::3 2/2 
Two approaches to semantic interfaces 
in text generation 
Christian Matthiessen 
Department of Linguistics, 
Sydney University 
This paper is a contribution towards the exploration of semantic interfaces in text generation 
systems. It suggests a general interpretation of semantics for the purpose of text generation as 
an interlevel between lexicogrammar (the resources of grammar and vocabulary) and higher 
levels of organization (knowledge base, user model, text planning, and so on). Two approaches to 
the design of this interlevel that have been implemented in generation systems are then presented 
-- chooser & inquiry semantics and situation-specific semantic systems. They are compared and 
contrasted to bring out their relative merits. 
1. The role of semantics in text generation: semantics as an inter-level 
Text generation is the creation of text, typically by means of a computer, in response to 
some well-specified need for text such as the need to report on tomorrow's weather or to 
define a particular term for somebody (e.g. Kempen, 1986; McKeown & Swartout, 1987). 
That is, a text has to be created step by step from the initial specification of a need for a 
text to a final output, either in writing or in speech; minimally we can specify the initial 
need for text at one end and lexicogrammatical representation & orthographic 
representation at the other. The organization of the process of generating a text and of the 
resources that are activated in the course of this process can be seen as stratification or 
the arrangement into levels of successive orders of symbolic abstraction. So given the task 
of text generation, the most productive interpretation of the semantic system is a stratal 
one, more specifically, semantics can be seen as the interlevel or interface between the 
linguistic resources of the system and the higher-level, non-linguistic ones. For instance, 
if there is a contextual specification that a service of some kind is needed, this can be 
addressed linguistically by choosing some semantic strategy such as pleading, requesting, 
or ordering; which strategy is selected will again depend on contextual factors such as the 
nature of the relationship between speaker and listener. The semantic selections are re- 
expressed lexicogrammatically and then again graphically or phonologically. 
While there are many possible conceptions of semantics, it is this interpretation of 
semantics as a strategic interlevel for accomplishing tasks linguistically that is 
central to text generation; and it is the conception of semantics that we find in systemic- 
functional linguistics. 1 The strategies can be represented as a set of inter-related 
options by means of the system network of systemic theory. This corresponds to 
McDonald's (1980) characterization of text generation as being organized around the 
notion of choice. That is, generating a text is essentially a process of selecting among all 
the various alternatives available at different levels of abstraction. As Patten (1988) has 
pointed out, there is a significant parallel between Halliday's (e.g., 1973, 1978) 
emphasis on the process of choice and the organization of language as inter-related options 
in systemic linguistics and the paradigm of problem solving in Artificial Intelligence, 
involving the process of searching for solutions from among the options in a solution space. 
2. Two approaches to the design of the semantic interlevel 
Given that we interpret semantics as an interlevel, one central question is how we can 
explore the organization of this interlevel. Since semantics faces upwards, towards higher 
levels of organization as well as downwards, towards lower levels within the linguistic 
system, there are two stratal approaches that can be adopted in exploring the organization 
~nd categories of semantics, (i) one from below; and (ii) one from above. 2 
(i) We can explore semantics from below, starting with lexicogrammar (the unified 
resource of grammar and vocabulary) -- what might be called a decoding or 
interpretive approach, since it works by decoding or interpreting lexicogrammar in 
semantic terms. 
(ii) Alternatively, we can explore it from above, from outside the linguistic system -- 
what might be called an encoding approach since it looks at semantics as an encoding 
strategy and explores how contextual categories are encoded semantically. 
Both of these approaches have been used in text generation systems. I will discuss one 
example of the decoding approach in more detail in Section 3, chooser & inquiry semantics 
developed for and used in the Penman text generation system (e.g., Mann, 1982; 
Matthiessen, 1988), and one example of the encoding approach in Section 4, the theory of 
situation-specific semantic systems (Halliday, 1973) modelled in Patten's (1988) SLANG 
generator. 3 I should emphasize that the decoding and encoding approaches should not be 
seen as mutually exclusive alternatives. Rather, the assumption is that they can be 
reconciled into one account of the semantic interlevel that brings out how it relates its two 
interfaces. 
3. Approaching semantics from below: chooser & inquiry semantics 
Approaching semantics from below means taking lexicogrammar as the point of departure 
in modelling semantics. 
3.1 The organization of the level below: the grammatical system network 
The nature of the model of semantics that results from a lexicogrammatical point of 
departure will obviously be determined to a large extent by the nature of the theory of 
grammar. If the focus of the grammar is on structure, the semantics will essentially be a 
