Discourse Representation Theory and the Semantics of Natural 
Languages: 
Contribution to a Panel on "Discourse Theory and Speech Acts" TINLAP3 
F. Guenthner 
Universit~it Tiibingen/FNS 
0. Some Questions 
The present panel was asked to elaborate on the following series of questions: 
\[1\] Is there yet any serious discourse theory with testable computational and empirical consequences? 
\[2\] What phenomena ought a processing theory of discourse understanding/generation to address itself that are not already being 
attended to currently? 
\[3\] What aspects of discourse are language problems and which are general AI/KR problems? 
\[4\] What makes a theory of discourse a processing theory? 
\[5\] Does spoken language affect one's theory of discourse? 
\[6\] Is there any real hope that we will be able to recognize the plans/goals etc. of a speaker? 
\[7\] How much of conversation is carried on through the linguistic window anyway? 
\[8\] Do current theories of text and dialogue mesh, and should they? 
But one additional question before we try to answer some of the above questions: 
\[9\] How would the answers to questions 1 - 8 differ if instead of"discourse" or "text" and 
• "dialogue" (or whatever) had we been asked these questions in terms of just the concepts of 
"sentence" or "language" ? 
Let me try and answer this question before coming to (some of) the others. 
For anyone who is not convinced that the definition of linguistic meaning must be based on 
something else than the "sentence", the answer to Question 9 must necessarily and lapidarily be 
something like "in nothing" - and with good reason. Of course this answer is not right: not because 
there is a panel that is supposed to give more interesting answers, but because the facts of language 
elicit another answer. 
The basic "unit" of linguistic meaning is the discourse and the basic concept of meaning is one that 
involves quite different semantic ingredients than those that previous attempts to formulate a 
semantics for natural language have made us familiar with. We take it that the concept of linguistic 
meaning should characterize a relation between "information ~tate~" (the term - as well as many 
ideas in this note - originates from two papers by Richard Smaby \[1978\] and \[1981\]). In other 
words we view the (linguistic) meaning of a discourse - and obviously also that of a single sentence 
- as something like a function which transforms information states into (usually but not always) 
105 
new information states. In fact, every linguistic construct below the discourse level, be it a phrase 
or even a basic expression, should ultimately be viewed in this manner. A systematic 
characterization of this relation will reveal that for just about every linguistic construct we are led to 
assume a much more intricate account of the "meaning" they convey than the typical "satisfaction" 
and "truth clauses" that we once assumed in the context of theories of meaning whose main goal 
was the systematic charactization of the models which verified isolated sentences. Traditional 
formal (or model theoretic) semantics is by no means falsified by this view of discourse semantics; 
on the contrary: what is made clear by the new account is that so-called truth relations (as applied to 
sentences or stretches of discourse) are no longer the primary relations but rather derived from a 
more fundamental kind of semantic relation. 
The answers to the above questions will as a result be couched in terms of how information states 
can be changed by the incorporation of new bits of information. 
1. Basic Features of a Theory of Discourse Meaning (or a partial answer to Q 1) 
1J What is "Linguistic Meaning"?: 
Even if it is more than a truism to assert that one of the basic functions of linguistic communication 
is to convey "information" we should not be misled into assuming too strict a link between the two. 
Certainly the traditional notion of "information equivalence" in terms what models (in the logical 
sense) are characterized by an utterance or sets of utterances is a most useful one. But this notion is 
too weak as a characterization of linguistic meaning; there are in general many sentences that - 
given a model - turn out to be equivalent, even though their linguistic properties are quite different. 
We claim that (logically) equivalent sentences can play very different roles in the setting of 
discourse. Observe to begin with that for many sentences the traditional approach in terms of truth 
conditions does not make much sense. Sentences with pronouns or various "elliptical" 
constructions for example cannot even be evaluated (i.e. semantically characterized) on their own. 
Nevertheless they have a linguistic meaning; but their meaning must be reconstructed as something 
else than simple truth conditions. Their meaning is inherently their effect - i.e. the change they can 
induce - on an antecedently given information state. 
1.2 Expressions, Representations, Models 
To be more precise we should say that questions of meaning should be treated at three different 
levels and in a particular order: 
- the linguistic meaning of an expression (be it an atomic expression, a phrase, a sentence or a bit of 
106 
discourse) is what determines its contribution to information states, or better, the conditions of its 
contribution to an information state. 
