The Multex generator and its environment: application 
and development 1 
Christian MATI'HIESSEN*, ZENG Licheng*, Marilyn CROSS**, Ichiro KOBAYASHI***, Kazuhiro 
TERUYA**** & WU Canzhong* 
*Macquarie University, Sydney, Australia; **DSTO, Canberra; ***Hosei University, Tokyo, Japan; 
***'~ University of New South Wales, Sydney 
1. Multex a multimodal and multilingual generation system 
Theaim of this paper is to report on generation-oriented research conducted by the Systemic Meaning 
Modelling Group based at Macquarie University, Sydney, and involving researchers-at other institutions in 
Australia, Germany and Japan. We will describe the core generation system Multex, which is a system for 
generating multilingual and multimodal presentations, based on the principles of systemic functional 
theory. We will also describe one of its application environments, the HINTS system. Together Multex 
and its "environment" constitute a long-term NLP research program based on systemic functional theory. 
Our work is thus relevant to current efforts in work on generation to link it as a component capability to a 
more comprehensive system (as in the work on MT in the Pangloss project, in the work on building a 
generation knowledge source in the PROFILE system \[Radev & McKeown, 1997\], in the work on 
generating multimodal healthcare briefings in the MAGIC system \[McKeown, Jordan & Allen, 1997\], or 
in the work on linking the KPML generator to an editor's workbench \[Bateman & Teich, 1995\]). 
2. Multex ,. a generation system in the Penman tradition 
2.1 Background: the Penman tradition 
Muitex is a generator in thePenman tradition. The Penman tradition started in 1980 with the initial 
development of the Penman generator (Mann & Matthiessen, 1985). The design of Penman is in many 
respects a systemic functional one (see Matthiessen & Bateman, 1991, for the systemic conception of text 
generation); that is, the system is based on a theoretical model of language developed within systemic 
functional linguistics (eg. Halliday 1994) and Multex shares this base with Penman, drawing on central 
categories that have been important in text generation such as that of the system network. Penman 
includes: 
(i) a systemic functional generation gr .a.mmar, the Nigel grammar, organized around system 
networks (Matthiessen 1983; cf. also Matthiessen, 1995), 
(ii) an interface between the grammar and higher-level components (e.g. the knowledge base and 
the text planner), the chooser & inquiry interface (Matthiessen & Bateman, 1991), 
(iii) a knowledge base that is organized under the "'Upper Model" (Matthiesse n & Bateman, 1991; 
cf. also Halliday & Matthiessen, in press) and 
I We gratefully acknowledge support for the research reported here from DSTO and the Australian Research 
Council. Multex is being developed by Zeng Licheng and SysAm by Wu Canzhong within the Systemic Meaning 
Modelling Group at Macquarie University. HINTS is being, developed at DSTO, Canberra, under the direction of 
Marilyn Cross. We would also like to thank Philip Collier for thorough comments of this paper. 
228 
(iv) a text planner that is based on RST (Rhetorical Structure Theory, Mann & Thompson, 1988; 
Hovy, 1993). 
The Penman system itself represents the first generation of systems in the Penman tradition; it has been 
distributed widely and is still being used at various research sites. It has been used for haany generation 
tasks, eg. generation in machine translation within the Pangloss project. Multex can be characterized as a 
third-generation system • within the Penman tradition. The most notable second-generation system is 
KPML (Komet-Penman MultiLingual), developed by John Bateman and his group, first at GMD/IPSI in 
Germany and now at Stirling University, Scotland (see Bateman, 1996). KPML includes Penman, but it 
goes considerably beyond the original Penman system, eg. in its modelling of multilinguality (cf. Bateman 
et al, 1991). In addition, KPML is equipped with a workbench interface for maintaining and developing its 
linguistic resources m one of the few or only such tools for generation systems. As a result, KPML is 
easier to use than Penman. 
2.2 Multex as a third-generation system in the Penman tradition 
In this section, we will outline some similarities and difference between Penman/KPML and Multex. 
Similarities. MulteX inherits major design themes from Penman, which we believe underlie the 
strength and longevity of Penman and the wide acceptance of KPML within the research community. 
These design features include: 
• grammar-centric generation. Instead of using a phrasal lexicon that directly maps concepts to 
predefined lexical-phrasal templates, both Penman/KPML and Multex have full knowledge of 
grammar, and build complex lexicogrammatical structures by "executing".the grammar. 
