ON MOVING ON ON ONTOLOGIES: 
MASS, COUNT AND LONG THIN THINGS 
Robin P. Fawcett 
Computational Linguistics Unit, University of Wales, Cardiff CF1 3EU, UK 
e-mall: fawcett @cardiff.ac.uk 
Abstract: This paper discusses the principles that should govern the construction of two components of a system for natural 
language generation (NLG): (1) the ontology - or, rather, as the paper argues, the 'ontological' aspects of a belief system -and (2) 
the semantic representation of noun senses. It is an interesting fact that many ontologies bear a striking resemblance to a system 
network, as used in systemic functional grammar (SFG). Furthermore, two major current research efforts in the field of ontology- 
building are designed to run with a SFG generator: Pangloss, where the generator is Penman, and COMMUNAL, where the generator 
is GENESYS. It is therefore important to establish a principled approach to the 'division of labour' between the ontology and the 
equivalent aspects of the model of language - here a system network for the 'meaning potential' of English nouns. (However, the 
general principles should be relevant to ANY model of language.) The paper summafises (a)the purposes and (b)the structure of (1) 
a system network for noun senses and (2) the equivalent ontology (based on what we in the COMMUNAL Project judge is required in 
the next generation of belief systems for NLG). Examples are given of current work on the relevant system network and, more 
briefly, of the equivalent ontological aspects of the belief system. In particular, reasons are given why it would be inappropriate to 
give a primary place to the 'mass' vs. 'count' distinction in an 'interlingua' ontology - and even, surprising though it may seem, in a 
language-specific semantics for English. Finally, it turns out that, in the new perspective presented here, there is no 'component' of 
the belief system that is 'the ontology', and the reasons for this apparently anomalous position are given. 
Keywords: ontology, system network, belief system, knowledge base, semantics, noun senses, natural language generation 
1 Some current issues in modelling 'ontologies' 
'Or" I One of the =lvens of Computational Linguistics (CL) - 
which is taken here in a sense that includes Machine 
Translation (MT) - is that any such system needs an 
ontology, l But what, precisely, is an ontology? It seems 
to be one of those concepts which everyone who works 
with it instinctively feels tfiey understand, so that the basic 
assumptions are seldom made explicit. In practical terms, 
there is a fairly general assumption that an ontology is 
closely related to, and perhaps isomorphic with, the 
'meanings' of the nouns of a language - or of a set of 
languages, the maximal set being all human languages. But 
in buildin,q a theoretically satisfactory overall model one 
discovers tI~at there are serious problems with this position, 
as will be shown in this paper. 
There is a long history of work on ontolo,~ies for CL, 
including/he important work at Carnegie-Meffbn Univers- 
ity over many years, that of Dahlgren and her collea,~ues 
(e.g. Dahlgren 1988) and the current Pangloss ProjecT, as 
described in Hovy and Nirenburg 1992, Hovy and Knight 
1993 and Knight 1993. We in the COMMUNAL Project 
have been considering alternative approaches to this aspect 
of what we term the belief system, and I would like to 
present here, for discussion by the wider NLG community, 
the principles that we have established, often after years of 
experimentation, as they relate to these matters. We are 
currently implementing a system based on these principles. 
In many respects, of course, our assumptions are 
similar to those of others working in this area. But our 
view is that the next generation of systems in Artificial 
Intelligence (AI) - and possibly also in MT - will require, as 
central components, belief systems that represent 
knowledge (or, more accurately, beliefs) of more types 
and in a more complex manner than in some current 
systems. As we shall see by the end of this paper, the 
phenomena that are often handled in terms of an ontolo~ov 
look somewhat different in this new perspective. In relati~ia 
to some of the issues to be discussed here, then, we are 
constructing a different overall model from that which 
appears to underlie much other current work. The purpose 
of this paper is to set out these ideas, to compare them with 
those of other researchers on whose work we are seeking to 
build, and to give some explanation of why we are 
following the direction that we are. We hope that this will 
open up further discussion about the next generation of 
belief systems. . . 
The first step is to be clear about what the issues are. 
They are (1) issues of levels (which we shall here assume to 
mean levels of language), (2) issues of components 
of the overall system, and (3) issues of the structure and 
contents of part of the largest component of the overall 
system, namely the belief system. We shall focuspartic- 
ularly on the types of relations that need to be recogmsed as 
holding between the 'concepts' in an ontology or rather, 
between the gene,ric, objects (in contrast with specific 
objects), such as dog, realized as dogs, as in I like dogs.2 
2 Why the discussion remains open 
Hovy and Nirenburg (1992), in clearing the ground for 
their discussion of the principles that should guide the 
construction of an ontology, suggest that 
most ontologies and domain models to date have been 
assembled based primarily on introspection, and often 
reflect the idiosyncrasies of the builders more than the 
requirements of the application (such as MT). Lacking 
well-founded guided principles, the ontology builder is 
working in the dark. 
This judgement seems a little hyperbolic, in view of the 
fact that a number of recent ontology-builders have 
explained their principles as fully as Hovy and Nirenburg 
do. Thus, while there are aspects of Dahlgren s frameworI( 
(1988:46f.) that are open to criticism, she in fact gives as 
detailed an account of her principles as do Hovy and 
Nirenburg. Bateman (1990, Bate, man et al 19,90) similarly 
explains the ideas underlyin,q the uover model in Penman. 
In fact, as Hovy and KnigSht (19~) state, the 'ontology 
base' in Pangloss is in part based on Penman (being a 
merger of this, the semantic categories from the Longman Dictionary of Contemporary English 
(LDOCE), which is 
intended as a taxonomy for the nouns of English), and 
ULTRA (Nirenburg and Defrise 1992)- which itself draws 
on the LDOCE categories). Nonetheless the main thrust of 
71 
7th International Generation Workshop • Kennebunkport, Maine ° June 21-24, 1994 
Hovy and Nirenberg's comment is surely right, in that the 
discussion of the principles governing the building 
onto!ogies is far from over. Indeed, it may be that the 
reqmrements of MT and AI (where NL is used to 
communicate with a problem solving system) are different 
(at least at this stage of the development of MT). 
Most - and perhaps all - ontology-builders accept that 
they are in fact working on the basis of natural language- 
and, typically, that thelanguage is English. In Bateman's 
words (1990:56) 'the Upper Model represents the speaker's 
,e, xpenence in terms of ,generalized lingu!sfically-motivated 
ontological" categories. He states that it is essential that 
constraints should be found for what an upper model should 
contain and how it should be organised \[his italics\], and he 
then goes on to suggest that it is the aspects of meaning 
found in the experiential meta-function in a systemic 
functional grammar that should guide the construction of 
the ontology. This application of SFG concepts provides a 
helpful framework in which to approach the problem. (We 
might note that distinctions that depend on register, e.g. on 
tenor (formality), such as that between cigarette and fag are 
irrelevant to the ontology.) The reason for depending on 
NL, then, seems to be that, without it, we would have no 
guidelines as to how to structure the ontology. 
Let us assume for the moment that this position is 
justified (though I shall suggest some problems with it in 
sections 5 and 6). This brings us to an important question 
for those who use the Upper Model or any derived ontology 
such as Pangloss. This is: if there, is to be this strong 
connection to the linguistic system (Bateman 1990:57), 
what part of it should that connection be to? 
And here we come to an apparentproblem for this line 
of argument. This is that the Nigel- model of language 
around which Penman is constructed lacks a specific 
network for noun senses. So what are the principles on 
which these aspects of the Upper Model are constructed? We 
shall return to this matter after the next section. 
3 Two extraneous factors that may lead to 
differing research assumptions 
First, however, let us be explicit about two of the 
factors that differentiate various research projects in this 
area, and so, perhaps, the assumptions underlying various 
conceptual frameworks that have been adopted. The first is 
the disciplinary coverage of ,researchers on a project - 
including the 'home discipline, as it were, of the leading 
researcher. Those such as myself whose starting point was 
linguistics are sometimes shocked at the cavalier way (note 
the loaded language to express the viewpoint!) in which non- 
linguists simply adopt the 'senses' of the nouns of English 
as the starting point for an ontology. AI-minded computer 
scientists may be equally shocked at the pussyfooting way 
in which many linguists refuse to recognise the need to 
move outside the semiotic system of language, even though 
this is patently necessary in order to budd adequate models 
of how language is used. Often neither really addresses the 
issue of where the semiotic system of a natural language 
ends (i.e. the semantics of the language) and where the 
',concepts,' (or whatever is assumed as the category for 
thinking ) begins. In this paper I shall set out a clear and, I 
believe, defensible position on these issues. 