- the l'¢pzesentational meaning of an expression is the particular "change" the expression effects on a 
given information state. 
- the model-theoretic meaning of an expression are the satisfiability conditions induced by the 
change given a particular class of models for the information states. 
There is a lot to be said about how these "levels" of meaning interact; for a discussion of this issue 
cf. Guenthner \[1985a\]. What should be pointed out here - albeit briefly - is that the general theory 
of meaning should characterize the representational and model-theoretic properties of expressions in 
terms of the linguistic meaning and not the other way around. In other words, it is the 
(characterization of the) linguistic meaning which gives rise to its model-theoretic (or truth 
conditional) properties and not vice-versa (as most recent work in formal semantics might have led 
us to believe). 
13 Two Types of Semantic Relations 
The above distinction between different levels of "meaning" is closely related to a distinction - first 
drawn in a systematic way by Hans Kamp in the framework of Discourse Represention Theory - 
between two kinds of semantic relations. Prior to Kamp's work most if not all semantic properties 
and relations were def'med in terms of truth-conditional terms, ultimately therefore in terms of the 
most basic model-theoretic relation: logical consequence. We shall call semantic relations definable 
in terms of consequence "T(ruth)-relations" and we shall oppose them to a perhaps more interesting 
class of semantic relations that we shall call "D(iscourse)-relations". T-relations (most notably 
consequence, truth and consistency) are relations that characterize the relation of an information 
state and a model, but T-relations are not the only kind of relations between informations and 
models. D-relations on the other hand characterize relations between information states. Both types 
of relations are in general applied to sentences and stretches of discourse, but they can be 
generalized to expressions below as well as above the sentence. Typical D-relations are 
presupposition, ambiguity and coherence. But there are many more. The topic of D-relations is 
pursued in more detail in Guenthner \[1985b\]. 
1.4 Aspects of the Theory of Communication 
All of the above makes a lot more sense once one considers the role of information states in the 
setting of a theory of communication. Sentences but more generally stretches of spoken and written 
107 
discourse are used to "convey" information. In other words they are used to transform information 
states. But how and with what "intention"? A simple but plausible answer (reminiscent of Kamp 
\[1985\]) goes as follows: in communicating a speaker has in general to circumscribe a 
sub-information state from his own overall representational set-up in such a way that there is a 
"choice of words" which allows the linguistic encoding of that sub-information state; the decoding 
of the resulting discourse by the hearer should then result in the installation of an "equivalent" 
sub-information state in the overall representational set-up of the hearer. The rules which govern 
both the encoding and the decoding of discourses are called discourse structure construction rules in 
Karnp \[1981\]. It is this algorithm which defines the concept of "linguistic meaning" mentioned 
above. We claim that the regularities of linguistic meaning are exactly captured by this algorithm. 
1.5 Understanding "Understanding" 
This is a far cry from a theory of understanding, but there is at least something to be said which 
distinguishes "understanding" from other concepts introduced so far. On our view the result of the 
successful decoding of a discourse results in an "interpretation" of the discourse with respect to the 
given antecedent information state. And "understanding" begins where interpretation ends: how one 
understands depends to a high degree what other available information the result of interpreting a 
sentence or discourse can interact with. Since no two hearers are exactly in the same information 
state it can hardly come as a surprise that the ramifications of the incorporated information will not 
play exactly the same role in their mental states. This is true even for rather banal utterances (like ".it 
is four o'clock") and obviously much. more for utterances that involve more intricate conceptual 
structures be these of the scientific, literary, sociological or whatever sort. Moreover, it seems clear 
that there can be effects of "understanding" without interpretation. For the inability to provide an 
interpretation also has side-effects; if the transfer of information fails (e.g. because the hearer 
simply doesn't understand the language that is used) or because central presuppositions axe not in 
place, the hearer can very well draw conclusions which have little or nothing to do with the 
linguistic meaning of the discourse. And it would be a great mistake to identify (or even associate) 
those conclusions with the linguistic meaning of the utterances employed. (For a more extensive 
discussion of these matters, of. Chapter 6 in Guenthner & Sabatier (to appear).) 
