Representing Knowledge for Planning Multisentential Text 
Jose Coch, Raphael David 
GSI-Erli / 1, place des Marseillais 
94227 Charenton-le-Pont Cedex, FRANCE 
e-maih jose.coch@erli.gsi.fr, raphael.david@erli.gsi.fr 
Abstract 
The AlethGen tool has been used to build a 
system for automatically generating replies to 
complaints, for a European mail-order company. In 
the current version, letters are generated in French. 
AlethGen consists of several modules: the text 
deep structure planner (or conceptual planner), the 
text surface structure planner (or rhetorical planner), 
the computation of pronouns and quantifiers, and the 
linguistic generator proper (globally inspired by the 
Meaning-Text Theory) which determines the 
concrete realization of sentences. 
The paper mainly describes AlethGen's 
conceptual and rhetorical planners and focuses on 
the knowledge they use and produce. 
1. Goals and general approach 
The main function of the texts we had to generate 
consisted indeed in replying to a complaint, thus in 
arguing and justifying a decision. This implies, on the 
one hand, a good management of the argumentative 
aspects, and on the other hand, the need to generate 
rather complex texts in a perfectly consistent style. 
The general approach of the conceptual planning is 
relation-based, as described in the Rhetorical Structure 
Theory or RST (Mann & Thompson, 1988). 
However, as shown in (Elhadad & Mc Keown, 1990), 
previous works in text generation systems (including 
RST) "have generally used a notion similar to rhetorical 
relations to describe the connection between 
propositions. They make a one-to-one mapping from 
these relations to connectives for generation (for 
example, the "opposition" relation would be realized by 
the connective "but"). In this approach it is difficult to 
distinguish between similar connectives (e.g. \[...\] but vs. 
although)." 
In our work, we assume that there are two different 
levels of text structure. The deep level is represented by 
unordered basic relations, and the surface level is 
represented by (ordered) lists of atomic events and 
rhetorical operators. 
2. Example 
Here is a letter generated by AlethGen and its 
translation in English: 
Cher Monsieur, 
J'ai bien re~u votre message minitel du 22 mars 
1993 concernant la commande du disque XXCC et du 
logiciel YYBB. Le disque a dtd exp~did le 3 f#vrier 1993 
par la Poste, il aurait donc dft parvenir it votre domicile. 
Je vais me renseigner pour savoir ce qui a pu se 
produire. 
Je ne peux pas renouveler cette commande car 
malheureusement, nous ne pouvons pas vous livrer avant 
le mariage. Dans votre intdr~t, je pr~f~re annuler cette 
commande. Elle ne fera l'objet d'aucun remboursement 
puisque vous aviez pr~vu de r~gler it la livraison. 
Je vous prie, Cher Monsieur, ... 
Dear Sir, 
I have received your Minitel message, dated 
March 22, 1993, concerning the order of an XXCC disc 
and a YYBB software. The disc was sent by post on 
February 3, 1993; you should therefore have received it 
by now. I will make inquiries to find out what may have 
happened. 
Unfortunately I cannot renew the order, as we 
are unable to deliver before the wedding. In your 
interest, it would be preferable that I cancel your order. 
It will not be refunded as you had decided to pay on 
delivery. 
Yours faithfully, 
In fact, the customer has sent a complaint letter 
because he has not received the items he had ordered. 
However, he maintains his order provided he receives the 
items before his wedding. The items that were sent are 
considered lost because of the dates. As the wedding is to 
be celebrated soon, it will not be possible to get the items 
before that date. The order must therefore be cancelled. 
The deep text structure is a list of two trees (in fact 
the introduction and the last phrases of the letters are 
computed from the input with a trivial system): 
• CONSEQUENCE ( 
BECAUSE ( delivery_parceh modality=past 
counterfactual, 
parcel_sent: lost), 
inquiry) 
• CONSEQUENCE ( 
condition_fulfillment_impossible, 
CONSEQUENCE ( 
AND (NOT renewal_of_the_order, 
cancellation), 
norefund)) 
The surface text structure is a list of events with 
rhetorical features and surface operators (note that the 
English (human) translation of the letter has a slightly 
different surface text structure). 
parcel_sent: 
203 
lost 
parcel_delivery: 
modality 1 =past counterfactual 
operator="donc" ("therefore/so") 
