They say it's a new sort of engine: 
but the SUMP's still there 
Karen Sparck Jones 
Computer Laboratory, University of Cambridge 
Corn Exchange Street, Cambridge CB2 3QG, England 
sparckjones%cl.cam.ac.uk@cs.ucl.ac.uk 
I shall lump the specific semantic formalisms currently touted together as 
manifestations of logicism, because the issue is whether the logicist approach to 
language processing is the right one, not whether one particular formalism is better 
than another. 
The \[ogicist model of language processing is essentially as follows. We use a 
phrase structure grammar, heavily laced with features, for syntactic parsing; in 
analysis syntactic processing drives semantic interpretation strictly compositionally, 
to build a logical form representing the literal meaning of a sentence. This logical 
form is further processed, both in discourse operations of a larger scale linguistic 
character, as in (some) pronoun resolution, and, more importantly, in inference on 
global and local world knowledge, to obtain a filled-out utterance interpretation. 
Logical form plays a key role, motivated as much by the need for reasoning to 
complete interpretation as by the need to supply an appropriate input to further 
reactive problem solving. The use of a logical form to represent the meaning of an 
input naturally fits the use of a logical formalism to characterise general and specific 
knowledge of the world to which a discourse refers and in which it occurs. 
Whether the logicist position is psychologically plausible is not an issue here: it 
can be adopted as a base for language processing without a commitment to, say, 
FOPC as a vehicle for human thought, and indeed can be adopted in a 
psychologically implausible form, for example with complete syntactic processing 
followed by semantic processing. The fact that a good deal of development work is 
being done using a case frame approach is no problem for the logicist either: case 
frames are either an alternative notation or can only work for the discourse of very 
restricted applications. 
The logicists' problems are those of getting an expressive and tractable enough 
logic: what sort of logic is powerful enough to capture linguistic constructs, 
characterise the world, and support common-sense reasoning; and is this logic 
computationally practical? Determining and representing the knowledge required for 
a particular system application is not necessarily more of a problem for the logicist 
than for anyone else. The threat to the logicist's moral, or at least mental, purity is 
Whether a logic which will do the job is really a logic at all. If the world is 
heterogeneous, our thinking sloppy, and our language uncertain, whatever closely 
reflects these may be barely worthy of the label "logic". 
This is not to rake up the old semantic nets versus predicate logic controversy, or 
its analogues. We may have a formalism with axioms, rules of inference, and so forth 
which is quite kosher as far as the manifest criteria for logics go, but which is a logic 
only in the letter, not the spirit. This is because, to do its job, it has got to absorb 
140 
the ad hoc miscellaneity that makes language only approximately systematic. 
Broadly speaking, it can do this in two ways. It can achieve its results through some 
proliferation of rules, weakening the idea of inference. Or it can achieve them 
essentially by following the expert system path, retaining a single rule of inference at 
the cost of very many specific, individual axioms. It is at least arguable that if the 
stock of initial propositions is a vast heap of particulars defining idiosyncratic local 
relationships, the fact that one is technically applying some plain rule of inference to 
follow a chain of argument is not that impressive: conciseness and generality, which 
at least some expect a logic to have, are not much in evidence. Precision and clarity 
may be equally unattainable, at any rate in practice. 
I believe the second possibility is already with us, masquerading in the respectable 
guise of meaning postulates, and that whatever precise view is taken of meaning 
postulates, they sell the logicists' pass. 
This is well illustrated by following through the implications of the processor 
design adumbrated in a recent SR\[ report. (As I am one of the authors of this 
report, I should make it clear that \[ am using this design as a vehicle for discussion, 
and not in a particularist critical spirit.) The report proposal adopts the logicist 
approach outlined earlier, for the purpose of building language processing interfaces 
to, for example, advisor systems. The design is for two processors. The first, the 
linguistic processor proper, is a general-purpose, application-independent component 
for syntactic analysis and the correlated construction of logical forms. The output of 
the linguistic processor is then fully interpreted (progressing from a representation of 
a sentence to that of an utterance) in relation to a discourse and domain context. 
The semantic operations of the linguistic processor proper deal, respectively, with 
the logical correlates of linguistic terms and expressions, and with the application of 
selection restrictions. The logical structures for linguistic expressions ~re determined 
by the domain-independent properties of items like articles and moda\[s and of 
syntactic constructs like verb phrases, and by the formal characterisation of domain 
lexical items primarily as predicates of so many arguments. The lexical information 
about sorts, which supports the selection restrictions, is functionally distinct, as its 
role is simply to eliminate interpretations. 
In this scheme of things the semantic information given for lexical items, and 
especially for 'content words', in the processor's output sentence representation, is 
fairly minimal. It is sparse, abstract, and opaque. The assumption is that predicates 
corresponding to substantive words are primarily given meaning by the domain 
description, and hence by the world which models this. Within the linguistic 
processor one sense of "supply", call it 'supply1', just means SUPPLY, where 
SUPPLY is an undefined predicate label. The sortal information bearing on 
predicate arguments which is exploited via selection restrictions does not appear in 
the linguistic processor's output meaning representation. 