semantics of grammatical structure, possibly cast in some form of predicate logic. 
However, if the grammar is paradigmatically organized -- i.e., if the theory takes choice 
as the basic organizing principle, as systemic theory does -- the semantics will 
essentially be one of choice as well. We can call this model of meaning choice semantics. 
A number of generation systems have used systemic-functional grammar (e.g., Davey, 
1978 \[Proteus\]; Mann & Matthiessen, 1985 \[Nigel\]; Bateman el al, 1987 \[the Kyoto 
grammar\]; Fawcett and Tucker, 1989 \[COMMUNAL\]; Patten, 1988 \[SLANG\]). The central 
organizing principle is the system network; for example, the system network below is a 
fragment of the grammar of MOOD in English (the lower part of Figure 1). 
3.2 The semantic control of a system: choosers and inquiries 
Now, we can organize a semantic interface in terms of the system network, which is what 
Bill Mann and I did in the development of the chooser and inquiry framework for one 
particular systemic generation grammar, the Nigel grammar developed at USC/ 
Information Sciences Institute (Matthiessen, 1981; Mann, 1982; Matthiessen, 1988); 
this framework was then adopted and extended for the systemic generator of Japanese and 
Chinese by John Bateman and associates at Kyoto University (Bateman & Matthiessen, 
forthcoming). Each system in the system network is equipped with a chooser--a 
semantic procedure for ascertaining the information needed to choose the appropriate 
feature in the system. The chooser achieves this by presenting one or more inquiries to 
the higher-level contextual systems above the grammar and its semantic interface.4 
Choosers can be added to the system network as shown schematically in the top layer of 
Figure 1 (choosers are represented by circles at the semantic level above the grammatical 
system network). 
(the MOOD fragment above is a simple taxonomy; but system networks in general allow for 
simultaneous systems and disjunctive entry conditions. -i"he former property is important for 
multifunctionality and parallel generation algorithms; cf. Tung, Matthiessen & Sondheimer, 1988) 323 
An inquiry is simply a demand for information -- e.g., 'is the current speech function a command, 
i.e. a demand on the addressee to perform a service or to provide goods?' (Command?); 'is the 
current speech function a question, i.e. a demand on the addressee to supply inforrnation?' 
(Question?); etc. -- and the context has to return a response. The chooser then acts according to 
the response, either by presenting another inquiry if more information is needed or by selecting 
one of the grammatical features of its system if it has enough information. To take a very simple 
example, the chooser of the system INDICATIVE TYPE has the task of choosing between 
'declarative' and 'interrogative'. It does this by presenting the inquiry Question? to the context -- 
an inquiry asking whether the current speech function is a demand for information, i.e. a question, 
or not. If the response is positive, the chooser selects 'interrogative'; if not, it selects 
'declarative'. The chooser thus treats 'declarative' as the default option of the system. 
clause 
indicative 
interrogative 
declarative 
variable?/ 
wh- 
yes~no 
imperative 
> 
('Z'z:~ re.z~2" d.z'.t'3e~t.z'ozz.k 
Fig. 1 : System network (grammar) with choosers (semantics) 
The inquiry has two or more possible responses; since these responses define branches in the 
organization of the chooser, this type of inquiry is called a branching inquiry (there is one 
other type, the so-called identifying inquiry, used to bind variables in inquiries to instantial 
values). This is an example of a minimal chooser: it consists of just one inquiry. However, a 
chooser may consist of more than one inquiry; if there is more than one, they are organized into a 
decision tree -- an Inquiry tree (for examples, see Mann, 1982; Matthiessen, 1983; 1988). 
The response to one inquiry simply leads to another inquiry. During generation, an inquiry tree is 
simply stepped through one inquiry at a time, until the process reaches a response that is a 
terminal branch in the tree and leads to the choice of one of the options the chooser is associated 
with. 5 
324 
To sum up, as the grammatical system network is traversed, systems in the network are 
reached and their choosers are called into action, The chooser of a system has the task of 
making an appropriate selection in its system and it does this by presenting one or more 
inquiries to the higher-level systems of the text generator. Consequently, the problem of 
controlling the grammatical resources in a purposeful way is decomposed into a number of 
very simple demands for information. These demands can be taken as the basis for 
specifying what kind of organization is needed to support the generation process: see 
Matthiessen (1987) and Bateman & Matthiessen (1989, forthcoming). While there are 
considerable advantages in assuming that the semantic interfaces is simply a collection of 
choosers, the work on English and Japanese generation points towards an organization of 
inquiries that is more global than choosers local to grammatical systems: see Bateman & 
Matthiessen (1989). 
4. Approaching semantics from above: context-based semantics 
When we approach semantics from above it is the interface between context and language 
that is highlighted. The role of semantics can be stated with respect to context as follows: 
semantics is the set of strategies for construing contextual meanings as linguistic meanings 