• clean modularity between linguistic resources and the processes that operate on the resources. 
Penman has an independent linguistic resource module which includes the Nigel grammar and the 
upper-model semantic framework. All the processes such as the text generation algorithm are defined 
based on the operations that are explainable by these linguistic resources. As we will show later, 
Multex not only inherits this design but also elevates the modularity between the linguistic resources 
and the NLP processes to a higher degree. 
• systemic functional theory. Penman, KPML and Multex are based on the same linguistic theory 
systemic functional theory. This compatibility makes it relatively easy for the three systems to share 
resources between them while each system may focus on some particular aspect of the theory. 
Differences. In terms of the system architecture and implementation, Multex differs considerably from 
Penman and KPML. This implementation is completely new. It is designed to be a small and practical 
content server to dynamically create multilingual and multimodal contents for different applications. To 
fulfil this design, Multex would need to interoperate with commercial applications, be object-oriented and 
be able to access mainstream computing resources eg. multimedia packages and databases. Multex, in its 
current form, can run as a standalone Java application , as a component in a CORBA environment, as a 
Web application and as a business object in a three tier client-server environment. 
In terms of the text generation technology, Multex •differs from the Penman/KPML tradition in the 
following aspects: 
• Multex is designed from the ground up to be a multilingual and multimodal content generator. At 
present it produces text in English and in limited Chinese and Japanese, It also dynamically constructs 
labelled maps, charts and tables. 
• Multex is discourse-oriented. That is, Multex generates on texts rather than only sentences basis as 
• Penman and KPML do. 
229 
• Multex has a higher degree of modularity in the organization of its linguistic resources. It has a self- 
contained linguistic engine, called the Meaning Base, which manages multilingual lexicogrammar and 
semantics, muitimodal resources and multiple domain models. The meaning base publishes an 
extensive Application Programming Interfaces (APIs), which NLP processes can use to access and to 
reason about the vast resources managed in the meaning base. 
• Multex uses a completely different generation algorithm from that of Penman/KPML. Chooser- 
inquiries and system network traversals are absent in Multex. Multex uses interstratal mapping 
patterns to efficiently draft semantic and lexicogrammatical plans, and it uses constraint-posting and 
plan-criticising methods to refine, reject and regenerate the drafted plans. 
• Multex uses a modular approach to text generation. The generation algorithm consists of a plar~ 
controller that controls the execution of a number of small content-creation agents, called Meaning 
Agents. Each Meaning Agent specialises in generating a particular kind of meaning and modality. 
Meaning agents for new specific modalities can be easily integrated into this general planning process. 
• Multex plans text in two modes: automatic and cooperative. In the cooperative mode, the generation 
process guides an end-user in their decision-making process, in a fashion similar to wizards in many 
Windows applications. 
2.3 An overview of the Multex architecture 
Figure 1 presents an overview of the Multex architecture. The role of the information producer is to 
creat e or manage data sources that supply useful data to the information consumer. The information 
consumer is an agent who needs data from the information producer but cannot process the data directly; 
the information consumer needs to consume the data in a processed form. Multex's function is to process 
the data into meaningful 
:'~- ~ information in the form 
of multilingual and 
multimodal contents. 
I Internally, Multex 
has two major 
components, a Meaning 
I1 Base and a set of NLP 
, processes. The Meaning 
Base contains the 
following resource 
modules: multiple 
linguistic systems (eg. 
the lexico-grammar and 
Meaning base(MB) j semantics of English, 
Multex Chinese, Japanese), 
Figure 1. Multex architecture multiple semiotic 
modalities (eg. the 
resources needed for creating charts and maps) , multiple domain models (eg. the knowledge about 
• tourism and aboutcommunicable disease). 
In addition, the Meaning Base supplies a full range of methods to populate itself with linguistic 
resources, as well as the methods to access and reason about the resources it maintains. An NLP process is 
an application or a service that performs some NLP functionality for the information consumer by drawing 
on the resources in the Meaning Base. Text generator is a NLP process of Multex, another partially 
implemented NLP process is a visual navigator of the Meaning Base. In future we can add a text 
understander to Multex as another "Multex-compliant" NLP process. 