The domain description gives meaning to the predicates through the \[ink provided 
by meaning postulates. These establish relations between predicates of the domain 
description language. But they are in a material sense part of the domain 
description, since these names are also used in the description of the properties of 
the domain world. Broadly speaking, the meaning postulates form part of the 
axiomatic apparatus of the domain description. Thus from a conventional point of 
view the lexicon says rather little about meaning: it merely points into a store of 
information about the world about which the system reasons, both to understand 
141 
what is being said and to react to this both in task appropriate actions and more 
specifically in linguistic response. The system structure is thus a particular 
manifestation of AI's emphasis on world knowledge and inference on this. The fact 
that meaning postulates are also the source of the sortal information applied through 
selection restrictions underlines the somewhat ambiguous character that meaning 
postulates have; but as noted, this sortal information does not figure as part of the 
information supplied in the representation of input text items in the output of the 
strictly linguistic processor. However the predicate labels of the meaning postulates 
may be word sense names so, e.g. 'supplyl' is directly mapped onto 'provide3': this 
suggests that the boundary between semantic information in some narrow linguistic 
sense which refers to the content of the lexicon that is transmitted by the first 
processor, and semantic or conceptual information in the broader sense of the 
knowledge about the world that is incorporated in the non-linguistic domain 
description, has no theoretical but only operational status. 
The immediate motivation for the system design just outlined is a very practical 
one, that of maximising system portability. Given our current inability to handle 
more than a very small universe of discourse computationally, we have to allow for 
the fact that some of the particular domain information appropriate to one specific 
application may be unhelpful or even confusing for another, and that the system 
design should therefore clearly separate the body of information which is general to 
language use and so should be transportable from that which is not. In the scheme 
presented the domain dependent information is confined to the lexical entries for the 
application vocabulary, and to the domain description. 
But logicists also appear to advocate this form of processor as a matter of 
principle. Setting aside the question of whether the control structure of the processor 
is psychologically plausible (because it would be perfectly possible to apply syntactic 
and semantic, and linguistic and non-linguistic, operations concurrently), there is 
still the question of whether a viable general-purpose computational language 
processing system can be built with a strategy that treats meaning in the way the 
design described does, with so little information about it in the lexicon and so much 
in the knowledge base. The strategy implies both that there are no particular 
processing problems which would stem from the need to include both common and 
specialised knowledge, and perhaps several areas of specialised knowledge, in the 
knowledge base and, more importantly that none follow from the comparative lack of 
semantic information of the conventional kind found in ordinary dictionary 
definitions in the lexicon used for the purely linguistic processes, i.e. that there is 
only sortal information for selection restriction purposes. The first problem is not 
unique to the logicist position: any attempt to use information about the world, as 
all systems must, has to tackle the problem of arbitrarily related subworlds. The 
second problem seems to be more narrowly one for the logicist. The point here is not 
so much that, in staged processing, the attempt to avoid duplicating information 
means that information is unhelpfully withheld from earlier processes in favour of 
later ones. The point is rather whether, even in a situation where concurrent 
processing is done, providing much of the information germane to word meanings via 
domain descriptions is the right way to do semantics. It may be a mistake to regard 
linguistic meaning and world reference as the same; it is possible that some 
information about meaning has to be supplied, for representational use, in a form 
exclusively designed for strictly linguistic processing. 
142 
But all this is speculation. What is clear, on the other hand, is that the meaning 
postulates strategy, even if it does not involve the problems just mentioned, will, 
when applied to a non-trivial universe of discourse, imply vast amounts of very 
miscellaneous stuff. If language is intrinsically complex, simplicity at one point 
simply pushes all the complexity (or mess) somewhere else. In more exclusively 
linguistic approaches to processing this tends to take the form of putting all the 
detail in the lexicon: then the grammar can be nice and straightforward. Maybe one 
can get away with a simple syntactic analyser and semantic interpreter; but only by 
supplying all the specialised matter they need to work effectively on the various 
texts they will encounter, through lexical entries. And if simplicity in one place is 
found at the expense of complexity in another, it does not obviously follow that the 
system as a whole has the elegance of its simpler part. In the same way, in the 
logicist approach, the set of meaning postulates required may turn out to be such a 
huge heterogenous mass as to suggest, to the disinterested observer, that the purity 
the use of logic might imply has been compromised. From this point of view, indeed, 
whether the kind of information captured by meaning postulates is deemed to be 
part of the domain description, or is deemed part of the lexicon and is even 
expressed in the representation delivered by the linguistic processor, is irrelevant. 
Either way, the logicist approach suffers from the Same UnManageable Problem of 
miscellaneous linguistically-relevant detail as every other approach to language 
processing. 
This is without considering the proposition that there is a much larger problem 
for which the logicists have so far offered us no real solutions: how to capture 
language use as this is a matter of salience, plausibility, metaphor, and the like. But 
whether or not the logicists can solve this, the real problem, we should not assume 
that they have got more mundane, literal matters of language taped. 
H. Alshawi, R.C. Moore, S.G. Pulman and K. Sparck Jones, Feasibility study for 
a research programme in natural-language processing, Final Report Project ECC- 
1437, SR\[ International, Cambridge, August 1986. 
J.R. Hobbs and R.C. Moore (edsi, Formal theories of the commonsense world, 
Norwood, N J: Ablex, 1985. 
143 