A second factor which undoubtedly affects the 
conceptual framework used in any given research project is 
the time scale set by the sponsors of one's research. 
Practically all sponsors of CL research expect products that 
are at least potentially 'applicable' (though often in a sense 
that is not well defined). For some researchers the time 
scale may be such that they must work with existing data 
bases, such as the machine readable version of the Longman 
Dictionary of Contemporary English (LDOCE) (Longman 
Group 1978) or 'Wordnet' (Miller 1990). This seems to be 
the case with the Pangloss project (Hovy and Nirenburg 
1992, Hovy and Knight 1993, Knight 1993), whose 
explicit goal is to combine the best features of these 
sources. This is an ambitious goal and I wish the 
researchers well. However the wen-known fact must be 
pointed out that, in the last decade and a half, many other 
researchers have put in a lot of work in trying to make 
LDOCE usable in a number of ways - and yet so far as I am 
aware no one yet has found a way to use these data as part 
of a belief system without an enormous amount of hand- 
editing. There are important questions that need to be asked 
about the relationship of these data to ontology building. 
The answers should relate to an integrated framework that 
provictes appropriately for at least the two linguistic levels 
of meaning and form, and, outside language, the categories 
of a belief system ('concepts'). 
Researchers who are workingto a less constrained time- 
table are perhaps more able to ask such questions. There are 
arguments for and against each approach, and it is not the 
purpose of this paper to criticize the work of those who 
seek directly to exploit existing data bases. Indeed, it may 
be that such work will in time produce solutions to the 
problems to be discussed here, by developing EVOLUTION- 
ARILY into more advanced models. Alternatively, it may 
be that a significantly different framework is required in 
order to achieve optimal representations of belief systems; 
both lines of inquiry should be pursued. 
In the COMMUNAL Project our task is to think 
speculatively about the next generation of belief systems, 
and about the components, relations and procedures that will 
be required in it. It is in the nature of research that we shall 
almost certainly have overlooked some aspects that will 
strike future researchers as important, but the enterprise is 
nonetheless worth attempting. Here, however, we shall not 
tU/to provide a complete overview, even very briefly, of all 
ot the components that we believe to be necessary in a 
belief system (for which see Fawcett 1993), but just those 
aspects that pertain to the concept of the 'ontology'. 
4 The intertwined concepts of 'ontology' and 
'system network': a brief history 
4.1 Halliday's proposal as a starting point 
Since ontologies display many of the characteristics of a 
s,~stem network, let us begin with Halliday's seminal paper 
Categories of the theory of grammar (1961). In it he 
proposed the concept of lexis 'as most delicate grammar'. 
He envisaged a model of language in which the earlier 
choices in a system network would be realised 
grammatically, i.e. in the structures of clause and group 
syntax and in grammatical items - and where these 
earlier choices would lead on to more 'delicate' choices that 
would be realised as iexical items. 
In the following decades Halliday and others did a great 
deal of work to develop the grammatical aspects of the 
model - including the important step of integrating 
intonat!on with grammar; But what of integrating lexis~. 
While grammatical items such as modal verbs and various 
types o~ determiner were modelled in system networks, the 
concept of system networks for lexical items lexis remained 
largely unexplored until the mid-70s to the mid-80s. In that 
penoa there were several small studies by systemic linguists 
(though not by Halliday himself) which implemented the 
concept of lexis in networks (Berry 1977, Fawcett 1980, 
Hasan 1987), but they were simply illustrative and there 
was no attempt to explore the implications of a 
comprehensive treatment of the original concept. (These 
'implementations' were linguistic descriptions, not 
computer implementations.) 
Meanwhde, Halliday had added the important notion of 
, - Oi i . meanm~ to that of choice at the heart of what now came 
to be called systemic functional grammar (SFG), so 
that the networks of Berry, Fawcett and Hasan were 
72 
7th International Generation Workshop ° Kennebunkport, Maine • June 21-24, 1994 
explicitly proposed as networks of features intended to 
capture the meaning potential of a language. It has 
always seemed clear to me that, since the networks specify 
meaning potential, the features in them can be appropriately 
regardedas semantic features (Fawcett 1973/81, 1980, 
etc). Halliday himself sometimes seems to agree with this 
position and sometimes to claim that the networks are at 
the level of form. (See the discussions of his variable 
position in Butler 1985 and, for example, Fawcett, Tucker 
and Lin 1993.) For the fullest and best account of the 
developments in SFG in modelling lexis, we must await 
Tucke?s PhD thesis, currently in preparation. 
4.2 Leech's logical semantics 
To the best of my belief, the first example of the 
intertwining of the two concepts ot an ontology and a 
system network occurred in the mid 1960s. It was at about 
this time that the system networks of SFG were beginning 
to be semanticized as the meaning potential of a language. 
Linguists working in the Chomskyan tradition had recently 
introduced the concept of selectional restrictions - which 
effectively presuppose semantic features. At this juncture a 
British hnguist, Geoffrey Leech, who was at University 
College London at the same time as Halliday was develop- 
ing SFG there in the 1970s, made the interesting experi- 
ment of combining the two concepts of system networks 
and semantic features in his Towards a Semantic Descript- 
ion of English, published in 1969 (and later incorporated in 
his standard text book Semantics (Leech 1974/81). This, 
then, was an early attempt to provide an ontology-based 
logic for reasoning, and it was done at a level that Leech 
assumed to be the level of semantics, i.e., presumably, a 
level within language. (In the view taken in this paper and 
to be expanded later, reasoning is i'n fact better modelled as 
taking place at a level outside and above' language - while 
being heavily influenced by languag,e.) Leecn's model 
included a taxonomy of 'types of object, and it had most of 
the characteristics of current ontologies (which we shall 
summarise in Section 6). This work was, in effect, Leech's 
attempt to combine the relevant parts of a system network 
with the demands of a reasoning system. It was related to a 
fairly standard logic, so that the taxonomy could be used for 
simple reasoning tasks, in some of the ways that current 
ontologies,are expected to.3 The crucial point, however, is 
that Leech s network was not in fact a system network, as 
the term is used in SFG, but an ontology. 
One striking v, isual icon of, this difference is the contrast 
between Halliday s and Leech s notations. They both use 
the usual systemi,c ,notation of a right-opening square 
bracket to ,indicate or and a curly right-opening bracket to 
mean and. But in Halliday's diagrams there is always an 
arrow pointing right, i.e. from the term (or terms) that 
is/are the entry condi,tion to the system towards the system 
itself. And in Leech s diagrams the arrows point from the 
system to the entry condition. In other words, the two 
apparently similar structures are to be used in different 
ways. As we shall see in Section 6, an ontology is 
typically (but not necessarily exclusively) traversed from 
right to left, which explains the direction of Leech's arrows. 
But a system network is intended to be traversed from left to 
,right, 1"., e. from the less 'delicate' choices to the more 
delicate ; see further below. (Some later systemic linguists, 
including myself, have followed Winograd (1972) in 
dispensing with the arrow. 4) 
4.3 Dahlgren's ontology 
Another ontology with interesting similarities to a system network is found in the important work of Dahl~ren 
(1988). Her description of her 'category cuts' caff be expressed directly as a system network, with 'entity as the 
entry condition that leads immediately to two simultaneous 
Stems; 'individual' vs. 'collective' and 'abstract' vs. 'real'. 
rther del:~ndent systems, some of which are entered simul- 
tane, ously (i.e. as 'cross-classifications'), introduce a total of 
37 category cuts' which, when all possible combinations 
are counted, generate 4272 potential ~combinations'. (This 
assumes that 'collective' leads on to 'mass' vs. 'set' vs. 
'structure', as implied at one point; at another it does not.) 
