4 stops), andthe last one (5 stops) are expressed respectively by an 
independent clause, a bracketed apposition and a relative clause. 
Example 6: partial distances = 3, names of the lines = 0 
(((entity . line)(direction . "Honor~-Beaugrand")(dimension . 1)(degree-dim .3)) 
((entity. 
((entity . 
((entity . 
((entity . 
((entity . 
station) (name . "Lionel-Groulx") (order . 1) (degree-order . 3)) 
line) (direction . "CSte-Vertu,) (dimension . 4) (degree-dim . 1)) 
station) (name . "Snowdon") (order . 4) (degree-order . 1)) 
line) (direction . "Saint-Michel") (dimension . 5) (degree-dim . 2)) 
station) (name . "Acadie") (order . S) (degree-order . 2))) 
partir de Charlevoix, prendre la direction Honor@-Beaugrand et changer ~ la station 
suivante qui est Lionel-Groulx. A Lionel-Groulx, prendre la direction CSte-Vertu et 
s'arrgter ~ Snowdon .(4-~me station). Changer pour prendre la direction Saint-Michel et 
descendre ~ Acadie qui est la 5-~me station. 
4 Conclusion and perspectives 
Analyzing a corpus of route descriptions we have found correlations between the relative importafice 
• of a given piece of information and its linguistic counterparts: according to its relative importance, 
an attribute is either expressed by an independent clause (salience high) or by a subordinate 
structure (salience low). We have also noticed that different subordinate devices vary in terms of 
their status of importance. For example, relative clauses seem to be used for signaling information. 
of higher prominence than bracketed appositions. 
We have applied the results of our analysis to a generator that can automatically produc e subway 
• route descriptions in French. By taking into account the relative importance of information, we 
are able to get a better :ontrol than otherwise over the use of linguistic resources observed in the 
66 
I 
I 
I 
I 
i 
il 
! 
corpus. This enables us to generate route descriptions which are in line with the ones produced by 
people. 
While our work is not the only one dealing with route descriptions (see, for example, \[Kle82, 
WR82, May90, Maa93, GL95\]), it is original in that it is, according to our knowledge, the only 
: attempt to show how the notion of "relative importance" of route information maps onto linguistic 
form. \ 
The results obtained so far are encouraging. Yet, they have to be evaluated i)y users, in order 
for us to see to what extent and with respect to what particular needs different versions of texts 
• produced by our generator are helpful for navigating in the subway. We will also investigate possible 
• extensions of the generator to other types of routes. 
Acknowledgment 
We would like to thank the Programme Franco-Qugbdcois de coopdration scientifique et technique, 
th&ne: "Inggnierie linguistique de la connaissance'for their financial Support, and Wojtek Fraczak 
for helping us during the implementation. 

References 
\[AP89\] J.-M. Adam and A. Petitjean. Le texte descriptif. Nathan, Paris, 1989. 
\[Bar66\] R. Barthes. Introduction gl'analyse structurale des r~cits. Communications, 8:1-27, 1966. 
\[Com92\] B. Combettes. L'organisation du texte. Didactique des textes. Universit~ de Metz, 1992. 
\[GL95\] A. Gryl and G. Ligozat. Route descriptions: a stratified approach. In IJCAI'95 Workshop on 
spatial and temporal reasoning, pages 57-64, Montreal, Canada, 1995. 
\[Kle82\] W. Klein. Local deixis in route directions. In R. J. Jarvella and W. Klein, ediiors, Speech, place, 
and action, pages 161-182. John Wiley & Sons, Ltd., 1982. 
\[Maa93\] W. Maat3. A cognitive model for the process ofmultimodal, incremental route descriptions. In 
• Spatial information theory, volume 716 of LNCS, pages 1-13. Springer-Verlag, 1993. 
\[May90\] M.-.Maybury. Planning multisententia! English text using communicative acts. PhD thesis, Univer- 
sity of Cambridge, 1990. 
\[vD77\] T.A. van Dijk. Text and context. Longman, 1977. 
\[WR82\] D. Wunderlich and R. Reinelt. How to get there from here. In R.J. Jarvella and W. Klein, editors, 
Speech, Place and Action, pages 183-201. John Wiley & Sons, Ltd. i 1982. 
\[Zoc96\] M. Zock. Sentence generation by pattern matching: the problem of syntactic choice. In R. Mitkov 
• and N. Nicolov, editors, Recent advances in Natural Language'Processing, pages 317-352. J. Ben- 
jamins, 1996. 