230 
I 
I 
I 
I 
LI 
I 
I 
I 
3. The Meaning Base 
Recent advances in Systemic Functional theory attempts to define the entire systemic linguistic model 
in a relatively small set of theoretical concepts.. This set of theoretical concepts, known as the systemic 
metalanguage, outlines the structure of a linguistic system and provides principles and methodologies for 
modularizing the linguistic resources, for analysing and interpreting language instances with grammar, 
and for modelling linguistic processes such as understanding and generation (cf. Matthiessen & Nesbitt 
1996, Halliday & Matthiessen in press: Section 1.9, for the conception of the metalanguage). Figure 2 
shows a simple taxonomy of the systemic metalanguage. The notion of metalanguage can be usefully 
applied to NLP systems for two reasons: (1) it provides a comprehensive and theory-motivated map of the 
resources available in a linguistic system. Resource developers can use this map to structure and develop 
fragments of linguistic resources, and to reason about the properties of the linguistic resources; (2) 
linguistic processes can be defined with respect to the necessary resources it draws on. 
• ~ System 
Conslrueing Instantiati0 . Instantiation-p attern >F patterns 
t_. Realization-patterns 
Instance 
Figure 2. A taxonomy of the systemic metalanguage. 
The meaning base is the 
implementation of the systemic 
metalanguage plus the linguistic 
• resources maintained by Multex. It is 
the linguistic engine of Multex. The 
• metalanguage concepts are made 
operational by being implemented as Java classes. Access to linguistic resources and all reasoning about 
the linguistic resources are defined as methods in the classes representing the metalanguage concepts. A 
Java-based Meaning Base Application Programming Interface (MB API), which consists of around 60 
metalanguage concepts and over 400 methods, is available for programmers to create NLP processes. In a 
sense, the systemic metalanguage is the protocol a NLP process talks with the meaning base. 
Linguistic resources are specified in a formalism called the Meaning Base Modelling Language 
(MBML). When the meaning base is being loaded, the linguistic resources are compiled, optimised and 
stored as objects in the meaning base. Space does not permit us to provide the details of MBML, but we 
will present some examples. 
Fie idType { 
ID (communicable-disease-outbreak) 
isa (violent-social -event ) 
// define slots 
slot (disease :type disease) 
slot(cause :type animal) 
slot (range : type place) 
slot(cases :type human :unify-with victim) 
slot (fatality :type human :unify-with victim 
slot (medical-investigate : type investigate) 
slot (trend : type disease-outbreak-profile) 
ConstrueStrategy Brief-report() ( 
ideationObj { ID(?general-report) type(addition) 
slots(:nuclear ?report-incidence :satellite *.trend) } 
ideationObj { ID(?report-incidence) type (addition) 
slots(:nuclear ?report-outbreak :satellite *.medical-investigate) } 
) 
ConstrueStrategy Detailed-report (report-source, confidence-level) { 
..... // semantic structure for detailed report } 
// other construe strategies 
} 
Figure 3. Definition of the FieldType communicable-disease-outbreak. 
This example defines a domain concept called communicable-disease-outbreak. In addition to 
specifying ISA relations and slots, one can define any number of construe strategies for a domain concept. 
231 
A construe strategy is in fact a set of parameterized semantic objects that construe a given domain 
situation as meaning in a specific communicative context. Let's consider another example: 
RealizeAs { 
realizing(cause-effect) 
register(communicable-disease-report) 
Language(ENGLISH) 
// meaning pattern 
" - ..IdeationObj { ID(?causation) type(cause-effect) slots(:cause ?cause :effect ?effect) } 
IdeationObj { ID(?cause) type(figure) slots(:time-loc ?Time :space-loc ?Place) } 
// wording pattern . 
LexgrmrObj { 
ID(?relational) 
type(relational relational-cause) 
slots(:identified ?x :identifier ?y :time-loc ?PPI :space-loc ?PP2) 
map(?cause ?X) map(?effect ?y) map(?causation ?relational) 
map(?time ?ppl) map(?place ?pp2) 
) 
} 
Figure 4. •Definition of a Realization pattern for the concept cause-effect. 
• Figure 4 defines an interstratal mapping pattern. It encodes the following linguistic knowledge: for the 
register communicable disease report, the conjunctive relation cause-effect at the semantic level can be 
realized lexicogrammatically as !'X causes Y" in English, .the temporal and spatial circumstance of the 
cause event should be realized as the Time-loc and Space,loc functions of the "X causes F' clause. 
LexgrmrType { 
Language(English) 
ID(outbreak) 
ISA(nominalize-neutral-material) 
.FromAbove(break-out) 
slot(thing :value Sword("outbreak")) 
. slot(qualifier :map-from actor) } 
Figure 5. Definition of the lexicogrammatical 
type. outbreak. 