and thus moving into the linguistic system,, Or~ if we focu,~ on the notion of ~o~.I ir 
particular, semantics is the set of strategies for achieving some goal through symbolic 
activity. This is a functional approach to semantics: it interprets semantics in terms of the 
uses it has evolved to serve in different communicative contexts. This functional approach 
has a number of consequences for semantics; I will mention three here, the second of which 
I will pursue in Section 4.1: 
(i) Semantic categories have to be sensitive not only to the downward interface to 
lexicogrammar, but also to the upward interface to context; they have to show how it is 
that semantic strategies can play a role in context. 
(ii) Since communicative contexts are highly diversified, we have to show how 
semantics can be responsible across these various contexts; one way of modelling this is 
to treat semantics itself as diversified into a number of semantic systems 'tailored' to 
specific communicative demands. 
(iii) Semantics has to be concerned with language functioning in context rather than 
any unit defined by lexicogrammar; consequently, the basic unit of semantics is text o- 
language functioning in context ~° rather than propositions or predications. 
4.1 Specificity of semantic categories -- functional diversification of 
semantics 
Since the approach from above takes context as its starting point, it is likely to yield 
situation-specific semantic systems: we project a variety of different uses onto 
semantics, giving us semantic interpretations of contextual categories; for example, 
'behavioural control of child' is semanticized as 'appeal to authority figure', 'threat of 
physical punishment', 'threat of loss of privilege', and the like, whereas 'behavioural 
control of student' is semanticized as 'warning about fees', 'threat of expulsion from 
programme', and the like. The notion of function reflected in this kind of semantics is thus 
use in context and there will be a large number of different uses.6There are at least two 
basic types of motivation for exploring and writing context-based semantic systems, (i) 
bridging and (ii) compilation. 
(i) 8ridging. The orientation towards context serves to bridge the gap between 
linguistic categories and higher-level categories. Within sociology, Halliday's concept of 
semantics is motivated partly because it can act as an interface between language and the 
rest of the social system. Turner (1973: 195) comments 
\[Halliday's concept of meaning potential\] should enable researchers to integrate sociological 
concepts and linguistic concepts. The sociological theory identifies the socially significant 
meanings. Once these are specified, their grammatical and lexical realizations are also 
capable of specification. 
Within computational linguistics and AI, it is possible to make similar observations: 
situation-specific semantic systems may serve to relate nomlinguistic categories to 
linguistic ones. 
(ii) Compilation. Furthermore, a situation-specific semantics can be seen as a set of 
strategies developed to deal efficiently with the specific, limited set of communication 
problerns inherent in that context of situation. We can find this consideration in 
computational linguistics and AI. As noted earlier, Patten (1988) has shown that the 
approach of situation-specific semantics can be motivated in AI terms as well as in 
linguistic terms. Patten treats text generation as problem solving and shows that there is a 
striking similarity between the AI problem-solving framework and Halliday's systemic 
approach to language. The similarity is all the more interesting because the two traditions 
have developed independently of one another. 
The lexicogrammatical system network can be seen as the space of inter-related 
alternatives for solving a communicative problem. There are different ways of searching 
the system network for appropriate feature selections. One way is to traverse the network 
from left to right and to reason about each systemic alternative by means of choosers (cf. 
Figure 1 above). Patten argues that it is potentially costly to do this kind of reasoning from 
basic principles. Another way is to rely on a strategy that has already been developed for a 
325 
326 
particular problem ("compiled knowledge") and this is what Patten takes a situation 
specific semantics to be. That is, for a given register there is a particular semantic 
strategy for traversing the lexicogrammatical system network. If we are faced with a novel 
generation task which does not correspond to a recognized register, we will have to revert 
to basic principles. 
4.2 An example: a semantics of control in a regulatory context 
Let's consider an example of the semantics of a particular register. Assume that we are 
building a generator for mother-child control, the situation Turner (1973) did 
sociological research on, Halliday (1973) uses as an example of sociological semantics, 
and Patten (1988) takes over for text generation.7 The situation is the following 
(Halliday, 1973: 65): 
\[A\] small boy has been playing with the neighbourhood children on a building site, and has come 
home grasping some object which he has acquired in the process. His mother disapproves, and 
wishes both to express her disapproval and to prevent him from doing the same thing again. 
The question is what her linguistic strategies are. The answer lies in the semantic system 
network; most generally, she can threaten, warn, or appeal to her son, issue a rule, etc.. 
The semantic network consists of systems like threat / warning, physical punishment / 