Interestingly, Dahlgren advises at one point that 'care 
must be taken when adding cross-classificatmns at a hi,gher 
node ..... to avoid proliferating empty terminal nodes ..... 
\[because this\] is a sur, e sign that the proposed cross- 
classification is spurious. If this were indeed to be accepted 
as a major critenon, one would unfortunately have to give 
her ontology rather low marks. This is because, of her 
4272 'terminal nodes', most are unused (all but 57 on one 
count).5 In fact, as we shall see in Section 5, Dahlgren's 
advice is fully appropriate for a systemic linguist who is 
using the network as a generator - but it is certainly much 
less relevant to an ontology. Why should this be so. -9 
The answer is that Dahlgren's taxonomy is not intended 
as a system network, and so it should not be thought of as 
operating by being traversed from left to right (e.g. as if to 
generate noun senses). An ontology is in fact typically used 
deductively, i.e. working from right to left TO REASON 
ABOUT OBJECTS (as Leech's arrows remind us). So the 
supposed 'overgeneration' of Dahlgren's ontology has no 
practical consequence - unless one wishes to use the 
ontology inductively. (In other words, many systems 
regularI~, use reasoning of the nature of 'If X is a rose then 
X is a flower, while there seems to be less demand for 
inductive reasoning. We shall return to the topic of 
reasoning in Section 6. 
I cite this case in order to show the need for clear 
principles in the construction of both (1) system networks 
at the level of semantics, i.e. within language, and (2) 
ontological relations in a belief system. 
We turn now to natural language generation systems 
that are based on SFG. There have been many of these, and 
many others that have incorporated significant aspects of the 
SFG approach to language (see Matthiessen and Bateman 
1991 and Fawcett, Tucker and Lin 1993 for overviews of 
the use of SFGs in NLG). Here we shall consider just the 
two major SFG natural language generators: Penman, 
which is to be used as the generator for the current Pangloss 
project, and GENESYS, which is the generator in the 
COMMUNAL Project. Given the significant place of SFG 
generators in NLG, it is important to be clear about the 
theoretical framework for what such enterprises are 
attempting. In particular, it is important to understand the 
relation between the system network-and the ontology. 
4.4 The Penman Upper Model 
The Penman Project was the first lar,~e SFG-based 
generator, and the mare work that establishes the structure 
and nature of what Halliday has called the lexieogrammar 
was done in the very early eighties (based fairly closely on 
Halliday's work of the seventms). In that period there was much more emphasis in linguistics as a whole on syntax 
and much less on lexis than there is now, and Halhday's 
programmatic use of the term 'lexicogrammar' was a far- 
sighted pointer to where work would be needed in the future. 
But Halliday himself has always approached language from 
the grammatical rather than the lexmal end and, under his 
guidance, Mann and Matthiessen naturally worked first on 
the grammatical structures and items. Unfortunately, when 
the time came to extend the model to lexis, the sponsors of 
the project (working within the traditional framework of the 
'grammar-vocabulary' distinction) required the Penman team ,NOT to imple,ment Halliday's concept of an integrated 
lexicogrammar, but to build instead a traditional lexicon which could be shared with the parser that was being 
contributed by another research team (working at Bolt, 
73 
7th Intemational Generation Workshop * Kennebunkport, Maine • June 21-24, 1994 
Beranek and Newman). There were some very interesting 
consequences of this decision. 
First, recall that the choices in the system networks 
were at this time being increasingly seen as the 'meaning 
potential, and thus as semantic choices. In much of the 
detailed description in Halliday 1985, for example, the term 
semantic abounds. And the concept of 'realization' itself- 
the process through which the choices in the system 
network are realized as st~ctures 7 is ~ne t, hat explicitly 
invokes the two levels of meaning and form. Thus if one 
does not represent the lexicon in system network form - as 
of course Penman does not - there is a major gap in the 
overall system network of the language at just tile place 
where one would wish to locate a system network of noun 
senses (to complement the list of noun forms). For/he 
same reason thePenman network lacks the meanings of all 
the other major word classes. This has made it harder than 
it is in the COMMUNAL framework (see below) to follow 
the increasingly strong trend in linguistics to give a more 
central place to lexis than it had in the 60s and 70s. 
(Fawcett Tucker and Lin 1993 show how lexis is handled in 
COMMUNAL, and important current modifications are 
exploiting yet further the value of having integrated 
lexicogrammatically realized networks of mean'ings - e.g. in 
verb complementation, and the many other cases where 
syntactic realizations depend upon lexical choice.) 
The Penman response to the position in which they 
found themselves was a sensible compromise. The 
development work that might have gone into the meaning 
potential of lexical forms went partly into the separate 
lexicon (building into it the relevant features from the 
grammar), and partly into the Upper Model (UM). The main, 
function of the UMis to serve as an abstraction hierarchy 
(Bateman 1990:57), and so to provide the usual functions of 
an ontology for property inheritance, etc. But at the same 
time,each concept in it is 'known to Penman', in the sense 
that it is possible to state for each concept in the UM the 
fragments of gran~matical or lexical realization that will be 
used to realize it (Bateman et al. 1990:5). 
However, the section of the UM corresponding to noun 
senses seems surprisingly small for this task - at least, as 
described in Bateman et al 1990. It con, tains just over 30 
categories (half being specialised within spatial-temporal'). 
The idea is that these are 'common core' concepts and that 
users of the UM will add to these as required.6 (It is 
interesting to compare these with the early semantic 
COMMUNAL categories, summarized in Section 5.) 
In Penman, then, the ontological categories map 
directly to (1)grammatically realizedoptions in the system 
network, and (2) the standard lexicon. The job does get 
done. But the great advantage of a SFG - namely that it 
can equally easily capture generalisations across large and 
small classes (including one-member classes) - is not ex- 
ploited. Thus Penman fails to utilize the many advantages 
of a fully integrated semantic network that provides for 
direct andintegrated realizations in grammar and lexis (and 
indeed also in intonation or punctuataon). 
At the time when it was developed, in the very early 
80s, Penman was a pioneering breakthrough. Its influence 
has changed the face of NLG. But the fact remains that it 
does not contain a lexicogrammar, in the SFG sense - just a 
grammar. The historic significance of Penman must not 
distract us from recognising that it has not yet explored Hal- 
liday s original proposal to integrate grammar and lexis in 
one great net-work. Yet this, as we in COMMUNAL have 
found, is a concept which - with the important modification 
that we do not restrict lexical meanings to the 'most 
delicate' parts of the network - brings enormous advantages 
of power and flexibility to the modelling of language. 
4.5 Alternative research strategies 
If one's goals are (1) to build a SFG generator and (2) to 
relate it out to a representation of those ontological 
relation, s required for reasoning, the following question 
arises: Is it (a) necessary and (b) desirable to have two 
separate layers of network, representing (1) the semantic 
stratum within language and (2) the ontological relations in 
the belief system?' 
There are two research routes that may lead to a sound 
answer. The first is to try having just one layer of network, 
and then to move on to two it it turns out to be desirable or 
necessary. The second route to an answer is to start with 
two networks, one for each level, and then, if there turns 
out to be inadequate motivation for maintaining two, to 
abandon one or conflate/he two of/hem. In effect, Penman 
had the first strategy forced upon them, while we in 
COMMUNAl_, have followed the second. As we considered 
the purposes - and so the desirable structural characteristics - 
of the two potential components, it became increasingly 
clear that there are advantages in including them both. 
The next two sections therefore set out the purposes and 
consequent structural organisation of (1) the system network 
for noun senses currently being implemented on a very large 
scale in COMMUNAL, and (2) the ontological aspects of 
the matchingpart of the belief system. While the discuss- 
ion is naturally exemplified from the COMMUNAL Pro- 
ject, the principles are of general relevance to any researcher 
working m/his,area. H, ere we shall restrict ourselves to the 
core area of objects, including abstract and event-like 
objects, i.e. that part of the belief system that corresponds 
to the senses of nouns. 
5 Thepurposes and structure of a system net- 
work for noun senses 
5.1 The purposes of a system network for noun senses 
My work in linguistics and NLP over the last couple of 
decades has taught me that one of the most important 
lessons to learn, when trying to model language, is: 
DO NOT TRY TO DO TOO MUCH WORK AT ANY ONE LEVEL. 
Thus the key to modelling language successfully is to have 
a sufficiently holistic theory and, to be able to recognize the 
appropriate level - or component - at which each particular 
type of work should be done. 
