TECHDOC: Multilingual generation of 
online and offline instructional text 
Dietmar RSsner 
FAW Ulm 
P.O. Box 2060, 89010 Ulm, Germany 
roesner@faw.uni-ulm.de 
Manfred Stede 
(University of Toronto and) FAW Ulm 
P.O. Box 2060, 89010 Ulm, Germany 
stede@faw.uni-ulm.de 
Project idea 
Supplying technical documentation accompanying a 
product in multiple languages is a growing prob- 
lem, particularly in Europe with its legislation on 
the common market. A huge amount of translation 
work needs to be done when creating and updat- 
ing technical documentation. In response to these 
needs, the TECHDOC project aims at supporting 
the creation and maintenance of technical documen- 
tation by knowledge-based, multilingual generation. 
The idea is to build up a knowledge base that in- 
cludes a model of the product in question, and to 
produce documentation in multiple languages auto- 
matically. At present, the system produces mainte- 
nance instructions in English, German, and French. 
The application domain of TECHDOC is techni- 
cal manuals, where one has to do with "real-world" 
texts: the domain is practical instead of a "toy 
world". At the same time, the language that is used 
in such manuals tends to be relatively simple; one 
mostly finds straightforward instructions that have 
been written with the intention to produce text that 
can be readily understood by a person who is ex- 
ecuting some maintenance activity. Moreover, the 
structure of manual sections is largely uniform and 
amenable to formalization. 
System architecture 
The system is based on a KB encoding knowledge 
about the technical domain and about schematic 
text structure in LOOM, a KL-ONE dialect (LOOM 
1991). A typical manual section first describes the 
location of the object to be repaired/maintained and 
then lists possible replacement parts/substances; 
next, the activities are described, which fall into 
the three general categories of checking some at- 
tribute (e.g., a fluid level), adding a substance and 
replacing a part/substance. These actions are repre- 
sented as plans in the traditional AI sense, i.e. with 
pre- and postconditions, and with recursive struc- 
ture (steps can be elaborated through complete re- 
finement plans). We call the schematic ordering of 
a manual section its macrostructure. 
For a language-independent text representation, a 
level was sought that captures the commonalities of 
the correponding sections of the German, English 
and French texts, i.e. that is not tailored towards one 
of the specific languages. Rhetorical Structure The- 
ory (RST) (Mann and Thompson 1987) turned out 
to be a useful formalism: for almost every section we 
investigated in our extensive corpus studies of mul- 
tilingual manuals, the RST analyses for the different 
language versions were identical. The system imple- 
ments this finding by mapping the schematic plan 
structure to a tree that captures the microstructure 
of documents by means of rhetorical relations hold- 
ing among elementary propositions or sub-trees, and 
a number of specific annotations. See (RSsner and 
Stede 1992) for details. 
This document representation is successively 
transformed into a sequence of sentence plans (to- 
gether with formatting instructions in a selectable 
target format; SGML, ISTEX, Zmacs and - for screen 
output - formatted ASCII are currently supported), 
which are handed over to sentence generators. For 
English, we use Penman (Penman 1991) and it s 
sentence planning language (SPL) as input terms. 
To produce German text, we have implemented 
a (partial) German version of Penman's grammar 
(NIGEL), which is enhanced by a morphology mod- 
ule, and a fragment of a French grammar in the same 
style. 
Interactive, multimodal instructions 
The interactive variant of the system, TECHDOC-I 
(Peter and R6sner 1994), produces instructions on- 
line, by engaging in a dialog with the user. On 
the basis of initial questions, users are assigned to 
a stereotypical class (ranging from 'novice' to 'ex- 
pert'), which subsequently influences the degree of 
detail in which instructions are produced. 
The generated texts on the screen are mouse- 
sensitive, so that the user can click on words or 
phrases (in any of the languages displayed) and re- 
quest additional information. Depending on the unit 
selected, a menu offers a range of applicable items 
for further inquiries. For example, the location of 
an object referred to in the text can be clarified by 
209 
an image in which that part is highlighted. Or, if an 
action is selected, the system offers to show a short 
video sequence on the screen, which illustrates how 
that action is to be performed. 
In essence, these facilities have paved the way to 
move from static, inactive strings as output to an ac- 
tive and dynamic interface for the associated knowl- 
edge sources and their various presentation modal- 
ities. The key is that all kinds of information (lex- 
emes in various languages, images and object's lo- 
cation therein, and video sequences) are associated 
with the underlying KB objects. 
Document authoring support 
Apart from supporting the end user with an inter- 
active system, the system is now being enhanced on 
the "opposite end", where the input to the docu- 
ment generation process is created. A first version 
of an authoring tool has been designed and imple- 
mented and tested with a number of users. This tool 
allows to interactively build up knowledge base in- 
stances of maintenance plans, including the actions 
and objects involved, and to convert them imme- 
diately into multilingual documents. The process 
is menu-driven, and the available options are deter- 
mined dynamically in accordance with the current 
state of the knowledge base -- in other words, the 
knowledge already encoded in the KB constrains the 
range of possible instances that are to be built with 
the authoring tool. 
Portable KB: the "middle model" 
Our KB design focuses on the targets that the rep- 
resentations can be used for more than language 
generation (tutoring, diagnosis systems, etc.) and 
that they should be portable across different techni- 
cal domains. We have collected recurring knowledge 
about various technical devices like types of connec- 
tions, about switches, about objects that turn up in 
a variety of designs and functions, for example fluid 
tanks, etc., in a middle modelthat can be carried, to- 
gether with the basic ontology (the so-called 'upper 
model' (Bateman 1990)), from one technical domain 
to the next. As the first practical domain extension, 
we moved from automobile manuals to several sec- 
tions of aircraft maintenance instructions, where the 
model of the specific objects and devices was sub- 
sumed under the middle model, in the place where 
the automobile model previously resided. 
The middle model is a set of quite abstract con- 
cepts; separating lexical from conceptual informa- 
tion allows us to account for lexical differences be- 
tween the languages; in our corpus studies of au- 
tomobile manuals, we found, for instance, that En- 
glish often uses the general remove where German 
verbs characterize the physical action more specif- 
ically (abziehen, abnehmen, herausschrauben, her- 
ausziehen, etc.). Our lexicalization procedure cap- 
tures these language-specific differences by exploring 
subsumption (see (Stede 1993)); the important point 
to note here is that such lexical matters are indepen- 
dent of the "conceptual kernel", which is represented 
in a language-independent way and allows for shar- 
ing as much information as possible. 
Perspective: From research prototype to 
practical usage 
TECHDOC is being developed into two complemen- 
tary directions: toward online, interactive instruc- 
tion dialog on the one hand, and toward production 
of fully formatted paper documents on the other. 
Each is suitable for its particular purposes; both are 
important. Online systems, for example, will make 
sense in settings where the system can be given sen- 
sory data from the device (temperature, pressure 
values, etc.) and thus exactly those instructions can 
be generated that are relevant in the current situa- 
tion. Paper documents, on the other hand, will still 
be used for getting familar with a product prior to 
using it, as well as for reference purposes -- they 
won't vanish altogether. 