mental punishment / restraint on behaviour, and so on. Semantic features are realized by 
preselections of grammatical features. For example, the semantic feature 'threat' is 
realized by selection of the grammatical feature 'declarative'. In general, delicate 
grammatical features are preselected and the less delicate features they presuppose can 
then be chosen automatically by moving from right to left by backward chaining in the 
system network rather than by explicit preselection. This method makes good use of the 
'logic' of the lexicogrammatical system network; see Figure 2 below. 
As the diagram indicates, there is a tendency for the situation-specific semantics to be more 
delicate than the non-specific lexicogrammar. This is to be expected, particularly in the fairly 
restricted registers that have been attempted in text generation: only a restricted subset of the 
lexicogrammatical resources will be employed and the semantics can simply 'turn off' certain 
parts of the grammar by never preselecting grammatical features in these parts. 
To extend Patten's line of research further within text generation, it is important to describe 
the semantic systems of a variety of situation types; for instance, Marilyn Cross, at Macquarie 
University, is currently working on descriptions of the water cycle for different addressees. 
There is good reason to think that the approach of situation-specific semantic systems will yield 
interesting results. The types of situation for which we can attempt to write semantic networks 
would also seem to be the types that can be addressed in text generation at present. 
Now, the example of regulatory semantics has been discussed in terms of a network of semantic 
systems such as threat / warning. To relate these systems upwards, we can assume realizations 
of contextual features of interaction by means of preselections of semantic features. 
Alternatively, we can easily turn the systems into inquiries that demand information from 
context. For instance, the system threat / warning can be re-represented as an inquiry concerned 
with the basic strategy of control: is the child to be controlled by appealing to authority, by 
threatening him with punishment or restraint on behaviour if he carries on; or by appealing to the 
dangers of the world, by warning him that his behaviour will harm him? 
Although the systemic semantics used by Patten (1988) is context based, the texts that 
can be generated do not extend beyond the clause complex. There is, however, every reason 
to expect a text semantics rather than only a lexicogrammatical semantics -- i.e. a 
semantics that is concerned with text as a semantic unit, the basic unit of communication. 
To develop the notion of text semantics further, we would need to examine proposals for 
how to organize text, since they would provide us with structures we can interpret as text 
semantic structures. The two types of approach that have been developed for text 
generation are (i) McKeown's Rhetorical Schemas (McKeown, 1982; Paris & McKeown, 
1987) (ii) and, within the Penman project, Rhetorical Structure Theory (see e.g., Mann 
& Thompson, 1987; Matthiessen & Thompson, 1988; Mann, Matthiessen & Thompson, 
1989). McKeown's work is very similar to systemic work on generic structures by Hasan 
(1978, 1984, etCo) and others. In either case, the structures they operate with can serve 
to realize semantic features in a text semantic system network° 
,% 0 
oJ 
~4 
o 
...o o {IJ 
Vl 
L'=a 
oJ 
5. Conclusion: the two approaches re-considered 
To recapitulate, taking the basic systemic posiiion that semantics is an interlevel 
between higher-level contextual systems and the purely language-internal level of 
lexicogrammar, I have suggested that we can approach it from either of the two semantic 
interfaces -- from above, from context, or from below, from lexicogrammar -= and that 
we find both approaches modelled in text-generation systems using systemic-functional 
grammar. The chooser and inquiry interface, built from below, has the advantage that it is 
fairly easy to develop once there is a significant systemic-functional grammar to base it 
on; it can be developed as 'semantic glosses' on the organization already embodied in the 
grammatical system network. It does not change the basic principle of generation supported 
by the grammar: the grammatical system network guides the generation process, which is 
essentially a traversal of the network, and choosers are activated in the course of this 
process. The collection of inquiries can be used as design requirements in the development 
of the organization of the context of the generation system. The approach from above has 
other advantages. It enables the semantics to refer to different grammatical contexts in 
realization, as in the case of requests being realized both by selections in MOOD and 
selections in MODALITY. Furthermore, it allows us to adapt the semantics to contextual 
requirements. This adaptation may take the form of a diversification of semantics into a 
range of situation-specific semantic systems. Such systems have the added advantage that 
they represent 'compiled knowledge': they allow the generation system to take advantage of 
the semantic strategies that have evolved for a particular communication task rather than 
having to solve the problem from first principles. This means, among other things, that 
only those parts of the lexicogrammar that are relevant in that situation have to be 
ex ~lored and others are simply 'blocked off' by preselections from the semantics. 
f 
+~ ~rz~.z'~ . . . 
physical punishm ent 