Once one commits oneself to having a separate 
component to handle reasoning (includingproperty inherit- 
ance, etc) that is OUTSIDE LANGUAGE(even though, as 
I have always insisted, its internal structures are strongly 
INFLUENCED by language), then it becomes immediately 
clear that the system networks inside the language system 
are NOT in fact well-suited for use in reasoning. The reason 
for this is very simple: the design features that are required 
in the structure of-the relevant parts of the network are 
different from the design features required for ontologies. 
So what are the purposes of/his part of the system 
network? Its primarypurpose is very simple. Just as the 
well known systems for transitivity, mood, and theme, etc 
generate clauses, so it is the purpose of this network to 
generate the nouns which will expound the heads of 
nominal groups.7 There is an important subsidiary purpose 
for those, noun senses to which participant roles (argument 
structure to some) are attached-, i.e. as with verb senses; 
compare the roles attached to die and death, to ascend and 
ascent, etc, but we cannot discuss these here. The other 
subsidiary purposes will be introduced in considering why 
the structure should have the form proposed below. 
5.2 The structure of a system network for noun senses 
What, then, should the internal structure of a system 
network for generatino_ nouns be like? The answers given 
here are derived from t~e experience of developing very full 
74 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
sub-networks for large areas of the network for the 'cultural 
classification' of Ythings' in English - i.e. for the 
classification of objects provided by the noun senses of 
English. As an indication of the co, verage, ,consider the 
fairly well-developed sub-network for artefacts that are for 
human consumption' (i.e. food and drink): it has 137 
systems and it generates 330 noun senses. Another quite 
well-developed area is that of plants, where there are 130 
systems that generate 246 noun senses. (This reflects a 
layman's taxonomy of different trees and flowers; a 
specialist could well have three to s!x times as many.) An 
area currently under development, use of land, includes 
'built up area', 'countryside, 'for_trav~lin-g', 'for_recrea- 
tion',-'w~teland' and 'for_dividing_land, and it has so far 
135 sub-systems, generating 370 noun senses. Many other 
areas are of similar size• 
We can give some idea of the overall taxonomic 
structure,by say!ng that 'use_of l,and' is one of fourteen 
types of artefact. The substantial (see below) features in 
the first few systems ,are as follows (where a system is 
represented as x -> y/z ): thing -> physical_thing / abstract 
thing / event thing; physical thing -> living thing / non 
living_thing;-living_thlng -~7 plant / creatu~; creature-> 
human_cr / non_human_cr; human_cr -> individual_hum / 
group hum; non living thing -> thing_as object/general 
_physTcal_phenoTnenon7 thing_as_substanCe; thing_as ob- 
ject -> artefact / natural object; general physical phen- 
omenon r> energy / weatffer / colour; eve~_thing -5 event 
_as_process / complex_event. The network is already very 
large, and it is growing all the time. 
What, then, are the pri,nciples on which it is construct- 
ed? The simplest 'network for the noun senses of English 
would consist of just one large system, containing one long 
list of the noun senses, each realized by a noun form - i.e. a 
single, massive system. Why not do this? The main 
linguistic evidence is the existence of hyponymic and 
contrastive relations between words - or more strictly, 
between word senses. Thus the word-form cat is a 
hyponym of mammal, because the ,meaning, 'cat' is 
systemically dependent on the meaning mammal'. And a 
system network is equally appropriate for giving formal 
expression to another major type of relationship between 
words, e.g. as specified in stand~d works on semantics such 
as Lyons 1977: that of 'contrast; the sense of dog is in 
contrast with that of cat The evidence that such hypon~/mic 
and contrastive relations are anc,hored ,firmly within 
language - and not simply in some higher component of 
the behef system, is the existence at the level oi ~ form of 
nouns such as thing, object, animal, mammal, person, etc; 
see the note on the as such type of feature in 5.5. The overall structure is therefore taxonomic. Within 
• tcr • • i • • this framework, we turn now to a ~mdehne cntermn that 
derives directly from the practical experience of building 
ontologies. 
Our experience in COMMUNAL suggests that it is 
advisable to avoid s, imultaneous entry conditions, i.e. right- 
opening curly 'and brackets. Here Dahlgren's point about 
overgeneration (4.3 above) DOES apply. (We shall meet the 
formalism when we consider ontological relations in 
Section 6, because in that type of network they do have a 
role to play.) Theproblem in generation is that such 
parallel systems lead to parallel pathways through the 
network, and so, almost inevitably, to permitting more co- 
selections than are in f,act possible. For example, the 
distinction between 'male and'female' is realized lexically 
in the case of only some animals. In COMMUNAL we 
simply allow the repetition of such systems, or gather the 
relevant features together into a disjunctive entry condition (as in Figure 2), whichever is easier in practical terms for 
the maintenance of the network as it is expanded. The key 
point is that, if a network designed to generate nouns allows one to go downpathways - and so to choose a set of 
features - FOR WHICH THERE IS NO REALIZATION, it 
is a bad system network.9 (But note that, while simultan- 
eity tends to lead to problems in networks for lexis, it has a 
valuable role in modelling g(ammatically-realized meanings - 
and, as we shall see in Section 6, ontological relations.) 
The large network for noun senses in COMMUNAL 
therefore has essentially the structure illustra,ted, in Figure 1. 
Neutrally, this structure can be described as: If a' then 'b' or 
'c'., ,M. ore typically, in a SFG framework, we express it as: 
'If a is chosen (by a speaker or a computer generator) then 
either 'b' or 'c' must be chosen.' Each feature may become 
the entry condition to a subsequent system, thus building 
up a ,s~,stem,network, so it might cog, tl,nue: 'If 'b' is chosen, 
then d' or 'e must be chosen, and if c is chosen then f or 
'g' or 'h' must be chosen .... ' and so on. Ii 
f I: 
Figure 1" A simple system network 
However, there is one type of complex entry condition 
which we have found to be occasionally useful in the 
network for noun senses: the disjunctive entry condition. 
This has the form shown in Figure 2: 
d 
c \[: 
Ii 
Figure 2: A disjunctive entry condition 
Here is a (simplified) example of this phenomen,,on,,from the 
current network (where "/" signals or' and -> signals 
'entry to a system'): 
tomato_plant_c / pepper plant_c / cucumber plant c / 
strawberry_plant_c -> 
95% plantness_explicit (66.325) 
5% plantness_implicit. 
This representation of a system network (in the 
COMMUNAL format) is to be read as follows: 
'If you select either 'tomato plant', etc (whic,h come from 
the ve etable sectmn of the ~,round crops network)or 
, d~ v • -- o • v strawberry_plant (which comes from the fruit sub- 
net,work), then there is a choice between having the concept 
of plantness' made specific and leaving it implicit.' So, if 
the choice is to make it explicit, forms such as tomato- 
plant will be generated, an, d, if implicit, forms such as 
tomato - but in the gardeners sense of 'tomato plant, as in 
l've watered the tomatoes this morning. Another valuable characteristic of the COMMUNAL 
system networks is the use of probabilities. As ,you can 
see, the weighting is 95% to 5% towards explicit (unless 
this is overruled by the discourse planner). The reason for 
this strong weighting is that there is another type of object 
75 
7th International Generation Workshop • Kennebunkport, Maine ° June 21-24, 1994 
that is commonly referred to by the same word-form, 
namely the fruit of the tomato plant. (The choice of 
'plantness explicit' triggers Realisation Rule 66.325, which 
adds to the head of the nominal group the item plant. ) 
A third criterion used in constructing the network relates 
to its value in specifying what in Chomskyan linguistics 
are termed 'selectiona\] restrictions'. Some might wish to 
argue that these really belong in the belief system rather 
than in language itself. However, it turns out that it is 
quite simple, if one constructs the network on appropriate 
criteria, to enable very many of the types of restriction that 
it was hoped to capture in transformational models of the 
1960s and 1970s to be captured in a system functional 
grammar. Moreover, since we use probabilities in the 
grammar in any case (for purposes that we cannot go into 
here), we can ,capture selection restrictions as relatively 
strong or weak preferences'. Thus there seems to be no reas- 
on why we should not c.apture such phenomena in both the 
system network within language and in the belief 
system. If we can do this, it brings the advantage, when 
TESTING the system, that the lanzua~e comoonents can be 
un independently of the behef system. They can be set to 
generate sentences randomly as we test the lexicogrammar, 
and still generate sentences that sound fairly plausible.10 
The way this works is, in outline form, as follows. 