Figure 5 defines a lexicogrammatical system 
outbreak. The Fro~ove clause indicates •that this 
• system prototypically realizes the semantic concept 
break-out. The :map-~rom actor term specifies that the 
Qualifier function is mapped from the Actor slot of 
• the break-out event (i.e. what breaks out). 
The meaning base has many other important 
features, eg. management of multilingual and multimodal resources. In addition, each metalanguage 
concept is associated with a visualizer (although this is only partially implemented), which enables a 
meaning base visualization tool to be easily constructed. 
4. The Text Planner 
4.1 The text planning architecture 
The Multex text planner is structured into two layers~ the plan control layer and the meaning agent 
layerl The plan control layer implements a general-purpose constraint-based planner, and the meaning 
agent layer maintains a number of processes specializing in-creating certain kinds of meaning. This 
architecture is inspired by the work on meta-planning (Hayes-Roth & Hayes-Roth 1978, Stefik 1981). 
The plan control layer consists of the processes for: (1) creating goals to be solved by the meaning 
agents; (2) spawning, scheduling and starting meaning agents; (3) introspecting on the local plans 
generated by the.meaning agents. Plan introspection includes deciding when the planning should stop, 
assimilating local plans into a global plan, posting constraints entailed by sub-plans to the global plan. 
Herep/an andsub-plan refer to a set of partially specified linguistic objects generated by meaning agents. 
The meaning agent layer consists of a number of meaning agents. A meaning agent is a self-contained 
• process that creates a specific kind of meaning in the form of semantic and lexicogrammatical objects by 
instantiating resources in the meaning base. Table 4.1 summarises the meaning agents available in Multex 
as well the the meaning base resources they rely on. 
232 
I 
I 
I 
I 
I 
Table 4.1Meanin i1||~~ ~~]~ ~= 
• .................. 
VisualConstrue ' the same as Construe,.except.that it uses a The same as above. • 
~ " .. , :. ,visual:tool to inte~et~with:eodusers. ..... . , 
• " ~ 7" " • ........ ? ~" ...... .... 
: : . ; :~/:~'".*;~'.:,'?,. ;;:,;:~':-.:~: : '~... :~,'~G~)%~ ~`~s~ ;i" .G ~ ~, ~ +",-~';5~ ;'~,,'i ;!'~::~ 
2':~i "'2~.?~,' ~"~ 2:" :: '~ '-:~: ~'~ Y?L;;;~ " '- G~5;~::?. L!~, ':i" .'-:~ ~:" ; :'~-':.C~.~'..G 
~,' " :~:~#FleJd~ypes,~Const~e:~Strate~es~,~: 
In fact, the meaning agents Construe and Realize alone suffice for the whole text generation process, 
because from a theoretical point of view, text generation consists of exactly two steps: mapping contextual 
situation to meaning and mapping meaning to wording. The first step is carried out by the Construe agent 
and the second step, by the Realize agent. The rest of the meaning agents provide additional functionality 
that is designed for specific applications. 
Moreover, a meaning agent has to implement a protocol, called meaning agent protocol, in order to be 
administered by the plan control layer. The protocol includes methods for determining whether goal has 
achieved, for inferring more goals to achieve, for searching the meaning base for appropriate resources 
and for turning them into a plan, etc. 
4.2 An example of text planning 
Here we will give a brief example of text planning in Multex. The input to the text generator is 
provided as a meaning request, which is passed from the information consumer either as a stream, 
or as an object. Figure 6 shows a meaning request in the form of a stream. 
MeaningRequest { 
MeaningService(brief-rep0rt) 
register(OommunicableDiseaseReport) 
Topics(diarrhoea-outbreak) 
Template { 
ID(diarrhoea-outbreak) 
type(communicable-disease-outbreak) 
slots( :disease <ebola-disease non-specific> 
:cause <rat plural> 
:time-loc outbreak-time 
:space-loc outbreak-place 
:cases human-i 
..... ) 
) 
233 
template { ID(0utbreak-time) type(in-the-time) slot(anchor april-1995)} 
// Other specification of contextual objects. 
Figure 6. An example of meaning request. 
The space limit does not permit us to give a •detailed trace of the text planning process. We 
can only list some salient points in the generation process in Table 2, 
2ritical steos in the text 
.... ..,... 