three, t ----~ mental p~ulis~ment 
indicative ~~ wh- 
imperative yes/no 
L ..... _. __ ................... _ ..... 
Fig. 2: Generation with situation-specific semantics 
The two approaches have been used in different generation systems, but they have not 
been brought together into one system and I have not explored the question whether this 
would be possible or not. We can obviously say that one chooses one approach or the other 
depending on the nature of the generalion task. For instance, more closed registers 
(specific sublanguages) might favour situation-specific semantic systems whereas more 
open registers might favour the use of a general semantic system. However, in the long 327 
run, such a position would clearly be unsatisfactory since it would commit a generation 
system to one type of generation task or another. The most clearly differentiated positions 
that are theoretically possible would thus be: 
(i) Situation-specific semantic systems are essentially based on different principles 
of organization, creating semantic potential for a given situation type, and cannot be 
drawn from one general semantic system; and 
(ii) Situation-specific systems are merely abbreviations of one general semantic 
system, 'blocking off' semantic potential that is not needed in a given situation type. 
This needs a long separate discussion and I will leave the issues at this point for now; for 
further discussion, see Matthiessen (1989) and cf. Bateman & Paris' (1989) approach to 
register by means of chooser and inquiry semantics. 
1 Traditionally, semantics has tended to be modelled from the point of view of comprehension, 
by reference to rules for interpreting syntactic structures. 
2 The two directions pertain to the design of the semantics, not to the direction in the flow of 
control. Encoding and decoding are thus not to be equated with generative and interpretive 
semantics. Both generative and interpretive semantics are essentially decoding in that they 
reflect the categories of grammar rather than contextual categories. 
3 Tile survey here is thus not exhaustive; in text generators, we also find the use of parallel, 
co-ordinated taxonomies (as in Jacobs, 1985), unification of semantic and grammatical 
information (cf. McKeown, 1982), and augmented phrase structure rules (cf. Sowa, 1983). 
The system network guides the generation process. In the course of generation, the system 
network is traversed from left to right, that is, from more general options towards the more 
specific ones that become reachable once the more general ones have been chosen (see the 
'traversal direction' in Figure 1 below). Any feature may have a realization statement associated 
with it; that is, a statement that specifies how the choice of the feature is realized structurally 
(no realization statements are shown in the network in Figure 1). For instance, the feature 
'declarative' is realized (in English) by the relative ordering of Subject before Finite (pigs can 
fly), while 'yes/no interrogative' is realized by the relative ordering of Finite before Subject 
(can pigs fly). As an option is chosen, any realization statements associated with it are 
executed, which means that a fragment is added to the grammatical structure being built as a 
realization of the selections. As the system network is traversed from left to right, structural 
specifications are accumulated until the network has been fully traversed and the structure fully 
specified by the realization statements that have been encountered and executed along the way. 
4 If inquiries are interpreted as being concerned with choice conditions (cf. Matthiessen, 
1988), we can see that these choice conditions are comparable to Fawcett's (1984: 166; 1983: 
section 3.2) procedural felicity conditions in his systemic model. 
5 i have glossed inquiries by using names such as Question? and Precede? and by using informal 
English questions, which is helpful in developing a design of a large system such as the Nigel 
generator. However, as part of an automatic text-generation system, inquiries are also 
Implemented: the steps for testing an inquiry to see which response is appropriate are spelled 
out in the generation programme (for more details on generation, see Nebel & Sondheimer, 1986). 
Thus, for instance, the source of the response to 'Posssesion? (Henry horse)' -o 'is the 
relationship between Henry and horse one of generalized possession, i.e. one of ownership, 
meronymy, close social association, etc.'-- might be derived ultimately from a relation in a data 
base. One important point here is that it is possible to specify different implementations of 
inquiries reflecting different types of representation of the information that will be the basis of 
the responses. This information might be represented in say an extended predicate calculus 
notation or in terms of some kind of frame-based network. In the current version of the Penman 
system of which the Nigel grammar and its chooser and inquiry interface form one part, there is a 
special simple notation for specifying the kinds of information the inquiries need (SPL or sentence 
planning language). 
6 This does not mean that the various situation-specific semantic systems cannot be derived 
from a generalized semantic systems, but I won't discuss this important issue here. 
7 We might undertake this project to test the model, possibly as a pilot for future work. The 
computer would simulate the mother. In other text generation situations, such as expert system 
explanation, the computer's social role is more likely to be that of a computer. 
329 