Suppose we have just generated a clause, with an Agent as 
Subject and ask as the main verb. We want to restrict the 
choices when the network is re-entered, so that, if a nominal 
gAroup with a noun at its head is to be genemt~ to fill the 
gent, it will have one that is plausible as an asker. We 
do this by specifying that, on re-entry to the network, the 
following features are chosen: \[thing, concrete, living, 
creature, 99.9% human / 0.1% non_human, whole_hum, 
99% individual_hum / 1% group_hum\]. ,At two points, 
you will notice, there is not an absolute preselection' of 
just one of the features in the system, but a preference for 
one over another. Thus the feature \[human\] Is shown to be 
a thousand times more probable than \[non_human\] - thus 
allowing for talking computers and talking animals, such as 
the white rabbit in Alice in Wonderland. Rather similarly, 
individual humans are shown as being a thousand times 
more likely to be the Agents in a process of 'asking' than 
are groups of humans - though groups, such as a 
committee, may well ask questions on occasions. 
We turn now to an important issue that arises as a 
result of taking this position, and we shall illustrate it first from English. 
5.3 Problem case 1: the 'count' versus 'mass' contrast 
One important effect of deciding to organize the network 
around meanings (rather than around grammatical correlates 
,of wor, d forms) is to put in perspective the contrast between 
• count and 'mass' noun senses that is so dominant in 
English (and many other European languages). If, for 
example, we are stating preferences for the Affected entity in 
a process of 'eating', it is not important whether what is 
eaten is mass or count. The COMMUNAL network for food 
classifies types of food on semantic criteria - so that, for 
example, vegetables that typically occur with ,a meat course 
are placed together, with the mass noun sense cabbage' next 
to the count noun senses of 'potato' or 'pea'. Moreover, the 
network also includes a way of showing that, while potato' 
occurs regularly as either singular or plural, it is rather 
unusual, to talk of a single pea. The result is that the 
system, knows that it is a tho, usand times more likely that 
"peas will be generated than pea'. (Note that the system 
networks in COMMUNAL not only have probabilities, but 
the grammar can chan,~e these when required. For a more 
detailed account of t~is, together with a full worked 
example, see Fawcett, Tucker and Lin 1993.) 
raowever, there is not in fact a neat 'count' versus 'mass' 
distinction in English at all. It is one of those useful gener- 
alisations which hold for 95% of the time, but which, if 
one commits oneself to it, leads to considerable trouble 
when the idea is extended to the whole of langua,~e. It is 
certainly not a category that can be extended, on thgbasis of 
a system network for English, even to a close European 
!anguag,e such as French. The illogicality - in physical 
number terms - of items such as furniture and cutlery is 
just the well-known visible tip of quite a large iceberg. 
First, note the 'plural-only' nouns, such as police, staff and 
contents. Theft there are the 'plural-prefern'ng' items,--with 
varying strengths of preference, as for example betwee,n pea 
ana sprout, and pebble and leaf. There are also the pair- 
only' nouns such as trousers, scissors, and binoculars - and, 
even though the word denotes two garments, of which only 
the bottom half conforms to the pattern, there is pyjamas. 
As an example of the dissonance within one relativel3; smal! 
semantic field between the grammatical criterion of 'number 
and the semantic classification of noun senses, consider the 
field of clothing. Suppose the problem is that of stating the 
preferences for the entity that is to complete a clause such 
as He went home and put on ..... It could be a nominal 
group with, at its head, (1) a mass noun such as some warm 
clothing, or (2) a plural-only noun as in some warm 
clothes, or (3) a singular noun such as a warm jersey, or (4) 
a pair-only noun such as some warm trousers. 
The COMMUNAL solution to this problem is as 
follows. Every feature in the syste, m, network f o, r which 
there is a realization is given a suffix c (for 'count thinzs) 
or ~_m' (for 'mass' things,) or '_pl' (fo~ plural only' thinks) 
or _pair' (for pair only' things). Those labelled' c' lead 
into the system for NUMBER, where there is achoice 
between \[singular\] and \[plural\], while for the others there is 
no choice: they are either \[mass\] or \[plural\]. Those for 
which a RELATIVE preference for \[plural\] has been stated 
will enter a version oT the NUMBER system for which the 
probabilities have been re-set according to the strength of 
the preference associated with than noun sense, i.e. the 
lexicogrammar ,'knows' that for 'pea' the probabilities of 
choosing 'plural are very much greater than for 'cabbaze'. 
Consider too the question of where to place th~ two 
senses - realized as count and mass nouns - of cloud. In this 
case the differences between the two appear to derive mainly 
from the fact that one is an individual entity and the other is 
a non-discrete mass of 'stuff; they are not differentiated by 
function. So the system network shows them to be 
distinguished as follows: 
cloud-> cloud as individual/cloud as mass. 
Notice that this is a rather different matter from 'oak' as 
'individual' and as 'mass'; in the first case the referent is a 
tree (with oak-tree as a possible variant, if 'treeness explicit' 
is selected), and it is located with the rest of the tr-eegas a 
 uP e of plant, while in its other sense it is a material whose nction is to be used for making things, and it is located 
with other types of wood, fairly c ose to 'iron', etc. 
In the COMMUNAL system it is even possible to 
build in the preference for meanings realized in expressions 
such as a pair, followed by of, as in a pair of trousers. This 
is because trousers rather thanpair is treated as the head of 
the nominal group. The words a pair are generated as a 
special nominal group expressing 'quantity' that fills the 
quantifying determiner. It is because the noun sense of the 
nominal .qroup is generated on the first pass through the 
network ~or the object i.n question - I.e. because the 
lexically realized meaning is integrated with the 
grammatically realized meaning in C ONIMUNAL- that part 
of the realization of the choice of trousers' can be to 
express a preference for the way in which the auantifvin,~ 
determiner is filled - here by the embedded nominal grou'p o-f 
a parr. 
Thus the COMMUNAL way of handlinz these various 
aspects of 'number' in English can reflect-accurately the 
76 
7th International Generation Workshop ° Kennebunkport, Maine • June 21-24, 1994 
true, complex nature of 'number' and 'quantity' in English, 
while at the same time maintaining the semantic relation- 
ships of hyl~onymy and contrast in the network, and so 
making possible the expression of preferences. 
5.4 Problem case 2: long thin things and other such gram- 
matically realized categories 
The issue is in fact much wider than that of whether 
'count' vs. 'mass' should be a primary distinction in English 
and related !anguages. How strong a candidate it is for a 
generalized interlingua' ontology that is to accommodate all 
the languages of the world? Consider Chinese, with its 
well-known classifier system, in which the mass-count 
distinction plays no part. Then think about Swahili, with 
its ki- vi- class of non-living things, its m- wa- class for 
humans, its u- class for abstract things, etc. Japanese, as it 
happens, has a special set of cardinal determiners, whose 
form depends on the semantic class of the noun: i.e. 
whether the object is human or a small thing or even, it 
would seem, a long thin thing. Thus, "if the thing 
concerned is a flower (hana), a tree (ki), a pen (pen), a 
pencil (enpitsu) or a,rive, r (kawa) - all longthin things - the 
determiner meaning one will be ippon. But if the thing is 
a human it is hitori, and if a non-human creature it is ipipi - 
and so on, for many more classes of thing and for many 
more cardinal determiners. The semantic generalisation that 
unites those things that require ippon appears to be simply 
that they are all "long thin things'. If this seems strange to 
the investigator from a European background, consider how 
odd it must seem to the investigator of English from 
outside Europe who finds that, in English and related 
languages, there exists a basic distinction which affects 
many aspects of the semantics and syntax of the nominal 
group, between things that are and things that are not 
'countable'. 
However, it must be emphasised again that, important 
as the distinction is at many points, it is not the basic 
organising p.rinciple of the semantics of English noun 
senses. While there is, of course, a distinction between 
'substances' and 'objects' in the taxonomy implied in the 
COMMUNAL network, and while all substances are 'mass', 
it is not the case that all 'objects' are 'count' - as we saw in 
section 5.3. 
In COMMUNAL, then, our semantic system network 
gives no weight to grammatically realized contrasts such as 
count' versus 'mass', and we use instead the semantic 
criteria such as those that help us to state the preferences 
associated with a given participant role such as agent. 