..... :., ] create file first me~ing agent: Construe(diarrhdea:outb~:bfief-mpoi't0) 
• . =.. . ..... :: ...... • ..... . . 
{ 
addition 
} 
. 0 , , , , !! 
:?Report-outbreak: { cause-effect } VqHO-investigate: { investigate } ::: 
~?Disease-outbreak: { break-out } ?Cases-and-fatahty: { casualty-report }~i 
i. . ] Find the object 9cases-and-fatality to be a domain concept that needs to be construed 
ntrospect I . " ,, 
• I as meamn~ as weu. 
spawn 
spawn a Realize meaning agent for each semantic object in the semantic network. 
The text planner performs a topological sorting on the semantic network so that the 
less dependent nodes get realized first, eg. the decision for realizing ?Report- 
outbreak is made earlier than the decision for realizing ?Disease-outbreak and 
?cases=and-fatality " 
R ..... The realization pattern in Figure 4 is instantiated. An "Xcauses Y" clause is added to 
.e::t~e~^:~eP° , the partially generated text. 
- briar,) 
Multex finally generates the •following passage from the meaning request in Figure 6: 
"An outbreak of ebola disease, which was caused by rat, in Kikwit, Zaire has led to 189 cases 
and 59 deaths in April 1995. The world health organization investigated the disease on 10 May 
1995. Incidence of ebola disease increased in 1995." 
234 
Multex's generation is robust. For example, all the slots in Figure 6, except the disease slot, can be 
totally or partially omitted, and Multex can still produce coherent text. If all optional slots are missing, 
Multex generates the text there is an outbreak of disease. 
5. Multex working with production applications: HINTS 
Multex has been designed to be able to work together with other NLP systems in an integrated system 
capable of various "information processing" tasks in addition to generation. One such integrated system is 
the HINTS system currently being developed by DSTO, Canberra, with contributions by the Systemic 
Meaning Modelling • Group a t Macquarie University and by the team working on the Fact Extractor at 
DSTO, Adelaide. (HINTS may be compared with MAGIC, a system capable of generating multimodal 
healthcare briefings \[McKeown, Jordan & Allen, 1997\]; but whereas MAGIC is intended to produce 
multimodal briefings about particular patients for "time-pressured caregivers", HINTS is a resource for 
health officers who monitor communicable diseases around the world based on collected documents. 
HINTS is a system developed to process information concerning communicable diseases, it has been 
designed for health officers of various kinds to help them cope with the fast flow of information and the 
daunting demand for regular reports and briefings of various kinds. HINTS integrates a number of Systems 
that it can make demands on for different kinds of information processing services. From the point of view 
of Multex, HINTS constitutes an information production, for which Multex provides a service in the form 
of multimodal communicable disease reports. In addition, Multex provides a resource that is used by other 
components of the HINTS system -- the meaning base. 
Let us describe HINTS first in terms of the general work flow and then discuss its significance for the 
Multex generator. Users interact with HINTS through friendly GUIs; they have all been designed jointly 
with representative users. A user will start a HINTS session by retrieving documents according to a certain 
retrieval template -- at present, this is just a collection of key words. For example, the user might want to 
retrieve all documents that are concerned with (outbreaks of) Ebola in a certain region over a certain 
period of time. These documents are retrieved either from an existing collection of documents or from on- 
line sources via the Internet. 
Once the relevant documents have been retrieved, the user can ask HINTS for a summary. The 
summarizer that HINTS calls upon at present operates at fairly low levels of abstraction; it relies on 
aspects of the layout of a document (eg. the subject header of e-mail messages), on paragraph initial 
placement, on conjunctive markers such as in summary, and the like. it does not engage in any 
lexicogrammatical or semantic processing of the texts. 
The user can also ask HINTS to extract "facts" from the collection of documents. HINTS uses the Fact 
Extractor (FE) developed by Peter Wallis and his team (e.g., Wallis & Chase, 1997). FE operates with a 
set of templates for extracting information about communicable diseases. These templates include dates, 
locations, cases and disease outbreaks. They consist of slots or roles tha~ have to be filled by FE with 
values extracted from the collection of documents. They are all derived from Multex's meaning base and 
are represented within FE by means of regular expressions. Once FE has extracted the relevant values, it 
fills in "'forms" based on the templates and if the user wants to generate reports based on the information 
extracted, a meanin.g request is generated from the templates and passed over to Multex. 