References 

Bateman, J., Gen-ichirou, & Atsuchi TabuchL 1987. Designing a compulational sysmmic 
grammar for text generation: a progress report. Department of Electrical Engineering, Kyoto 
University. 

Bateman, J. & C. Matthiessen. 1989. The text-base uncovered. In H. Bluhme ted), \[Selected 
papers from the leternational Conference on Language and Text Research, Xi'an Jiaotong 
University, Xi'an, P.R.C., March 89\]. To appear. 

Bateman, J. & C. Matthiessen. forthcoming. Systemic Linguislics and Text Generation: 
Experiences from Japanese and English. London: Pinter. 

Sateman, J. & C. Paris. 1989. Constraining the deployment of lexicogrammatical resources 
during text generation: towards a computational instantiation of register theory. Paper 
presented at Sixteenth International Systemic Congress. MS, USCi\[SL To appear in E. Vento\[a 
ted), Selected Papers from the Sixteenth International Systemic Congress. 

Davey, A. 1978. Discourse Production. Edinburgh: Edieborgh University Press. 

Fawcett, R. 1983. Language as a semiological system: a re-interpretation of Saussure. In J 
MorreaU (ed.). The Ninth LACUS Forum, 1982. Columbia: Bornbe~,m Press. 

Fawcett, R. 1984. System networks, codes, and knowledge of the universe. In R. Fawcett, 
M.A.K. Halliday, S. Lamb, & A Makkai (eds.). The Semiotic of Language and Culture. Volume 1: 
Language as Social Semiotic. Volume 2: Language and other Semiotic Systems of Culture. 
London: Frances P~nter. 