Is there a price to be paid for all of the advantages 
outlined in the, preceding sections? The answer seems to be 
that there isn t. If the mechanism for handling NUMBER 
in a way that is dependent on noun senses (in 5.3) required 
one to work from a visual representation of the wiring, it 
would be tedious in the extreme But in the computer it ~s a 
simple matter - and this has in turn suggested a simple 
representation for the written version. 
5.5 'Special features' in the system network 
In a full system network for nou, n senses, the relations 
between features are not all of the subcategorization type 
that the systemic notation typically signifies. The 
extensive work that has been done in COMMUNAL over 
the ,last few years has produced a small set of supplementary 
(or special ),features whose function is to express relations 
between the substantial' features that represent noun senses. 
In principle, all of those shown in Figure 3 can occur 
between any two substantive features. Here, then, 'x' stands 
for a feature such as 'human', and 'a', 'b' and 'c' are the next 
'substantive' features. Possible realizations of the selection of the feature (in some cases in features dependent on 
substantive features) are given in Figure 3. 
X~ 
F x_as_such 
- whole x-- 1 
- t_ x_specified I 
-part_of x 
group_of_x 
E! 
(e.g. person) 
(e.g. adult) 
(e.g. adolescent) 
(e.g. child) 
(e.g. hand) 
(e.g. committee) 
Figure 3: Some types of features that are not related by 'be- 
type'relations 
'Substantive' features are those that translate types of 
object in the belief system. In most cases none, or only 
one, of the special features are needed bet w, een substantive 
features. Note in particular the 'xas-such feature. This 
provides for the generation of those nouns that are used for 
Iexical substituUon such as thing, stuff and, more delic- 
ately, plant and animal. It is the existence of such forms 
that provides the evidence that such less delicate meanings 
exist in natural languages, and that the intra-linguistic level 
of semantics does indeed require a network for it to be 
adequately represented. 
There are possible variations of the feature \[xspecified\] 
which it is important to note. The type of specification 
may be spelled out more fully, with a feature corresponding 
to each type that leads to its own dependent network. Thus 
in some cases (such as \[human\]) we may find Ix_as_such\] 
vs. Ix_specified_by_form\] vs. Ix_specified by_role\]. 
The somewhat dense description given above is intended 
to give the flavour of the criteria that are guiding the 
construction of the very large system network for noun 
senses in COMMUNAL. 
There is one last benefit that this approach brings. It is 
a practical rather than a theoretical one. This is that the 
process of constructing the network, and so of deciding 
which types of 'special feature' are needed, goes some way 
in preparing the ground for constructing the equivalent 
ontological aspects of the belief system. And it is this to 
which we now turn. 
6 The purposes and structure of the ontological 
aspects of a belief system 
6.1 The purposes of an ontology 
What are ontologies for? They are as they are, of 
course, because of the functions that they are required to 
perform. In the case of an ontology, the purpose is NOT to 
represent the meanings of the nouns of a language, but to 
facilitate reasoning. Thus lexically prominent register 
distinctions such as that between fag and cigarette would be 
modelled as de,noting t,he same generic object, and so 
share the same concept. This is because the same events 
Or 'propositions') are attached to it, whatever degree of 
rmality is selected in the TENOR system. It is the latter 
that requires recognition in the belief system, and not a 
difference of concept. 
Ontologies are used in two principal ways, the second 
beinz deoendent on the first. The first is for the type of 
reaso-nin~known as entailment (see, for example, Leech 
1969 and 1974/81). Through entailment a reasoner can 
infer from the belief (or proposition) that Object 79 is a 
member of the set, of objects denoted by the form dog (and 
by the sense 'dog) that it is also a member of the set of 
objects to which the sense 'mammal' pertains. In layman's 
lan_oua2e, A dog is a mammal. The crucial point Is that, 
typl~caffy, the directionality of the reasoning is from 
the more delicate category to the less delicate. For examp!e, 
in terms of Figure 4 below, if an object is 'd' it is also b' 
77 
7th International Generation Workshop • Kennebunkport, Maine ° Jmae 21-24, 1994 
and 'a'. This directionality is of course different from the 
typical use of a system network in generation. 
The second main use of ontologies is an extension of 
this. It is for the inheritance of what are commonly 
called 'properties'. (However, as we shall see, the term 
'oroperty' is potentially misleading). Thus, if such 
'iarop'erti'es' are attached to a category at one node of the 
ontology, they can then be assumed to hold for any given 
object to which a logically dependent category applies., So, 
if we build into a belief system the proposition that land 
mammals typically have four limbs', then we can infer that, 
because a dog is a land mammal, it too typically has four 
limbs - and so on. (More strictly, as we shall shortly see, 
the relationship is between the generic thing that corres- 
ponds to that node; in the COMMUNAL model of logic dogs 
is a referring expression as much as is this dog.) 
These types of entailment-based reasoning are 
important. But it is equally important to recognize that 
they are only one of a variety of types of reasoning that are 
regularly used in real life situations. Moreover, the concept 
that all such beliefs about categories of object are handled 
by inheritance is not a matter that is entirely beyond 
dlspute; It is arguable, for example, that the category of 
human is so prominent in our perception of the world that 
we build sets of beliefs around it - andthat we do not use an 
ontolo,qical structure in order to inherit, every time that we 
refer tGa human, all of the many propositions that relate to 
the many other, ever more general, categories that are 
superordmate to 'human' - such as 'mammal', 'creature', 
'living' and 'concrete'. And, once one allows for this 
possil~ility, -there is no clear way of knowing where to stop. 
There may be many such nodes that laave attached to them 
large sets of beliefs (or 'propositions ) that are held by the 
system, and which may indeed involve the redundant 
repetition of beliefs that are attached to less delicate 
concepts. Who can say whether it is more economical (or 
more ele_~ant more efficient, or whatever) to store a large 
(but not l~fir~ite) set of propositions many times over at tile 
various nodes where they are most often needed - the basic 
types', in the terms of Rosch (1978) - or to store each of 
them just once but to have to perform a multiple act of 
entailment reasoning, involving multiple searches back 
down the tree, every time one u,ses the belief that a dog 
needs air to breath, or is a 'creature, or is a 'concrete' object? 
A particular dog-lover, for example, may have his/her set of 
primary beliefs about spaniels attached to 'spaniel', rather 
than to 'dog', and so on (cp. Reiter and Dale 1992). In other 
words, while we are undoubtedly capable of performing the 
quite complex type of reasoning involved in inheritance, it 
may be that it is not the backbone of all reasoning that it 
has sometimes been assumed to be. 
The ar_oument is not that inheritance has no place in our 
reasoning,'but (1) that it may have a much less central place 
than has generally been supposed, and (2) that other types of 
reasoning are probably equally or mo, re important. Abelief 
system of the type assumed here is object-oriented in the 
sense that it consists of a vast number of specific 
objects and generic objects (with each of the former 
linked by a 'be-an-instance-of relationship to one or more of 
the latter) such that one of the many things that the system 
believes about any generic object is what other generic 
object - or objects - that generic object is itself a type of. 
In the above discussion, we have been assuming that we 
are considerin,~ generic objects, e.g. cats in general. I 
assume that t~ere would be general agreement that the 
relationship between 'mammal' and 'cat' is what we might 
term a 'be-a-type-of relationship, while that between a 
specific instance of a cat, such as our family cat Timmy, is 
a 'be-an-instance-of relationship. Specific objects are relat- 
ed to concepts via the belief that 'Timmy is an instance of a 
cat'. In other words, we need to distinguish between (1) 
'Timmy is a cat' and (2) 'Cats are mammals. 
I suggested earlier that the use of the term 'properties' to 
refer to the 'propositions' attached to categones can be mis- 
leading. Consider the simplest mod, el of ,what 'properties' 
shouldbe attached to categories (the frame approach). All 
there is space to say here is that there are clearly limitations 
as to what can be expressed in such limited structures, and, 
like others, we in COMMUNAL are exploring richer 
alternatives. In our case we are experimenting with a 
specially developed logical form for the representation of 
both (I) complex beliefs about, say, the mating habits of 
dogs, and (2) the belief that dogs are mammals - and indeed 
the belief that dogs are typically pets ('multiple 
inheritance'). 