2. A Computational Account of DRT 
22. The Fragment in Kamp \[1981\]: Some Remarks 
An important step towards formulating a theory of discourse analysis was taken in Kamp \[1981\] 
108 
and later work by Kamp and others within the framework of Discourse Reprepresention Theory 
(DRT). As presented in that paper, the fragment has a straightforward computational interpretation: 
let d be a discourse consisting of the sentences sl,...,sn, then the Discourse Representation 
Structure (DRS) associated with d (via the so-called DRS construction algorithm A, the core part of 
DRT) is A(d), i.e. A(s) o A(sl) .... o A(sn), where the "result" of applying A to a sentence s 
depends on the DRS derived from the previous sentences. 
An extension of the original fragment of DRT was first implemented in Prolog in Kolb \[1985\]; this 
implementation also had the merit of providing a deductive account (restricted to a generalized 
syllogistic language) for DRSs. 
2.2 Implementing DRT in Prolog 
Prolog lends itsself in a natural way as an implementation language for DRT. (A faithful and at the 
same time optimal implementation of the fragment in Kamp \[1981\] takes up about 11/2 pages of 
Prolog code.) Several alternatives to the implementation of DRT (in Prolog) have been investigated 
(cf. Guenthner & Sabatier \[to appear\] Chapter 6); the most recent implementation takes as it 
top-level predicate the relation mentioned above between an antecedent context, a resulting context 
and a discourse. A context is taken to consist of a DRS together with an (ordered) list of possible 
antecedents for pronouns. (Similar predicates apply to all syntactic categories in the fragment.) The 
implementation is reversible and thus generates DRSs for bits of discourse as well as bits of 
discourse from DRSs. Among the new features of this implementation, we should cite among 
others the treatment of: disjoint reference, possessive and reflexive pronouns, forward anaphora 
(e.g. the Migs & Pilots variety) as well as a simple but effective treatment of quantifier scope 
ambiguities. For instance, the scope effects of sentences likes every man loves a woman are 
dealt with by treating the "weakest" reading first; if later sentences force a stronger reading, 
backtracking will induce a DRS where the indefinite noun phrase is given a "topmost" 
interpretation, i.e. by introducing the discourse reference of that noun phrase in the principal DRS 
(and nowhere else). There is some linguistic evidence that for this kind of quantifier interaction 
there are no other plausible possibilities. In future versions of this implementation we shall exploit 
the fact that the programm is reversible for certain applications requiring generation. 
Finally, we can extract information from DRSs by translating them into predicate logic; a theorem 
prover based on Smullyan's tableaux system has successfully been used to derive answers from 
DRSs in a deductive manner. In ,addition to semantic evaluation (with respect to a model or 
database) we thus have another way of using DRSs for interactive information processing. 
109 
References 
Barwise, J. & J. Perry Situations and Attitudes, MIT Press, 1983. 
Guenthner, F. \[1985a\] "Linguistic Meaning in Discourse Representation Theory", FNS-Bericht- 
85-4, to appear in Synthese. 
Guenthner, F. \[ 1985b\] "Zwei Typen yon semantischen Relationen", FNS-Bericht-85-5, to appear 
in O. Herzog,Wissensbasierte Systeme und Natfirlich Sprache, Springer. 
Guenthner, F. & P. Sabatier An Introduction to Natural Language Processing in 
Prolog, to appear. 
Kamp, H. \[1981\] "A Theory of Truth and Semantic Representation", in J. Groenendijk et al. (eds.) 
Formal Methods in the Study of Languages~ Foris, 1984, 
Kamp, H. \[1983\] "SID without Time and Questions", CSLI-Report, to appear. 
Kamp, H. \[1985\] "Context, Thought and Communication", Aristotelian Society, New Series, 
1984/1985. 
Kolb, H.P. \[1985\] Aspekte der Implementation der Diskursrepr~entationstheorie, 
FNS-Skript-85-1, Universit~t Tiibingen. 
Lewis, D. \[1979\] "Score-Keeping in a Language Game", Journal of Philosophical Logic, 
1979 
Smaby, R. \[1978\] "Ambiguous Co-reference with Quantifiers" in: F. Guenthner & S.J. Schmidt 
(eds.) Formal Semantics and Pragmatics for Natural Languages, Reidel, 1978. 
Smaby, R. \[1981\] "Pronouns and Ambiguity: A Sample Case", in U. M6nnich (ed.) Aspects of 
Philosophical Logic, Reidel, 1981. 
ii0 