Fawcett, R. & G. Tucker. 1999. Prototype Generators 1 and 2. COMMUNAL Report Number 10, 
Computational Linguistics Unit, University of Wales College of Cardiff. 

Halliday, M.A.K 1973. Explorations in the functions of language. London: Edward Arnold. 

Halliday, M.A.K. 1978. Language as social semiotic. The social interpretation of language and 
meaning. London: Edward Arnold. 

Hasan, R. 1978. Text in 1he systemic-functiona$ model. In W. Dressier ted.). Current Trends in 
Text Linguistics. Berlin: de Gruyter. 

Hasan, R. 1984. The Nursery Tale as a Genre, Nottingham Linguistic Circular , No. 13. Special 
Issue on Systemic Linguistics. 

Kempen, G 1986. Language Generation Systems. In I. Batori, W. Lenders & W. Putschke teals.), 
Computational Linguistics. An International Handbook on Computer Oriented Language Research 
and Applications. Berlin/Now York: Walter de Gruyter 

Kempen, G. ted.). 1987. Natural Language Generation. Dordrecht: Martinus Nijhoff. 

McDonald, D. 1980. Language Production as a Process of Decision-making under Constraints MIT 
Ph.D. Dissertation. MIT Report. 

McKeown, K. 1982. Generating Natural Language Text in F\]esponse to Queslions about Database 
Struclure. Ph.D. Dissertation, Universily of Pennsylvania. 

McKeown, K. & W, Swartout. 1987. Language generation and explanation. Annual Review of 
Computer Science. 2: 401-49. 

Mann, W. 1982. The Anatomy of a Systemic Choice USC ISI/RR-92-104 Also in Discourse 
Processes. 

Mann, W. & C. Matthiessen. 1985. Dernonslration of the Nigel Text Generation Computer 
Program. In J. Benson & W. Greaves (eds), Syslemic Perspectives on Discourse, Volume 1. 
Norwood: Ablex. 

Mann, W., C Matlhiessen & S. Thompson. 1989. Rhetorical Structure \]beery and Text Analysis 
In Mann & Thompson (eds). Text description: diverse analyses of a fund raisi~g text 
Amsterdam: Benjamins. Forthcoming. 

Matthiessen, C. 1981. A Grammar and a Lexicon for a Te×t~Production System In The 
Proceedings of the 19th Annual Meeting of the Association lor Computational Lillguistics 

Matthiessen, C. 1983. Choosing primary lense in English. Foundations of Language, 73 

Mattbiessen, C. 1987 Notes on the organization of the environnlent of a text generation 
grammar. In Kempen ted.). Also as USC ISI/RS-87-177. 

Matthiessen, C. 1989. A systemic semantics: the chooser and inquiry framework. It; 8enson, 
Cummings & Greaves (eds.), Systemic Functional Approaches to Discourse: Selected Papers 
from the Twelfth International Systemic Workshop No,wood: Ablex. Also as USC ISI/RS-87-189. 

Matthiessen, C. 1889. Systemic theory and text generation: some cenha\[ design considerations 
In Proceedings of Australla-Japan Joinl Symposium on Naturai Language processing, 
Melbourne University, Melbourne, Nov. 27-g, 1989. 

Malthiessen, C. 8, S. Thompson. 1987. The Structure of Discourse and "Subordination". USC 
ISIIRS-87.~83. Also in Haiman & Thompson (eds.), 1989 Clause combining in grammar and 
discourse. Amsterdam: Beniamins. 

Paris, C. & K. McKeown. 1987. Discourse strategies for describing complex physical objecls In 
G. Kempen ted.). 

Patten, T. 1998. Systemic text generation as problem solving. Cambridge: Cambridge University 
Press. 

Tung, Y-W, C. Matthiessen & N. Sondheimer. 1988. On Parallelism and the Penman Natural 
Language Generation System. USC, ISI/RR-98-195. 

Turner, G. 1973. Social class and children's language of control at age five and age seven. In 
Bernstein ted.). Class, Codes and Control 2: appfied studies towards a sociology of language 
London: Routledge & Kegan Paul. 