Multex then construes the information in terms of its domain model of communicable diseases. Since 
the templates used by FE are derived from the Multex meaning base, all the information they provide Can 
be classified according to existing domain types. However, Multex will have to draw on domain 
knowledge to expand the information to the point where it can support generation. Once Multex has 
processed the information it receives from FE, it starts the incremental generation process sketched above. 
This will include not only decisions controlling Multex's generation process but also opportunities for the 
user to include quoted material from any of the documents that have been retrieved and to add his/her own 
235 
text. The latter option means that users can add information that embodies a fair amount of interpretation. 
In the register of communicable disease reports, this information has an interpersonal orientation: either it 
represents the user's expert evaluation of the information produced automatically by Multex ('how 
common?, how likely?') or it represents the user's recommendation (actions that should be taken by health 
authorities based on the information produced by Multex). This is a case where the prototypical Construe 
meaning agent is not •adequate, a VisualConstrue meaning agent is hence supplied to meet the additional 
demand of HINTS. When the user is satisfied that the generation process has finished, Multex produces a 
document in HTML format and it is handed over to a browser for display. 
AS this brief description indicates, HINTS is an interesting, information-rich environment for exploring 
multimodal generation. In particular, it is worth noting that Multex receives information that has been 
extracted from written• documents, but it produces presentations that may include charts and labelled 
maps. For example, Multex is able to retrieve a relevant map from the meaning base based on the spatial 
information in the meaning request and then label some hot spots on the map with the text fragments it 
produces. It is .also worth noting that in the HINTS environment, the •process of generation is very much a 
collaborative effort. The user exercises control over information sources and s/he can make decisions 
during the incremental generation process. This means that in this environment Multex functions as a 
writer's tool; and it can be used in preparing regular briefings or web pages. Further, although Multex's 
multilingual capability is not presently deployed in HINTS, Multex is able to generate the multimodal 
reports in languages other than English. For example, users could extract information from English 
documents and use Multex to generate• a multimodal report in Chinese. This capability can be of 
considerable value, as is demonstrated by the TREE project (Somers et al, 1997): the TREE system can 
search the Intemet for job ads and then summarize these in various languages. 
6. Conclusion 
In this paper, we have attempted to give a sense our "contextualized" approach to text generation. By 
contextualized, we mean that we are taking an ecological approach to text generation where it is located 
in relation to application environments and is linked to an environment of support tools. This move 
towards contextualized text generation capabilities resonates with the field in general and reflects the 
growing maturity of the "art" of text generation. Having discussed the two kinds of context -- application 
and development -- and having presented the architecture of Multex, we can now be more precise about 
the way in which they relate to the core Multex system than we were able to be in our introduction. Both 
application and development systems relate to resources within Multex; but at •present they relate to 
different levels within Multex. HINTS and similar systems relate to the highest levels of resources -- to 
context. 
Application relates to the highest level of organization in the first instance -- to context. As we have 
noted, this is the level where Multex can work with a complete model of its operational environment. 
Context is thus the appropriate level to capture generalizations about systems of different kinds -- systems 
suchas the Fact Extractor of HINTS or image interpretation systems in other environments. 
We have only been able to provide a brief sketch, leaving out many operational details. In conclusion 
we would just liketo add something about developments in the near future. We will add spoken output to 
Multex by linking it to a speech synthesizer; this work will be based on the flexible design of the 
expression level that we noted earlier and it will be guided by previous work by John Bateman and Elke 
Teich linking KPML to a speech synthesizer (Teich et al, 1997). We will also develop an interface system 
that is capable of translating KPML specifications into Multex specifications and vice versa. Further, we 
plan to integrate a development and reference workbench that we are building, SysAm (for Systemic 
Amanuensis) and Multex so as to provide a homogeneous generation and development environment of the 
kind currently offered by KPML. 
236 
I 
ii 
!1 
il 
ii 
ii 
ii 
il 
1 
! 
I 
I 
I 
I 
I 
I 
! 
! 

References 
Bateman, John A., Christian Matthiessen, Keizo 
Nanri & Licheng Zeng. 1991. The rapid 
prototyping of natural language generation 
components: an application of functional typology. 
Proceedings of the 12th international conference 
on artificial intelligence, Sydney, 24-30 August 
1991. Sydney. San Mateo, CA: Morgan Kaufman. 
Bateman, John A. & Elke Teich. 1995. Selective 
• information presentation in an integrated 
publication system: an application of genre-driven 
text generation. Information Processing and 
Management 31.5: 753-768. 