References 

John A. Bateman. Upper modeling: A level of semantics for natural language processing. In Proc. 
of the 5th International WS on Natural Language 
Generation, Pittsburgh, PA, 1990. 

The LOOM Knowledge Representation System. 
Documentation package, USC/ISI, Marina Del 
Rey, CA, 1991. 

William C. Mann and Sandra A. Thompson. 
Rhetorical structure theory: A theory of text or- 
ganization. In L.Polanyi, editor, The Structure of 
Discourse. Ablex, Norwood, NJ, 1987. 

The Penman documentation package. USC/ISI, Ma- 
rina Del Rey, CA, 1991. 

Gerhard Peter and Dietmar Rosner. User-model-driven generation of instructions. To appear 
in User Modeling and User-Adapted Interaction, 
1994. 

Dietmar Rosner and Manfred Stede. Customizing 
RST for the automatic production of technical 
manuals. In R. Dale, E. Hovy, D. RSsner, and 
O. Stock, editors, Aspects of Automated Natural Language Generation - Proc. of the 6th International WS on Natural Language Generation. 
Springer, Berlin/Heidelberg, 1992. 

Manfred Stede. Lexical options in multilingual gen- 
eration from a knowledge base. In: Working notes 
of the 4th European Workshop on Natural Lan- 
guage Generation.