inquiry 
<new paragraph> 
non renewal_of_the_order 
"car" ("because/as") 
condition_fulfillment_impossible 
modality2=ne_pas_pouvoir (to be unable) 
modality l=malheureusement (unfortunately) 
cancellation 
modality 1 =pr6f'ere (better/preferable) 
no_refund 
3. Knowledge 
3.1. Rules for conceptual planning 
If p and q are events, one can have a rule (cause p 
consq q) in the knowledge base. If the input has p & q, 
one can generate something like "p caused q" or "q, 
because p". All the possible values are given in the 
following table: 
p q one can ~enerate 
T T "q because p" 
F F nothing (does not apply) 
F T nothing (does not apply) 
T F ""<1 though p" 
The last line is the "exception" to the rule. 
Rules can be more complex, and have a condition in 
addition to the cause: (cause p cond r consq q). For 
example, 
cause: addressee_wants cancel (his order) 
cond: not_delivered (the articles have not been...) 
consq: cancellation (of the order) 
Clearly, not_delivered is not a cause for cancellation 
but only a condition, because the reason for cancellation 
is addressee_wants_cancel. 
p r q one can ~enerate 
T T T "q because p" 
T F T "q because p though -r" 
F T/F T/F nothing (does not apply) 
T T F "~q though p" 
T F F "~q though p because -r" 
So the knowledge for planning the deep structure of 
the text is a set of domain-dependent conceptual rules. 
3.2. Rhetorical rules 
Given a simple relational tree consisting of an 
operator as well as two events, it is easy to see that 
several surface orders and several surface operators are 
possible. For instance, the tree BECAUSE (You ask me 
for X, I see to it so that X) may have the following 
linguistic realizations: 
- I see to it so that X, because you asked me for Y 
- You asked me for Y, so I see to it so that X. 
The rhetorical module chooses concrete operators, 
modalities, and order according to certain attributes, e.g. 
the fact that the addressee is aware of an event or not, the 
fact that it is in favour of the addressee or not, etc. 
4. Implementation 
The system was written in C++ under Unix and 
effectively runs on RS/6000 and Sun SparcStation. 
In the current version, there are 130 conceptual rules 
for 50 models of events, and 75 rhetorical rules. 
In any case, the system generates a whole letter on a 
complex problem (15-20 sentences) in less than 2,5 
seconds. 
5. Conclusion 
The distinction between deep and surface structure of 
the text allows the system to generate relevant letters 
from an argumentative point of view. 
Besides, letters may deal with a rather large number 
of various topics (payment of lost or damaged items, 
etc.). Our concern was therefore to factor the information 
as much as possible, in other words we wanted to make 
certain pieces of knowledge portable. For that purpose, it 
is interesting here to remind the distinction between 
domain knowledge and domain communicative 
knowledge (DCK) proposed in (Kittredge and al., 1991). 
For example a police crime report and a detective novel 
can use the same domain knowledge but the DCK is 
quite different. We think that it is preferable to call this 
knowledge "Style Dependent Knowledge", because two 
detective novels in two different "domains" (for example 
a political murder and a crime of passion) could have the 
same style. 
In our system, conceptual rules are clearly domain- 
dependent. The communicative knowledge used by the 
rhetoric module to compute the surface structure of the 
text is style-dependent but it is domain-portable. 

References 

Cerbah Farid (1992). "Integrating qualitative reasoning 
and text planning to generate causal explanations" in 
Proceedings of 1992 International Computational 
Linguistics Conference. 

Coch Jos4, and David RaphaSl (1994). "Une application 
de g6n6ration de textes" in Actes de TALN'94. France. 

Elhadad Michael, and Mc Keown, Kathleen. R. (1990). 
"Generating connectives" in Proceedings of 1990 
International Computational Linguistics Conference. 

Elhadad Michael (1992). "Generating coherent 
argumentative paragraphs" in Proceedings of 1992 
International Computational Linguistics Conference. 

Kittredge Richard, Korelsky Tanya, and Rainbow Owen 
(1991) "On the need for domain communication 
knowledge" in Computational Intelligence V7, N4, 
Canada. 

Mann William C. and Thompson Sandra. A. (1988). 
"Rhetorical Structure Theory: Towards a functional 
theory of text organization" in Text 8(3). 