The purpose of ontological relations, then, is to facil- 
itate reasoning. But our prediction is that in future there 
will be less emphasis on inheritance and more on other 
logical relationships. Given these purposes, we turn now 
to the question of the structure of ontologies. 
6.2 The structure of ontologies 
There seems to be a general agreement that an ontology 
has the general form of a taxonomy' or 'hierarchy' (in one 
sense of that overused term) or 'tree (which is equal!.y over- 
used). The type of 'tree' required here is thus a paradigmatic 
tree (rather as is a system network) where, in the simplest 
model, a pathway that consists of a list of features chosen 
as one traverses the network corresponds to any one (as p,ect 
of) an object. In t h,e simplest type of taxonomy, then, a is 
subcategofized as b or 'c ~, 'b' as' d' or 'e', 'c' as 'f, 'g' or 'h', 
and so on; see Figure 4. 
d 
a~ b~ef 
.h 
Figure 4 : Relations in the simplest type of ontology 
Such a network show, s that, if an object satisfies the con- 
ditions for being an h ,it follows that it is also a 'c', and so 
also an a'. So far, this structure is similar to the simplest 
type of system network, as ,shown in Figure 1. 
This simplest structure for an ontology is in practice 
au,~mented in various ways in all of the ontologies that I 
kn~w of. The essential addition is as in Figure 5: 
h j 
k 
d 
e 1 
f m 
Figure 5: Simultaneity in ontological relationships 
This is to be read as, 'a' is subcategorized as 'b', 'c' or 
'd', and also as 'e' or 'f; 'b' is s, ubcategonzed as 'g' or 'h', and 
'd' as T, 'j' or 'k', and also as i or m. 
Here the right-opening curly brackets do NOT signify 
that you must follow all the designated paths, as they would 
in a system network - because this type of network is 
desig.ned to be used from right to left, for the types of 
entadment and inheritance outlined above. 
Dahlgren (1988:46f.) discusses the mathematical 
78 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24; 1994 
properties of ontologies, and she quite rightly points out 
that, while binary ontologies 'have simplifying math- 
ematical properties, they are not likely to be .psycho- 
logically reap (e.g. suggesting that we operate with 'fish' 
vs. 'bird' vs. 'mammal', etc). She also points out the need 
for cross-classification, in that we classify animals, let us 
say, both in terms of their 'types' (a concept that itself needs 
to be unpacked, at least for some purposes), but a!so in 
terms of their 'role' in humankind s scheme of things ( tame' 
vs. 'wild' animals, probably with other such subcategories 
as well)• Thus Dahlgren and others operate with the 
formalism shown in Figure 5. 
In the cases of other ontologies there is even more free- 
dom, in that any feature can be linked by a line that 
represents a 'be-type' relationship to any feature to its left. 
And in some ontolbgies, of course, other relationships such 
as 'be-a-part-of' are used. 
However, there is an important sense in which the 
above diagrams are misleading - at least in relation to the 
current COMMUNAL framework. This is because we 
consider that the relations do not hold, in fact, between the 
'concepts' in the ontolo~ies, but between generic obiects. 
The relationship that we"have shown in Figures 4 and5 by 
a line can also be expressed by the following: 
e (el 11, \[ca:o222, pr:bejype, at:o333\]). 
o(o222, \[cc:dog, qt:all\]). 
0(0333, \[cc: mammal\])• 
This, states that event no. 111 has a carrier (a 'participant 
role), which is object no. 222, a predicate, which is 
'be_type', and an attribute (another participant role), which 
!s objectno. 333• Object 222 is then defined as the class of 
O" " " " ' ' all do=s, while object no. 333 is defined as mammals. 
There are several aspects of the meaning of this 
representation of a belief which are assigned by default - 
,most importantly, that the time position of the event is 
past, present and future', that the' 'confidence level' is so 
high as to be interpretable as 100% confident'. Taken 
together, they state, m a natural language translation: All 
dogs are mammals. 
This event-based representation is used in order to 
enable the system to carry out reasoning on inputs to and 
outputs from the system that have the sorts of annoyingly 
messy complexity associated with natural language - 
representations of time position, usuality and quantification, 
for example• In the belief systems of the future it will no 
longer be possible, in our view, to depend on simple data 
structures such as frames. The need to incorporate more 
sophisticated representations of time, of modality and of 
other such phenomena related to events demands that 
infor,mation be stored in the form of some type of 'predicate 
logic, e.g. as in the (simplified) example above• 
The key point is that the representation, like the 
minimal operational syntactic unit otthe clause, is based on 
the concept of an EVENT (our equivalent term to 'propos- 
ition' and 'eventuality' in other frameworks). It uses categ- 
ories that reflect idealised aspects of systemic functional 
grammar, so it is a 'systemic functional logical form' 
(SFLF). 
Note, finally, that the relationship that it expresses is 
NOT one that holds between two concepts, but between two 
referring expressions, each with its SFLF structure, and 
each of which refers to a generic object• 
To adopt this position leads to a reappraisal of the 
status of dmgrams representing ontological relationships 
such as those in Figures 4 and 5. Those relations are, m 
the approach advocated here, simply one event type among 
many within the mass of belief, s that the system holds 
about dogs• In other words, the fact' (i.e. the confidently 
held belief) that dogs are mammals is just one of many 
things that the system assumes that it 'knows' about the 
generic object of 'dogs. 
7 Summary and conclusions 
The COMMUNAL Project (Fawcett, Tucker and Lin 
1993) has demonstrated the immense advantages that follow 
from having a unified system network for all meanings, 
whether reafized grammatically or lexically. Thus Hallida~'s 
original 1961 insight was well-founded, the only major 
modification needed being that lexically realized meanings 
are not necessarily most delicate', in the sense of 'at or near 
the terminal leaves of the system network. (Meanings 
realized in intonation should similarly be integrated into the 
overall network, and the way in which thls is done in 
COMMUNAL is described in Fawcett 1980.) In other 
approaches some equivalent intralinguistic level of 
'linguistic meaning' would surely be of similar value, 
It would be interesting to compare the COMMUNAL 
'division of labour in modelling lexical meaning with other 
current proposals, such as those of Reiter and Pustejovsky, 
but that is beyond the scope of this paper. What is clear is 
that information about the purposes oil entities, how they 
come into being, etc is all handled through events attached 
to the equivalent genetic object in the belief system• 
The two types of network - the system network of noun 
senses and the ontological relations in the belief system - 
serve different purposes. One is to generate nouns (and to 
control relateOstructures such as those associated with 
participant roles) and the other is to facilitate reasoning. In 
the first it isprefer,able to avoid right-opening brackets - but 
in the second such cross-classifications' (or their equivalents 
• I ' 0" " " m an event-based lo=lcal form) are wtal to certain types of 
reasoning• Since the two structures serve different 
functions, both have their place in a holistic model of 
language. (I say 'structures, because it is probably unhelp- 
ful, in the new framework, to think of the e, quivalen, ts of 
ontological relations in the belief system as a network .) 
In the previous section we have seen how the 
COMMUNAL Project has begun to develop a new way of 
representing - and so thinking about - the relations found in 
standard ontologies. Since these are now seen as a part of 
the belief system, the suggestion !s that we no longer need 
to think in terms of 'an ontology, in the sense of a taxo- 
nomy of 'concepts' - but rather in terms of relations between 
genetic objects. We think that these fit more naturally into 
the more complex, event-based reasoning that we judge will 
be needed for the systems of the future• So perhaps there is 
no such thing as an ontology, but instead a set otmutually 
referrin,2 beliefs, each represented as an event? 
If tKis approach is on the right lines, or even if it is 
only partly on the ri,2ht lines, it is time to be moving on to 
the exploration of t~e next generation of belief systems. 
These will make different assumptions about the nature of 
the relationship of word forms to word senses, and word 
senses to the related aspects of the belie, f system - and 
perhaps too about the nature of 'ontological relations. 
Notes 
! The research reported here was supported by grants from 
DRA Malvem (under contract no. ER1/9/4/2181/23), from the 
University Research Council of International Computers Ltd, 
and from Longman, and by the University of Wales, Cardiff. I 
am grateful to my colleagues Gordon Tucker, Yuen Lin and 
Ulrich Gysel for the many useful discussions which have 
contributed to the emergence of the view presented here. The 
main debt is to Gordon Tucker; together we have hammered out 
the sense in which we now interpret Halliday's original 
challenging concept of 'lexis as most delicate grammar'• I am 
also grateful to the three anonymous reviewers of this paper; 
the remaining infelicities being, as always, the author's. 