Bateman, John A. 1996. KPML Development 
Environment: multilingual linguistic resource 
development and sentence generation. Release 0.9, 
March 1996.• IPSI/ GMD, 15 Dolivostrasse, 
Darmstadt, Germany. 
Hall iday, M.A.K. 1994. An introduction to functional 
grammar. London: Edward Arnold. Second. 2nd 
revised edition• 
Halliday, M.A.K. & C. Matthiessen, in press. 
Construing experience through meaning: a 
language-based approach to cognition. London: 
Cassell. 
Hayes-Roth, B. and Hayes-Roth, F. (1978). "A 
Cognitive Model of Planning", In Allen, Hendler 
and Tate (eds) (1990). Readings in Planning. 245- 
262. San Mateo, CA: Morgan Kaufmann 
Publishers, Inc. 
Hovy, Eduard H. 1993. Automated Discourse 
Generation Using Discourse Structure Relations. 
Artificial Intelligence 63(1-2) Special Issue on 
Natural Language Processing (341-386). 
Kittredge, R. 1987. The significance of sub!anguage 
for automatic translation. S. Nirenburg (ed.), 
Machine translation: theoretical and 
methodological issues. Cambridge: Cambridge 
University Press. 
McKeown, Kathleen R., Desmond A. Jordan & Barry 
A. Allen. 1997. Language generation for 
multimedia healthcare briefings. In Proceedings of 
the Fifth Conference on Applied Natural Language 
Processing, 31 March - 3 April 1997, Washington 
Marriott Hotel, Washington, DC, USA. Association 
for Computational Linguistics. pp. 277-82. 
Mann, William C. & Christian• M.I.M. Matthiessen.. 
1985. Demonstration of the Nigel Text Generation 
Computer Program. Benson & Greaves (ed.), 
Systemic Functional Approaches to Discourse. 
Norwood: Ablex. 
Mann, William C. and Sandra A. Thompson. 1988. 
Rhetorical Structure Theory: Toward a Functional 
Theory of Text Organization. Text 8.3:243-281. 
Matthiessen, Christian M.I.M.- 1983. Systemic 
grammar in computation: the Nigel case. The First 
Annual Conference of the European Chapter of the 
Association for Computational Linguistics. Pisa. 
Matthiessen, Christian M.I.M. 1995. 
Lexicogrammatical cartography: English systems. 
Tokyo: International Language Sciences 
Publishers. 
Matthiessen, Christian M.I.M & Christopher Nesbitt. 
1996. On the idea of theory-neutral descriptions. 
Hasan, R., C. Cloran & D.G. Butt (eds.). 
Functional descriptions: theory in practice. 
Amsterdam: Benjamins.. 
Matthiessen, Christian M.I.M. & John A. Bateman. 
1991. Systemic linguistics and text generation: 
experiences from Japanese and English. London: 
Frances Pinter. 
Radev, Dragomir R. & Kathleen R. McKeown. 1997. 
Building a generation knowledge source using 
internet-accessibie newswire. In Proceedings of the 
Fifth Conference on Applied Natural Language 
Processing, 31 March - 3 April.1997, Washington 
Marriott Hotel, Washington, DC, USA. Association 
for Computational Linguistics. pp. 221-8. 
Somers, Harold, Bill Black, Joakim Nivre, TorbjOrn 
Lager, Annrosa Multari, Luca Gilardoni, Jeremy 
Ellman & Alex Rogers. 1997. Multilingual 
generation and summarization of job adverts: the 
TREE project. In Proceedings of the Fifth 
Conference on Applied Natural Language 
• Processing, 31 March - 3 April 1997, Washington 
Marriott Hotel, Washington, DC, USA. Association 
for Computational Linguistics. pp. 269-76. 
Stefik, M. (1981). "Planning and Meta-Planning 
(MOLGEN: Part 2)", Artificial Intelligence (16) 
141-170. 
Teich, Elke, E. Hagen, B. Grote & J.A. Bateman. 
1997. From communicative context to speech: 
integrating dialogue processing, speech production 
and natural language generation. Speech 
Communication 21.I~2: 73-99. 
Wallis, Peter & Greg Chase. 1997. An Information 
Extraction System• In 1997 Australasian Natural 
Language Processing Summer Workshop, 
Macquarie University, February 1997. 
