Schank/Riesbeck vs. Norman/Rumelhart: What's the Difference? 
Marc Eisenstadt 
The Open University 
Milton Keynes, ENGLAND 
This paper explores the fundamental differences between 
two sentence-parsers developed in the early 1970's: 
Riesbeck's parser for $chank's'conceptual dependency' 
theory (4, 5), and the 'LNR' parser for Norman and 
Rumelhart's 'active :~emantic network' theory (3). The 
Riesbeck parser and the I,NR parser share a common goal - 
that of trsnsforming an input sentence into a canonical 
form for later use by memory~inference~paraphrase 
processes, l,'or both parserz, this transformation is the 
act of 'comprehension', although they appear to go about 
it in very (Jifferent ways. Are these differences real 
or apparent? 
Riesbeck's parser i~ implemented as n production system, 
in which input text can either ssti~{y the condition 
side of any production rule within ~ packet of 
currently-active rules, or else interrupt processing by 
disabling the current packet of rules and enabling 
('triggering') a new packet of rules. In operation, the 
main verb of each segment of text is located, and a 
pointer to its lexical decomposition (canonical form) is 
established in memory. The surrounding text, primerily 
noun phrases, is then systematically mapped onto vacant 
case frame slots within the memory representation of the 
decomposed verb. Case information is signposted by a 
verb-triggered packet of production rules which expects 
certain cldsses of entity (e.g. animate recipient) to be 
encountered in the text. Phrase boundaries are handled 
by keyword-triggered packets of rules which initiate and 
terminate the parsing of phrases. 
In contrast to this, the LNR parser is implemented as an 
augmented transition network, in which input text can 
either satisfy a current expectation or cause back- 
tracking to a point at which an alternative expectation 
can be satisfied. In operation, input text is mapped 
onto a surface case frame, which is an n-ary predicate 
containing a pointer to the appropriate code responsible 
for decomposing the predicate into canonical form. Case 
information is signposted by property-list indicators 
stored in the lexical entry for verbs. These indicators 
act as signals or flags which are inspected by augmented 
tests on PUSH NP and PUSH PP arcs in order to decide 
whether such transitions are to be allowed. Phrase 
boundaries are handled by the standard ATN PUSH and POP 
mechanisms, with provision for backtracking if an 
initially-fulfilled expectation later turns out to have 
been incorrect. 
In order to determine which differences are due to 
notational conventions, I have implemented versions of 
both parsers in Kaplan's General Syntactic Processor 
(GSP) formalism (2)~ a simple but elegant generalization 
of ATNs. In GSP terms, Riesbeck's active packets of 
production rules are grammar states, and each rule is 
represented as a grammar arc. Rule-packet triggering is 
handled by storing in the lexicon the GSP code which 
transfers control to a new grammar state when an 
interrupt is called for. Each packet is in effect a 
sub-grammar of the type handled normally by an ATN PUSH 
and POP. The important difference is that the expensive 
actions normally associated with PUSH and POP (e.g. 
saving registers, building structures) only occur after 
it is safe to perform them. That is, bottom-up 
interrupts and very cheap 'lookahead' ensure that waste- 
ful backtracking is largely avoided. 
Riesbeck's verb-triggered packet of rules (i.e. the 
entire sub-grammar which is entered after the verb is 
encountered) is isomorphic to the LNR-style use of 
lexical flags, which are in effect 'raised' and 
'lowered' ~olely for the benefit of augmented tests on 
verb-independent ~rcs. Where Riesbeck depicts a 
'satisfied expectation' by deleting the relevant 
production rule from the currently-active packet, LNR 
achieves the same effect by using augmented tests on 
PUSH NP and PUSII PP arcz to determine whether a 
particular case frame Slot has already been filled. 
Both approaches are handled with equal ease by GSP. 
In actual practice, Riesbeck's case frame expectations 
are typically tests for simple selectional restrictions, 
whereas LNR's case frame expectations are typically 
tests for the order in which noun phrases are encounter- 
ed. Prepositions, naturally, are used by both parsers 
as important case frame clues: Riesbeck has a verb- 
triggered action alter the interrupt code associated 
with prepositions so that they 'behave' in precisely 
the right way; this is isomorphic to LNR's flags which 
are stored in the lexical entry for a verb and examined 
by augmented tests on verb-independent prepositional 
phrase arcs in the grammar. 
The behaviour of Riesbeck's verb-triggered packets 
(verb-dependent sub-grammars) is actually independent of 
when a pointer to the lexical decomposition of the verb 
is established (i.e. whether a pointer is added as soon 
as the verb is encountered or whether it is added after 
the end of the sentence has been reached). Thus, any 
claims about the possible advantages of 'early' or 
'instantaneous' decomposition are moot. Since 
Riesbeck's cases are filled primarily on the basis of 
fairly simple selectional restrictions, there is no 
obvious reason why his parser couldn't have built some 
other kind of internal representation, based on any one 
of several linguistic theories of lexical decomposition. 
Although Riesbeck's decomposition could occur after the 
entire sentence has been parsed, LNR's decomposition 
must occur at this point, because it uses a network- 
matching algorithm to find already-present structures in 
memory, and relies upon the arguments of the main n-ary 
predicate of the sentence being as fully specified as 
possible. 
Computationally, the major difference between the two 
parsers is that Riesbeck's parser uses interrupts to 
initiate 'safe' PUSHes and POPs to and from sub-gra,s,ars, 
whereas the L~R parser performs 'risky' PUSHes and POPs 
like any purely top-down parser. Riesbeck's mechanism 
is potentially very powerful, and the performance of the 
LNR parser can be improved by allowing this mechanism to 
be added automatically by the compiler which transforms 
an LNR augmented transition network into GSP ~chine 
code. Each parser can thus be mapped fairly clesJnly 
onto the other, with the only irreconcilable difference 
between them being the degree to which they rely on 
verb-dependent selectional restrictions to guide the 
process of filling in case frames. This character- 
ization of the differences between them, based on 
implementing them within a common GSP framework, is 
somewhat surprising, since (a) the differences have 
nothing to do with 'conceptual dependency' or 'active 
septic networks' s~ud (b) the computational difference 
between them immediately suggests a way to auton~tically 
incorporate bottom-up processing into the LNR parser to 
improve not only its efficiency, but also its 
psychological plausibility. A GSP implementation of a 
'hybrid' version of the two parsers is outlined in (I). 
15 
REFERENCES 
(1) Eisenstadt, M. Alternative parsers fur conceptual 
dependency: getting there is half the fun° 
Proceedings of the sixth international 
~oint conference on artifici%l 
intelli~ence, Tokyo, 1979. 
(2) Kaplan, R.M. A general syntactic processor. In 
R. Ruetin (Ed.) Natural language 
processing. Englewood Cliffs, N.J.: 
Prentice-Hal1, 1973. 
(3) 
(5) 
Norman, D.A., Rumelhart, D.E., and the LNR 
Research Group. Explorations in 
cognition. San Francisco: W.h. Freeman 
1975. 
Riesbeck, C.K. Computational understanding: 
analysis of sentences and context. 
Working paper 4, Istituto per gli Studi 
Semantici e Cognitivi, Castab~nola, 
Switzerland, 1974. 
Schank, R.C. Conceptual dependency: a theory of 
natural language understanding. 
Co~rtitive Psychology, vol. 3, no. 4, 1972. 
16 