2 The term 'typic' has been used in preference to 'generic' i n 
some of our earlier writings because we could give to this new 
term a has a clear definition - leaving 'generic' for 'generic (= 
79 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
genre) structure'. Here 'genetic' will serve equally well. 
3 The fact that Leech could not propose a workable way to 
map the structures of language onto those that he inherited from 
standard logic is significant, but he can hardly be censured for 
failing to solve a problem that has also defeated all others who 
have attempted it. 
4 There are occasions in natural language processing when it 
may be desirable and/or necessary to use system networks by 
working from right to left, e.g. as proposed (1) by Patten 
(1988) for some aspects of generation and (2) by O'Donoghue 
(1994) for the higher stages of parsing. But these are 
adaptations of the central concept for specific purposes, and 
they do not affect the main point that system networks are 
designed to be used in generation, working from left to right. 
5 While Dahlgren has only 57 'category cuts', some of her 
terminal nodes such as 'animal' bring together a number of 
these and then lead off into further subsystems - and so 
effectively to further 'terminal nodes', such as 'vertebrate' vs, 
non-vertebrate', and then 'mammal' vs. 'bird' vs.'fish'. Thus 
there are in practice a few more than 57 'terminal nodes'. But 
this does not affect the more general point being made here. 
6 Our experience in COMMUNAL- working at both the 
semantic and the belief system levels - suggests that the 
development of even quite delicate areas of the network raises 
problems requiring a high level of expertise and experience for 
a satisfactory solution, and we believe that this experience is 
mirrored elsewhere. One can foresee that clients might be wise 
to subcontract the work of ontology-building back to the 
Penman team. 
7 Contrary to what is still a quite widespread assumption, 
there is no.need for a separate network from that for the 
nominal group to deal with the rank of the word; the nouns at 
the head of a nominal group are simply the realization of one 
part of its meaning (see Fawcett, Tucker and Lin 1993). 
8 These have ontological equivalents in the COMMUNAL 
belief system, so that interested persons may wish to compare 
this slicing of the universe of objects with those assumed for 
other ontologies. It would be an interesting exercise to get the 
creators of alternative taxonomies to explain their reasons for 
foregrounding their primary distinctions. 
9 A pseudo-remedy that some systemicists use is to add 
complex wiring that prevents the realization of all but those 
combinations of features for which there is a realization (e.g. 
'dog' plus 'female' but not 'hare' plus 'female'). But this to try to 
correct poor systemic modelling at too late a stage; it is the 
task of a system network to constrain possible co-selections. 
l0 The random generation of sentences is frowned upon by 
some NLG researchers, but it serves a valuable role in 
developing large generators, because it tests the availability of 
the lexicogrammatical resources, and so has a role in solving 
the 'expressibility' problem (Meteer 1992). 

References 
Bateman, J.A., 1990. 'Upper modelling: organizing language 
for natural language processing'. In Procs of 5th 
International Workshop on Natural Language Generation, 
Pittsburgh, pp. 54-61. 
Bateman, J.A., Kaspar, R.T., Moore, J.D., and Whitney, R.A.. 
1990. A General Organization of Knowledge for Natural 
Language Processing:the Penman Upper Model. Technical 
Report, USC/ISI Information Sciences Institute, Marina del 
Rey, California. 
Berry, M., 1977. Introduction to Systemic Linguistics, Vol 2: 
Levels and Links. London: Batsford. 
Butler, C.S., 1985. Systemic Linguistics: Theory and Applic- 
ation. London: Batsford. 
Dahlgren, K, 1988. Naive Semantics for Natural Language. 
Boston: Kluwer. 
Fawcett, R.P., 1973/81. 'Generating a sentence in systemic 
functional Grammar'. University College London (mimeo). 
Reprinted in Halliday, M.A.K., and Martin, J.R., (eds.) 
1981 , Readings in Systemic Linguistics. Batsford, pp. 
146-83. 
Fawcett, R.P., 1980. Cognitive Linguistics and Social 
Interaction: Towards an Integrated Model of a Systemic 
Functional Grammar and the Other Components of an 
Interacting Mind. Heidelberg: Julius Groos. 
Fawcett, R.P., 1990. 'The computer generation of speech with 
semantically and discoursally motivated intonation'. In 
Procs of 5th International Workshop on Natural Language 
Generation, Pittsburgh, pp. 164-73a. 
Fawcett, R.P., 1993. 'Language as program: a reassessment of 
the nature of descriptive linguistics'. In Language Sciences 
14.4: pp. 623-57. 
Fawcett, R.P., 1994. 'A generationist approach to grammar 
reversibility in natural language processing'. In 
Strzalkowski, T., (ed.), Reversible Grammar in Natural 
Language Generation., Dordrecht: Kluwer, pp. 365-413. 
Fawcett, R.P., Tucker, G.H., .and Lin, Y.Q., 1993. 'How a 
systemic functional grammar works: the role of realization 
in realization'. In Horacek, H., and Zock, M., (eds.), 1993, 
New Concepts in Natural Language Generation, London: 
Pinter, pp. 114-86. 
Halliday, M.A.K., 1961. 'Categories of the theory of grammar'. 
In Word 17, pp. 241-92. 
Hasan, R., 1987. 'The grammarian's dream: lexis as most 
delicate grammar'. In Halliday, M.A.K., and Fawcett, R.P. 
(eds.) 1987, New Developments in Systemic Linguistics, 
Vol 1: Theory and Description, London: Frances Pinter, 
pp. 184-211. 
Hovy, E.H., and Knight, K., 1993. 'Motivating shared 
knowledge resources: an example from the PANGLOSS 
collaboration'. In Proceedings of the IJCAI Workshop on 
Shared Knowledge, Chambery, France. 
Hovy, E.H., and Nirenburg, S., 1992. 'Approximating an 
interlingua in a principled way'. In Proceedings of the 
DARPA Speech and Natural Language Workshop, 
Hawthorne, NY. 
Knight, K., 1993. 'Building a large ontology for machine 
translation'. In Proceedings of DARPA Human Language 
Conference, March 1993. 
Leech, G.N., 1969. Towards a Semantic Description of 
English. London: Longman. 
Leech, G.N., 1974/81. Semantics. Harmondsworth: Penguin. 
Longman Group, 1978. Longman Dictionary of Contemporary 
English. London: Longman. 
Lyons, J., 1977. Semantics (Vols 1 & 2). Cambridge: Cam- 
bridge University Press. 
Matthiessen, C.M.I.M., and Bateman, J.A., 1991. Text Genera- 
tion and Systemic Functional Linguistics. London: Pinter. 
Meteer, M., 1992. Expressibility and the Problem of Efficient 
Text Planning. London: Pinter. 
Miller, G., 1990. 'Wordnet: an on-line lexical database'. In 
International Journal of Lexicography 3(4) (Special Issue). 
Nirenburg, S., and Defrise, C., 1992. 'Application-oriented 
computational semantics'. In Johnson, R., and Johnson, 
M., (eds.) Computational Linguistics and Formal 
Semantics. Cambridge: Cambridge University Press. 
O'Donoghue, T.F., 1994. 'Semantic interpretation in a 
systemic grammar'. In Strzalkowski, T., (ed.), Reversible 
Grammar in Natural Language Generation., Dordrecht: 
Kluwer, pp. 415-447. 
Patten, T., 1988. Systemic Text Generation as Problem 
Solving. Cambridge: Cambridge University Press. 
Reiter, E., and Dale, R., 1992. 'A fast algorithm for the 
generation of referring expressions'. In Proceedings of 
COLING-92: 14th International Conference on 
Computational Linguistics (Nantes). 1992. Morristown 
NJ: Bell Communications Research. 
Rosch, E., 1978. 'Principles of classification'. In Rosch, E., 
and Lloyd, B.B., (eds.). Cognition and Categorization. 
New Jersey: Erlbaum. 
Winograd, T., 1972. Understanding Natural Language. 
Edinburgh: Edinburgh University Press. 
